Foil Branch Log

Snapshot -- an offline archive of commit messages from the eight foil/* branches of cheyras/deckpal. These branches are being deleted; nothing on this page survives anywhere else in the live repository.

DeckPal — Foil Branch Log

The foil track was developed on eight parallel branches that were never merged into main. Their files are redundant — every branch carries a near-identical tree, and no branch holds a file the others lack. Their commit messages are not: they carry the measurements, the refusals and the reasoning behind the pattern taxonomy, the mask pipeline and the resolver, and roughly none of it is restated in DECISIONS.md.

So the branches go and the writing stays. This page is that writing, complete and unfiltered.

Archived 2026-08-31. The branches were still present on origin when this page was written; their deletion is a separate, later step. Everything below is the pre-deletion state.

What is here

• 465 commits across 8 branches, carrying 76,291 words of commit message.
• 262 of them have a body longer than ~100 words. Those are reproduced in full, grouped by branch, chronological within branch.
• The remaining 203 — chores, rebases, one-line fixes — are listed in a compact table at the end. Deprioritized, not discarded.
• Every commit on all eight branches is here. Nothing was filtered out.

The eight branches share no commits with each other beyond main: each is an independently rebased line, so a piece of work that was carried across branches appears once per branch with a different SHA each time. Sole author throughout: cheyras <cheyras@gmail.com>.

Pre-deletion branch tips

Branch	Tip SHA	Commits not on main	Words
foil/analysis-source	65a683a8a4692f85fbaba041bbe00c0c5089260e	62	10,533
foil/main	ceb0ab1a812845a3f35397b4fe51b20025929450	63	10,724
foil/mask-edgetrace	91615360d19105389f480e8d1c5a0699bf7ac0c5	53	8,204
foil/mask-edgetrace-batch	c611e3f88df4f456eb7c67249cff1edcd0c7f21f	54	8,577
foil/mask-learn	d1dd25f85276516af9923f41ceda82ac3fdfc497	57	9,267
foil/mask-straighten	9db4e93eebd4bef87896f1b15ae1636ce9a37a66	51	7,803
foil/mask-vector	bf1ff3f219aa7a102c6de8ba5083b694c32cadfb	62	10,533
foil/tooling-research	85af395c28d2c4f3b9bf43d0fc6293c9951272d4	63	10,650
main (for reference, at archive time)	3e702637013c8e50f78b97770697757e63b7a988	—	—

These SHAs resolve against the offline mirror deckpal-mirror-2026-08-31.git, taken with git clone --mirror on 2026-08-31 and verified git fsck-clean with all nine tips matching origin at that moment. The mirror is the archive; no tags were created, deliberately — a tag would keep the objects reachable, which is exactly what the deletion is meant to end.

origin's history is expected to be rewritten after this page was written (an unrelated media purge on main), so SHAs quoted here may not resolve against the live repository. Treat them as identifiers into the mirror, not as links.

Branches

Branch	Full messages below	In the table
foil/analysis-source	35	27
foil/main	35	28
foil/mask-edgetrace	30	23
foil/mask-edgetrace-batch	31	23
foil/mask-learn	32	25
foil/mask-straighten	29	22
foil/mask-vector	35	27
foil/tooling-research	35	28

---

foil/analysis-source

Tip 65a683a8a4692f85fbaba041bbe00c0c5089260e · 62 commits · 2026-08-01 → 2026-08-09 · 10,533 words · 35 reproduced below.

f652fd1 — 2026-08-01 — foil/main: workbench v1 — three.js viewer, 5-pattern starter library, era layout masks

Quarantined foil workbench at /pokedex/foil-lab (URL-only, linked from nowhere, chrome-free): real owned-card scans from the image cache, tilt via pointer/gyro(iOS permission)/manual with reduced-motion respected, uniform sliders, pattern + mask-scope overrides, mask overlay toggle, copy-recipe JSON. Pattern library v1 tuned against owned eras (Base/WOTC, SV, Mega Evolution): Starlight, Cosmos, SV default holo, SV reverse sheet, Cracked Ice — uniform contract (core + uP0..3) designed for the full 15-20 set. Era layout spec as data (era-layouts.json, rects measured on real scans); resolver v1 (series, rarity, variant kind) -> pattern + mask scope. SKILL.md x2 (foil-effects, mask-pipeline stub). main.tsx gains a dev-only double-createRoot guard (Vite HMR entry-invalidation footgun, see DECISIONS).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

f98cf4d — 2026-08-01 — foil/main v2: Pencil mask editing, comment queue, iPad two-column, branch api

Apple-Pencil hand-mask drawing on the workbench (MaskEditor: card-aligned canvas overlay, pen+mouse default w/ finger toggle, pressure brush, eraser, 12-step undo, touch-action none, starts from layout prior); masks persist as committed ground-truth artifacts data/foil-masks//.png+json (derivation_method hand) via the env-gated branch api (POKEDEX_FOIL_LAB=1, port 3712, PGPOOL_MAX=1 — inert in prod), auto-load and beat the layout tier (uMaskTex contract; CanvasTexture flipY=false, exactly one V flip). Comment button captures card/variant/pattern/scope/mask state + all slider values to issues/foil// (report.md + context.json), out of the fix-issues sweep. Era-grouped owned-card picker (WOTC / SWSH / SV+ME headings); two-column layout (viewer | controls) from 700px — verified at 390px, 744x1133, 1133x744, 1440x900 with simulated pen strokes. SKILL.md x2 updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

77813a1 — 2026-08-02 — foil: sidecar v2 (prior+diff), artwork-keyed masks, comment linkage, codify ritual

• Every hand-mask save records its starting prior (era/scope/resolver rect + RESOLVER_VERSION) and the api renders .prior.png + .diff.png (green = human added, red = removed) + Jaccard agreement, server-side via a pure-JS PNG codec (no native addons). Saves without a parsable prior 400.
• Machamp mask backfilled (deterministic prior wotc/window/v1): agreement 0.6409 — his correction is the subject silhouette, the rule's first lesson.
• Artwork keying: variants of one cardId provably share one scan, so mask GETs alias to a sibling variant's mask when prior.scope matches (X-Foil-Mask- Alias-Of). Cross-card reprints unprovable from the catalog -> per-card only.
• Comments auto-link the exact saved mask state (maskFile/maskSavedAt/ maskAliasOf/maskHasPriorDiff).
• mask-pipeline SKILL: sidecar v2 storage, identity rule + limits, and the codify ritual; first worked example data/foil-masks/codified/wotc.md (n=1: WOTC window = art-window rect minus subject silhouette; noted in era-layouts.json wotc.notes).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

83b51bc — 2026-08-02 — foil: any-card picker — full-catalog browse/search, Owned-only as a toggle

Chey: pull up ANY card in the catalog and tilt it, owned or not. Picker now browses the whole catalog by era -> set (paged strip, promo sets run 300+), adds full-catalog name search, and keeps Owned-only as a filter toggle (default on = prior behavior). Zero new endpoints: existing /series, /series/:slug, /sets/:setId (paged), /search, /cards/:cardId serve it all; unowned cards list their catalog variants (quantity 0) and resolve the same base-guess pattern path. Selection is never clobbered by filters; unmapped series get an honest "Other eras" bucket. Verified on unowned neo2-13 Umbreon (Auto - Starlight WOTC) at 390px + both iPad viewports.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

6f617a9 — 2026-08-02 — foil: taxonomy integration — all 39 pattern types in the app, cited resolver

Mislabel corrections from the video-reference research, applied to apps/ (research docs untouched):

• sv-holo renamed vertical-sheen (it always rendered the HGSS-XY vertical bands, not SV's). PATTERN_ALIASES keeps old slugs resolving forever — sidecars, comment context.json, Copy-recipe JSON never orphan.
• Sheen family = ONE generator at four rotations + stripe option: five honest implemented slugs (vertical / horizontal=TRUE SV default "Mirage" / diagonal right "/" / diagonal left "" / striped "Line").
• Cosmos label drops "/ Galaxy" (Galaxy is a starlight synonym; synonym moved to the starlight entry). reverse-sheet description notes it ≈ pokeball-masterball stamped sheet.
• Vintage narrowed: starlight = Base/Jungle/Fossil only; Base Set 2 → Call of Legends holos = cosmos.

Pattern dropdown carries the FULL 39-slug vocabulary, grouped: implemented recipes first, then every gap type rendering via its nearest recipe and labeled "approx via ..." (amber note in the UI). Unimplemented != hidden.

Resolver v2: base guess reads the cited usage table (usage-index.json, derived from research/foil-pattern-usage.json by tools/foil/build-usage-index.mjs; set-name match -> era/series token match -> era-default heuristics). The winning row's confidence + citation hosts surface in the workbench ("Guess: set-level citation, high confidence (bulbapedia..., collexy...)"). Reverse rows citing 'mirror' are deprioritized per the research known-gap (ink-design evidence, not foil-pattern evidence). RESOLVER_VERSION -> 2 (scope semantics unchanged from v1).

Verified: typecheck + build green; Playwright at 390/744x1133/1133x744 (~/.pokedex-dev-hub-legacy/foil-shots/w2/) — Machamp auto=starlight set-level-high with his hand mask still resolving, SV holo auto=horizontal-sheen, SV reverse auto=energy-symbols-ii, fireworks approx labeling.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

69dd9fb — 2026-08-02 — foil: fix diffuse darkening — scan sampled sRGB-decoded but never re-encoded (issue ls9u0y)

The face texture was uploaded SRGB8_ALPHA8 (hardware sRGB->linear decode at sample time) while the foil ShaderMaterial writes gl_FragColor raw — three.js only appends the linear->sRGB colorspace_fragment chunk to built-in materials — so every scan rendered in linear values displayed as sRGB (measured flat 184->123 / 199->147, exactly the sRGB->linear curve). Sample the scan undecoded (NoColorSpace): the compositing model is authored in display space and pattern=none at rest is now pixel-comparable to the flat view (mean abs luma diff 44.55 -> 2.16, mapping identity +/-2). Also zero pattern-none's default specular so the baseline is truly the plain scan.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

eb22559 — 2026-08-02 — foil: workbench split — canon pattern lab (A) | card adjustment surface (B)

Chey's comment issues/foil/2026-08-02_12-59-52-368_4aq756 (resolved here):

• /pokedex/foil-lab/canon (CanonLab.tsx): bare holofoil pattern on a blank card (no ink; black/dark/silver/white flat-tone base via the normal texture path) beside the REAL reference tilt clip + 8 keyframes streamed from research/foil-video-reference/ by new env-gated dev routes (byte-range via res.sendFile for iOS Safari). Full 39-slug picker, sliders, tilt, and Save canon -> data/foil-canon/.json (full uniform snapshot that replaces recipe code defaults as the baseline).
• /pokedex/foil-lab (FoilLab.tsx reframed): per-card masks/comments unchanged, plus per-card uniform overrides saved as a SPARSE diff vs the canon baseline -> data/foil-overrides//.json (untouched uniforms keep tracking canon). Off-canon sliders get a dot; saving with sliders matching canon deletes the override.
• Layering model in foil/canon.ts (code defaults < canon < override < live); shared atoms + surface tabs in foil/ui.tsx; routes wired in main.tsx; AppShell chrome-free check widened to /foil-lab/*; both data dirs re-included in .gitignore; ports 5184/3713 recorded in ORCHESTRATION.md.

Verified in-browser at 390x844 / 744x1133 / 1133x744 (clip playing beside the bare cosmos render; Machamp base1-8 hand mask still resolves) — shots in ~/.pokedex-dev-hub-legacy/foil-shots/canon-lab/. Typecheck + build green (web, api). patterns.ts untouched (R0 re-tunes concurrently).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9b7bda4 — 2026-08-02 — foil R0: re-tune implemented recipes per Gemini verification notes

cosmos: full activation-model rewrite — dark field, smaller/denser orbs mostly near-invisible, cluster pops via low-freq noise over cell ids, pinprick 4-point twinkles (was: dense wall of large saturated orbs). starlight (+ starlight-ii via shared GLSL): parallax architecture untouched; tighter visibility lobe (pow 5->9, floor .18->.08), sharper glyph cores + narrower flares, smaller blobs, saturated star color, galaxy wash halved; starlight-ii art gate lowered for the mid-orange Evolutions field. cracked-ice: authored intra-shard grain removed per Chey's accuracy ruling — facets flash as solid smooth planes; whitening at peak reduced, uSat up. sheen family: generator grows sharp/beam/barcode options. vertical-sheen -> new barcode variant (thin spectral lines of varying width, CD-groove slide) with tamed beam + art gate. diagonals: sharp 3.0, beam 0.55, specular 0.35 (center blow-out), diagonal-sheen-right uP0 2->7 (same-sheet consistency). striped-vertical-sheen: stripes finer + more blended (90->130, floor .40). horizontal-sheen: untouched (20/20).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

1c46a0b — 2026-08-02 — foil R0: rounds 2-3 re-tunes + verification report, field notes, decisions

starlight: longer thin-armed 4-point glyphs + 8-point-burst subset, tighter pop lobe (pow 11, floor .04), full-saturation star color, parallax default 2.4, gain 3.6 — best of 3 rounds 11/20 (still-frame judging cannot see the parallax; recorded as honest residual with a video-judge recommendation). starlight-ii: 20/20 (uP2 pinned so base-default drift can't invalidate it). cracked-ice: facet gain 1.1 -> 0.55 (art visible through flashes), density 10, less whitening — 15/20 yay; shard-geometry uniformity queued for R1. Verification report gains the R0 section (before->after per pattern, the diag-left judge-noise event, diffuse-fix supersession note); foil-effects SKILL.md gains the R0 field notes; DECISIONS entry for root cause + wave.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

c573622 — 2026-08-02 — foil: cited per-card pattern assignments across the whole catalog (resolver v3)

Chey's mlmwmp comment: the library was funneling to era defaults ("cosmos"). Full-catalog enumeration (20,964 cards / 37,627 variants, via the branch read api) + an 18-lane Ringer research swarm (foil-card-assignments, GLM-5.2, executed selector-integrity checks) produced:

• research/foil-card-assignments.json — 111 set/rarity/kind/card-level rows, 19 catalog foil-facet mappings (holo-foil-cosmos, reverse-foil-masterball, normal-foil-galaxy...), 55 honest residuals; per-lane provenance + method.
• tools/foil/build-assignments-index.mjs -> apps/web/src/foil/assignments-index.json (trimmed for the bundle, duplicate-facet guard).
• resolver v3 (RESOLVER_VERSION=3): assignment tier above the v2 usage table; specificity cardIds > variantKinds > facet > rarities > set > v2 > heuristic; normal-kind variants with a declared foil facet now render window foil (previously scope none); FoilLab passes cardId.
• era-layouts: wotc seriesSlugs 'ecard' -> 'e-card', + 'legendary-collection' (both had fallen through to modern-sv rects and heuristics).

Coverage (foil-bearing variants): 79.5% set-cited / 20.4% era-token / 0.2% heuristic -> 93.4% set / 4.3% facet-or-card / 2.4% era-token / 0 heuristic. 13 in-browser spot-checks in ~/.pokedex-dev-hub-legacy/foil-shots/assignments/. Issue mlmwmp resolved. Weak spots + curation ledger in DECISIONS.md.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

e55586e — 2026-08-02 — foil R1: dedicated recipes for the twelve owned-era gap patterns (10/12 verified)

Chey's comment dml369 ("develop out the rest of the missing holofoil patterns"): dedicated GLSL for fireworks, ace-spec, energy-symbols-ii, rainbow-glitter-sheen, ex-starfoil, prismatic-pokeball, tinsel-ii, cosmos-iii-smooth, pokeball-masterball, radiant, rainbow-glitter, confetti — implemented flags flipped, approx labels dropped, slugs stable. Gemini-verified 10/12 match over 3 rounds (verdict table + honest residuals in research/foil-verification.md R1 section). Honest nays: tinsel-ii (procedural static plateaus at judge score 2) and prismatic-pokeball (screen-only blend cannot darken a white reverse body — structural, Chey's call).

R0 residuals cleared in the same pass: cracked-ice anisotropic shattered-glass metric (re-judged yay 14/20); starlight milky wash + per-layer visibility curves with starlight-ii re-banked clean 20/20 through both shared-GLSL changes; starlight parallax judged from a new 16-frame fine sweep and still scored flat — verified real by eye, recorded as judge-blindness.

Field notes appended to foil-effects SKILL.md; dated DECISIONS entry; issue dml369 resolved.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

4ade910 — 2026-08-02 — foil R2: uDarken blend-model term — prismatic-pokeball 17/20 yay, tinsel-ii 16/20 yay

One opt-in substrate-darkening term in the shared composite (main()): body = scan * (1 - uDarken * mask * gate) before the foil screen-blends. Default 0 is bit-identical to the old render; absent canon/override keys seed 0. Full 21-pattern regression sweep on provably identical renders: no true regressions (and a measurement of Gemini single-roll noise, ±3-6; diag-right + pokeball-masterball double-nays recorded as unstable verdicts on geometry proof). prismatic-pokeball rebuilt on the term (dark-mirror base, facet-quantized flash lobe, ball watermark as ink-overprint suppression): structural nay -> 17/20 yay. tinsel-ii one-line opt-in (uDarken 0.4): static plateau -> 16/20 yay. 20 of 21 implemented types now match; sole nay starlight (judge parallax blindness).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

8c275b8 — 2026-08-02 — foil R2 stage 3: thirteen dedicated recipes (gap patterns + RC dots) + subset-commons resolver tier

• patterns.ts: real recipes for mirror, rainbow-mirror, vertical-sheen-rainbow (dark-mirror family on uDarken), pokeball-hologram, prism, water-web, energy-symbols, pinwheel, ex-emerald, tinsel, cosmos-ii-pixel, crosshatch, radiant-collection-dots — implemented: true, approx labels dropped, slugs stable. Unused plain SHEEN_V generator removed (last fallback users gone).
• resolver v4: card-level 'normal'-class assignment rows consulted before the scope-none early return — subset cards the catalog declares plain but which physically carry foil (RC commons dot overprint) resolve their subset's pattern with scope full.
• foil-card-assignments.json: two card-level rows (bw11-RC1..20, g1 25 non-ultra RC ids) with video citation (non-holo RC Teddiursa keeps the dot overprint); builder enforces cardIds on 'normal' rows; index regenerated.

Verified in-browser: g1-RC1 + bw11-RC1 render the RC dot treatment via Auto (guess card|medium). Gemini verdicts in research/foil-verification.md (R2 recipe-wave section, next commit).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

c23e4a6 — 2026-08-02 — foil R2b: four §40–43 vocabulary recipes (all match) + resolver v5 scope overrides

The last lane of the wave — dedicated recipes for the vocab-extension types: gold-secret (warm-locked goldRamp field + chromatic glitter pops + SWSH burst rays — 19/20 yay), vstar-pearl (near-white pearl on uDarken 0.3, warm-led diagonal wash, frame-line streaks — 13/20 yay), shiny-vault (silvery field + band-keyed sparkle-glyph amplifiers, uDarken 0.15 — 11 nay → 16/20 yay r2), detective-pikachu (diagonal sheen × photo-luminance via the documented uFace contract exception — 20/20 yay). 38/43 taxonomy types real, 34 hold match.

Resolver v5: assignment rows carry an optional per-row scope override (window/ full/sheet), honored by the set-row tier and the cls-'normal' tier — baby shinies render window despite the 'Shiny rare' full-foil rarity, VSTARs full despite plain holo kinds, det1 window despite 'ultra rare'.

Assignment research: +2 facet rows (gold, gold-jumbo → gold-secret; retires both facet residuals, 1-card metal collateral flagged) and +10 cited rows — sma 51 baby/35 GX/8 gold splits, swsh4.5sv 104 baby + 18 full-face (incl. the black shiny Eternatus pair, name-verified not gold), sv04.5 rarity split, VSTAR via the dedicated 'Holo Rare VSTAR' rarity, det1 ultra-rare + cls-normal rows (all 18 cards window). Residuals annotated with resolutions.

build-usage-index.mjs: tolerate era-wide rows with no scope.sets (the vocab gold rows crashed the build — the 122-row usage index never compiled until now). Verdicts/frames under ~/.pokedex-dev-hub-legacy/foil-verify/ (verify-

-r2b*); shots in foil-shots/r2b/. Docs: foil-verification.md R2b section, SKILL R2b field notes, DECISIONS entry.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

37b6105 — 2026-08-03 — foil R3: sheen-family streak-field rework — all 8 of Chey's canon-lab comments resolved, 7/7 Gemini 20/20

sheenGlsl is now a streak field (Chey's critique = the spec): sparse irregular streaks, per-streak lean following card tilt (crisscross; converging pairs terminate where they meet), stretched-ellipse taper, hue running ALONG each strip as well as across. striped-vertical-sheen rebuilt from a fresh Gemini corpus re-spec run with Chey's description embedded (grouped reveal via wide sweeping windows over a subtle bottom-converging fan; gemini-spec-r3.md); ex-starfoil rebased on the reworked diagonal base + fine CD-line layer; vstar-pearl rebuilt from the horizontal-sheen streak field, pearl kept.

Canon migrations (full-snapshot rule, DECISIONS): striped + horizontal uDarken 0 -> 0.32 (bright-substrate legibility, third confirmation). All other canon values carry over; uniform semantics unchanged. validate_spec.py frame regex single-digit fix (undercounted two-digit citations, failed an honest worker).

Verdicts (Chey's words folded into every judge prompt as acceptance criteria): vertical/right/left/vstar-pearl 20/20 first roll; striped + ex-starfoil 20/20 round 2; horizontal 20/20 round 3. Table + residuals in research/foil-verification.md R3.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

85a3b68 — 2026-08-03 — foil R3-MOTION: axis-split starlight, per-dot cosmos, hologram radiant, the rainbow-glitter-sheen delta

Chey's six motion-model canon-lab comments resolved (his eye = ground truth):

• starlight (+ starlight-ii via shared GLSL): top-to-bottom hue BANDING (soft-quantized bands; his canon 0.62/0.6 lands the reference order blue->green->red on the R->B->G cosine ramp) + AXIS-SPLIT tilt: vertical slides the whole field (~10% card height), horizontal drives the per-star random fade; his 3-layer uP1 parallax rides on top untouched. Judge: starlight 17/20 yay round 2 — the FIRST starlight yay (standing parallax nay broken); starlight-ii 20/20.
• cosmos: whole-field slide + per-dot random response (independent random tilt axes for brightness and hue per orb — no shared sweep axis); at tilt 0 his canon appearance is pixel-identical. Judge: 20/20.
• radiant: discrete hologram steps (half-cell up/down, 60% hold / 40% crossfade — one position fades out as the next fades in), uP1 wired as Hologram travel (canon migrated 0 -> 2.2; was an unused placeholder), grid lines ~45% thicker. Judge nay 13/20 across 3 rounds on a pixel-refuted "slides continuously" claim — recorded as still-frame motion blindness (same class as starlight parallax); adjacent-frame crops prove discrete positions + crossfade. Chey's live tilt is the tiebreak.
• rainbow-glitter-sheen: dedicated Gemini delta-articulation pass first (he couldn't name the difference), pixel-verified, then fixed: narrow laser-saturated striped chevron (sigma 0.010->0.0035, hue traversal 5x->9x, pow-deepened at low gain), fainter +-0.8 repeats, chevron angle 1.3->1.9, denser colored always-visible glitter, uDarken 0->0.4 (4th legibility-physics data point). Judge: 19/20 yay.

Docs: research/foil-verification.md R3-MOTION section, DECISIONS entry, foil-effects SKILL field notes. Frames/verdicts in ~/.pokedex-dev-hub-legacy/foil-verify (jobs -r3m, manifest-r3m{,2,3}, delta-rainbow-glitter-sheen); gate shots in foil-shots/r3-motion.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

34e6108 — 2026-08-03 — foil mask-refine: window handles → flatten → hand refine

Chey's requested workflow on the Card-adjust surface: draggable corner/edge handles adjust the layout window rect per card (WindowEditor.tsx, 44px touch targets, finger-drag allowed), the adjusted geometry persists as artwork-keyed scope-agnostic sidecars (data/foil-windows//.json, /foil-lab/windows routes), and Flatten bakes it through the STANDARD hand-mask save path (sidecar prior stays the era rule; new optional prior.window records the adjustment) then opens the existing paint tooling.

• rasterizeWindowRect shared by Flatten + loadLayoutRect (identical bakes)
• saveMask prior = layoutMask (rule), never the effective mask
• windows GET aliases across variants of the same scan, newest savedAt
• verified in-browser desktop + 390px: sma-SV1/sma-SV2 (SM baby shiny, borrowed modern-sv rect) and me01-034 (zero-coverage gap); base1-8 + base1-5 hand-mask regression clean; test artifacts deleted (corpus rule)
• ports: vite 5186 / api 3714 recorded in ORCHESTRATION.md

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

f1abc0c — 2026-08-03 — foil R3-GLYPH: glyph-pattern reworks per Chey's 7 comments + the drop-in glyph slot for his SVGs

Behavior (his notes verbatim = acceptance criteria): reverse-sheet rainbow only on grainy glyphs over a neutral sheet; energy-symbols checkerboard banks that swap with tilt; energy-symbols-ii sporadic random banks (~half faint at any time); ace-spec diamonds grow/shrink with tilt + a touch of blur; prismatic-pokeball ball never darkens (white-mixed pale response + coherent plane flash, suppression deleted); radiant-collection-dots shapes catch a traveling rainbow band (16/20 YAY — standing R2 nay broken).

Glyph slot: research/foil-glyphs//glyph[-N].svg -> lab-gated api routes -> atlas rasterizer (foil/glyphs.ts) -> uGlyphTex contract + glyphTex() helper; ~2.5s auto-pickup on save, procedural fallback when absent, README tells Chey exactly where his four promised files go.

Canon migration: reverse-sheet uP1 0 -> 0.6 (dead placeholder now drives his requested glyph grain). Verdicts + pixel-refutations recorded in research/foil-verification.md (R3-GLYPH); lessons in foil-effects SKILL.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

fc6e1bb — 2026-08-03 — foil R3-MISC: the uTint ink-metallic blend term + Chey's twelve remaining canon-lab comments (12/12 yay)

The reverse-holo fix (his chat report): screen-blending achromatic foil light over colored art compresses chroma — "dull and grayish, rather than making the color look metallic." New core uniform uTint (default 0 = bit-exact legacy) multiplies the foil layer + in-mask specular by the luminance-normalized scan chroma SQUARED (the double ink pass of a real mirror reflection); neutral on blank cards, so no canon-lab appearance moved. Reverse-family recipes opt in; verified before/after at 5 tilt angles on Victini ball-reverse, Crystal Energy, Pineco. uP4/uP5 param slots added in the same contract change.

The twelve comments, each mechanism-mapped in its issue frontmatter and the verification doc: cracked-ice triangulated + half-off gate; fireworks grid lattice + center-emanating radial hue; rainbow-mirror rebuilt as mirror + hue-banded spotlight; tinsel-ii silver lines + vertical rainbow band (band lives only on the line-work); prism rebuilt as pinwheel-kin from the corpus-vs-corpus delta pass; cosmos-ii pixel-quantized stamp edges; crosshatch thinned; gold-secret adjustable burst origin (uP4/uP5, center default, per-card via overrides) + holographic grain; dedicated recipes for sequin (glyph-family glints), tcg-classic (flat starlight + rainbow glitter), acid-wash (water-web kin), disco (perfect disc lattice, galaxy ignition) — zero approxVia fallbacks remain.

Canon migrations (appearance-preserving unless he asked otherwise): fireworks uP0 3→4.5 (dead two-octave structure), gold-secret uP4/uP5 0.5 added, sequin uP0-uP3 re-keyed to the new recipe. Verdicts: research/foil-verification.md "R3-MISC" — 12/12 Gemini yay incl. the tcg-classic ink-scope re-roll.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9849952 — 2026-08-03 — foil R4-COMPOSITE: the ink-density invariant — foil adds pop, never washes text or mutes ink

Chey's two me05 comments (7rtnzx mirror reverse "blows out the darks/text", 19mo4l holo "darkens/muddies the colors") ruled a CORE COMPOSITE INVARIANT, engineered in shader.ts main(): ink density estimated from the scan (inkDark = darker-than-local-8-tap-field average, exactly 0 on flat blank bases; inkColor = chroma above the canon-lab-tone floor) now gates every layer — dark ink blocks the flash + shields the specular (text crisp at every tilt), ALL ink is exempt from uDarken substrate attenuation (printed color never muted; the mirror stays dark only where the mirror is visible), colored ink auto-tints its flash (max(uTint, inkColor)) and uInkPop pumps chroma under the flash so colors read MORE saturated and metallic.

New core uniforms uInkGuard (default 1) / uInkPop (default 0.5), both 0 = bit-exact legacy composite; sliders in both labs; canon files untouched (absent keys inherit the new defaults — catalog-wide fix, zero migrations).

Proof: canon-lab blank-card pairs (mirror dark+silver, cosmos, horizontal-sheen) pixel-identical (compare -metric AE 0) under a frame-stepped rAF harness with a same-settings AE-0 control, plus a Gemini sanity pass (all pairs "identical"). On-card before/after pairs across the 8-card sample (both issue cards, Victini, Pineco, Darkrai, Cetitan, Machamp WOTC, Hydreigon SIR; 3 tilts, 390px + desktop) in ~/.pokedex-dev-hub-legacy/foil-shots/r4-composite/. Both issues resolved with mechanism-mapped notes; composite contract documented in the foil-effects SKILL; DECISIONS + research/foil-verification.md updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

c4a0fef — 2026-08-04 — foil R4b: scan-additive composite (rest parity by construction) + canon-lab card preview

Chey's Grubbin (me05-002, Pitch Black) review refuted R4's ink heuristics: the reverse still muddied the card at rest (measured ΔC +32, green→yellow), darkened the header under tilt (−22 L), and blew out text (glyph contrast −26%). New law on real scans (uScanBase 1): the scan is a photograph of the card at rest — foil is additive dynamic light only, tilt-gated to ZERO at rest, luminance-headroom + per-channel clamped (text sacred at every angle, uArtGate opens the clamp where dark scan pixels ARE the foil), chroma- preserving with a pastel-safe tint ramp and an L²-gated chroma pump. Canon lab blank bases run the classic composite textually unchanged: AE-0 proven (CDP frame-stepped harness, control + knob pairs; Gemini pass agreed). uInkGuard 0 still reproduces the pre-R4 composite exactly.

Canon-lab card preview (his request verbatim): blank/on-card/↻-another chips render the live sliders on a random catalog card the resolver assigns the pattern to — server-sampled from the baked inversion (tools/foil/ build-pattern-cards.mts → data/foil-pattern-cards.json, gitignored; GET /foil-lab/pattern-cards/:patternId), honoring hand masks + adjusted windows; empty-pool patterns say "no catalog cards". Also commits Chey's live mirror-canon tweak from the Grubbin session. Full metrics + verdicts: DECISIONS 2026-08-04.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

c32c7ac — 2026-08-04 — foil R4c: onset ease — the sheen leans in instead of switching on

Chey on R4b's rest gate: "it's like it isn't there at all until I've tilted the card a tiny bit and then it's like it just suddenly appears." Measured: the headroom clamp saturated R4b's smoothstep(.02,.28) gate, so ALL added light arrived inside tilt 0.03-0.13 (0 -> 16.1 dL mean) then froze flat to tilt 1.0 — a step smeared over 4% of the tilt range.

Now: one wide ease ramp = smoothstep(0, .45, |tilt|) (full at 0.45 — every R4b tilt-0.5 metric unchanged), flash gated by ramp^1.5 (cubic-slow start), specular by ramp^2.5 (the broad gloss band now trails the flash instead of leading it), plus a REST=0.006 sub-JND living rest sheen blended into both gates (rest +0.43 dL card-mean, max 2/255 — visually still the scan). After-curve: smooth S over 0->0.25, no dead zone, no knee; text clamp untouched (glyph contrast 33.6 -> 33.6 at flash peak); blank-card canon lab bit-level unchanged (fuzz-1% AE 0 vs control floor).

Harness + curves + filmstrips: ~/.pokedex-dev-hub-legacy/foil-shots/r4c/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ca7dbf1 — 2026-08-04 — foil R4d: motion-first onset — tiny gestures move the sheen; glow arrives late; Chey gets the dials

Chey rejected R4c too ('It still lights up pretty noticeably when I tilt the card a tiny bit. I don't like it.') — measured the gesture→tilt mapping first this time: pointer is 0.0051 tilt/px at 390px and gyro is 0.036 tilt/deg, so a 30px thumb drag (tilt 0.154) was already getting HALF the full glow under R4c. The curve was never the felt variable.

• shader.ts: onset gates now uniform-driven — uOnsetRange (glow full at |tilt|, default 0.5) and uOnsetCurve (flash exponent, default 3.5; specular trails at +1). (0.45, 1.5) reproduces R4c exactly (verified to the third decimal); defaults are strictly softer at every tilt and equal from 0.5 up, so all R4b/R4c peak metrics are unchanged (Grubbin 17.976 ΔL / text 34.02, Machamp 18.035 / text 45.6 — exact).
• FoilLab + CanonLab: 'Onset range' / 'Onset curve' sliders, canon- stored like every other uniform; absent canon keys inherit defaults.
• patterns.ts: CoreDefaults admits the two keys per recipe.

Pointer-path proof (real pointermove, 390px, 30px drag, first 300ms): before mean ΔL climbs to 6.6 and rising; after stays within ±0.08 while per-pixel movement reaches 6.1 mean / 18.9 p95 — moves, doesn't light up. Blank canon lab at fuzz-1% AE 0 vs control. Artifacts in ~/.pokedex-dev-hub-legacy/foil-shots/r4d/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

a1526c6 — 2026-08-05 — foil R5: the metallic composite — always on the card, energy redistributed, Chey's compositing dials

Chey's ruling (2026-08-05, verbatim in DECISIONS): no onset range at all — the effect is ALWAYS on the card; dial in how it composites onto the ink; metallic and spectral, not brightening.

• Onset machinery removed (R4c/R4d ramp, REST floor, uOnsetRange/uOnsetCurve sliders on both surfaces). No canon/override file ever stored the keys; CardViewer skips unknown uniforms, so stale keys anywhere are inert.
• R5 metalness law (scan path only): scan = albedo; the pattern field splits around PIVOT 0.40 — above = tinted-specular highlight, below = bounded multiplicative depth (mirror-turned-away darks, sqrt-concave). Energy redistributes; contrast + chroma carry the metallic read.
• Soft-knee headroom: the R4b hard clamp plateaued strong fields into a flat wash (the "brightened" read); the budget now compresses monotonically so pattern structure survives. Per-channel no-clip caps stay hard; glyph ink (inkDark) exempt from highlight AND depth — text contrast preserved (Grubbin 86->97 at peak, Machamp flat).
• New canon-stored dials on both surfaces: Metallic 0.6, Sheen strength 0.55, Sheen tint 0.5 (= the R4 chroma law exactly), Depth 0.55, Texture 1. uMetal 0 / uIntensity 0 / uInkGuard 0 all still yield the plain scan.
• Blank canon lab pixel-unchanged: CDP frame-stepped before/after AE 0 at fuzz 1% (mirror/cosmos/horizontal-sheen/starlight, tilted + rest); one Gemini vision pass agrees (all pairs identical). Shots + per-card rest pairs: ~/.pokedex-dev-hub-legacy/foil-shots/r5/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

8b45995 — 2026-08-07 — foil R5b: metalness is MIRROR ONLY — every other pattern gets its own colour back

Chey: "You applied the metallic treatment to EVERY HOLO PATTERN when I only wanted it for mirror. So now, every single holo pattern looks like mirror and you can barely see the pattern at all, it's like every single one is being applied in a way that it adds no rainbow color or anything to the card."

R5 shipped its metalness dials as GLOBAL_DEFAULTS, so all 43 recipes inherited a law built for one of them. uMetal is now the scan-path LAW SELECTOR and defaults to 0 — nothing can inherit metalness; mirror is the only recipe that opts in, in its own patterns.ts defaults.

• Metalness branch is the R5 code textually unmoved, so mirror's arithmetic is unchanged by construction. Proven: his hand-tuned canon on Grubbin me05-002
  • Tropius me05-001, rest + tilt 0.2/0.35/0.5, 390px and desktop — 16/16 frames AE 0 at fuzz 1%, same floor as the same-build control pair.
• uMetal 0 runs the ADDITIVE law: the pattern's own emission (uSheen gain) screen-combined per channel over a substrate-attenuated body, tinted toward the ink only as far as uSheenTint asks. Three fixes to what R5 broke — the scalar no-clip cap let the artwork's brightest channel throttle the whole flash; the quartic ink budget (shaped to starve mid-dark glyphs) starved ordinary mid-tone art instead; PIVOT ate the field before anything flashed. Cosmos measured 0.005% of art-window pixels changed >8% under R5.
• uDepth decoupled from uDarken: uDarken is the blank-canon substrate, frozen by 30 saved canon files, and 0 on exactly the pale reverse sheets that need one. The substrate follows the pattern's dark half so it buys contrast, not a uniform dim (a flat 0.34 cost mean dL -21.7 and 29% of text contrast).
• Per-recipe families in patterns.ts (PARTICLE / SHEET / WASH / PEARL) plus seven by-eye corrections, each judged on a card the resolver assigns.
• Onset machinery stays deleted. Text guarantee re-verified: glyphs matte and crisp at peak tilt on a reverse whose foil covers the text box.
• Blank-card canon appearance unchanged: 8 patterns AE 0 (6 direct; cracked-ice and tinsel-ii against the pre-R5b code with canon held constant, since Chey re-saved those two mid-wave).
• SKILL.md composite contract rewritten around per-recipe selection.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

e25bc24 — 2026-08-07 — foil R6: composite defaults re-derived from Chey's canons; the guard learns to tell text from artwork

Chey: "i re-adjusted cracked-ice. take a look at it, and then make the changes based on everything you're observing to make all of this better - i want it so that i have to do as little of work personally on all the rest as possible."

WHAT HIS CANONS TAUGHT

uSheenTint 0 on all three non-mirror canons -> 0 for every non-mirror family (was 0.15/0.5/0.85/0.6). Ink-tinting a pattern's own highlight is exactly "it adds no rainbow color or anything to the card". uSheen he raised all three -> families now 1.2-4.2 (were 1.0-2.6). uDepth NOT universal: 0 on cracked-ice, 1 on tinsel-ii and rgs. It darkens the card WHERE THE PATTERN IS NOT, so its cost scales with 1 - duty. Measured every recipe's duty cycle; his three choices line up with it, and it predicted both dials for the only canon that exercises the defaults. uSat 1 + wide spread -> shipped on the 13 recipes with no canon.

Families are renamed for how light LANDS (flash / line / stamp / field / pearl / metal) and stated explicitly as FoilPattern.family. The stamp tier exists because the first pass got it wrong: energy-symbols-ii lights 5.8% of the face, so a field substrate darkened the other 94% and the card read "someone dimmed this".

THE INK GUARD: SPLIT, DON'T WEAKEN

inkDark is a local-contrast measure, so it fired on every dark mark - including every outline and shading edge in the ILLUSTRATION (measured: 86-89% of strong hits inside the art window are artwork, not text). The additive law spends it as a hard coverage multiply, so guard 1 punched the pattern full of holes. That is why he kept pulling it to ~0.51/0.56 on sheet/wash patterns.

Now split by glyphness into inkGlyph (sacred, zero flash, unchanged) and inkDetail (artwork darks keep the pattern on a tightened budget). inkAny is byte-for-byte the old inkDark and every pre-R6 consumer still reads it.

Success test met: at guard 1 under R6, tinsel-ii and rgs look BETTER than his 0.51/0.56 do today - tinsel-ii's artwork is visible through the streaks instead of greyed out. Both canons migrated to 1, reason recorded in each file's note; every other value in them is his, untouched.

SLIDER RANGES (his second ask)

Four dials sat exactly on their cap: uIntensity 2->4, uSheen 3->6, uSat 1->2, uHueSpread 1.5->3. Not extended: uInkGuard (his carve-out), uMetal (a law selector), the 0..1 mix weights, and uDarken/uDepth (already fully spent at 1).

uSheen needed a shader change to be honest: the additive budget's ceiling was reached at uSheen 2.0, so the top third of the slider bought almost nothing - very likely why he parked two canons at the ceiling. It now resumes climbing above 3, leaving every stored value <= 3 bit-for-bit identical.

The 19 duplicated lines across the two surfaces are one CORE_SLIDERS table; uGrain is dimmed and labelled "metal law only" where it does nothing.

NEW: "Apply composite -> family" copies only the on-card dials to siblings, never pattern shape.

PROVEN, NOT ASSUMED (AE vs a base build compiled from the previous commit)

• blank canon room: AE 0 across all 44 implemented recipes x 4 tones x 3 tilts
• mirror: AE 0 on the card path too - pixel-frozen
• text at flash peak: +0.1 to -1.1 points of glyph/paper contrast on every reverse whose foil covers the text box; the library's two largest costs (rainbow-glitter, water-web) toned down per-recipe and re-checked by eye
• every implemented recipe eyeballed on 1-2 real assigned cards at rest and two tilts; shots in ~/.pokedex-dev-hub-legacy/foil-shots/r6/

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

68653c4 — 2026-08-07 — Mask provenance: sidecar v3, the AI learning loop, and the collapse safeguard

Chey (2026-08-07): mark AI-generated masks, let an AI learn from his hand-done ones, then let him correct the AI's mask so it "can observe the diff and continue to improve without me having to hand-paint all the masks."

derivation_method existed in v2 but the API HARDCODED it to 'hand' — harmless while only a Pencil touched the corpus, actively wrong once mask-refine's Flatten started writing machine geometry through the same save path.

Taxonomy (5 values, 4 cases that must never blur): layout-flatten machine geometry, unpainted weight 0 hand human, from a layout prior weight 1 hand-refined human, on an existing mask weight 1 ai generator, UNREVIEWED weight 0 ai-corrected AI proposed, human edited weight 0.6

The label is never taken from the client. The client reports only what the canvas was SEEDED with; writeMaskRecord derives the method by diffing saved pixels against what that seed rasterizes to. A machine label needs a full GeneratorIdentity, which HTTP callers cannot supply. One write path serves the route and generators both.

• Anti-feedback-collapse enforced in selectExemplars(), not by convention: unreviewed ai is weight 0 at any corpus size. Codify now routes through it.
• Corrections keep the parent PNG + change map + metrics (incl. a 4x4 grid of where corrections land) — the supervised pairs a generator lane consumes via the self-describing tuples manifest.
• Seam rule, measured not guessed: canvas roundRect vs the server SDF disagree on 389/330260 px, ALL in the 1px AA band. Without tolerating it every unpainted Flatten would stamp hand.
• Guard: if the file at the target path is an unreviewed ai mask it IS the parent, whatever the client claims — AI ancestry cannot be laundered. That guard caught a real bug during this build (ActionBtn passed React's MouseEvent into saveMask's optional arg, blanking derivation); derivation is now required on the PUT so it fails loudly instead.
• v1/v2 load forever (normalize-on-read); Chey's two masks migrated additively and still read HAND-PAINTED, agreement 0.641, pixels untouched.
• Trial batch wotc-window-trial-1: 6 WOTC Base holo rares from window-artgate@1, all ai/unreviewed, reversible. Leave-one-out on the human corpus beat the rect baseline (0.588 -> 0.688) — real, but n=2 same-set, so explicitly not validation.

Surfaces: provenance badge + detail in the workbench, Mask corpus panel, GET /foil-lab/masks/corpus (+?tuples/?exemplars), artifact routes, and corpus.ts report|exemplars|tuples|migrate. Pure tests added to CI (test:foil). Browser-verified at 390px and desktop. See DECISIONS.md 2026-08-07 and the mask-pipeline SKILL.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b915cea — 2026-08-08 — foil mask-straighten: read the INTENT out of Chey's hand mask, and make the undo real

Chey, 2026-08-08: "it's impossible to get the lines really straight so I'm hoping you can get computer vision on the mask and card art in tandem to really see my intent there, and make the mask nice and crisp and straight on the lines I was trying to draw along. Will you try that? And then I can either approve or revert."

Smoothing a wobbly boundary gives a smoothly wobbly boundary. The premise here is that a hand-drawn foil boundary is an ATTEMPT TO TRACE A PRINTED EDGE — frame, art box, species strip, stage tag — and those are dead straight in the scan. So: contour the mask, cut each loop into near-axis runs (PCA orientation, wobble gaps bridged), robust-fit each, hunt the scan's directional gradient with a local Hough over (angle, offset) plus a proximity prior, replace, intersect adjacent lines into real corners, rasterize matching his own AA character.

The refusals are the feature, and each is a recorded param: weak evidence, or an ambiguous band of comparable ridges, may nudge but never relocate; with no artwork edge it straightens only to HIS own fit; short and curved runs pass through untouched. A first cut "closed" the Tropius top-left corner — which turned out to be him deliberately tracing around the BASIC stage tag. The stricter rules keep it.

• apps/api/src/foil/line-snap.ts the geometry, pure, no I/O
• generator.ts line-snap@1, a REFINER (requiresSource)
• generate-masks.ts run --refine, archives, restoring revert
• provenance.ts supersedes + verbatim sha256 undo archive

A generator write onto an existing mask now THROWS without explicit supersede { runId }; with it, every artifact the mask had is archived verbatim under superseded/./ with a self-describing manifest, and revert --run-id verifies the whole archive BEFORE deleting anything live, then restores byte-for-byte. supersedes is deliberately NOT correction: one is a human editing a machine (training signal), the other a machine replacing a human with nobody's agreement. --refine also refuses a source with exemplar weight 0, so a refiner can never eat its own unreviewed output.

Tropius me05-001/37184 (run straighten-tropius-1): 8 runs snapped to printed edges, 1 self-straightened, 13 left exactly as drawn, 8 corners closed, 2 specks dropped. Max correction 3.74px, 1.06% of the face changed, Jaccard 0.9798, confidence 0.557. Honest residual: the card's bottom border was too faint to accept, so that run self-straightened and the report says so.

test:foil 25/25 (9 geometry + 5 supersede/restore, incl. byte-for-byte round trip, a deliberate curve surviving, an ambiguous band refused, and a corrupt archive aborting before deletion). Revert exercised twice end to end, sha256sum clean. Shots: ~/.pokedex-dev-hub-legacy/foil-shots/mask-straighten/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id straighten-tropius-1

d3f95e3 — 2026-08-08 — foil R7: his eye overturns two banked mechanisms; "no catalog cards" gets a reason

Chey's 15 canon-lab comments (2026-08-07), all resolved in issues/foil/.

Eleven shader verdicts. Two overturn mechanisms built from earlier corpus — energy-symbols' parity CHECKERBOARD (R3-GLYPH, from his own words) and prism's solid-facet grid (R3-MISC, from a Gemini delta pass). Having seen both rendered he rejected them. A description is a hypothesis about the render, not a spec.

cosmos the 0.085 UV field slide he read as parallax is now uP4 "Surface drift" at 0.012; zero at rest, canon unchanged striped-v-sheen grey lines: the unconditional silver term is gone AND the second-order lobe cut 0.4 -> 0.16 (it drew a whole second, near-achromatic line population on the far side) rainbow-mirror hue axis is card X, blue -> green -> red, on BOTH the sheen and the spotlight; core whitening 0.85 -> 0.32 energy-symbols checkerboard gone: per-symbol response AXIS + PHASE + transition WIDTH over a low-freq cluster field energy-symbols-ii same treatment (it shared the two-bank mechanism) pokeball-hologram flat SDF icon -> shaded 3D sphere (hemisphere normal, tilt-following light, belt, hub, two-tone shell) tinsel the "blinds" were a literal full-card striation floor; demoted to uP2, default 0 (his canon stores 0) tinsel-ii line-work gated to the colour band, silver 0.60 -> 0.10, band width exposed as uP4 prism pinwheel's spoke fan WITHOUT the per-wedge phase that makes a pinwheel spin; density 33 -> 18 radiant hologram step 0.5 cell -> uP4 0.17, drive rescaled so total travel per unit tilt is unchanged rainbow-glitter invisible on card: bigger white-lifted flakes AND the family fixed (flash -> field; no substrate, no contrast)

Canon policy held: shader.ts / canon.ts / CardViewer.tsx / data/foil-canon have a ZERO diff, sheenGlsl() is byte-identical (all four sheen canons frozen), and every new behaviour went into a NEW uniform old canon files inherit. No stored value migrated. One recommendation left for him, not applied: tinsel-ii's starkness is substantially his own canon (uDepth 1 at ~135 lines).

"Why are there no catalog cards?" — asked four times, three answers. The baked inversion is now v2 and emits a machine-readable diagnosis the lab renders: outranked diagonal-sheen-left — 1,818 SM reverses ARE cited for it, but energy-symbols wins them (the emblem layer vs the sheet rotation; both true). A capped SECONDARY pool is now baked via a new citedFoilPatterns() export and served as via:'cited', so the preview works. Never flip a winner to fill a pool. class-absent pinwheel, pokeball-hologram — the catalog has reverse variants for ex1-ex5 ONLY; the whole late-EX reverse era is missing upstream. The UI now says so instead of "no catalog cards". fixed vertical-sheen-rainbow 0 -> 12 cards. known_residuals recorded twice that the 6 Holo Rare Energies of ex13/ex16 use Sheen but could not be targeted without a card list; the catalog names exactly six per set. Two cardId rows added citing the same Collexy quotes, both residuals closed.

Typecheck (all workspaces) + web build + foil tests green; 14 recipes compile clean; before/after shots on blank + assigned card at 390px and desktop in ~/.pokedex-dev-hub-legacy/foil-shots/r7/. SKILL R7 field notes + DECISIONS appended.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

35d72c9 — 2026-08-08 — foil edge-trace: the artwork holds the geometry, his mask says which of it to follow

Chey, after reviewing line-snap: "Really just 'straighten' isn't quite enough, really just detecting edges and actually tracing accurately around them is the real move." So the premise moves. line-snap treats his strokes as the geometry and nudges near-axis runs onto printed lines. edge-trace treats his mask as a STATEMENT OF INTENT and derives the boundary from the card's own edges.

apps/api/src/foil/edge-trace.ts (pure, no I/O): Di Zenzo colour structure tensor (Scharr per channel) -> NMS -> hysteresis; colour because a green field against a silver border reads 0.13 luminance/px and 44.3 tensor/px at the same edge. Then livewire: a corridor around his boundary (a hard bound, not a suggestion), anchors at fixed spacing PLUS every hard turn in his own line, Dijkstra on ridge + linked + proximity^2 + direction, sub-pixel refinement onto the |dI/dn| peak, and only THEN a straight fit where the traced path really is straight.

His own correction said the same thing before he did: 114 of his 627 correction px sit in one 40x50 box around the species strip's swept tail, which line-snap had squared off by intersecting two lines across it.

Two traps, both now locked by tests:

• the MAD-trim trap: a robust fit discards a small feature as an outlier, reports "straight to 0.2px RMS", and crisping then flattens what it ignored. straightMaxDevPx requires every point in a run to be close.
• the half-pixel frame: mask space centres pixels at x+0.5, a gradient array at integer x. All artwork lookups convert in one place.

Measured, not asserted. generate-masks.ts adherence scores any set of masks by distance to the NEAREST printed edge maximum, with line-snap's own luminance Sobel as the default probe so the incumbent gets the home field: his hand mask 50.7% within 1px / 0.847px mean line-snap@1 74.8% / 0.545px his correction 75.4% / 0.548px edge-trace@1 77.1% / 0.199px (colour probe: 98.4% / 0.082px)

Applied to Tropius me05-001/37184 as run edgetrace-tropius-1, ai/unreviewed, his ai-corrected mask preserved as parent + verbatim supersede archive. Revert proven byte-identical twice before shipping.

test:foil 35/35 (10 new). Typecheck all workspaces, api+mcp+web builds green. Visuals in ~/.pokedex-dev-hub-legacy/foil-shots/mask-edgetrace/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id edgetrace-tropius-1

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

35fcbd7 — 2026-08-08 — Mask editor: pan + pinch-zoom that keeps the strokes honest, and no more "selected" card

Chey, before a batch of hand masks: "something that would be nice is to be able to pan and pinch to zoom while editing masks. In addition, sometimes the card image gets 'selected' when I'm trying to draw which is annoying."

Zoom is a camera view offset, not a CSS scale. camera.setViewOffset(Wz, Hz, x, y, W, H) re-rasterizes the card at the zoomed size, so 4x zoom is 4x real detail — the point of zooming is tracing a printed edge, and a magnified 390px framebuffer would not help. The overlays get the same transform on one wrapper div, written imperatively (no re-render per frame). Because that is a pure crop of a linearly-scaled render, u lands at u*z - offset in both places, so getBoundingClientRect already reports the truth and the pointer->texel math is unchanged. Verified numerically: a 3.375x corner stroke lands at texels [84,180,108,196] vs a predicted [85,180,109,196]; touch at 3x lands within one texel at 390 and 744.

Brush is constant in SCREEN px (lineWidth = brushSize / zoom): zooming in must buy finer control, not a fatter-looking stroke. Two fingers pinch/pan; one finger pans unless it is a drawing finger or on a window handle; a pen owns the surface (palm rejection); a second finger mid-stroke aborts AND rolls the stroke back. Desktop gets wheel/trackpad zoom at the cursor, middle- or Space-drag pan, and +/-/0. HUD shows the zoom and fits back in one tap. Pan is clamped so the card cannot be lost. Zoom is live only while editing, resets on exit, and never reaches the artifact — a mask saved at 3.375x round-trips as 490x674 with an identical bbox. WindowEditor divides drag deltas by zoom and counter-scales its handles, so they stay 44px on screen (44.0 -> 44.0 measured at 2.25x).

Selection fix: user-select/touch-callout/user-drag none, draggable=false, and preventDefault on dragstart/selectstart and the paint pointerdown across both canvases and the viewer column; touch-action stays none so the two fixes do not fight. iOS also needs the gesturestart/change/end veto — touch-action alone does not stop Safari page zoom.

Cost the most time: window listeners registered in the controller factory were killed by StrictMode's simulated unmount, so the first pinch started and never ended and every later gesture and stroke was dead. They are now a symmetric attach()/detach() pair driven by an effect. See DECISIONS.md.

44/44 browser checks at 1280 / 744 / 390 (touch via CDP), his committed masks still load, scratch card artifacts removed. Shots in ~/.pokedex-dev-hub-legacy/foil-shots/mask-ux/.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9eb88fb — 2026-08-08 — foil region-learn: the REGIONS come from his masks, the geometry from the artwork

The last lane measured why rule-seeded edge-trace failed on every me05 reverse: 99.7% of the gap between the layout rule and Chey's intent is REGION decisions, and a tracer can only move a boundary, never add or remove a region. This is the missing half.

region-learn@1 partitions a card face into five structural classes computed from its OWN printing (border / furniture / frameBody / windowBackground / windowSubject), reads off his masks WHICH of them carry foil, applies that to a new card, and only then hands the boundary to edge-trace. Nothing about Pokemon is hard-coded — the same code reads OPPOSITE policies from the two classes in the corpus, and a test asserts the votes, not just the verdict.

The policy, in his pixels: modern-sv/sheet foil = the COLOURED FRAME BODY, nothing else. Species strip (tails included), stage tag, evolution medallion + its sprite, "evolves from" bar, copyright footer: all excluded. And THE SILVER BORDER RING IS NOT FOIL — 0.2/0.2/0.4% across his three masks. That was flagged as "his call, not a measurement"; it is now a measurement, and unanimous. wotc/window foil = the illustration's own background, minus the subject silhouette, minus the stage box where it overlaps. 4/4.

Bar written down before any number existed: mean IoU >= 0.90, no held-out card below 0.85. Leave-one-out: modern-sv/sheet 0.7566 -> 0.9691 (worst 0.9519) PASS -> batch of 8 shipped wotc/window 0.6146 -> 0.8971 (worst 0.8599) FAIL by 0.0029 -> no batch

The two residuals differ in kind, and that is the finding: modern-sv's is a 1-2px registration hairline; wotc's is 8,343px of Machamp's body read as background because his blue-grey sits inside the colour distance of the teal holo field. Every frame-level decision on WOTC is right. The subject silhouette is not.

Also found and NOT changed (era-wide, his call): wotc.artWindow is 11px too high at the top and 6px too wide at the right — his four masks and each card's own printed edge agree, and base1-8 is the card the era was "measured on". Same pattern as me04-007 for modern-sv.

Also: the bevel side. An illustration box has TWO parallel printed edges; which one the foil stops at is which side the foil is on, not which is strongest. Picking the strongest put base1-6's window top 8px wrong and scored it 0.0001; picking the foil side took it to 0.8789.

Batch regionlearn-me05-1 (8 cards, 5 frame colours, 4 with medallions, all ai / unreviewed) applied and proven reversible byte-for-byte — including a revert bug fixed here: it left the empty card directory behind, so a reverted card still looked masked to anything that walks data/foil-masks.

Recommend reverting wotc-window-trial-1: those five base1 masks were made by window-artgate@1, which scores 0.7694 mean / 0.7330 worst on this same test.

test:foil 46/46 (11 new, including the anti-collapse rule locked three ways).

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

5f1909a — 2026-08-08 — foil vector-template@1: the layout is the artifact, and it is analytic geometry

Chey: "what I draw is intent... generated masks should feel like they're derived from clean vectors" — and "we don't need 3,454 vector masks, all of these share the same 2 layouts really."

Mined his 8 correction diffs first. They say the regions were nearly right (0.12% to 2.23% changed) and that the errors were structural, not stylistic:

• the coloured sliver pinched off between the evolution medallion and the border ring IS foil (he added it on all 4 Stage-1 cards). This was the open question flagged in codified/modern-sv.md as "Chey has never ruled on that sliver". It is ruled on now.
• achromatic marks printed ON the coloured field carry foil: the colourless energy symbols, the retreat symbols, the regulation-mark box, the name text. region-learn@1 carved holes around all of them; he filled every one.
• the left/right window edge sat on the bevel's INNER line; he moved it out.
• and, decisively for this lane: where the print is straight, HIS line is straighter than region-learn@1's traced one. The machine was the thing that wobbled, so a smoother tracer was never the fix.

Geometry, measured over his 11 masks: 50.05% of the face is foil on ALL of them, 47.18% bare on all, 2.77% contested — and the Basic-vs-Stage1 difference is a SINGLE 70x85 blob. So it is not 2 layouts, it is one layout plus one optional element, which the fitter discovers from the corpus itself (bimodal split, 80.5pp separation) rather than being told a medallion exists.

vector-template@1 fits lines + circular arcs + exact corners to the consensus of his masks and rasterises that. 58 primitives, 13.2 KB of committed JSON, applied at render time — instead of 3,454 rasters.

Leave-one-out vs the bar committed in a14a163 BEFORE any number existed (mean >= 0.94, none below 0.90): mean 0.9904, worst 0.9838. PASS. region-learn@1 on the same 11: 0.9757 / 0.9519. Boundary p95 1px on 10 of 11 (was 2-72px).

Vector-ness at 0.35px tolerance on SUB-PIXEL contours (measuring on traceLoops is a trap — crack following returns a rectilinear staircase and rates a hand blob as perfectly axis-aligned): primitives/kpx region-learn@1 33.7 HIS 21.3 vector-template@1 10.7 This is the finding: his hand masks are ALREADY more vector-like than the generator that was supposed to be learning from them.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

7a7429b — 2026-08-08 — foil vector-template: make the emitted path actually analytic (two real bugs)

Writing the round-trip test found two things the IoU table was hiding.

1. JOINS WERE RAW CONTOUR POINTS. vectorizeLoop fitted lines, axis-snapped them, and then took each run's endpoint straight off the sampled contour, only substituting an intersection for line/line corners. So the emitted "vertical" left edge of a rounded rect ran from x=25.00 to x=25.66 — a polyline through contour samples with arcs bridging them, wearing a vector's clothes. The straightness this lane exists to deliver was measured but not emitted. Joins are now the actual intersection of the two primitives that meet there (line/line, line/arc, arc/arc all solved), with joinMaxMovePx rejecting a join that lands far from where the contour really turned.

2. ARC SWEEP WAS INVERTED. cross < 0 ? 1 : 0 — with y down, a positive cross product is a clockwise turn on screen and sweep=1 is clockwise, so every fitted arc bulged the wrong way. Round-trip IoU on a rounded rect: 0.9482 before, 0.9994 after. It only cost ~0.5% on the real template because that template is 56 lines and 2 arcs, which is exactly why a synthetic test found it and the corpus numbers did not.

Template refitted. Nothing moved much, because the bugs were small in this geometry — but the artifact now says what it means: in-sample mean IoU 0.9908 -> 0.9911 vector-ness prim/kpx region-learn@1 33.72 | HIS 21.35 | mine 10.34 long-primitive fraction 0.9047 | 0.9679 | 0.9861 axis-aligned fraction 0.7875 | 0.8554 | 0.9321 correction margin: closer to what he drew than to what he rejected on all 10.

test:foil 56/56.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2de2deb — 2026-08-08 — foil: serve the template as a mask, and write down what his corrections ruled

API: GET /foil-lab/masks/:cardId/:variantId?era=&scope=&evolves=1 now falls back to the vector template when no hand mask resolves. A hand mask ALWAYS answers first, so this cannot shadow Chey's work — it only fills what used to 404 — and it is labelled X-Foil-Mask-Review-Status: unreviewed with the generator identity, run id and exemplar count.

The cache is TWO PNGs, not 3,454. A template's rasterisation depends only on (era, scope, hasOptionalElement, w, h) and not on the card, so every Basic reverse gets byte-identical pixels. ~17 KB each under IMAGE_CACHE_ROOT/foil-templates/, mirroring the image-cache contract: disk-only, outside git, safe to delete, rebuilds in milliseconds. Committing per-card masks would have been ~100-350 MB against the template's 13.3 KB.

evolves is an INPUT, not something the route detects, for two structural reasons: these routes have no DB and the server never shells out, so the process can neither read card.stage nor decode card art. The caller already knows. This turned out to be the right call for a second reason — the artwork probe agrees with the catalog on 89.3% of Pokemon but misfires on pale frames (it called sv09-120 Dunsparce, a Basic, "medallion present" at 0.0% chroma).

codified/modern-sv.md pass 2 (n=11) records the three rulings his corrections make, with numbers:

• THE SLIVER IS FOIL. Pass 1 flagged it as "Chey has never ruled on that sliver". region-learn@1 scored 0.0% there on all four Stage-1 cards; he added it on all four. "Largest coloured component" was the wrong rule; the right one is every coloured region that reaches the border ring. The medallion's own sprite stays out because it is fully enclosed — topology, not size.
• Achromatic ink printed ON the coloured field carries foil: colourless energy symbols, retreat symbols, the regulation box, the name text. Furniture is distinguished topologically, not by chroma.
• The window edge sits on the bevel's OUTER line.

And what is NOT covered, said out loud: Trainer (431 variants) and Energy (104) are a different layout with zero exemplars — visibly wrong, probe agrees with the catalog on 0/49 of them. Pale frames cannot self-verify (45.5% of Lightning and 41.0% of Colorless flagged vs 0.0% of Fire/Psychic/Darkness) — a contrast effect, calibrated against his own hand masks scoring 0.720 on the same probe.

Revert proven, not asserted: a 2-card run reverts to a byte-identical tree (git sees nothing), and a supersede round-trip restores the earlier mask with matching sha256 on both the PNG and the sidecar.

DECISIONS also records an operational gotcha I caused: pkill -f on "apps/api/dist/index.js" matches the deployed pm2 pokedex-api too. pm2 healed it in seconds (prod 200, unstable restarts 0), but kill dev instances by PID.

typecheck clean across all workspaces; test:foil 56/56; api + web builds green.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

---

foil/main

Tip ceb0ab1a812845a3f35397b4fe51b20025929450 · 63 commits · 2026-08-01 → 2026-08-09 · 10,724 words · 35 reproduced below.

db82e3f — 2026-08-01 — foil/main: workbench v1 — three.js viewer, 5-pattern starter library, era layout masks

Quarantined foil workbench at /pokedex/foil-lab (URL-only, linked from nowhere, chrome-free): real owned-card scans from the image cache, tilt via pointer/gyro(iOS permission)/manual with reduced-motion respected, uniform sliders, pattern + mask-scope overrides, mask overlay toggle, copy-recipe JSON. Pattern library v1 tuned against owned eras (Base/WOTC, SV, Mega Evolution): Starlight, Cosmos, SV default holo, SV reverse sheet, Cracked Ice — uniform contract (core + uP0..3) designed for the full 15-20 set. Era layout spec as data (era-layouts.json, rects measured on real scans); resolver v1 (series, rarity, variant kind) -> pattern + mask scope. SKILL.md x2 (foil-effects, mask-pipeline stub). main.tsx gains a dev-only double-createRoot guard (Vite HMR entry-invalidation footgun, see DECISIONS).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

e3dc8be — 2026-08-01 — foil/main v2: Pencil mask editing, comment queue, iPad two-column, branch api

Apple-Pencil hand-mask drawing on the workbench (MaskEditor: card-aligned canvas overlay, pen+mouse default w/ finger toggle, pressure brush, eraser, 12-step undo, touch-action none, starts from layout prior); masks persist as committed ground-truth artifacts data/foil-masks//.png+json (derivation_method hand) via the env-gated branch api (POKEDEX_FOIL_LAB=1, port 3712, PGPOOL_MAX=1 — inert in prod), auto-load and beat the layout tier (uMaskTex contract; CanvasTexture flipY=false, exactly one V flip). Comment button captures card/variant/pattern/scope/mask state + all slider values to issues/foil// (report.md + context.json), out of the fix-issues sweep. Era-grouped owned-card picker (WOTC / SWSH / SV+ME headings); two-column layout (viewer | controls) from 700px — verified at 390px, 744x1133, 1133x744, 1440x900 with simulated pen strokes. SKILL.md x2 updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

b98e8f5 — 2026-08-02 — foil: sidecar v2 (prior+diff), artwork-keyed masks, comment linkage, codify ritual

• Every hand-mask save records its starting prior (era/scope/resolver rect + RESOLVER_VERSION) and the api renders .prior.png + .diff.png (green = human added, red = removed) + Jaccard agreement, server-side via a pure-JS PNG codec (no native addons). Saves without a parsable prior 400.
• Machamp mask backfilled (deterministic prior wotc/window/v1): agreement 0.6409 — his correction is the subject silhouette, the rule's first lesson.
• Artwork keying: variants of one cardId provably share one scan, so mask GETs alias to a sibling variant's mask when prior.scope matches (X-Foil-Mask- Alias-Of). Cross-card reprints unprovable from the catalog -> per-card only.
• Comments auto-link the exact saved mask state (maskFile/maskSavedAt/ maskAliasOf/maskHasPriorDiff).
• mask-pipeline SKILL: sidecar v2 storage, identity rule + limits, and the codify ritual; first worked example data/foil-masks/codified/wotc.md (n=1: WOTC window = art-window rect minus subject silhouette; noted in era-layouts.json wotc.notes).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

4548fca — 2026-08-02 — foil: any-card picker — full-catalog browse/search, Owned-only as a toggle

Chey: pull up ANY card in the catalog and tilt it, owned or not. Picker now browses the whole catalog by era -> set (paged strip, promo sets run 300+), adds full-catalog name search, and keeps Owned-only as a filter toggle (default on = prior behavior). Zero new endpoints: existing /series, /series/:slug, /sets/:setId (paged), /search, /cards/:cardId serve it all; unowned cards list their catalog variants (quantity 0) and resolve the same base-guess pattern path. Selection is never clobbered by filters; unmapped series get an honest "Other eras" bucket. Verified on unowned neo2-13 Umbreon (Auto - Starlight WOTC) at 390px + both iPad viewports.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

dc8e2f8 — 2026-08-02 — foil: taxonomy integration — all 39 pattern types in the app, cited resolver

Mislabel corrections from the video-reference research, applied to apps/ (research docs untouched):

• sv-holo renamed vertical-sheen (it always rendered the HGSS-XY vertical bands, not SV's). PATTERN_ALIASES keeps old slugs resolving forever — sidecars, comment context.json, Copy-recipe JSON never orphan.
• Sheen family = ONE generator at four rotations + stripe option: five honest implemented slugs (vertical / horizontal=TRUE SV default "Mirage" / diagonal right "/" / diagonal left "" / striped "Line").
• Cosmos label drops "/ Galaxy" (Galaxy is a starlight synonym; synonym moved to the starlight entry). reverse-sheet description notes it ≈ pokeball-masterball stamped sheet.
• Vintage narrowed: starlight = Base/Jungle/Fossil only; Base Set 2 → Call of Legends holos = cosmos.

Pattern dropdown carries the FULL 39-slug vocabulary, grouped: implemented recipes first, then every gap type rendering via its nearest recipe and labeled "approx via ..." (amber note in the UI). Unimplemented != hidden.

Resolver v2: base guess reads the cited usage table (usage-index.json, derived from research/foil-pattern-usage.json by tools/foil/build-usage-index.mjs; set-name match -> era/series token match -> era-default heuristics). The winning row's confidence + citation hosts surface in the workbench ("Guess: set-level citation, high confidence (bulbapedia..., collexy...)"). Reverse rows citing 'mirror' are deprioritized per the research known-gap (ink-design evidence, not foil-pattern evidence). RESOLVER_VERSION -> 2 (scope semantics unchanged from v1).

Verified: typecheck + build green; Playwright at 390/744x1133/1133x744 (~/.pokedex-dev-hub-legacy/foil-shots/w2/) — Machamp auto=starlight set-level-high with his hand mask still resolving, SV holo auto=horizontal-sheen, SV reverse auto=energy-symbols-ii, fireworks approx labeling.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

44e79fb — 2026-08-02 — foil: fix diffuse darkening — scan sampled sRGB-decoded but never re-encoded (issue ls9u0y)

The face texture was uploaded SRGB8_ALPHA8 (hardware sRGB->linear decode at sample time) while the foil ShaderMaterial writes gl_FragColor raw — three.js only appends the linear->sRGB colorspace_fragment chunk to built-in materials — so every scan rendered in linear values displayed as sRGB (measured flat 184->123 / 199->147, exactly the sRGB->linear curve). Sample the scan undecoded (NoColorSpace): the compositing model is authored in display space and pattern=none at rest is now pixel-comparable to the flat view (mean abs luma diff 44.55 -> 2.16, mapping identity +/-2). Also zero pattern-none's default specular so the baseline is truly the plain scan.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

861404e — 2026-08-02 — foil: workbench split — canon pattern lab (A) | card adjustment surface (B)

Chey's comment issues/foil/2026-08-02_12-59-52-368_4aq756 (resolved here):

• /pokedex/foil-lab/canon (CanonLab.tsx): bare holofoil pattern on a blank card (no ink; black/dark/silver/white flat-tone base via the normal texture path) beside the REAL reference tilt clip + 8 keyframes streamed from research/foil-video-reference/ by new env-gated dev routes (byte-range via res.sendFile for iOS Safari). Full 39-slug picker, sliders, tilt, and Save canon -> data/foil-canon/.json (full uniform snapshot that replaces recipe code defaults as the baseline).
• /pokedex/foil-lab (FoilLab.tsx reframed): per-card masks/comments unchanged, plus per-card uniform overrides saved as a SPARSE diff vs the canon baseline -> data/foil-overrides//.json (untouched uniforms keep tracking canon). Off-canon sliders get a dot; saving with sliders matching canon deletes the override.
• Layering model in foil/canon.ts (code defaults < canon < override < live); shared atoms + surface tabs in foil/ui.tsx; routes wired in main.tsx; AppShell chrome-free check widened to /foil-lab/*; both data dirs re-included in .gitignore; ports 5184/3713 recorded in ORCHESTRATION.md.

Verified in-browser at 390x844 / 744x1133 / 1133x744 (clip playing beside the bare cosmos render; Machamp base1-8 hand mask still resolves) — shots in ~/.pokedex-dev-hub-legacy/foil-shots/canon-lab/. Typecheck + build green (web, api). patterns.ts untouched (R0 re-tunes concurrently).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

eabb72a — 2026-08-02 — foil R0: re-tune implemented recipes per Gemini verification notes

cosmos: full activation-model rewrite — dark field, smaller/denser orbs mostly near-invisible, cluster pops via low-freq noise over cell ids, pinprick 4-point twinkles (was: dense wall of large saturated orbs). starlight (+ starlight-ii via shared GLSL): parallax architecture untouched; tighter visibility lobe (pow 5->9, floor .18->.08), sharper glyph cores + narrower flares, smaller blobs, saturated star color, galaxy wash halved; starlight-ii art gate lowered for the mid-orange Evolutions field. cracked-ice: authored intra-shard grain removed per Chey's accuracy ruling — facets flash as solid smooth planes; whitening at peak reduced, uSat up. sheen family: generator grows sharp/beam/barcode options. vertical-sheen -> new barcode variant (thin spectral lines of varying width, CD-groove slide) with tamed beam + art gate. diagonals: sharp 3.0, beam 0.55, specular 0.35 (center blow-out), diagonal-sheen-right uP0 2->7 (same-sheet consistency). striped-vertical-sheen: stripes finer + more blended (90->130, floor .40). horizontal-sheen: untouched (20/20).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9bdf318 — 2026-08-02 — foil R0: rounds 2-3 re-tunes + verification report, field notes, decisions

starlight: longer thin-armed 4-point glyphs + 8-point-burst subset, tighter pop lobe (pow 11, floor .04), full-saturation star color, parallax default 2.4, gain 3.6 — best of 3 rounds 11/20 (still-frame judging cannot see the parallax; recorded as honest residual with a video-judge recommendation). starlight-ii: 20/20 (uP2 pinned so base-default drift can't invalidate it). cracked-ice: facet gain 1.1 -> 0.55 (art visible through flashes), density 10, less whitening — 15/20 yay; shard-geometry uniformity queued for R1. Verification report gains the R0 section (before->after per pattern, the diag-left judge-noise event, diffuse-fix supersession note); foil-effects SKILL.md gains the R0 field notes; DECISIONS entry for root cause + wave.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

50906c4 — 2026-08-02 — foil: cited per-card pattern assignments across the whole catalog (resolver v3)

Chey's mlmwmp comment: the library was funneling to era defaults ("cosmos"). Full-catalog enumeration (20,964 cards / 37,627 variants, via the branch read api) + an 18-lane Ringer research swarm (foil-card-assignments, GLM-5.2, executed selector-integrity checks) produced:

• research/foil-card-assignments.json — 111 set/rarity/kind/card-level rows, 19 catalog foil-facet mappings (holo-foil-cosmos, reverse-foil-masterball, normal-foil-galaxy...), 55 honest residuals; per-lane provenance + method.
• tools/foil/build-assignments-index.mjs -> apps/web/src/foil/assignments-index.json (trimmed for the bundle, duplicate-facet guard).
• resolver v3 (RESOLVER_VERSION=3): assignment tier above the v2 usage table; specificity cardIds > variantKinds > facet > rarities > set > v2 > heuristic; normal-kind variants with a declared foil facet now render window foil (previously scope none); FoilLab passes cardId.
• era-layouts: wotc seriesSlugs 'ecard' -> 'e-card', + 'legendary-collection' (both had fallen through to modern-sv rects and heuristics).

Coverage (foil-bearing variants): 79.5% set-cited / 20.4% era-token / 0.2% heuristic -> 93.4% set / 4.3% facet-or-card / 2.4% era-token / 0 heuristic. 13 in-browser spot-checks in ~/.pokedex-dev-hub-legacy/foil-shots/assignments/. Issue mlmwmp resolved. Weak spots + curation ledger in DECISIONS.md.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

77b6646 — 2026-08-02 — foil R1: dedicated recipes for the twelve owned-era gap patterns (10/12 verified)

Chey's comment dml369 ("develop out the rest of the missing holofoil patterns"): dedicated GLSL for fireworks, ace-spec, energy-symbols-ii, rainbow-glitter-sheen, ex-starfoil, prismatic-pokeball, tinsel-ii, cosmos-iii-smooth, pokeball-masterball, radiant, rainbow-glitter, confetti — implemented flags flipped, approx labels dropped, slugs stable. Gemini-verified 10/12 match over 3 rounds (verdict table + honest residuals in research/foil-verification.md R1 section). Honest nays: tinsel-ii (procedural static plateaus at judge score 2) and prismatic-pokeball (screen-only blend cannot darken a white reverse body — structural, Chey's call).

R0 residuals cleared in the same pass: cracked-ice anisotropic shattered-glass metric (re-judged yay 14/20); starlight milky wash + per-layer visibility curves with starlight-ii re-banked clean 20/20 through both shared-GLSL changes; starlight parallax judged from a new 16-frame fine sweep and still scored flat — verified real by eye, recorded as judge-blindness.

Field notes appended to foil-effects SKILL.md; dated DECISIONS entry; issue dml369 resolved.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

7f2a101 — 2026-08-02 — foil R2: uDarken blend-model term — prismatic-pokeball 17/20 yay, tinsel-ii 16/20 yay

One opt-in substrate-darkening term in the shared composite (main()): body = scan * (1 - uDarken * mask * gate) before the foil screen-blends. Default 0 is bit-identical to the old render; absent canon/override keys seed 0. Full 21-pattern regression sweep on provably identical renders: no true regressions (and a measurement of Gemini single-roll noise, ±3-6; diag-right + pokeball-masterball double-nays recorded as unstable verdicts on geometry proof). prismatic-pokeball rebuilt on the term (dark-mirror base, facet-quantized flash lobe, ball watermark as ink-overprint suppression): structural nay -> 17/20 yay. tinsel-ii one-line opt-in (uDarken 0.4): static plateau -> 16/20 yay. 20 of 21 implemented types now match; sole nay starlight (judge parallax blindness).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

55415d2 — 2026-08-02 — foil R2 stage 3: thirteen dedicated recipes (gap patterns + RC dots) + subset-commons resolver tier

• patterns.ts: real recipes for mirror, rainbow-mirror, vertical-sheen-rainbow (dark-mirror family on uDarken), pokeball-hologram, prism, water-web, energy-symbols, pinwheel, ex-emerald, tinsel, cosmos-ii-pixel, crosshatch, radiant-collection-dots — implemented: true, approx labels dropped, slugs stable. Unused plain SHEEN_V generator removed (last fallback users gone).
• resolver v4: card-level 'normal'-class assignment rows consulted before the scope-none early return — subset cards the catalog declares plain but which physically carry foil (RC commons dot overprint) resolve their subset's pattern with scope full.
• foil-card-assignments.json: two card-level rows (bw11-RC1..20, g1 25 non-ultra RC ids) with video citation (non-holo RC Teddiursa keeps the dot overprint); builder enforces cardIds on 'normal' rows; index regenerated.

Verified in-browser: g1-RC1 + bw11-RC1 render the RC dot treatment via Auto (guess card|medium). Gemini verdicts in research/foil-verification.md (R2 recipe-wave section, next commit).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

798ed3d — 2026-08-02 — foil R2b: four §40–43 vocabulary recipes (all match) + resolver v5 scope overrides

The last lane of the wave — dedicated recipes for the vocab-extension types: gold-secret (warm-locked goldRamp field + chromatic glitter pops + SWSH burst rays — 19/20 yay), vstar-pearl (near-white pearl on uDarken 0.3, warm-led diagonal wash, frame-line streaks — 13/20 yay), shiny-vault (silvery field + band-keyed sparkle-glyph amplifiers, uDarken 0.15 — 11 nay → 16/20 yay r2), detective-pikachu (diagonal sheen × photo-luminance via the documented uFace contract exception — 20/20 yay). 38/43 taxonomy types real, 34 hold match.

Resolver v5: assignment rows carry an optional per-row scope override (window/ full/sheet), honored by the set-row tier and the cls-'normal' tier — baby shinies render window despite the 'Shiny rare' full-foil rarity, VSTARs full despite plain holo kinds, det1 window despite 'ultra rare'.

Assignment research: +2 facet rows (gold, gold-jumbo → gold-secret; retires both facet residuals, 1-card metal collateral flagged) and +10 cited rows — sma 51 baby/35 GX/8 gold splits, swsh4.5sv 104 baby + 18 full-face (incl. the black shiny Eternatus pair, name-verified not gold), sv04.5 rarity split, VSTAR via the dedicated 'Holo Rare VSTAR' rarity, det1 ultra-rare + cls-normal rows (all 18 cards window). Residuals annotated with resolutions.

build-usage-index.mjs: tolerate era-wide rows with no scope.sets (the vocab gold rows crashed the build — the 122-row usage index never compiled until now). Verdicts/frames under ~/.pokedex-dev-hub-legacy/foil-verify/ (verify-

-r2b*); shots in foil-shots/r2b/. Docs: foil-verification.md R2b section, SKILL R2b field notes, DECISIONS entry.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b8c3b4f — 2026-08-03 — foil R3: sheen-family streak-field rework — all 8 of Chey's canon-lab comments resolved, 7/7 Gemini 20/20

sheenGlsl is now a streak field (Chey's critique = the spec): sparse irregular streaks, per-streak lean following card tilt (crisscross; converging pairs terminate where they meet), stretched-ellipse taper, hue running ALONG each strip as well as across. striped-vertical-sheen rebuilt from a fresh Gemini corpus re-spec run with Chey's description embedded (grouped reveal via wide sweeping windows over a subtle bottom-converging fan; gemini-spec-r3.md); ex-starfoil rebased on the reworked diagonal base + fine CD-line layer; vstar-pearl rebuilt from the horizontal-sheen streak field, pearl kept.

Canon migrations (full-snapshot rule, DECISIONS): striped + horizontal uDarken 0 -> 0.32 (bright-substrate legibility, third confirmation). All other canon values carry over; uniform semantics unchanged. validate_spec.py frame regex single-digit fix (undercounted two-digit citations, failed an honest worker).

Verdicts (Chey's words folded into every judge prompt as acceptance criteria): vertical/right/left/vstar-pearl 20/20 first roll; striped + ex-starfoil 20/20 round 2; horizontal 20/20 round 3. Table + residuals in research/foil-verification.md R3.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9223975 — 2026-08-03 — foil R3-MOTION: axis-split starlight, per-dot cosmos, hologram radiant, the rainbow-glitter-sheen delta

Chey's six motion-model canon-lab comments resolved (his eye = ground truth):

• starlight (+ starlight-ii via shared GLSL): top-to-bottom hue BANDING (soft-quantized bands; his canon 0.62/0.6 lands the reference order blue->green->red on the R->B->G cosine ramp) + AXIS-SPLIT tilt: vertical slides the whole field (~10% card height), horizontal drives the per-star random fade; his 3-layer uP1 parallax rides on top untouched. Judge: starlight 17/20 yay round 2 — the FIRST starlight yay (standing parallax nay broken); starlight-ii 20/20.
• cosmos: whole-field slide + per-dot random response (independent random tilt axes for brightness and hue per orb — no shared sweep axis); at tilt 0 his canon appearance is pixel-identical. Judge: 20/20.
• radiant: discrete hologram steps (half-cell up/down, 60% hold / 40% crossfade — one position fades out as the next fades in), uP1 wired as Hologram travel (canon migrated 0 -> 2.2; was an unused placeholder), grid lines ~45% thicker. Judge nay 13/20 across 3 rounds on a pixel-refuted "slides continuously" claim — recorded as still-frame motion blindness (same class as starlight parallax); adjacent-frame crops prove discrete positions + crossfade. Chey's live tilt is the tiebreak.
• rainbow-glitter-sheen: dedicated Gemini delta-articulation pass first (he couldn't name the difference), pixel-verified, then fixed: narrow laser-saturated striped chevron (sigma 0.010->0.0035, hue traversal 5x->9x, pow-deepened at low gain), fainter +-0.8 repeats, chevron angle 1.3->1.9, denser colored always-visible glitter, uDarken 0->0.4 (4th legibility-physics data point). Judge: 19/20 yay.

Docs: research/foil-verification.md R3-MOTION section, DECISIONS entry, foil-effects SKILL field notes. Frames/verdicts in ~/.pokedex-dev-hub-legacy/foil-verify (jobs -r3m, manifest-r3m{,2,3}, delta-rainbow-glitter-sheen); gate shots in foil-shots/r3-motion.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b3c4770 — 2026-08-03 — foil mask-refine: window handles → flatten → hand refine

Chey's requested workflow on the Card-adjust surface: draggable corner/edge handles adjust the layout window rect per card (WindowEditor.tsx, 44px touch targets, finger-drag allowed), the adjusted geometry persists as artwork-keyed scope-agnostic sidecars (data/foil-windows//.json, /foil-lab/windows routes), and Flatten bakes it through the STANDARD hand-mask save path (sidecar prior stays the era rule; new optional prior.window records the adjustment) then opens the existing paint tooling.

• rasterizeWindowRect shared by Flatten + loadLayoutRect (identical bakes)
• saveMask prior = layoutMask (rule), never the effective mask
• windows GET aliases across variants of the same scan, newest savedAt
• verified in-browser desktop + 390px: sma-SV1/sma-SV2 (SM baby shiny, borrowed modern-sv rect) and me01-034 (zero-coverage gap); base1-8 + base1-5 hand-mask regression clean; test artifacts deleted (corpus rule)
• ports: vite 5186 / api 3714 recorded in ORCHESTRATION.md

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d547992 — 2026-08-03 — foil R3-GLYPH: glyph-pattern reworks per Chey's 7 comments + the drop-in glyph slot for his SVGs

Behavior (his notes verbatim = acceptance criteria): reverse-sheet rainbow only on grainy glyphs over a neutral sheet; energy-symbols checkerboard banks that swap with tilt; energy-symbols-ii sporadic random banks (~half faint at any time); ace-spec diamonds grow/shrink with tilt + a touch of blur; prismatic-pokeball ball never darkens (white-mixed pale response + coherent plane flash, suppression deleted); radiant-collection-dots shapes catch a traveling rainbow band (16/20 YAY — standing R2 nay broken).

Glyph slot: research/foil-glyphs//glyph[-N].svg -> lab-gated api routes -> atlas rasterizer (foil/glyphs.ts) -> uGlyphTex contract + glyphTex() helper; ~2.5s auto-pickup on save, procedural fallback when absent, README tells Chey exactly where his four promised files go.

Canon migration: reverse-sheet uP1 0 -> 0.6 (dead placeholder now drives his requested glyph grain). Verdicts + pixel-refutations recorded in research/foil-verification.md (R3-GLYPH); lessons in foil-effects SKILL.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

73be25c — 2026-08-03 — foil R3-MISC: the uTint ink-metallic blend term + Chey's twelve remaining canon-lab comments (12/12 yay)

The reverse-holo fix (his chat report): screen-blending achromatic foil light over colored art compresses chroma — "dull and grayish, rather than making the color look metallic." New core uniform uTint (default 0 = bit-exact legacy) multiplies the foil layer + in-mask specular by the luminance-normalized scan chroma SQUARED (the double ink pass of a real mirror reflection); neutral on blank cards, so no canon-lab appearance moved. Reverse-family recipes opt in; verified before/after at 5 tilt angles on Victini ball-reverse, Crystal Energy, Pineco. uP4/uP5 param slots added in the same contract change.

The twelve comments, each mechanism-mapped in its issue frontmatter and the verification doc: cracked-ice triangulated + half-off gate; fireworks grid lattice + center-emanating radial hue; rainbow-mirror rebuilt as mirror + hue-banded spotlight; tinsel-ii silver lines + vertical rainbow band (band lives only on the line-work); prism rebuilt as pinwheel-kin from the corpus-vs-corpus delta pass; cosmos-ii pixel-quantized stamp edges; crosshatch thinned; gold-secret adjustable burst origin (uP4/uP5, center default, per-card via overrides) + holographic grain; dedicated recipes for sequin (glyph-family glints), tcg-classic (flat starlight + rainbow glitter), acid-wash (water-web kin), disco (perfect disc lattice, galaxy ignition) — zero approxVia fallbacks remain.

Canon migrations (appearance-preserving unless he asked otherwise): fireworks uP0 3→4.5 (dead two-octave structure), gold-secret uP4/uP5 0.5 added, sequin uP0-uP3 re-keyed to the new recipe. Verdicts: research/foil-verification.md "R3-MISC" — 12/12 Gemini yay incl. the tcg-classic ink-scope re-roll.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

3bc5fe2 — 2026-08-03 — foil R4-COMPOSITE: the ink-density invariant — foil adds pop, never washes text or mutes ink

Chey's two me05 comments (7rtnzx mirror reverse "blows out the darks/text", 19mo4l holo "darkens/muddies the colors") ruled a CORE COMPOSITE INVARIANT, engineered in shader.ts main(): ink density estimated from the scan (inkDark = darker-than-local-8-tap-field average, exactly 0 on flat blank bases; inkColor = chroma above the canon-lab-tone floor) now gates every layer — dark ink blocks the flash + shields the specular (text crisp at every tilt), ALL ink is exempt from uDarken substrate attenuation (printed color never muted; the mirror stays dark only where the mirror is visible), colored ink auto-tints its flash (max(uTint, inkColor)) and uInkPop pumps chroma under the flash so colors read MORE saturated and metallic.

New core uniforms uInkGuard (default 1) / uInkPop (default 0.5), both 0 = bit-exact legacy composite; sliders in both labs; canon files untouched (absent keys inherit the new defaults — catalog-wide fix, zero migrations).

Proof: canon-lab blank-card pairs (mirror dark+silver, cosmos, horizontal-sheen) pixel-identical (compare -metric AE 0) under a frame-stepped rAF harness with a same-settings AE-0 control, plus a Gemini sanity pass (all pairs "identical"). On-card before/after pairs across the 8-card sample (both issue cards, Victini, Pineco, Darkrai, Cetitan, Machamp WOTC, Hydreigon SIR; 3 tilts, 390px + desktop) in ~/.pokedex-dev-hub-legacy/foil-shots/r4-composite/. Both issues resolved with mechanism-mapped notes; composite contract documented in the foil-effects SKILL; DECISIONS + research/foil-verification.md updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

07eb483 — 2026-08-04 — foil R4b: scan-additive composite (rest parity by construction) + canon-lab card preview

Chey's Grubbin (me05-002, Pitch Black) review refuted R4's ink heuristics: the reverse still muddied the card at rest (measured ΔC +32, green→yellow), darkened the header under tilt (−22 L), and blew out text (glyph contrast −26%). New law on real scans (uScanBase 1): the scan is a photograph of the card at rest — foil is additive dynamic light only, tilt-gated to ZERO at rest, luminance-headroom + per-channel clamped (text sacred at every angle, uArtGate opens the clamp where dark scan pixels ARE the foil), chroma- preserving with a pastel-safe tint ramp and an L²-gated chroma pump. Canon lab blank bases run the classic composite textually unchanged: AE-0 proven (CDP frame-stepped harness, control + knob pairs; Gemini pass agreed). uInkGuard 0 still reproduces the pre-R4 composite exactly.

Canon-lab card preview (his request verbatim): blank/on-card/↻-another chips render the live sliders on a random catalog card the resolver assigns the pattern to — server-sampled from the baked inversion (tools/foil/ build-pattern-cards.mts → data/foil-pattern-cards.json, gitignored; GET /foil-lab/pattern-cards/:patternId), honoring hand masks + adjusted windows; empty-pool patterns say "no catalog cards". Also commits Chey's live mirror-canon tweak from the Grubbin session. Full metrics + verdicts: DECISIONS 2026-08-04.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

176cfe8 — 2026-08-04 — foil R4c: onset ease — the sheen leans in instead of switching on

Chey on R4b's rest gate: "it's like it isn't there at all until I've tilted the card a tiny bit and then it's like it just suddenly appears." Measured: the headroom clamp saturated R4b's smoothstep(.02,.28) gate, so ALL added light arrived inside tilt 0.03-0.13 (0 -> 16.1 dL mean) then froze flat to tilt 1.0 — a step smeared over 4% of the tilt range.

Now: one wide ease ramp = smoothstep(0, .45, |tilt|) (full at 0.45 — every R4b tilt-0.5 metric unchanged), flash gated by ramp^1.5 (cubic-slow start), specular by ramp^2.5 (the broad gloss band now trails the flash instead of leading it), plus a REST=0.006 sub-JND living rest sheen blended into both gates (rest +0.43 dL card-mean, max 2/255 — visually still the scan). After-curve: smooth S over 0->0.25, no dead zone, no knee; text clamp untouched (glyph contrast 33.6 -> 33.6 at flash peak); blank-card canon lab bit-level unchanged (fuzz-1% AE 0 vs control floor).

Harness + curves + filmstrips: ~/.pokedex-dev-hub-legacy/foil-shots/r4c/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d8a8562 — 2026-08-04 — foil R4d: motion-first onset — tiny gestures move the sheen; glow arrives late; Chey gets the dials

Chey rejected R4c too ('It still lights up pretty noticeably when I tilt the card a tiny bit. I don't like it.') — measured the gesture→tilt mapping first this time: pointer is 0.0051 tilt/px at 390px and gyro is 0.036 tilt/deg, so a 30px thumb drag (tilt 0.154) was already getting HALF the full glow under R4c. The curve was never the felt variable.

• shader.ts: onset gates now uniform-driven — uOnsetRange (glow full at |tilt|, default 0.5) and uOnsetCurve (flash exponent, default 3.5; specular trails at +1). (0.45, 1.5) reproduces R4c exactly (verified to the third decimal); defaults are strictly softer at every tilt and equal from 0.5 up, so all R4b/R4c peak metrics are unchanged (Grubbin 17.976 ΔL / text 34.02, Machamp 18.035 / text 45.6 — exact).
• FoilLab + CanonLab: 'Onset range' / 'Onset curve' sliders, canon- stored like every other uniform; absent canon keys inherit defaults.
• patterns.ts: CoreDefaults admits the two keys per recipe.

Pointer-path proof (real pointermove, 390px, 30px drag, first 300ms): before mean ΔL climbs to 6.6 and rising; after stays within ±0.08 while per-pixel movement reaches 6.1 mean / 18.9 p95 — moves, doesn't light up. Blank canon lab at fuzz-1% AE 0 vs control. Artifacts in ~/.pokedex-dev-hub-legacy/foil-shots/r4d/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

6b865c8 — 2026-08-05 — foil R5: the metallic composite — always on the card, energy redistributed, Chey's compositing dials

Chey's ruling (2026-08-05, verbatim in DECISIONS): no onset range at all — the effect is ALWAYS on the card; dial in how it composites onto the ink; metallic and spectral, not brightening.

• Onset machinery removed (R4c/R4d ramp, REST floor, uOnsetRange/uOnsetCurve sliders on both surfaces). No canon/override file ever stored the keys; CardViewer skips unknown uniforms, so stale keys anywhere are inert.
• R5 metalness law (scan path only): scan = albedo; the pattern field splits around PIVOT 0.40 — above = tinted-specular highlight, below = bounded multiplicative depth (mirror-turned-away darks, sqrt-concave). Energy redistributes; contrast + chroma carry the metallic read.
• Soft-knee headroom: the R4b hard clamp plateaued strong fields into a flat wash (the "brightened" read); the budget now compresses monotonically so pattern structure survives. Per-channel no-clip caps stay hard; glyph ink (inkDark) exempt from highlight AND depth — text contrast preserved (Grubbin 86->97 at peak, Machamp flat).
• New canon-stored dials on both surfaces: Metallic 0.6, Sheen strength 0.55, Sheen tint 0.5 (= the R4 chroma law exactly), Depth 0.55, Texture 1. uMetal 0 / uIntensity 0 / uInkGuard 0 all still yield the plain scan.
• Blank canon lab pixel-unchanged: CDP frame-stepped before/after AE 0 at fuzz 1% (mirror/cosmos/horizontal-sheen/starlight, tilted + rest); one Gemini vision pass agrees (all pairs identical). Shots + per-card rest pairs: ~/.pokedex-dev-hub-legacy/foil-shots/r5/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

8458ab1 — 2026-08-07 — foil R5b: metalness is MIRROR ONLY — every other pattern gets its own colour back

Chey: "You applied the metallic treatment to EVERY HOLO PATTERN when I only wanted it for mirror. So now, every single holo pattern looks like mirror and you can barely see the pattern at all, it's like every single one is being applied in a way that it adds no rainbow color or anything to the card."

R5 shipped its metalness dials as GLOBAL_DEFAULTS, so all 43 recipes inherited a law built for one of them. uMetal is now the scan-path LAW SELECTOR and defaults to 0 — nothing can inherit metalness; mirror is the only recipe that opts in, in its own patterns.ts defaults.

• Metalness branch is the R5 code textually unmoved, so mirror's arithmetic is unchanged by construction. Proven: his hand-tuned canon on Grubbin me05-002
  • Tropius me05-001, rest + tilt 0.2/0.35/0.5, 390px and desktop — 16/16 frames AE 0 at fuzz 1%, same floor as the same-build control pair.
• uMetal 0 runs the ADDITIVE law: the pattern's own emission (uSheen gain) screen-combined per channel over a substrate-attenuated body, tinted toward the ink only as far as uSheenTint asks. Three fixes to what R5 broke — the scalar no-clip cap let the artwork's brightest channel throttle the whole flash; the quartic ink budget (shaped to starve mid-dark glyphs) starved ordinary mid-tone art instead; PIVOT ate the field before anything flashed. Cosmos measured 0.005% of art-window pixels changed >8% under R5.
• uDepth decoupled from uDarken: uDarken is the blank-canon substrate, frozen by 30 saved canon files, and 0 on exactly the pale reverse sheets that need one. The substrate follows the pattern's dark half so it buys contrast, not a uniform dim (a flat 0.34 cost mean dL -21.7 and 29% of text contrast).
• Per-recipe families in patterns.ts (PARTICLE / SHEET / WASH / PEARL) plus seven by-eye corrections, each judged on a card the resolver assigns.
• Onset machinery stays deleted. Text guarantee re-verified: glyphs matte and crisp at peak tilt on a reverse whose foil covers the text box.
• Blank-card canon appearance unchanged: 8 patterns AE 0 (6 direct; cracked-ice and tinsel-ii against the pre-R5b code with canon held constant, since Chey re-saved those two mid-wave).
• SKILL.md composite contract rewritten around per-recipe selection.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2baaf7e — 2026-08-07 — foil R6: composite defaults re-derived from Chey's canons; the guard learns to tell text from artwork

Chey: "i re-adjusted cracked-ice. take a look at it, and then make the changes based on everything you're observing to make all of this better - i want it so that i have to do as little of work personally on all the rest as possible."

WHAT HIS CANONS TAUGHT

uSheenTint 0 on all three non-mirror canons -> 0 for every non-mirror family (was 0.15/0.5/0.85/0.6). Ink-tinting a pattern's own highlight is exactly "it adds no rainbow color or anything to the card". uSheen he raised all three -> families now 1.2-4.2 (were 1.0-2.6). uDepth NOT universal: 0 on cracked-ice, 1 on tinsel-ii and rgs. It darkens the card WHERE THE PATTERN IS NOT, so its cost scales with 1 - duty. Measured every recipe's duty cycle; his three choices line up with it, and it predicted both dials for the only canon that exercises the defaults. uSat 1 + wide spread -> shipped on the 13 recipes with no canon.

Families are renamed for how light LANDS (flash / line / stamp / field / pearl / metal) and stated explicitly as FoilPattern.family. The stamp tier exists because the first pass got it wrong: energy-symbols-ii lights 5.8% of the face, so a field substrate darkened the other 94% and the card read "someone dimmed this".

THE INK GUARD: SPLIT, DON'T WEAKEN

inkDark is a local-contrast measure, so it fired on every dark mark - including every outline and shading edge in the ILLUSTRATION (measured: 86-89% of strong hits inside the art window are artwork, not text). The additive law spends it as a hard coverage multiply, so guard 1 punched the pattern full of holes. That is why he kept pulling it to ~0.51/0.56 on sheet/wash patterns.

Now split by glyphness into inkGlyph (sacred, zero flash, unchanged) and inkDetail (artwork darks keep the pattern on a tightened budget). inkAny is byte-for-byte the old inkDark and every pre-R6 consumer still reads it.

Success test met: at guard 1 under R6, tinsel-ii and rgs look BETTER than his 0.51/0.56 do today - tinsel-ii's artwork is visible through the streaks instead of greyed out. Both canons migrated to 1, reason recorded in each file's note; every other value in them is his, untouched.

SLIDER RANGES (his second ask)

Four dials sat exactly on their cap: uIntensity 2->4, uSheen 3->6, uSat 1->2, uHueSpread 1.5->3. Not extended: uInkGuard (his carve-out), uMetal (a law selector), the 0..1 mix weights, and uDarken/uDepth (already fully spent at 1).

uSheen needed a shader change to be honest: the additive budget's ceiling was reached at uSheen 2.0, so the top third of the slider bought almost nothing - very likely why he parked two canons at the ceiling. It now resumes climbing above 3, leaving every stored value <= 3 bit-for-bit identical.

The 19 duplicated lines across the two surfaces are one CORE_SLIDERS table; uGrain is dimmed and labelled "metal law only" where it does nothing.

NEW: "Apply composite -> family" copies only the on-card dials to siblings, never pattern shape.

PROVEN, NOT ASSUMED (AE vs a base build compiled from the previous commit)

• blank canon room: AE 0 across all 44 implemented recipes x 4 tones x 3 tilts
• mirror: AE 0 on the card path too - pixel-frozen
• text at flash peak: +0.1 to -1.1 points of glyph/paper contrast on every reverse whose foil covers the text box; the library's two largest costs (rainbow-glitter, water-web) toned down per-recipe and re-checked by eye
• every implemented recipe eyeballed on 1-2 real assigned cards at rest and two tilts; shots in ~/.pokedex-dev-hub-legacy/foil-shots/r6/

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

92f52aa — 2026-08-07 — Mask provenance: sidecar v3, the AI learning loop, and the collapse safeguard

Chey (2026-08-07): mark AI-generated masks, let an AI learn from his hand-done ones, then let him correct the AI's mask so it "can observe the diff and continue to improve without me having to hand-paint all the masks."

derivation_method existed in v2 but the API HARDCODED it to 'hand' — harmless while only a Pencil touched the corpus, actively wrong once mask-refine's Flatten started writing machine geometry through the same save path.

Taxonomy (5 values, 4 cases that must never blur): layout-flatten machine geometry, unpainted weight 0 hand human, from a layout prior weight 1 hand-refined human, on an existing mask weight 1 ai generator, UNREVIEWED weight 0 ai-corrected AI proposed, human edited weight 0.6

The label is never taken from the client. The client reports only what the canvas was SEEDED with; writeMaskRecord derives the method by diffing saved pixels against what that seed rasterizes to. A machine label needs a full GeneratorIdentity, which HTTP callers cannot supply. One write path serves the route and generators both.

• Anti-feedback-collapse enforced in selectExemplars(), not by convention: unreviewed ai is weight 0 at any corpus size. Codify now routes through it.
• Corrections keep the parent PNG + change map + metrics (incl. a 4x4 grid of where corrections land) — the supervised pairs a generator lane consumes via the self-describing tuples manifest.
• Seam rule, measured not guessed: canvas roundRect vs the server SDF disagree on 389/330260 px, ALL in the 1px AA band. Without tolerating it every unpainted Flatten would stamp hand.
• Guard: if the file at the target path is an unreviewed ai mask it IS the parent, whatever the client claims — AI ancestry cannot be laundered. That guard caught a real bug during this build (ActionBtn passed React's MouseEvent into saveMask's optional arg, blanking derivation); derivation is now required on the PUT so it fails loudly instead.
• v1/v2 load forever (normalize-on-read); Chey's two masks migrated additively and still read HAND-PAINTED, agreement 0.641, pixels untouched.
• Trial batch wotc-window-trial-1: 6 WOTC Base holo rares from window-artgate@1, all ai/unreviewed, reversible. Leave-one-out on the human corpus beat the rect baseline (0.588 -> 0.688) — real, but n=2 same-set, so explicitly not validation.

Surfaces: provenance badge + detail in the workbench, Mask corpus panel, GET /foil-lab/masks/corpus (+?tuples/?exemplars), artifact routes, and corpus.ts report|exemplars|tuples|migrate. Pure tests added to CI (test:foil). Browser-verified at 390px and desktop. See DECISIONS.md 2026-08-07 and the mask-pipeline SKILL.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

6b61213 — 2026-08-08 — foil mask-straighten: read the INTENT out of Chey's hand mask, and make the undo real

Chey, 2026-08-08: "it's impossible to get the lines really straight so I'm hoping you can get computer vision on the mask and card art in tandem to really see my intent there, and make the mask nice and crisp and straight on the lines I was trying to draw along. Will you try that? And then I can either approve or revert."

Smoothing a wobbly boundary gives a smoothly wobbly boundary. The premise here is that a hand-drawn foil boundary is an ATTEMPT TO TRACE A PRINTED EDGE — frame, art box, species strip, stage tag — and those are dead straight in the scan. So: contour the mask, cut each loop into near-axis runs (PCA orientation, wobble gaps bridged), robust-fit each, hunt the scan's directional gradient with a local Hough over (angle, offset) plus a proximity prior, replace, intersect adjacent lines into real corners, rasterize matching his own AA character.

The refusals are the feature, and each is a recorded param: weak evidence, or an ambiguous band of comparable ridges, may nudge but never relocate; with no artwork edge it straightens only to HIS own fit; short and curved runs pass through untouched. A first cut "closed" the Tropius top-left corner — which turned out to be him deliberately tracing around the BASIC stage tag. The stricter rules keep it.

• apps/api/src/foil/line-snap.ts the geometry, pure, no I/O
• generator.ts line-snap@1, a REFINER (requiresSource)
• generate-masks.ts run --refine, archives, restoring revert
• provenance.ts supersedes + verbatim sha256 undo archive

A generator write onto an existing mask now THROWS without explicit supersede { runId }; with it, every artifact the mask had is archived verbatim under superseded/./ with a self-describing manifest, and revert --run-id verifies the whole archive BEFORE deleting anything live, then restores byte-for-byte. supersedes is deliberately NOT correction: one is a human editing a machine (training signal), the other a machine replacing a human with nobody's agreement. --refine also refuses a source with exemplar weight 0, so a refiner can never eat its own unreviewed output.

Tropius me05-001/37184 (run straighten-tropius-1): 8 runs snapped to printed edges, 1 self-straightened, 13 left exactly as drawn, 8 corners closed, 2 specks dropped. Max correction 3.74px, 1.06% of the face changed, Jaccard 0.9798, confidence 0.557. Honest residual: the card's bottom border was too faint to accept, so that run self-straightened and the report says so.

test:foil 25/25 (9 geometry + 5 supersede/restore, incl. byte-for-byte round trip, a deliberate curve surviving, an ambiguous band refused, and a corrupt archive aborting before deletion). Revert exercised twice end to end, sha256sum clean. Shots: ~/.pokedex-dev-hub-legacy/foil-shots/mask-straighten/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id straighten-tropius-1

d627995 — 2026-08-08 — foil R7: his eye overturns two banked mechanisms; "no catalog cards" gets a reason

Chey's 15 canon-lab comments (2026-08-07), all resolved in issues/foil/.

Eleven shader verdicts. Two overturn mechanisms built from earlier corpus — energy-symbols' parity CHECKERBOARD (R3-GLYPH, from his own words) and prism's solid-facet grid (R3-MISC, from a Gemini delta pass). Having seen both rendered he rejected them. A description is a hypothesis about the render, not a spec.

cosmos the 0.085 UV field slide he read as parallax is now uP4 "Surface drift" at 0.012; zero at rest, canon unchanged striped-v-sheen grey lines: the unconditional silver term is gone AND the second-order lobe cut 0.4 -> 0.16 (it drew a whole second, near-achromatic line population on the far side) rainbow-mirror hue axis is card X, blue -> green -> red, on BOTH the sheen and the spotlight; core whitening 0.85 -> 0.32 energy-symbols checkerboard gone: per-symbol response AXIS + PHASE + transition WIDTH over a low-freq cluster field energy-symbols-ii same treatment (it shared the two-bank mechanism) pokeball-hologram flat SDF icon -> shaded 3D sphere (hemisphere normal, tilt-following light, belt, hub, two-tone shell) tinsel the "blinds" were a literal full-card striation floor; demoted to uP2, default 0 (his canon stores 0) tinsel-ii line-work gated to the colour band, silver 0.60 -> 0.10, band width exposed as uP4 prism pinwheel's spoke fan WITHOUT the per-wedge phase that makes a pinwheel spin; density 33 -> 18 radiant hologram step 0.5 cell -> uP4 0.17, drive rescaled so total travel per unit tilt is unchanged rainbow-glitter invisible on card: bigger white-lifted flakes AND the family fixed (flash -> field; no substrate, no contrast)

Canon policy held: shader.ts / canon.ts / CardViewer.tsx / data/foil-canon have a ZERO diff, sheenGlsl() is byte-identical (all four sheen canons frozen), and every new behaviour went into a NEW uniform old canon files inherit. No stored value migrated. One recommendation left for him, not applied: tinsel-ii's starkness is substantially his own canon (uDepth 1 at ~135 lines).

"Why are there no catalog cards?" — asked four times, three answers. The baked inversion is now v2 and emits a machine-readable diagnosis the lab renders: outranked diagonal-sheen-left — 1,818 SM reverses ARE cited for it, but energy-symbols wins them (the emblem layer vs the sheet rotation; both true). A capped SECONDARY pool is now baked via a new citedFoilPatterns() export and served as via:'cited', so the preview works. Never flip a winner to fill a pool. class-absent pinwheel, pokeball-hologram — the catalog has reverse variants for ex1-ex5 ONLY; the whole late-EX reverse era is missing upstream. The UI now says so instead of "no catalog cards". fixed vertical-sheen-rainbow 0 -> 12 cards. known_residuals recorded twice that the 6 Holo Rare Energies of ex13/ex16 use Sheen but could not be targeted without a card list; the catalog names exactly six per set. Two cardId rows added citing the same Collexy quotes, both residuals closed.

Typecheck (all workspaces) + web build + foil tests green; 14 recipes compile clean; before/after shots on blank + assigned card at 390px and desktop in ~/.pokedex-dev-hub-legacy/foil-shots/r7/. SKILL R7 field notes + DECISIONS appended.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

82b502c — 2026-08-08 — foil edge-trace: the artwork holds the geometry, his mask says which of it to follow

Chey, after reviewing line-snap: "Really just 'straighten' isn't quite enough, really just detecting edges and actually tracing accurately around them is the real move." So the premise moves. line-snap treats his strokes as the geometry and nudges near-axis runs onto printed lines. edge-trace treats his mask as a STATEMENT OF INTENT and derives the boundary from the card's own edges.

apps/api/src/foil/edge-trace.ts (pure, no I/O): Di Zenzo colour structure tensor (Scharr per channel) -> NMS -> hysteresis; colour because a green field against a silver border reads 0.13 luminance/px and 44.3 tensor/px at the same edge. Then livewire: a corridor around his boundary (a hard bound, not a suggestion), anchors at fixed spacing PLUS every hard turn in his own line, Dijkstra on ridge + linked + proximity^2 + direction, sub-pixel refinement onto the |dI/dn| peak, and only THEN a straight fit where the traced path really is straight.

His own correction said the same thing before he did: 114 of his 627 correction px sit in one 40x50 box around the species strip's swept tail, which line-snap had squared off by intersecting two lines across it.

Two traps, both now locked by tests:

• the MAD-trim trap: a robust fit discards a small feature as an outlier, reports "straight to 0.2px RMS", and crisping then flattens what it ignored. straightMaxDevPx requires every point in a run to be close.
• the half-pixel frame: mask space centres pixels at x+0.5, a gradient array at integer x. All artwork lookups convert in one place.

Measured, not asserted. generate-masks.ts adherence scores any set of masks by distance to the NEAREST printed edge maximum, with line-snap's own luminance Sobel as the default probe so the incumbent gets the home field: his hand mask 50.7% within 1px / 0.847px mean line-snap@1 74.8% / 0.545px his correction 75.4% / 0.548px edge-trace@1 77.1% / 0.199px (colour probe: 98.4% / 0.082px)

Applied to Tropius me05-001/37184 as run edgetrace-tropius-1, ai/unreviewed, his ai-corrected mask preserved as parent + verbatim supersede archive. Revert proven byte-identical twice before shipping.

test:foil 35/35 (10 new). Typecheck all workspaces, api+mcp+web builds green. Visuals in ~/.pokedex-dev-hub-legacy/foil-shots/mask-edgetrace/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id edgetrace-tropius-1

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

116427d — 2026-08-08 — Mask editor: pan + pinch-zoom that keeps the strokes honest, and no more "selected" card

Chey, before a batch of hand masks: "something that would be nice is to be able to pan and pinch to zoom while editing masks. In addition, sometimes the card image gets 'selected' when I'm trying to draw which is annoying."

Zoom is a camera view offset, not a CSS scale. camera.setViewOffset(Wz, Hz, x, y, W, H) re-rasterizes the card at the zoomed size, so 4x zoom is 4x real detail — the point of zooming is tracing a printed edge, and a magnified 390px framebuffer would not help. The overlays get the same transform on one wrapper div, written imperatively (no re-render per frame). Because that is a pure crop of a linearly-scaled render, u lands at u*z - offset in both places, so getBoundingClientRect already reports the truth and the pointer->texel math is unchanged. Verified numerically: a 3.375x corner stroke lands at texels [84,180,108,196] vs a predicted [85,180,109,196]; touch at 3x lands within one texel at 390 and 744.

Brush is constant in SCREEN px (lineWidth = brushSize / zoom): zooming in must buy finer control, not a fatter-looking stroke. Two fingers pinch/pan; one finger pans unless it is a drawing finger or on a window handle; a pen owns the surface (palm rejection); a second finger mid-stroke aborts AND rolls the stroke back. Desktop gets wheel/trackpad zoom at the cursor, middle- or Space-drag pan, and +/-/0. HUD shows the zoom and fits back in one tap. Pan is clamped so the card cannot be lost. Zoom is live only while editing, resets on exit, and never reaches the artifact — a mask saved at 3.375x round-trips as 490x674 with an identical bbox. WindowEditor divides drag deltas by zoom and counter-scales its handles, so they stay 44px on screen (44.0 -> 44.0 measured at 2.25x).

Selection fix: user-select/touch-callout/user-drag none, draggable=false, and preventDefault on dragstart/selectstart and the paint pointerdown across both canvases and the viewer column; touch-action stays none so the two fixes do not fight. iOS also needs the gesturestart/change/end veto — touch-action alone does not stop Safari page zoom.

Cost the most time: window listeners registered in the controller factory were killed by StrictMode's simulated unmount, so the first pinch started and never ended and every later gesture and stroke was dead. They are now a symmetric attach()/detach() pair driven by an effect. See DECISIONS.md.

44/44 browser checks at 1280 / 744 / 390 (touch via CDP), his committed masks still load, scratch card artifacts removed. Shots in ~/.pokedex-dev-hub-legacy/foil-shots/mask-ux/.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9127416 — 2026-08-08 — foil region-learn: the REGIONS come from his masks, the geometry from the artwork

The last lane measured why rule-seeded edge-trace failed on every me05 reverse: 99.7% of the gap between the layout rule and Chey's intent is REGION decisions, and a tracer can only move a boundary, never add or remove a region. This is the missing half.

region-learn@1 partitions a card face into five structural classes computed from its OWN printing (border / furniture / frameBody / windowBackground / windowSubject), reads off his masks WHICH of them carry foil, applies that to a new card, and only then hands the boundary to edge-trace. Nothing about Pokemon is hard-coded — the same code reads OPPOSITE policies from the two classes in the corpus, and a test asserts the votes, not just the verdict.

The policy, in his pixels: modern-sv/sheet foil = the COLOURED FRAME BODY, nothing else. Species strip (tails included), stage tag, evolution medallion + its sprite, "evolves from" bar, copyright footer: all excluded. And THE SILVER BORDER RING IS NOT FOIL — 0.2/0.2/0.4% across his three masks. That was flagged as "his call, not a measurement"; it is now a measurement, and unanimous. wotc/window foil = the illustration's own background, minus the subject silhouette, minus the stage box where it overlaps. 4/4.

Bar written down before any number existed: mean IoU >= 0.90, no held-out card below 0.85. Leave-one-out: modern-sv/sheet 0.7566 -> 0.9691 (worst 0.9519) PASS -> batch of 8 shipped wotc/window 0.6146 -> 0.8971 (worst 0.8599) FAIL by 0.0029 -> no batch

The two residuals differ in kind, and that is the finding: modern-sv's is a 1-2px registration hairline; wotc's is 8,343px of Machamp's body read as background because his blue-grey sits inside the colour distance of the teal holo field. Every frame-level decision on WOTC is right. The subject silhouette is not.

Also found and NOT changed (era-wide, his call): wotc.artWindow is 11px too high at the top and 6px too wide at the right — his four masks and each card's own printed edge agree, and base1-8 is the card the era was "measured on". Same pattern as me04-007 for modern-sv.

Also: the bevel side. An illustration box has TWO parallel printed edges; which one the foil stops at is which side the foil is on, not which is strongest. Picking the strongest put base1-6's window top 8px wrong and scored it 0.0001; picking the foil side took it to 0.8789.

Batch regionlearn-me05-1 (8 cards, 5 frame colours, 4 with medallions, all ai / unreviewed) applied and proven reversible byte-for-byte — including a revert bug fixed here: it left the empty card directory behind, so a reverted card still looked masked to anything that walks data/foil-masks.

Recommend reverting wotc-window-trial-1: those five base1 masks were made by window-artgate@1, which scores 0.7694 mean / 0.7330 worst on this same test.

test:foil 46/46 (11 new, including the anti-collapse rule locked three ways).

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

782c096 — 2026-08-08 — foil vector-template@1: the layout is the artifact, and it is analytic geometry

Chey: "what I draw is intent... generated masks should feel like they're derived from clean vectors" — and "we don't need 3,454 vector masks, all of these share the same 2 layouts really."

Mined his 8 correction diffs first. They say the regions were nearly right (0.12% to 2.23% changed) and that the errors were structural, not stylistic:

• the coloured sliver pinched off between the evolution medallion and the border ring IS foil (he added it on all 4 Stage-1 cards). This was the open question flagged in codified/modern-sv.md as "Chey has never ruled on that sliver". It is ruled on now.
• achromatic marks printed ON the coloured field carry foil: the colourless energy symbols, the retreat symbols, the regulation-mark box, the name text. region-learn@1 carved holes around all of them; he filled every one.
• the left/right window edge sat on the bevel's INNER line; he moved it out.
• and, decisively for this lane: where the print is straight, HIS line is straighter than region-learn@1's traced one. The machine was the thing that wobbled, so a smoother tracer was never the fix.

Geometry, measured over his 11 masks: 50.05% of the face is foil on ALL of them, 47.18% bare on all, 2.77% contested — and the Basic-vs-Stage1 difference is a SINGLE 70x85 blob. So it is not 2 layouts, it is one layout plus one optional element, which the fitter discovers from the corpus itself (bimodal split, 80.5pp separation) rather than being told a medallion exists.

vector-template@1 fits lines + circular arcs + exact corners to the consensus of his masks and rasterises that. 58 primitives, 13.2 KB of committed JSON, applied at render time — instead of 3,454 rasters.

Leave-one-out vs the bar committed in a14a163 BEFORE any number existed (mean >= 0.94, none below 0.90): mean 0.9904, worst 0.9838. PASS. region-learn@1 on the same 11: 0.9757 / 0.9519. Boundary p95 1px on 10 of 11 (was 2-72px).

Vector-ness at 0.35px tolerance on SUB-PIXEL contours (measuring on traceLoops is a trap — crack following returns a rectilinear staircase and rates a hand blob as perfectly axis-aligned): primitives/kpx region-learn@1 33.7 HIS 21.3 vector-template@1 10.7 This is the finding: his hand masks are ALREADY more vector-like than the generator that was supposed to be learning from them.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

15af7c5 — 2026-08-08 — foil vector-template: make the emitted path actually analytic (two real bugs)

Writing the round-trip test found two things the IoU table was hiding.

1. JOINS WERE RAW CONTOUR POINTS. vectorizeLoop fitted lines, axis-snapped them, and then took each run's endpoint straight off the sampled contour, only substituting an intersection for line/line corners. So the emitted "vertical" left edge of a rounded rect ran from x=25.00 to x=25.66 — a polyline through contour samples with arcs bridging them, wearing a vector's clothes. The straightness this lane exists to deliver was measured but not emitted. Joins are now the actual intersection of the two primitives that meet there (line/line, line/arc, arc/arc all solved), with joinMaxMovePx rejecting a join that lands far from where the contour really turned.

2. ARC SWEEP WAS INVERTED. cross < 0 ? 1 : 0 — with y down, a positive cross product is a clockwise turn on screen and sweep=1 is clockwise, so every fitted arc bulged the wrong way. Round-trip IoU on a rounded rect: 0.9482 before, 0.9994 after. It only cost ~0.5% on the real template because that template is 56 lines and 2 arcs, which is exactly why a synthetic test found it and the corpus numbers did not.

Template refitted. Nothing moved much, because the bugs were small in this geometry — but the artifact now says what it means: in-sample mean IoU 0.9908 -> 0.9911 vector-ness prim/kpx region-learn@1 33.72 | HIS 21.35 | mine 10.34 long-primitive fraction 0.9047 | 0.9679 | 0.9861 axis-aligned fraction 0.7875 | 0.8554 | 0.9321 correction margin: closer to what he drew than to what he rejected on all 10.

test:foil 56/56.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

bb2765c — 2026-08-08 — foil: serve the template as a mask, and write down what his corrections ruled

API: GET /foil-lab/masks/:cardId/:variantId?era=&scope=&evolves=1 now falls back to the vector template when no hand mask resolves. A hand mask ALWAYS answers first, so this cannot shadow Chey's work — it only fills what used to 404 — and it is labelled X-Foil-Mask-Review-Status: unreviewed with the generator identity, run id and exemplar count.

The cache is TWO PNGs, not 3,454. A template's rasterisation depends only on (era, scope, hasOptionalElement, w, h) and not on the card, so every Basic reverse gets byte-identical pixels. ~17 KB each under IMAGE_CACHE_ROOT/foil-templates/, mirroring the image-cache contract: disk-only, outside git, safe to delete, rebuilds in milliseconds. Committing per-card masks would have been ~100-350 MB against the template's 13.3 KB.

evolves is an INPUT, not something the route detects, for two structural reasons: these routes have no DB and the server never shells out, so the process can neither read card.stage nor decode card art. The caller already knows. This turned out to be the right call for a second reason — the artwork probe agrees with the catalog on 89.3% of Pokemon but misfires on pale frames (it called sv09-120 Dunsparce, a Basic, "medallion present" at 0.0% chroma).

codified/modern-sv.md pass 2 (n=11) records the three rulings his corrections make, with numbers:

• THE SLIVER IS FOIL. Pass 1 flagged it as "Chey has never ruled on that sliver". region-learn@1 scored 0.0% there on all four Stage-1 cards; he added it on all four. "Largest coloured component" was the wrong rule; the right one is every coloured region that reaches the border ring. The medallion's own sprite stays out because it is fully enclosed — topology, not size.
• Achromatic ink printed ON the coloured field carries foil: colourless energy symbols, retreat symbols, the regulation box, the name text. Furniture is distinguished topologically, not by chroma.
• The window edge sits on the bevel's OUTER line.

And what is NOT covered, said out loud: Trainer (431 variants) and Energy (104) are a different layout with zero exemplars — visibly wrong, probe agrees with the catalog on 0/49 of them. Pale frames cannot self-verify (45.5% of Lightning and 41.0% of Colorless flagged vs 0.0% of Fire/Psychic/Darkness) — a contrast effect, calibrated against his own hand masks scoring 0.720 on the same probe.

Revert proven, not asserted: a 2-card run reverts to a byte-identical tree (git sees nothing), and a supersede round-trip restores the earlier mask with matching sha256 on both the PNG and the sidecar.

DECISIONS also records an operational gotcha I caused: pkill -f on "apps/api/dist/index.js" matches the deployed pm2 pokedex-api too. pm2 healed it in seconds (prod 200, unstable restarts 0), but kill dev instances by PID.

typecheck clean across all workspaces; test:foil 56/56; api + web builds green.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

---

foil/mask-edgetrace

Tip 91615360d19105389f480e8d1c5a0699bf7ac0c5 · 53 commits · 2026-08-01 → 2026-08-09 · 8,204 words · 30 reproduced below.

c4afdea — 2026-08-01 — foil/main: workbench v1 — three.js viewer, 5-pattern starter library, era layout masks

Quarantined foil workbench at /pokedex/foil-lab (URL-only, linked from nowhere, chrome-free): real owned-card scans from the image cache, tilt via pointer/gyro(iOS permission)/manual with reduced-motion respected, uniform sliders, pattern + mask-scope overrides, mask overlay toggle, copy-recipe JSON. Pattern library v1 tuned against owned eras (Base/WOTC, SV, Mega Evolution): Starlight, Cosmos, SV default holo, SV reverse sheet, Cracked Ice — uniform contract (core + uP0..3) designed for the full 15-20 set. Era layout spec as data (era-layouts.json, rects measured on real scans); resolver v1 (series, rarity, variant kind) -> pattern + mask scope. SKILL.md x2 (foil-effects, mask-pipeline stub). main.tsx gains a dev-only double-createRoot guard (Vite HMR entry-invalidation footgun, see DECISIONS).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

899b5fe — 2026-08-01 — foil/main v2: Pencil mask editing, comment queue, iPad two-column, branch api

Apple-Pencil hand-mask drawing on the workbench (MaskEditor: card-aligned canvas overlay, pen+mouse default w/ finger toggle, pressure brush, eraser, 12-step undo, touch-action none, starts from layout prior); masks persist as committed ground-truth artifacts data/foil-masks//.png+json (derivation_method hand) via the env-gated branch api (POKEDEX_FOIL_LAB=1, port 3712, PGPOOL_MAX=1 — inert in prod), auto-load and beat the layout tier (uMaskTex contract; CanvasTexture flipY=false, exactly one V flip). Comment button captures card/variant/pattern/scope/mask state + all slider values to issues/foil// (report.md + context.json), out of the fix-issues sweep. Era-grouped owned-card picker (WOTC / SWSH / SV+ME headings); two-column layout (viewer | controls) from 700px — verified at 390px, 744x1133, 1133x744, 1440x900 with simulated pen strokes. SKILL.md x2 updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

77e1319 — 2026-08-02 — foil: sidecar v2 (prior+diff), artwork-keyed masks, comment linkage, codify ritual

• Every hand-mask save records its starting prior (era/scope/resolver rect + RESOLVER_VERSION) and the api renders .prior.png + .diff.png (green = human added, red = removed) + Jaccard agreement, server-side via a pure-JS PNG codec (no native addons). Saves without a parsable prior 400.
• Machamp mask backfilled (deterministic prior wotc/window/v1): agreement 0.6409 — his correction is the subject silhouette, the rule's first lesson.
• Artwork keying: variants of one cardId provably share one scan, so mask GETs alias to a sibling variant's mask when prior.scope matches (X-Foil-Mask- Alias-Of). Cross-card reprints unprovable from the catalog -> per-card only.
• Comments auto-link the exact saved mask state (maskFile/maskSavedAt/ maskAliasOf/maskHasPriorDiff).
• mask-pipeline SKILL: sidecar v2 storage, identity rule + limits, and the codify ritual; first worked example data/foil-masks/codified/wotc.md (n=1: WOTC window = art-window rect minus subject silhouette; noted in era-layouts.json wotc.notes).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

c1cd95f — 2026-08-02 — foil: any-card picker — full-catalog browse/search, Owned-only as a toggle

Chey: pull up ANY card in the catalog and tilt it, owned or not. Picker now browses the whole catalog by era -> set (paged strip, promo sets run 300+), adds full-catalog name search, and keeps Owned-only as a filter toggle (default on = prior behavior). Zero new endpoints: existing /series, /series/:slug, /sets/:setId (paged), /search, /cards/:cardId serve it all; unowned cards list their catalog variants (quantity 0) and resolve the same base-guess pattern path. Selection is never clobbered by filters; unmapped series get an honest "Other eras" bucket. Verified on unowned neo2-13 Umbreon (Auto - Starlight WOTC) at 390px + both iPad viewports.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

88f260a — 2026-08-02 — foil: taxonomy integration — all 39 pattern types in the app, cited resolver

Mislabel corrections from the video-reference research, applied to apps/ (research docs untouched):

• sv-holo renamed vertical-sheen (it always rendered the HGSS-XY vertical bands, not SV's). PATTERN_ALIASES keeps old slugs resolving forever — sidecars, comment context.json, Copy-recipe JSON never orphan.
• Sheen family = ONE generator at four rotations + stripe option: five honest implemented slugs (vertical / horizontal=TRUE SV default "Mirage" / diagonal right "/" / diagonal left "" / striped "Line").
• Cosmos label drops "/ Galaxy" (Galaxy is a starlight synonym; synonym moved to the starlight entry). reverse-sheet description notes it ≈ pokeball-masterball stamped sheet.
• Vintage narrowed: starlight = Base/Jungle/Fossil only; Base Set 2 → Call of Legends holos = cosmos.

Pattern dropdown carries the FULL 39-slug vocabulary, grouped: implemented recipes first, then every gap type rendering via its nearest recipe and labeled "approx via ..." (amber note in the UI). Unimplemented != hidden.

Resolver v2: base guess reads the cited usage table (usage-index.json, derived from research/foil-pattern-usage.json by tools/foil/build-usage-index.mjs; set-name match -> era/series token match -> era-default heuristics). The winning row's confidence + citation hosts surface in the workbench ("Guess: set-level citation, high confidence (bulbapedia..., collexy...)"). Reverse rows citing 'mirror' are deprioritized per the research known-gap (ink-design evidence, not foil-pattern evidence). RESOLVER_VERSION -> 2 (scope semantics unchanged from v1).

Verified: typecheck + build green; Playwright at 390/744x1133/1133x744 (~/.pokedex-dev-hub-legacy/foil-shots/w2/) — Machamp auto=starlight set-level-high with his hand mask still resolving, SV holo auto=horizontal-sheen, SV reverse auto=energy-symbols-ii, fireworks approx labeling.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

4351cb2 — 2026-08-02 — foil: fix diffuse darkening — scan sampled sRGB-decoded but never re-encoded (issue ls9u0y)

The face texture was uploaded SRGB8_ALPHA8 (hardware sRGB->linear decode at sample time) while the foil ShaderMaterial writes gl_FragColor raw — three.js only appends the linear->sRGB colorspace_fragment chunk to built-in materials — so every scan rendered in linear values displayed as sRGB (measured flat 184->123 / 199->147, exactly the sRGB->linear curve). Sample the scan undecoded (NoColorSpace): the compositing model is authored in display space and pattern=none at rest is now pixel-comparable to the flat view (mean abs luma diff 44.55 -> 2.16, mapping identity +/-2). Also zero pattern-none's default specular so the baseline is truly the plain scan.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

710d42b — 2026-08-02 — foil: workbench split — canon pattern lab (A) | card adjustment surface (B)

Chey's comment issues/foil/2026-08-02_12-59-52-368_4aq756 (resolved here):

• /pokedex/foil-lab/canon (CanonLab.tsx): bare holofoil pattern on a blank card (no ink; black/dark/silver/white flat-tone base via the normal texture path) beside the REAL reference tilt clip + 8 keyframes streamed from research/foil-video-reference/ by new env-gated dev routes (byte-range via res.sendFile for iOS Safari). Full 39-slug picker, sliders, tilt, and Save canon -> data/foil-canon/.json (full uniform snapshot that replaces recipe code defaults as the baseline).
• /pokedex/foil-lab (FoilLab.tsx reframed): per-card masks/comments unchanged, plus per-card uniform overrides saved as a SPARSE diff vs the canon baseline -> data/foil-overrides//.json (untouched uniforms keep tracking canon). Off-canon sliders get a dot; saving with sliders matching canon deletes the override.
• Layering model in foil/canon.ts (code defaults < canon < override < live); shared atoms + surface tabs in foil/ui.tsx; routes wired in main.tsx; AppShell chrome-free check widened to /foil-lab/*; both data dirs re-included in .gitignore; ports 5184/3713 recorded in ORCHESTRATION.md.

Verified in-browser at 390x844 / 744x1133 / 1133x744 (clip playing beside the bare cosmos render; Machamp base1-8 hand mask still resolves) — shots in ~/.pokedex-dev-hub-legacy/foil-shots/canon-lab/. Typecheck + build green (web, api). patterns.ts untouched (R0 re-tunes concurrently).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

3d6c64f — 2026-08-02 — foil R0: re-tune implemented recipes per Gemini verification notes

cosmos: full activation-model rewrite — dark field, smaller/denser orbs mostly near-invisible, cluster pops via low-freq noise over cell ids, pinprick 4-point twinkles (was: dense wall of large saturated orbs). starlight (+ starlight-ii via shared GLSL): parallax architecture untouched; tighter visibility lobe (pow 5->9, floor .18->.08), sharper glyph cores + narrower flares, smaller blobs, saturated star color, galaxy wash halved; starlight-ii art gate lowered for the mid-orange Evolutions field. cracked-ice: authored intra-shard grain removed per Chey's accuracy ruling — facets flash as solid smooth planes; whitening at peak reduced, uSat up. sheen family: generator grows sharp/beam/barcode options. vertical-sheen -> new barcode variant (thin spectral lines of varying width, CD-groove slide) with tamed beam + art gate. diagonals: sharp 3.0, beam 0.55, specular 0.35 (center blow-out), diagonal-sheen-right uP0 2->7 (same-sheet consistency). striped-vertical-sheen: stripes finer + more blended (90->130, floor .40). horizontal-sheen: untouched (20/20).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

cb27850 — 2026-08-02 — foil R0: rounds 2-3 re-tunes + verification report, field notes, decisions

starlight: longer thin-armed 4-point glyphs + 8-point-burst subset, tighter pop lobe (pow 11, floor .04), full-saturation star color, parallax default 2.4, gain 3.6 — best of 3 rounds 11/20 (still-frame judging cannot see the parallax; recorded as honest residual with a video-judge recommendation). starlight-ii: 20/20 (uP2 pinned so base-default drift can't invalidate it). cracked-ice: facet gain 1.1 -> 0.55 (art visible through flashes), density 10, less whitening — 15/20 yay; shard-geometry uniformity queued for R1. Verification report gains the R0 section (before->after per pattern, the diag-left judge-noise event, diffuse-fix supersession note); foil-effects SKILL.md gains the R0 field notes; DECISIONS entry for root cause + wave.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

63de48f — 2026-08-02 — foil: cited per-card pattern assignments across the whole catalog (resolver v3)

Chey's mlmwmp comment: the library was funneling to era defaults ("cosmos"). Full-catalog enumeration (20,964 cards / 37,627 variants, via the branch read api) + an 18-lane Ringer research swarm (foil-card-assignments, GLM-5.2, executed selector-integrity checks) produced:

• research/foil-card-assignments.json — 111 set/rarity/kind/card-level rows, 19 catalog foil-facet mappings (holo-foil-cosmos, reverse-foil-masterball, normal-foil-galaxy...), 55 honest residuals; per-lane provenance + method.
• tools/foil/build-assignments-index.mjs -> apps/web/src/foil/assignments-index.json (trimmed for the bundle, duplicate-facet guard).
• resolver v3 (RESOLVER_VERSION=3): assignment tier above the v2 usage table; specificity cardIds > variantKinds > facet > rarities > set > v2 > heuristic; normal-kind variants with a declared foil facet now render window foil (previously scope none); FoilLab passes cardId.
• era-layouts: wotc seriesSlugs 'ecard' -> 'e-card', + 'legendary-collection' (both had fallen through to modern-sv rects and heuristics).

Coverage (foil-bearing variants): 79.5% set-cited / 20.4% era-token / 0.2% heuristic -> 93.4% set / 4.3% facet-or-card / 2.4% era-token / 0 heuristic. 13 in-browser spot-checks in ~/.pokedex-dev-hub-legacy/foil-shots/assignments/. Issue mlmwmp resolved. Weak spots + curation ledger in DECISIONS.md.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

3f6f88c — 2026-08-02 — foil R1: dedicated recipes for the twelve owned-era gap patterns (10/12 verified)

Chey's comment dml369 ("develop out the rest of the missing holofoil patterns"): dedicated GLSL for fireworks, ace-spec, energy-symbols-ii, rainbow-glitter-sheen, ex-starfoil, prismatic-pokeball, tinsel-ii, cosmos-iii-smooth, pokeball-masterball, radiant, rainbow-glitter, confetti — implemented flags flipped, approx labels dropped, slugs stable. Gemini-verified 10/12 match over 3 rounds (verdict table + honest residuals in research/foil-verification.md R1 section). Honest nays: tinsel-ii (procedural static plateaus at judge score 2) and prismatic-pokeball (screen-only blend cannot darken a white reverse body — structural, Chey's call).

R0 residuals cleared in the same pass: cracked-ice anisotropic shattered-glass metric (re-judged yay 14/20); starlight milky wash + per-layer visibility curves with starlight-ii re-banked clean 20/20 through both shared-GLSL changes; starlight parallax judged from a new 16-frame fine sweep and still scored flat — verified real by eye, recorded as judge-blindness.

Field notes appended to foil-effects SKILL.md; dated DECISIONS entry; issue dml369 resolved.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b0be3a1 — 2026-08-02 — foil R2: uDarken blend-model term — prismatic-pokeball 17/20 yay, tinsel-ii 16/20 yay

One opt-in substrate-darkening term in the shared composite (main()): body = scan * (1 - uDarken * mask * gate) before the foil screen-blends. Default 0 is bit-identical to the old render; absent canon/override keys seed 0. Full 21-pattern regression sweep on provably identical renders: no true regressions (and a measurement of Gemini single-roll noise, ±3-6; diag-right + pokeball-masterball double-nays recorded as unstable verdicts on geometry proof). prismatic-pokeball rebuilt on the term (dark-mirror base, facet-quantized flash lobe, ball watermark as ink-overprint suppression): structural nay -> 17/20 yay. tinsel-ii one-line opt-in (uDarken 0.4): static plateau -> 16/20 yay. 20 of 21 implemented types now match; sole nay starlight (judge parallax blindness).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d8e46da — 2026-08-02 — foil R2 stage 3: thirteen dedicated recipes (gap patterns + RC dots) + subset-commons resolver tier

• patterns.ts: real recipes for mirror, rainbow-mirror, vertical-sheen-rainbow (dark-mirror family on uDarken), pokeball-hologram, prism, water-web, energy-symbols, pinwheel, ex-emerald, tinsel, cosmos-ii-pixel, crosshatch, radiant-collection-dots — implemented: true, approx labels dropped, slugs stable. Unused plain SHEEN_V generator removed (last fallback users gone).
• resolver v4: card-level 'normal'-class assignment rows consulted before the scope-none early return — subset cards the catalog declares plain but which physically carry foil (RC commons dot overprint) resolve their subset's pattern with scope full.
• foil-card-assignments.json: two card-level rows (bw11-RC1..20, g1 25 non-ultra RC ids) with video citation (non-holo RC Teddiursa keeps the dot overprint); builder enforces cardIds on 'normal' rows; index regenerated.

Verified in-browser: g1-RC1 + bw11-RC1 render the RC dot treatment via Auto (guess card|medium). Gemini verdicts in research/foil-verification.md (R2 recipe-wave section, next commit).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

a047891 — 2026-08-02 — foil R2b: four §40–43 vocabulary recipes (all match) + resolver v5 scope overrides

The last lane of the wave — dedicated recipes for the vocab-extension types: gold-secret (warm-locked goldRamp field + chromatic glitter pops + SWSH burst rays — 19/20 yay), vstar-pearl (near-white pearl on uDarken 0.3, warm-led diagonal wash, frame-line streaks — 13/20 yay), shiny-vault (silvery field + band-keyed sparkle-glyph amplifiers, uDarken 0.15 — 11 nay → 16/20 yay r2), detective-pikachu (diagonal sheen × photo-luminance via the documented uFace contract exception — 20/20 yay). 38/43 taxonomy types real, 34 hold match.

Resolver v5: assignment rows carry an optional per-row scope override (window/ full/sheet), honored by the set-row tier and the cls-'normal' tier — baby shinies render window despite the 'Shiny rare' full-foil rarity, VSTARs full despite plain holo kinds, det1 window despite 'ultra rare'.

Assignment research: +2 facet rows (gold, gold-jumbo → gold-secret; retires both facet residuals, 1-card metal collateral flagged) and +10 cited rows — sma 51 baby/35 GX/8 gold splits, swsh4.5sv 104 baby + 18 full-face (incl. the black shiny Eternatus pair, name-verified not gold), sv04.5 rarity split, VSTAR via the dedicated 'Holo Rare VSTAR' rarity, det1 ultra-rare + cls-normal rows (all 18 cards window). Residuals annotated with resolutions.

build-usage-index.mjs: tolerate era-wide rows with no scope.sets (the vocab gold rows crashed the build — the 122-row usage index never compiled until now). Verdicts/frames under ~/.pokedex-dev-hub-legacy/foil-verify/ (verify-

-r2b*); shots in foil-shots/r2b/. Docs: foil-verification.md R2b section, SKILL R2b field notes, DECISIONS entry.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

bc2ff24 — 2026-08-03 — foil R3: sheen-family streak-field rework — all 8 of Chey's canon-lab comments resolved, 7/7 Gemini 20/20

sheenGlsl is now a streak field (Chey's critique = the spec): sparse irregular streaks, per-streak lean following card tilt (crisscross; converging pairs terminate where they meet), stretched-ellipse taper, hue running ALONG each strip as well as across. striped-vertical-sheen rebuilt from a fresh Gemini corpus re-spec run with Chey's description embedded (grouped reveal via wide sweeping windows over a subtle bottom-converging fan; gemini-spec-r3.md); ex-starfoil rebased on the reworked diagonal base + fine CD-line layer; vstar-pearl rebuilt from the horizontal-sheen streak field, pearl kept.

Canon migrations (full-snapshot rule, DECISIONS): striped + horizontal uDarken 0 -> 0.32 (bright-substrate legibility, third confirmation). All other canon values carry over; uniform semantics unchanged. validate_spec.py frame regex single-digit fix (undercounted two-digit citations, failed an honest worker).

Verdicts (Chey's words folded into every judge prompt as acceptance criteria): vertical/right/left/vstar-pearl 20/20 first roll; striped + ex-starfoil 20/20 round 2; horizontal 20/20 round 3. Table + residuals in research/foil-verification.md R3.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

e3f389d — 2026-08-03 — foil R3-MOTION: axis-split starlight, per-dot cosmos, hologram radiant, the rainbow-glitter-sheen delta

Chey's six motion-model canon-lab comments resolved (his eye = ground truth):

• starlight (+ starlight-ii via shared GLSL): top-to-bottom hue BANDING (soft-quantized bands; his canon 0.62/0.6 lands the reference order blue->green->red on the R->B->G cosine ramp) + AXIS-SPLIT tilt: vertical slides the whole field (~10% card height), horizontal drives the per-star random fade; his 3-layer uP1 parallax rides on top untouched. Judge: starlight 17/20 yay round 2 — the FIRST starlight yay (standing parallax nay broken); starlight-ii 20/20.
• cosmos: whole-field slide + per-dot random response (independent random tilt axes for brightness and hue per orb — no shared sweep axis); at tilt 0 his canon appearance is pixel-identical. Judge: 20/20.
• radiant: discrete hologram steps (half-cell up/down, 60% hold / 40% crossfade — one position fades out as the next fades in), uP1 wired as Hologram travel (canon migrated 0 -> 2.2; was an unused placeholder), grid lines ~45% thicker. Judge nay 13/20 across 3 rounds on a pixel-refuted "slides continuously" claim — recorded as still-frame motion blindness (same class as starlight parallax); adjacent-frame crops prove discrete positions + crossfade. Chey's live tilt is the tiebreak.
• rainbow-glitter-sheen: dedicated Gemini delta-articulation pass first (he couldn't name the difference), pixel-verified, then fixed: narrow laser-saturated striped chevron (sigma 0.010->0.0035, hue traversal 5x->9x, pow-deepened at low gain), fainter +-0.8 repeats, chevron angle 1.3->1.9, denser colored always-visible glitter, uDarken 0->0.4 (4th legibility-physics data point). Judge: 19/20 yay.

Docs: research/foil-verification.md R3-MOTION section, DECISIONS entry, foil-effects SKILL field notes. Frames/verdicts in ~/.pokedex-dev-hub-legacy/foil-verify (jobs -r3m, manifest-r3m{,2,3}, delta-rainbow-glitter-sheen); gate shots in foil-shots/r3-motion.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d7f1636 — 2026-08-03 — foil mask-refine: window handles → flatten → hand refine

Chey's requested workflow on the Card-adjust surface: draggable corner/edge handles adjust the layout window rect per card (WindowEditor.tsx, 44px touch targets, finger-drag allowed), the adjusted geometry persists as artwork-keyed scope-agnostic sidecars (data/foil-windows//.json, /foil-lab/windows routes), and Flatten bakes it through the STANDARD hand-mask save path (sidecar prior stays the era rule; new optional prior.window records the adjustment) then opens the existing paint tooling.

• rasterizeWindowRect shared by Flatten + loadLayoutRect (identical bakes)
• saveMask prior = layoutMask (rule), never the effective mask
• windows GET aliases across variants of the same scan, newest savedAt
• verified in-browser desktop + 390px: sma-SV1/sma-SV2 (SM baby shiny, borrowed modern-sv rect) and me01-034 (zero-coverage gap); base1-8 + base1-5 hand-mask regression clean; test artifacts deleted (corpus rule)
• ports: vite 5186 / api 3714 recorded in ORCHESTRATION.md

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

837243c — 2026-08-03 — foil R3-GLYPH: glyph-pattern reworks per Chey's 7 comments + the drop-in glyph slot for his SVGs

Behavior (his notes verbatim = acceptance criteria): reverse-sheet rainbow only on grainy glyphs over a neutral sheet; energy-symbols checkerboard banks that swap with tilt; energy-symbols-ii sporadic random banks (~half faint at any time); ace-spec diamonds grow/shrink with tilt + a touch of blur; prismatic-pokeball ball never darkens (white-mixed pale response + coherent plane flash, suppression deleted); radiant-collection-dots shapes catch a traveling rainbow band (16/20 YAY — standing R2 nay broken).

Glyph slot: research/foil-glyphs//glyph[-N].svg -> lab-gated api routes -> atlas rasterizer (foil/glyphs.ts) -> uGlyphTex contract + glyphTex() helper; ~2.5s auto-pickup on save, procedural fallback when absent, README tells Chey exactly where his four promised files go.

Canon migration: reverse-sheet uP1 0 -> 0.6 (dead placeholder now drives his requested glyph grain). Verdicts + pixel-refutations recorded in research/foil-verification.md (R3-GLYPH); lessons in foil-effects SKILL.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

47039ff — 2026-08-03 — foil R3-MISC: the uTint ink-metallic blend term + Chey's twelve remaining canon-lab comments (12/12 yay)

The reverse-holo fix (his chat report): screen-blending achromatic foil light over colored art compresses chroma — "dull and grayish, rather than making the color look metallic." New core uniform uTint (default 0 = bit-exact legacy) multiplies the foil layer + in-mask specular by the luminance-normalized scan chroma SQUARED (the double ink pass of a real mirror reflection); neutral on blank cards, so no canon-lab appearance moved. Reverse-family recipes opt in; verified before/after at 5 tilt angles on Victini ball-reverse, Crystal Energy, Pineco. uP4/uP5 param slots added in the same contract change.

The twelve comments, each mechanism-mapped in its issue frontmatter and the verification doc: cracked-ice triangulated + half-off gate; fireworks grid lattice + center-emanating radial hue; rainbow-mirror rebuilt as mirror + hue-banded spotlight; tinsel-ii silver lines + vertical rainbow band (band lives only on the line-work); prism rebuilt as pinwheel-kin from the corpus-vs-corpus delta pass; cosmos-ii pixel-quantized stamp edges; crosshatch thinned; gold-secret adjustable burst origin (uP4/uP5, center default, per-card via overrides) + holographic grain; dedicated recipes for sequin (glyph-family glints), tcg-classic (flat starlight + rainbow glitter), acid-wash (water-web kin), disco (perfect disc lattice, galaxy ignition) — zero approxVia fallbacks remain.

Canon migrations (appearance-preserving unless he asked otherwise): fireworks uP0 3→4.5 (dead two-octave structure), gold-secret uP4/uP5 0.5 added, sequin uP0-uP3 re-keyed to the new recipe. Verdicts: research/foil-verification.md "R3-MISC" — 12/12 Gemini yay incl. the tcg-classic ink-scope re-roll.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b287fde — 2026-08-03 — foil R4-COMPOSITE: the ink-density invariant — foil adds pop, never washes text or mutes ink

Chey's two me05 comments (7rtnzx mirror reverse "blows out the darks/text", 19mo4l holo "darkens/muddies the colors") ruled a CORE COMPOSITE INVARIANT, engineered in shader.ts main(): ink density estimated from the scan (inkDark = darker-than-local-8-tap-field average, exactly 0 on flat blank bases; inkColor = chroma above the canon-lab-tone floor) now gates every layer — dark ink blocks the flash + shields the specular (text crisp at every tilt), ALL ink is exempt from uDarken substrate attenuation (printed color never muted; the mirror stays dark only where the mirror is visible), colored ink auto-tints its flash (max(uTint, inkColor)) and uInkPop pumps chroma under the flash so colors read MORE saturated and metallic.

New core uniforms uInkGuard (default 1) / uInkPop (default 0.5), both 0 = bit-exact legacy composite; sliders in both labs; canon files untouched (absent keys inherit the new defaults — catalog-wide fix, zero migrations).

Proof: canon-lab blank-card pairs (mirror dark+silver, cosmos, horizontal-sheen) pixel-identical (compare -metric AE 0) under a frame-stepped rAF harness with a same-settings AE-0 control, plus a Gemini sanity pass (all pairs "identical"). On-card before/after pairs across the 8-card sample (both issue cards, Victini, Pineco, Darkrai, Cetitan, Machamp WOTC, Hydreigon SIR; 3 tilts, 390px + desktop) in ~/.pokedex-dev-hub-legacy/foil-shots/r4-composite/. Both issues resolved with mechanism-mapped notes; composite contract documented in the foil-effects SKILL; DECISIONS + research/foil-verification.md updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

0485f55 — 2026-08-04 — foil R4b: scan-additive composite (rest parity by construction) + canon-lab card preview

Chey's Grubbin (me05-002, Pitch Black) review refuted R4's ink heuristics: the reverse still muddied the card at rest (measured ΔC +32, green→yellow), darkened the header under tilt (−22 L), and blew out text (glyph contrast −26%). New law on real scans (uScanBase 1): the scan is a photograph of the card at rest — foil is additive dynamic light only, tilt-gated to ZERO at rest, luminance-headroom + per-channel clamped (text sacred at every angle, uArtGate opens the clamp where dark scan pixels ARE the foil), chroma- preserving with a pastel-safe tint ramp and an L²-gated chroma pump. Canon lab blank bases run the classic composite textually unchanged: AE-0 proven (CDP frame-stepped harness, control + knob pairs; Gemini pass agreed). uInkGuard 0 still reproduces the pre-R4 composite exactly.

Canon-lab card preview (his request verbatim): blank/on-card/↻-another chips render the live sliders on a random catalog card the resolver assigns the pattern to — server-sampled from the baked inversion (tools/foil/ build-pattern-cards.mts → data/foil-pattern-cards.json, gitignored; GET /foil-lab/pattern-cards/:patternId), honoring hand masks + adjusted windows; empty-pool patterns say "no catalog cards". Also commits Chey's live mirror-canon tweak from the Grubbin session. Full metrics + verdicts: DECISIONS 2026-08-04.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b2058cf — 2026-08-04 — foil R4c: onset ease — the sheen leans in instead of switching on

Chey on R4b's rest gate: "it's like it isn't there at all until I've tilted the card a tiny bit and then it's like it just suddenly appears." Measured: the headroom clamp saturated R4b's smoothstep(.02,.28) gate, so ALL added light arrived inside tilt 0.03-0.13 (0 -> 16.1 dL mean) then froze flat to tilt 1.0 — a step smeared over 4% of the tilt range.

Now: one wide ease ramp = smoothstep(0, .45, |tilt|) (full at 0.45 — every R4b tilt-0.5 metric unchanged), flash gated by ramp^1.5 (cubic-slow start), specular by ramp^2.5 (the broad gloss band now trails the flash instead of leading it), plus a REST=0.006 sub-JND living rest sheen blended into both gates (rest +0.43 dL card-mean, max 2/255 — visually still the scan). After-curve: smooth S over 0->0.25, no dead zone, no knee; text clamp untouched (glyph contrast 33.6 -> 33.6 at flash peak); blank-card canon lab bit-level unchanged (fuzz-1% AE 0 vs control floor).

Harness + curves + filmstrips: ~/.pokedex-dev-hub-legacy/foil-shots/r4c/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

0ee7ba9 — 2026-08-04 — foil R4d: motion-first onset — tiny gestures move the sheen; glow arrives late; Chey gets the dials

Chey rejected R4c too ('It still lights up pretty noticeably when I tilt the card a tiny bit. I don't like it.') — measured the gesture→tilt mapping first this time: pointer is 0.0051 tilt/px at 390px and gyro is 0.036 tilt/deg, so a 30px thumb drag (tilt 0.154) was already getting HALF the full glow under R4c. The curve was never the felt variable.

• shader.ts: onset gates now uniform-driven — uOnsetRange (glow full at |tilt|, default 0.5) and uOnsetCurve (flash exponent, default 3.5; specular trails at +1). (0.45, 1.5) reproduces R4c exactly (verified to the third decimal); defaults are strictly softer at every tilt and equal from 0.5 up, so all R4b/R4c peak metrics are unchanged (Grubbin 17.976 ΔL / text 34.02, Machamp 18.035 / text 45.6 — exact).
• FoilLab + CanonLab: 'Onset range' / 'Onset curve' sliders, canon- stored like every other uniform; absent canon keys inherit defaults.
• patterns.ts: CoreDefaults admits the two keys per recipe.

Pointer-path proof (real pointermove, 390px, 30px drag, first 300ms): before mean ΔL climbs to 6.6 and rising; after stays within ±0.08 while per-pixel movement reaches 6.1 mean / 18.9 p95 — moves, doesn't light up. Blank canon lab at fuzz-1% AE 0 vs control. Artifacts in ~/.pokedex-dev-hub-legacy/foil-shots/r4d/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

11010e3 — 2026-08-05 — foil R5: the metallic composite — always on the card, energy redistributed, Chey's compositing dials

Chey's ruling (2026-08-05, verbatim in DECISIONS): no onset range at all — the effect is ALWAYS on the card; dial in how it composites onto the ink; metallic and spectral, not brightening.

• Onset machinery removed (R4c/R4d ramp, REST floor, uOnsetRange/uOnsetCurve sliders on both surfaces). No canon/override file ever stored the keys; CardViewer skips unknown uniforms, so stale keys anywhere are inert.
• R5 metalness law (scan path only): scan = albedo; the pattern field splits around PIVOT 0.40 — above = tinted-specular highlight, below = bounded multiplicative depth (mirror-turned-away darks, sqrt-concave). Energy redistributes; contrast + chroma carry the metallic read.
• Soft-knee headroom: the R4b hard clamp plateaued strong fields into a flat wash (the "brightened" read); the budget now compresses monotonically so pattern structure survives. Per-channel no-clip caps stay hard; glyph ink (inkDark) exempt from highlight AND depth — text contrast preserved (Grubbin 86->97 at peak, Machamp flat).
• New canon-stored dials on both surfaces: Metallic 0.6, Sheen strength 0.55, Sheen tint 0.5 (= the R4 chroma law exactly), Depth 0.55, Texture 1. uMetal 0 / uIntensity 0 / uInkGuard 0 all still yield the plain scan.
• Blank canon lab pixel-unchanged: CDP frame-stepped before/after AE 0 at fuzz 1% (mirror/cosmos/horizontal-sheen/starlight, tilted + rest); one Gemini vision pass agrees (all pairs identical). Shots + per-card rest pairs: ~/.pokedex-dev-hub-legacy/foil-shots/r5/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9c971c2 — 2026-08-07 — foil R5b: metalness is MIRROR ONLY — every other pattern gets its own colour back

Chey: "You applied the metallic treatment to EVERY HOLO PATTERN when I only wanted it for mirror. So now, every single holo pattern looks like mirror and you can barely see the pattern at all, it's like every single one is being applied in a way that it adds no rainbow color or anything to the card."

R5 shipped its metalness dials as GLOBAL_DEFAULTS, so all 43 recipes inherited a law built for one of them. uMetal is now the scan-path LAW SELECTOR and defaults to 0 — nothing can inherit metalness; mirror is the only recipe that opts in, in its own patterns.ts defaults.

• Metalness branch is the R5 code textually unmoved, so mirror's arithmetic is unchanged by construction. Proven: his hand-tuned canon on Grubbin me05-002
  • Tropius me05-001, rest + tilt 0.2/0.35/0.5, 390px and desktop — 16/16 frames AE 0 at fuzz 1%, same floor as the same-build control pair.
• uMetal 0 runs the ADDITIVE law: the pattern's own emission (uSheen gain) screen-combined per channel over a substrate-attenuated body, tinted toward the ink only as far as uSheenTint asks. Three fixes to what R5 broke — the scalar no-clip cap let the artwork's brightest channel throttle the whole flash; the quartic ink budget (shaped to starve mid-dark glyphs) starved ordinary mid-tone art instead; PIVOT ate the field before anything flashed. Cosmos measured 0.005% of art-window pixels changed >8% under R5.
• uDepth decoupled from uDarken: uDarken is the blank-canon substrate, frozen by 30 saved canon files, and 0 on exactly the pale reverse sheets that need one. The substrate follows the pattern's dark half so it buys contrast, not a uniform dim (a flat 0.34 cost mean dL -21.7 and 29% of text contrast).
• Per-recipe families in patterns.ts (PARTICLE / SHEET / WASH / PEARL) plus seven by-eye corrections, each judged on a card the resolver assigns.
• Onset machinery stays deleted. Text guarantee re-verified: glyphs matte and crisp at peak tilt on a reverse whose foil covers the text box.
• Blank-card canon appearance unchanged: 8 patterns AE 0 (6 direct; cracked-ice and tinsel-ii against the pre-R5b code with canon held constant, since Chey re-saved those two mid-wave).
• SKILL.md composite contract rewritten around per-recipe selection.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b8561ce — 2026-08-07 — foil R6: composite defaults re-derived from Chey's canons; the guard learns to tell text from artwork

Chey: "i re-adjusted cracked-ice. take a look at it, and then make the changes based on everything you're observing to make all of this better - i want it so that i have to do as little of work personally on all the rest as possible."

WHAT HIS CANONS TAUGHT

uSheenTint 0 on all three non-mirror canons -> 0 for every non-mirror family (was 0.15/0.5/0.85/0.6). Ink-tinting a pattern's own highlight is exactly "it adds no rainbow color or anything to the card". uSheen he raised all three -> families now 1.2-4.2 (were 1.0-2.6). uDepth NOT universal: 0 on cracked-ice, 1 on tinsel-ii and rgs. It darkens the card WHERE THE PATTERN IS NOT, so its cost scales with 1 - duty. Measured every recipe's duty cycle; his three choices line up with it, and it predicted both dials for the only canon that exercises the defaults. uSat 1 + wide spread -> shipped on the 13 recipes with no canon.

Families are renamed for how light LANDS (flash / line / stamp / field / pearl / metal) and stated explicitly as FoilPattern.family. The stamp tier exists because the first pass got it wrong: energy-symbols-ii lights 5.8% of the face, so a field substrate darkened the other 94% and the card read "someone dimmed this".

THE INK GUARD: SPLIT, DON'T WEAKEN

inkDark is a local-contrast measure, so it fired on every dark mark - including every outline and shading edge in the ILLUSTRATION (measured: 86-89% of strong hits inside the art window are artwork, not text). The additive law spends it as a hard coverage multiply, so guard 1 punched the pattern full of holes. That is why he kept pulling it to ~0.51/0.56 on sheet/wash patterns.

Now split by glyphness into inkGlyph (sacred, zero flash, unchanged) and inkDetail (artwork darks keep the pattern on a tightened budget). inkAny is byte-for-byte the old inkDark and every pre-R6 consumer still reads it.

Success test met: at guard 1 under R6, tinsel-ii and rgs look BETTER than his 0.51/0.56 do today - tinsel-ii's artwork is visible through the streaks instead of greyed out. Both canons migrated to 1, reason recorded in each file's note; every other value in them is his, untouched.

SLIDER RANGES (his second ask)

Four dials sat exactly on their cap: uIntensity 2->4, uSheen 3->6, uSat 1->2, uHueSpread 1.5->3. Not extended: uInkGuard (his carve-out), uMetal (a law selector), the 0..1 mix weights, and uDarken/uDepth (already fully spent at 1).

uSheen needed a shader change to be honest: the additive budget's ceiling was reached at uSheen 2.0, so the top third of the slider bought almost nothing - very likely why he parked two canons at the ceiling. It now resumes climbing above 3, leaving every stored value <= 3 bit-for-bit identical.

The 19 duplicated lines across the two surfaces are one CORE_SLIDERS table; uGrain is dimmed and labelled "metal law only" where it does nothing.

NEW: "Apply composite -> family" copies only the on-card dials to siblings, never pattern shape.

PROVEN, NOT ASSUMED (AE vs a base build compiled from the previous commit)

• blank canon room: AE 0 across all 44 implemented recipes x 4 tones x 3 tilts
• mirror: AE 0 on the card path too - pixel-frozen
• text at flash peak: +0.1 to -1.1 points of glyph/paper contrast on every reverse whose foil covers the text box; the library's two largest costs (rainbow-glitter, water-web) toned down per-recipe and re-checked by eye
• every implemented recipe eyeballed on 1-2 real assigned cards at rest and two tilts; shots in ~/.pokedex-dev-hub-legacy/foil-shots/r6/

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

142fe42 — 2026-08-07 — Mask provenance: sidecar v3, the AI learning loop, and the collapse safeguard

Chey (2026-08-07): mark AI-generated masks, let an AI learn from his hand-done ones, then let him correct the AI's mask so it "can observe the diff and continue to improve without me having to hand-paint all the masks."

derivation_method existed in v2 but the API HARDCODED it to 'hand' — harmless while only a Pencil touched the corpus, actively wrong once mask-refine's Flatten started writing machine geometry through the same save path.

Taxonomy (5 values, 4 cases that must never blur): layout-flatten machine geometry, unpainted weight 0 hand human, from a layout prior weight 1 hand-refined human, on an existing mask weight 1 ai generator, UNREVIEWED weight 0 ai-corrected AI proposed, human edited weight 0.6

The label is never taken from the client. The client reports only what the canvas was SEEDED with; writeMaskRecord derives the method by diffing saved pixels against what that seed rasterizes to. A machine label needs a full GeneratorIdentity, which HTTP callers cannot supply. One write path serves the route and generators both.

• Anti-feedback-collapse enforced in selectExemplars(), not by convention: unreviewed ai is weight 0 at any corpus size. Codify now routes through it.
• Corrections keep the parent PNG + change map + metrics (incl. a 4x4 grid of where corrections land) — the supervised pairs a generator lane consumes via the self-describing tuples manifest.
• Seam rule, measured not guessed: canvas roundRect vs the server SDF disagree on 389/330260 px, ALL in the 1px AA band. Without tolerating it every unpainted Flatten would stamp hand.
• Guard: if the file at the target path is an unreviewed ai mask it IS the parent, whatever the client claims — AI ancestry cannot be laundered. That guard caught a real bug during this build (ActionBtn passed React's MouseEvent into saveMask's optional arg, blanking derivation); derivation is now required on the PUT so it fails loudly instead.
• v1/v2 load forever (normalize-on-read); Chey's two masks migrated additively and still read HAND-PAINTED, agreement 0.641, pixels untouched.
• Trial batch wotc-window-trial-1: 6 WOTC Base holo rares from window-artgate@1, all ai/unreviewed, reversible. Leave-one-out on the human corpus beat the rect baseline (0.588 -> 0.688) — real, but n=2 same-set, so explicitly not validation.

Surfaces: provenance badge + detail in the workbench, Mask corpus panel, GET /foil-lab/masks/corpus (+?tuples/?exemplars), artifact routes, and corpus.ts report|exemplars|tuples|migrate. Pure tests added to CI (test:foil). Browser-verified at 390px and desktop. See DECISIONS.md 2026-08-07 and the mask-pipeline SKILL.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

c5ad7e2 — 2026-08-08 — foil mask-straighten: read the INTENT out of Chey's hand mask, and make the undo real

Chey, 2026-08-08: "it's impossible to get the lines really straight so I'm hoping you can get computer vision on the mask and card art in tandem to really see my intent there, and make the mask nice and crisp and straight on the lines I was trying to draw along. Will you try that? And then I can either approve or revert."

Smoothing a wobbly boundary gives a smoothly wobbly boundary. The premise here is that a hand-drawn foil boundary is an ATTEMPT TO TRACE A PRINTED EDGE — frame, art box, species strip, stage tag — and those are dead straight in the scan. So: contour the mask, cut each loop into near-axis runs (PCA orientation, wobble gaps bridged), robust-fit each, hunt the scan's directional gradient with a local Hough over (angle, offset) plus a proximity prior, replace, intersect adjacent lines into real corners, rasterize matching his own AA character.

The refusals are the feature, and each is a recorded param: weak evidence, or an ambiguous band of comparable ridges, may nudge but never relocate; with no artwork edge it straightens only to HIS own fit; short and curved runs pass through untouched. A first cut "closed" the Tropius top-left corner — which turned out to be him deliberately tracing around the BASIC stage tag. The stricter rules keep it.

• apps/api/src/foil/line-snap.ts the geometry, pure, no I/O
• generator.ts line-snap@1, a REFINER (requiresSource)
• generate-masks.ts run --refine, archives, restoring revert
• provenance.ts supersedes + verbatim sha256 undo archive

A generator write onto an existing mask now THROWS without explicit supersede { runId }; with it, every artifact the mask had is archived verbatim under superseded/./ with a self-describing manifest, and revert --run-id verifies the whole archive BEFORE deleting anything live, then restores byte-for-byte. supersedes is deliberately NOT correction: one is a human editing a machine (training signal), the other a machine replacing a human with nobody's agreement. --refine also refuses a source with exemplar weight 0, so a refiner can never eat its own unreviewed output.

Tropius me05-001/37184 (run straighten-tropius-1): 8 runs snapped to printed edges, 1 self-straightened, 13 left exactly as drawn, 8 corners closed, 2 specks dropped. Max correction 3.74px, 1.06% of the face changed, Jaccard 0.9798, confidence 0.557. Honest residual: the card's bottom border was too faint to accept, so that run self-straightened and the report says so.

test:foil 25/25 (9 geometry + 5 supersede/restore, incl. byte-for-byte round trip, a deliberate curve surviving, an ambiguous band refused, and a corrupt archive aborting before deletion). Revert exercised twice end to end, sha256sum clean. Shots: ~/.pokedex-dev-hub-legacy/foil-shots/mask-straighten/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id straighten-tropius-1

ae97986 — 2026-08-08 — foil R7: his eye overturns two banked mechanisms; "no catalog cards" gets a reason

Chey's 15 canon-lab comments (2026-08-07), all resolved in issues/foil/.

Eleven shader verdicts. Two overturn mechanisms built from earlier corpus — energy-symbols' parity CHECKERBOARD (R3-GLYPH, from his own words) and prism's solid-facet grid (R3-MISC, from a Gemini delta pass). Having seen both rendered he rejected them. A description is a hypothesis about the render, not a spec.

cosmos the 0.085 UV field slide he read as parallax is now uP4 "Surface drift" at 0.012; zero at rest, canon unchanged striped-v-sheen grey lines: the unconditional silver term is gone AND the second-order lobe cut 0.4 -> 0.16 (it drew a whole second, near-achromatic line population on the far side) rainbow-mirror hue axis is card X, blue -> green -> red, on BOTH the sheen and the spotlight; core whitening 0.85 -> 0.32 energy-symbols checkerboard gone: per-symbol response AXIS + PHASE + transition WIDTH over a low-freq cluster field energy-symbols-ii same treatment (it shared the two-bank mechanism) pokeball-hologram flat SDF icon -> shaded 3D sphere (hemisphere normal, tilt-following light, belt, hub, two-tone shell) tinsel the "blinds" were a literal full-card striation floor; demoted to uP2, default 0 (his canon stores 0) tinsel-ii line-work gated to the colour band, silver 0.60 -> 0.10, band width exposed as uP4 prism pinwheel's spoke fan WITHOUT the per-wedge phase that makes a pinwheel spin; density 33 -> 18 radiant hologram step 0.5 cell -> uP4 0.17, drive rescaled so total travel per unit tilt is unchanged rainbow-glitter invisible on card: bigger white-lifted flakes AND the family fixed (flash -> field; no substrate, no contrast)

Canon policy held: shader.ts / canon.ts / CardViewer.tsx / data/foil-canon have a ZERO diff, sheenGlsl() is byte-identical (all four sheen canons frozen), and every new behaviour went into a NEW uniform old canon files inherit. No stored value migrated. One recommendation left for him, not applied: tinsel-ii's starkness is substantially his own canon (uDepth 1 at ~135 lines).

"Why are there no catalog cards?" — asked four times, three answers. The baked inversion is now v2 and emits a machine-readable diagnosis the lab renders: outranked diagonal-sheen-left — 1,818 SM reverses ARE cited for it, but energy-symbols wins them (the emblem layer vs the sheet rotation; both true). A capped SECONDARY pool is now baked via a new citedFoilPatterns() export and served as via:'cited', so the preview works. Never flip a winner to fill a pool. class-absent pinwheel, pokeball-hologram — the catalog has reverse variants for ex1-ex5 ONLY; the whole late-EX reverse era is missing upstream. The UI now says so instead of "no catalog cards". fixed vertical-sheen-rainbow 0 -> 12 cards. known_residuals recorded twice that the 6 Holo Rare Energies of ex13/ex16 use Sheen but could not be targeted without a card list; the catalog names exactly six per set. Two cardId rows added citing the same Collexy quotes, both residuals closed.

Typecheck (all workspaces) + web build + foil tests green; 14 recipes compile clean; before/after shots on blank + assigned card at 390px and desktop in ~/.pokedex-dev-hub-legacy/foil-shots/r7/. SKILL R7 field notes + DECISIONS appended.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

5d9d070 — 2026-08-08 — foil edge-trace: the artwork holds the geometry, his mask says which of it to follow

Chey, after reviewing line-snap: "Really just 'straighten' isn't quite enough, really just detecting edges and actually tracing accurately around them is the real move." So the premise moves. line-snap treats his strokes as the geometry and nudges near-axis runs onto printed lines. edge-trace treats his mask as a STATEMENT OF INTENT and derives the boundary from the card's own edges.

apps/api/src/foil/edge-trace.ts (pure, no I/O): Di Zenzo colour structure tensor (Scharr per channel) -> NMS -> hysteresis; colour because a green field against a silver border reads 0.13 luminance/px and 44.3 tensor/px at the same edge. Then livewire: a corridor around his boundary (a hard bound, not a suggestion), anchors at fixed spacing PLUS every hard turn in his own line, Dijkstra on ridge + linked + proximity^2 + direction, sub-pixel refinement onto the |dI/dn| peak, and only THEN a straight fit where the traced path really is straight.

His own correction said the same thing before he did: 114 of his 627 correction px sit in one 40x50 box around the species strip's swept tail, which line-snap had squared off by intersecting two lines across it.

Two traps, both now locked by tests:

• the MAD-trim trap: a robust fit discards a small feature as an outlier, reports "straight to 0.2px RMS", and crisping then flattens what it ignored. straightMaxDevPx requires every point in a run to be close.
• the half-pixel frame: mask space centres pixels at x+0.5, a gradient array at integer x. All artwork lookups convert in one place.

Measured, not asserted. generate-masks.ts adherence scores any set of masks by distance to the NEAREST printed edge maximum, with line-snap's own luminance Sobel as the default probe so the incumbent gets the home field: his hand mask 50.7% within 1px / 0.847px mean line-snap@1 74.8% / 0.545px his correction 75.4% / 0.548px edge-trace@1 77.1% / 0.199px (colour probe: 98.4% / 0.082px)

Applied to Tropius me05-001/37184 as run edgetrace-tropius-1, ai/unreviewed, his ai-corrected mask preserved as parent + verbatim supersede archive. Revert proven byte-identical twice before shipping.

test:foil 35/35 (10 new). Typecheck all workspaces, api+mcp+web builds green. Visuals in ~/.pokedex-dev-hub-legacy/foil-shots/mask-edgetrace/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id edgetrace-tropius-1

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

---

foil/mask-edgetrace-batch

Tip c611e3f88df4f456eb7c67249cff1edcd0c7f21f · 54 commits · 2026-08-01 → 2026-08-09 · 8,577 words · 31 reproduced below.

76871d2 — 2026-08-01 — foil/main: workbench v1 — three.js viewer, 5-pattern starter library, era layout masks

Quarantined foil workbench at /pokedex/foil-lab (URL-only, linked from nowhere, chrome-free): real owned-card scans from the image cache, tilt via pointer/gyro(iOS permission)/manual with reduced-motion respected, uniform sliders, pattern + mask-scope overrides, mask overlay toggle, copy-recipe JSON. Pattern library v1 tuned against owned eras (Base/WOTC, SV, Mega Evolution): Starlight, Cosmos, SV default holo, SV reverse sheet, Cracked Ice — uniform contract (core + uP0..3) designed for the full 15-20 set. Era layout spec as data (era-layouts.json, rects measured on real scans); resolver v1 (series, rarity, variant kind) -> pattern + mask scope. SKILL.md x2 (foil-effects, mask-pipeline stub). main.tsx gains a dev-only double-createRoot guard (Vite HMR entry-invalidation footgun, see DECISIONS).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

b171fab — 2026-08-01 — foil/main v2: Pencil mask editing, comment queue, iPad two-column, branch api

Apple-Pencil hand-mask drawing on the workbench (MaskEditor: card-aligned canvas overlay, pen+mouse default w/ finger toggle, pressure brush, eraser, 12-step undo, touch-action none, starts from layout prior); masks persist as committed ground-truth artifacts data/foil-masks//.png+json (derivation_method hand) via the env-gated branch api (POKEDEX_FOIL_LAB=1, port 3712, PGPOOL_MAX=1 — inert in prod), auto-load and beat the layout tier (uMaskTex contract; CanvasTexture flipY=false, exactly one V flip). Comment button captures card/variant/pattern/scope/mask state + all slider values to issues/foil// (report.md + context.json), out of the fix-issues sweep. Era-grouped owned-card picker (WOTC / SWSH / SV+ME headings); two-column layout (viewer | controls) from 700px — verified at 390px, 744x1133, 1133x744, 1440x900 with simulated pen strokes. SKILL.md x2 updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

0f30c9f — 2026-08-02 — foil: sidecar v2 (prior+diff), artwork-keyed masks, comment linkage, codify ritual

• Every hand-mask save records its starting prior (era/scope/resolver rect + RESOLVER_VERSION) and the api renders .prior.png + .diff.png (green = human added, red = removed) + Jaccard agreement, server-side via a pure-JS PNG codec (no native addons). Saves without a parsable prior 400.
• Machamp mask backfilled (deterministic prior wotc/window/v1): agreement 0.6409 — his correction is the subject silhouette, the rule's first lesson.
• Artwork keying: variants of one cardId provably share one scan, so mask GETs alias to a sibling variant's mask when prior.scope matches (X-Foil-Mask- Alias-Of). Cross-card reprints unprovable from the catalog -> per-card only.
• Comments auto-link the exact saved mask state (maskFile/maskSavedAt/ maskAliasOf/maskHasPriorDiff).
• mask-pipeline SKILL: sidecar v2 storage, identity rule + limits, and the codify ritual; first worked example data/foil-masks/codified/wotc.md (n=1: WOTC window = art-window rect minus subject silhouette; noted in era-layouts.json wotc.notes).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

a64ad4b — 2026-08-02 — foil: any-card picker — full-catalog browse/search, Owned-only as a toggle

Chey: pull up ANY card in the catalog and tilt it, owned or not. Picker now browses the whole catalog by era -> set (paged strip, promo sets run 300+), adds full-catalog name search, and keeps Owned-only as a filter toggle (default on = prior behavior). Zero new endpoints: existing /series, /series/:slug, /sets/:setId (paged), /search, /cards/:cardId serve it all; unowned cards list their catalog variants (quantity 0) and resolve the same base-guess pattern path. Selection is never clobbered by filters; unmapped series get an honest "Other eras" bucket. Verified on unowned neo2-13 Umbreon (Auto - Starlight WOTC) at 390px + both iPad viewports.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

f95ccb2 — 2026-08-02 — foil: taxonomy integration — all 39 pattern types in the app, cited resolver

Mislabel corrections from the video-reference research, applied to apps/ (research docs untouched):

• sv-holo renamed vertical-sheen (it always rendered the HGSS-XY vertical bands, not SV's). PATTERN_ALIASES keeps old slugs resolving forever — sidecars, comment context.json, Copy-recipe JSON never orphan.
• Sheen family = ONE generator at four rotations + stripe option: five honest implemented slugs (vertical / horizontal=TRUE SV default "Mirage" / diagonal right "/" / diagonal left "" / striped "Line").
• Cosmos label drops "/ Galaxy" (Galaxy is a starlight synonym; synonym moved to the starlight entry). reverse-sheet description notes it ≈ pokeball-masterball stamped sheet.
• Vintage narrowed: starlight = Base/Jungle/Fossil only; Base Set 2 → Call of Legends holos = cosmos.

Pattern dropdown carries the FULL 39-slug vocabulary, grouped: implemented recipes first, then every gap type rendering via its nearest recipe and labeled "approx via ..." (amber note in the UI). Unimplemented != hidden.

Resolver v2: base guess reads the cited usage table (usage-index.json, derived from research/foil-pattern-usage.json by tools/foil/build-usage-index.mjs; set-name match -> era/series token match -> era-default heuristics). The winning row's confidence + citation hosts surface in the workbench ("Guess: set-level citation, high confidence (bulbapedia..., collexy...)"). Reverse rows citing 'mirror' are deprioritized per the research known-gap (ink-design evidence, not foil-pattern evidence). RESOLVER_VERSION -> 2 (scope semantics unchanged from v1).

Verified: typecheck + build green; Playwright at 390/744x1133/1133x744 (~/.pokedex-dev-hub-legacy/foil-shots/w2/) — Machamp auto=starlight set-level-high with his hand mask still resolving, SV holo auto=horizontal-sheen, SV reverse auto=energy-symbols-ii, fireworks approx labeling.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

2cef421 — 2026-08-02 — foil: fix diffuse darkening — scan sampled sRGB-decoded but never re-encoded (issue ls9u0y)

The face texture was uploaded SRGB8_ALPHA8 (hardware sRGB->linear decode at sample time) while the foil ShaderMaterial writes gl_FragColor raw — three.js only appends the linear->sRGB colorspace_fragment chunk to built-in materials — so every scan rendered in linear values displayed as sRGB (measured flat 184->123 / 199->147, exactly the sRGB->linear curve). Sample the scan undecoded (NoColorSpace): the compositing model is authored in display space and pattern=none at rest is now pixel-comparable to the flat view (mean abs luma diff 44.55 -> 2.16, mapping identity +/-2). Also zero pattern-none's default specular so the baseline is truly the plain scan.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

55f66ca — 2026-08-02 — foil: workbench split — canon pattern lab (A) | card adjustment surface (B)

Chey's comment issues/foil/2026-08-02_12-59-52-368_4aq756 (resolved here):

• /pokedex/foil-lab/canon (CanonLab.tsx): bare holofoil pattern on a blank card (no ink; black/dark/silver/white flat-tone base via the normal texture path) beside the REAL reference tilt clip + 8 keyframes streamed from research/foil-video-reference/ by new env-gated dev routes (byte-range via res.sendFile for iOS Safari). Full 39-slug picker, sliders, tilt, and Save canon -> data/foil-canon/.json (full uniform snapshot that replaces recipe code defaults as the baseline).
• /pokedex/foil-lab (FoilLab.tsx reframed): per-card masks/comments unchanged, plus per-card uniform overrides saved as a SPARSE diff vs the canon baseline -> data/foil-overrides//.json (untouched uniforms keep tracking canon). Off-canon sliders get a dot; saving with sliders matching canon deletes the override.
• Layering model in foil/canon.ts (code defaults < canon < override < live); shared atoms + surface tabs in foil/ui.tsx; routes wired in main.tsx; AppShell chrome-free check widened to /foil-lab/*; both data dirs re-included in .gitignore; ports 5184/3713 recorded in ORCHESTRATION.md.

Verified in-browser at 390x844 / 744x1133 / 1133x744 (clip playing beside the bare cosmos render; Machamp base1-8 hand mask still resolves) — shots in ~/.pokedex-dev-hub-legacy/foil-shots/canon-lab/. Typecheck + build green (web, api). patterns.ts untouched (R0 re-tunes concurrently).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

75b1fc7 — 2026-08-02 — foil R0: re-tune implemented recipes per Gemini verification notes

cosmos: full activation-model rewrite — dark field, smaller/denser orbs mostly near-invisible, cluster pops via low-freq noise over cell ids, pinprick 4-point twinkles (was: dense wall of large saturated orbs). starlight (+ starlight-ii via shared GLSL): parallax architecture untouched; tighter visibility lobe (pow 5->9, floor .18->.08), sharper glyph cores + narrower flares, smaller blobs, saturated star color, galaxy wash halved; starlight-ii art gate lowered for the mid-orange Evolutions field. cracked-ice: authored intra-shard grain removed per Chey's accuracy ruling — facets flash as solid smooth planes; whitening at peak reduced, uSat up. sheen family: generator grows sharp/beam/barcode options. vertical-sheen -> new barcode variant (thin spectral lines of varying width, CD-groove slide) with tamed beam + art gate. diagonals: sharp 3.0, beam 0.55, specular 0.35 (center blow-out), diagonal-sheen-right uP0 2->7 (same-sheet consistency). striped-vertical-sheen: stripes finer + more blended (90->130, floor .40). horizontal-sheen: untouched (20/20).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

e520d3a — 2026-08-02 — foil R0: rounds 2-3 re-tunes + verification report, field notes, decisions

starlight: longer thin-armed 4-point glyphs + 8-point-burst subset, tighter pop lobe (pow 11, floor .04), full-saturation star color, parallax default 2.4, gain 3.6 — best of 3 rounds 11/20 (still-frame judging cannot see the parallax; recorded as honest residual with a video-judge recommendation). starlight-ii: 20/20 (uP2 pinned so base-default drift can't invalidate it). cracked-ice: facet gain 1.1 -> 0.55 (art visible through flashes), density 10, less whitening — 15/20 yay; shard-geometry uniformity queued for R1. Verification report gains the R0 section (before->after per pattern, the diag-left judge-noise event, diffuse-fix supersession note); foil-effects SKILL.md gains the R0 field notes; DECISIONS entry for root cause + wave.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2b0c135 — 2026-08-02 — foil: cited per-card pattern assignments across the whole catalog (resolver v3)

Chey's mlmwmp comment: the library was funneling to era defaults ("cosmos"). Full-catalog enumeration (20,964 cards / 37,627 variants, via the branch read api) + an 18-lane Ringer research swarm (foil-card-assignments, GLM-5.2, executed selector-integrity checks) produced:

• research/foil-card-assignments.json — 111 set/rarity/kind/card-level rows, 19 catalog foil-facet mappings (holo-foil-cosmos, reverse-foil-masterball, normal-foil-galaxy...), 55 honest residuals; per-lane provenance + method.
• tools/foil/build-assignments-index.mjs -> apps/web/src/foil/assignments-index.json (trimmed for the bundle, duplicate-facet guard).
• resolver v3 (RESOLVER_VERSION=3): assignment tier above the v2 usage table; specificity cardIds > variantKinds > facet > rarities > set > v2 > heuristic; normal-kind variants with a declared foil facet now render window foil (previously scope none); FoilLab passes cardId.
• era-layouts: wotc seriesSlugs 'ecard' -> 'e-card', + 'legendary-collection' (both had fallen through to modern-sv rects and heuristics).

Coverage (foil-bearing variants): 79.5% set-cited / 20.4% era-token / 0.2% heuristic -> 93.4% set / 4.3% facet-or-card / 2.4% era-token / 0 heuristic. 13 in-browser spot-checks in ~/.pokedex-dev-hub-legacy/foil-shots/assignments/. Issue mlmwmp resolved. Weak spots + curation ledger in DECISIONS.md.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2e537da — 2026-08-02 — foil R1: dedicated recipes for the twelve owned-era gap patterns (10/12 verified)

Chey's comment dml369 ("develop out the rest of the missing holofoil patterns"): dedicated GLSL for fireworks, ace-spec, energy-symbols-ii, rainbow-glitter-sheen, ex-starfoil, prismatic-pokeball, tinsel-ii, cosmos-iii-smooth, pokeball-masterball, radiant, rainbow-glitter, confetti — implemented flags flipped, approx labels dropped, slugs stable. Gemini-verified 10/12 match over 3 rounds (verdict table + honest residuals in research/foil-verification.md R1 section). Honest nays: tinsel-ii (procedural static plateaus at judge score 2) and prismatic-pokeball (screen-only blend cannot darken a white reverse body — structural, Chey's call).

R0 residuals cleared in the same pass: cracked-ice anisotropic shattered-glass metric (re-judged yay 14/20); starlight milky wash + per-layer visibility curves with starlight-ii re-banked clean 20/20 through both shared-GLSL changes; starlight parallax judged from a new 16-frame fine sweep and still scored flat — verified real by eye, recorded as judge-blindness.

Field notes appended to foil-effects SKILL.md; dated DECISIONS entry; issue dml369 resolved.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

205f783 — 2026-08-02 — foil R2: uDarken blend-model term — prismatic-pokeball 17/20 yay, tinsel-ii 16/20 yay

One opt-in substrate-darkening term in the shared composite (main()): body = scan * (1 - uDarken * mask * gate) before the foil screen-blends. Default 0 is bit-identical to the old render; absent canon/override keys seed 0. Full 21-pattern regression sweep on provably identical renders: no true regressions (and a measurement of Gemini single-roll noise, ±3-6; diag-right + pokeball-masterball double-nays recorded as unstable verdicts on geometry proof). prismatic-pokeball rebuilt on the term (dark-mirror base, facet-quantized flash lobe, ball watermark as ink-overprint suppression): structural nay -> 17/20 yay. tinsel-ii one-line opt-in (uDarken 0.4): static plateau -> 16/20 yay. 20 of 21 implemented types now match; sole nay starlight (judge parallax blindness).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

4d72237 — 2026-08-02 — foil R2 stage 3: thirteen dedicated recipes (gap patterns + RC dots) + subset-commons resolver tier

• patterns.ts: real recipes for mirror, rainbow-mirror, vertical-sheen-rainbow (dark-mirror family on uDarken), pokeball-hologram, prism, water-web, energy-symbols, pinwheel, ex-emerald, tinsel, cosmos-ii-pixel, crosshatch, radiant-collection-dots — implemented: true, approx labels dropped, slugs stable. Unused plain SHEEN_V generator removed (last fallback users gone).
• resolver v4: card-level 'normal'-class assignment rows consulted before the scope-none early return — subset cards the catalog declares plain but which physically carry foil (RC commons dot overprint) resolve their subset's pattern with scope full.
• foil-card-assignments.json: two card-level rows (bw11-RC1..20, g1 25 non-ultra RC ids) with video citation (non-holo RC Teddiursa keeps the dot overprint); builder enforces cardIds on 'normal' rows; index regenerated.

Verified in-browser: g1-RC1 + bw11-RC1 render the RC dot treatment via Auto (guess card|medium). Gemini verdicts in research/foil-verification.md (R2 recipe-wave section, next commit).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

49548d4 — 2026-08-02 — foil R2b: four §40–43 vocabulary recipes (all match) + resolver v5 scope overrides

The last lane of the wave — dedicated recipes for the vocab-extension types: gold-secret (warm-locked goldRamp field + chromatic glitter pops + SWSH burst rays — 19/20 yay), vstar-pearl (near-white pearl on uDarken 0.3, warm-led diagonal wash, frame-line streaks — 13/20 yay), shiny-vault (silvery field + band-keyed sparkle-glyph amplifiers, uDarken 0.15 — 11 nay → 16/20 yay r2), detective-pikachu (diagonal sheen × photo-luminance via the documented uFace contract exception — 20/20 yay). 38/43 taxonomy types real, 34 hold match.

Resolver v5: assignment rows carry an optional per-row scope override (window/ full/sheet), honored by the set-row tier and the cls-'normal' tier — baby shinies render window despite the 'Shiny rare' full-foil rarity, VSTARs full despite plain holo kinds, det1 window despite 'ultra rare'.

Assignment research: +2 facet rows (gold, gold-jumbo → gold-secret; retires both facet residuals, 1-card metal collateral flagged) and +10 cited rows — sma 51 baby/35 GX/8 gold splits, swsh4.5sv 104 baby + 18 full-face (incl. the black shiny Eternatus pair, name-verified not gold), sv04.5 rarity split, VSTAR via the dedicated 'Holo Rare VSTAR' rarity, det1 ultra-rare + cls-normal rows (all 18 cards window). Residuals annotated with resolutions.

build-usage-index.mjs: tolerate era-wide rows with no scope.sets (the vocab gold rows crashed the build — the 122-row usage index never compiled until now). Verdicts/frames under ~/.pokedex-dev-hub-legacy/foil-verify/ (verify-

-r2b*); shots in foil-shots/r2b/. Docs: foil-verification.md R2b section, SKILL R2b field notes, DECISIONS entry.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

11ed90a — 2026-08-03 — foil R3: sheen-family streak-field rework — all 8 of Chey's canon-lab comments resolved, 7/7 Gemini 20/20

sheenGlsl is now a streak field (Chey's critique = the spec): sparse irregular streaks, per-streak lean following card tilt (crisscross; converging pairs terminate where they meet), stretched-ellipse taper, hue running ALONG each strip as well as across. striped-vertical-sheen rebuilt from a fresh Gemini corpus re-spec run with Chey's description embedded (grouped reveal via wide sweeping windows over a subtle bottom-converging fan; gemini-spec-r3.md); ex-starfoil rebased on the reworked diagonal base + fine CD-line layer; vstar-pearl rebuilt from the horizontal-sheen streak field, pearl kept.

Canon migrations (full-snapshot rule, DECISIONS): striped + horizontal uDarken 0 -> 0.32 (bright-substrate legibility, third confirmation). All other canon values carry over; uniform semantics unchanged. validate_spec.py frame regex single-digit fix (undercounted two-digit citations, failed an honest worker).

Verdicts (Chey's words folded into every judge prompt as acceptance criteria): vertical/right/left/vstar-pearl 20/20 first roll; striped + ex-starfoil 20/20 round 2; horizontal 20/20 round 3. Table + residuals in research/foil-verification.md R3.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

af6c32c — 2026-08-03 — foil R3-MOTION: axis-split starlight, per-dot cosmos, hologram radiant, the rainbow-glitter-sheen delta

Chey's six motion-model canon-lab comments resolved (his eye = ground truth):

• starlight (+ starlight-ii via shared GLSL): top-to-bottom hue BANDING (soft-quantized bands; his canon 0.62/0.6 lands the reference order blue->green->red on the R->B->G cosine ramp) + AXIS-SPLIT tilt: vertical slides the whole field (~10% card height), horizontal drives the per-star random fade; his 3-layer uP1 parallax rides on top untouched. Judge: starlight 17/20 yay round 2 — the FIRST starlight yay (standing parallax nay broken); starlight-ii 20/20.
• cosmos: whole-field slide + per-dot random response (independent random tilt axes for brightness and hue per orb — no shared sweep axis); at tilt 0 his canon appearance is pixel-identical. Judge: 20/20.
• radiant: discrete hologram steps (half-cell up/down, 60% hold / 40% crossfade — one position fades out as the next fades in), uP1 wired as Hologram travel (canon migrated 0 -> 2.2; was an unused placeholder), grid lines ~45% thicker. Judge nay 13/20 across 3 rounds on a pixel-refuted "slides continuously" claim — recorded as still-frame motion blindness (same class as starlight parallax); adjacent-frame crops prove discrete positions + crossfade. Chey's live tilt is the tiebreak.
• rainbow-glitter-sheen: dedicated Gemini delta-articulation pass first (he couldn't name the difference), pixel-verified, then fixed: narrow laser-saturated striped chevron (sigma 0.010->0.0035, hue traversal 5x->9x, pow-deepened at low gain), fainter +-0.8 repeats, chevron angle 1.3->1.9, denser colored always-visible glitter, uDarken 0->0.4 (4th legibility-physics data point). Judge: 19/20 yay.

Docs: research/foil-verification.md R3-MOTION section, DECISIONS entry, foil-effects SKILL field notes. Frames/verdicts in ~/.pokedex-dev-hub-legacy/foil-verify (jobs -r3m, manifest-r3m{,2,3}, delta-rainbow-glitter-sheen); gate shots in foil-shots/r3-motion.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9d74475 — 2026-08-03 — foil mask-refine: window handles → flatten → hand refine

Chey's requested workflow on the Card-adjust surface: draggable corner/edge handles adjust the layout window rect per card (WindowEditor.tsx, 44px touch targets, finger-drag allowed), the adjusted geometry persists as artwork-keyed scope-agnostic sidecars (data/foil-windows//.json, /foil-lab/windows routes), and Flatten bakes it through the STANDARD hand-mask save path (sidecar prior stays the era rule; new optional prior.window records the adjustment) then opens the existing paint tooling.

• rasterizeWindowRect shared by Flatten + loadLayoutRect (identical bakes)
• saveMask prior = layoutMask (rule), never the effective mask
• windows GET aliases across variants of the same scan, newest savedAt
• verified in-browser desktop + 390px: sma-SV1/sma-SV2 (SM baby shiny, borrowed modern-sv rect) and me01-034 (zero-coverage gap); base1-8 + base1-5 hand-mask regression clean; test artifacts deleted (corpus rule)
• ports: vite 5186 / api 3714 recorded in ORCHESTRATION.md

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

66a3210 — 2026-08-03 — foil R3-GLYPH: glyph-pattern reworks per Chey's 7 comments + the drop-in glyph slot for his SVGs

Behavior (his notes verbatim = acceptance criteria): reverse-sheet rainbow only on grainy glyphs over a neutral sheet; energy-symbols checkerboard banks that swap with tilt; energy-symbols-ii sporadic random banks (~half faint at any time); ace-spec diamonds grow/shrink with tilt + a touch of blur; prismatic-pokeball ball never darkens (white-mixed pale response + coherent plane flash, suppression deleted); radiant-collection-dots shapes catch a traveling rainbow band (16/20 YAY — standing R2 nay broken).

Glyph slot: research/foil-glyphs//glyph[-N].svg -> lab-gated api routes -> atlas rasterizer (foil/glyphs.ts) -> uGlyphTex contract + glyphTex() helper; ~2.5s auto-pickup on save, procedural fallback when absent, README tells Chey exactly where his four promised files go.

Canon migration: reverse-sheet uP1 0 -> 0.6 (dead placeholder now drives his requested glyph grain). Verdicts + pixel-refutations recorded in research/foil-verification.md (R3-GLYPH); lessons in foil-effects SKILL.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

1a91aa0 — 2026-08-03 — foil R3-MISC: the uTint ink-metallic blend term + Chey's twelve remaining canon-lab comments (12/12 yay)

The reverse-holo fix (his chat report): screen-blending achromatic foil light over colored art compresses chroma — "dull and grayish, rather than making the color look metallic." New core uniform uTint (default 0 = bit-exact legacy) multiplies the foil layer + in-mask specular by the luminance-normalized scan chroma SQUARED (the double ink pass of a real mirror reflection); neutral on blank cards, so no canon-lab appearance moved. Reverse-family recipes opt in; verified before/after at 5 tilt angles on Victini ball-reverse, Crystal Energy, Pineco. uP4/uP5 param slots added in the same contract change.

The twelve comments, each mechanism-mapped in its issue frontmatter and the verification doc: cracked-ice triangulated + half-off gate; fireworks grid lattice + center-emanating radial hue; rainbow-mirror rebuilt as mirror + hue-banded spotlight; tinsel-ii silver lines + vertical rainbow band (band lives only on the line-work); prism rebuilt as pinwheel-kin from the corpus-vs-corpus delta pass; cosmos-ii pixel-quantized stamp edges; crosshatch thinned; gold-secret adjustable burst origin (uP4/uP5, center default, per-card via overrides) + holographic grain; dedicated recipes for sequin (glyph-family glints), tcg-classic (flat starlight + rainbow glitter), acid-wash (water-web kin), disco (perfect disc lattice, galaxy ignition) — zero approxVia fallbacks remain.

Canon migrations (appearance-preserving unless he asked otherwise): fireworks uP0 3→4.5 (dead two-octave structure), gold-secret uP4/uP5 0.5 added, sequin uP0-uP3 re-keyed to the new recipe. Verdicts: research/foil-verification.md "R3-MISC" — 12/12 Gemini yay incl. the tcg-classic ink-scope re-roll.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

5d35aaf — 2026-08-03 — foil R4-COMPOSITE: the ink-density invariant — foil adds pop, never washes text or mutes ink

Chey's two me05 comments (7rtnzx mirror reverse "blows out the darks/text", 19mo4l holo "darkens/muddies the colors") ruled a CORE COMPOSITE INVARIANT, engineered in shader.ts main(): ink density estimated from the scan (inkDark = darker-than-local-8-tap-field average, exactly 0 on flat blank bases; inkColor = chroma above the canon-lab-tone floor) now gates every layer — dark ink blocks the flash + shields the specular (text crisp at every tilt), ALL ink is exempt from uDarken substrate attenuation (printed color never muted; the mirror stays dark only where the mirror is visible), colored ink auto-tints its flash (max(uTint, inkColor)) and uInkPop pumps chroma under the flash so colors read MORE saturated and metallic.

New core uniforms uInkGuard (default 1) / uInkPop (default 0.5), both 0 = bit-exact legacy composite; sliders in both labs; canon files untouched (absent keys inherit the new defaults — catalog-wide fix, zero migrations).

Proof: canon-lab blank-card pairs (mirror dark+silver, cosmos, horizontal-sheen) pixel-identical (compare -metric AE 0) under a frame-stepped rAF harness with a same-settings AE-0 control, plus a Gemini sanity pass (all pairs "identical"). On-card before/after pairs across the 8-card sample (both issue cards, Victini, Pineco, Darkrai, Cetitan, Machamp WOTC, Hydreigon SIR; 3 tilts, 390px + desktop) in ~/.pokedex-dev-hub-legacy/foil-shots/r4-composite/. Both issues resolved with mechanism-mapped notes; composite contract documented in the foil-effects SKILL; DECISIONS + research/foil-verification.md updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

73bf70f — 2026-08-04 — foil R4b: scan-additive composite (rest parity by construction) + canon-lab card preview

Chey's Grubbin (me05-002, Pitch Black) review refuted R4's ink heuristics: the reverse still muddied the card at rest (measured ΔC +32, green→yellow), darkened the header under tilt (−22 L), and blew out text (glyph contrast −26%). New law on real scans (uScanBase 1): the scan is a photograph of the card at rest — foil is additive dynamic light only, tilt-gated to ZERO at rest, luminance-headroom + per-channel clamped (text sacred at every angle, uArtGate opens the clamp where dark scan pixels ARE the foil), chroma- preserving with a pastel-safe tint ramp and an L²-gated chroma pump. Canon lab blank bases run the classic composite textually unchanged: AE-0 proven (CDP frame-stepped harness, control + knob pairs; Gemini pass agreed). uInkGuard 0 still reproduces the pre-R4 composite exactly.

Canon-lab card preview (his request verbatim): blank/on-card/↻-another chips render the live sliders on a random catalog card the resolver assigns the pattern to — server-sampled from the baked inversion (tools/foil/ build-pattern-cards.mts → data/foil-pattern-cards.json, gitignored; GET /foil-lab/pattern-cards/:patternId), honoring hand masks + adjusted windows; empty-pool patterns say "no catalog cards". Also commits Chey's live mirror-canon tweak from the Grubbin session. Full metrics + verdicts: DECISIONS 2026-08-04.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

fac400c — 2026-08-04 — foil R4c: onset ease — the sheen leans in instead of switching on

Chey on R4b's rest gate: "it's like it isn't there at all until I've tilted the card a tiny bit and then it's like it just suddenly appears." Measured: the headroom clamp saturated R4b's smoothstep(.02,.28) gate, so ALL added light arrived inside tilt 0.03-0.13 (0 -> 16.1 dL mean) then froze flat to tilt 1.0 — a step smeared over 4% of the tilt range.

Now: one wide ease ramp = smoothstep(0, .45, |tilt|) (full at 0.45 — every R4b tilt-0.5 metric unchanged), flash gated by ramp^1.5 (cubic-slow start), specular by ramp^2.5 (the broad gloss band now trails the flash instead of leading it), plus a REST=0.006 sub-JND living rest sheen blended into both gates (rest +0.43 dL card-mean, max 2/255 — visually still the scan). After-curve: smooth S over 0->0.25, no dead zone, no knee; text clamp untouched (glyph contrast 33.6 -> 33.6 at flash peak); blank-card canon lab bit-level unchanged (fuzz-1% AE 0 vs control floor).

Harness + curves + filmstrips: ~/.pokedex-dev-hub-legacy/foil-shots/r4c/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

bbb6476 — 2026-08-04 — foil R4d: motion-first onset — tiny gestures move the sheen; glow arrives late; Chey gets the dials

Chey rejected R4c too ('It still lights up pretty noticeably when I tilt the card a tiny bit. I don't like it.') — measured the gesture→tilt mapping first this time: pointer is 0.0051 tilt/px at 390px and gyro is 0.036 tilt/deg, so a 30px thumb drag (tilt 0.154) was already getting HALF the full glow under R4c. The curve was never the felt variable.

• shader.ts: onset gates now uniform-driven — uOnsetRange (glow full at |tilt|, default 0.5) and uOnsetCurve (flash exponent, default 3.5; specular trails at +1). (0.45, 1.5) reproduces R4c exactly (verified to the third decimal); defaults are strictly softer at every tilt and equal from 0.5 up, so all R4b/R4c peak metrics are unchanged (Grubbin 17.976 ΔL / text 34.02, Machamp 18.035 / text 45.6 — exact).
• FoilLab + CanonLab: 'Onset range' / 'Onset curve' sliders, canon- stored like every other uniform; absent canon keys inherit defaults.
• patterns.ts: CoreDefaults admits the two keys per recipe.

Pointer-path proof (real pointermove, 390px, 30px drag, first 300ms): before mean ΔL climbs to 6.6 and rising; after stays within ±0.08 while per-pixel movement reaches 6.1 mean / 18.9 p95 — moves, doesn't light up. Blank canon lab at fuzz-1% AE 0 vs control. Artifacts in ~/.pokedex-dev-hub-legacy/foil-shots/r4d/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

571bc99 — 2026-08-05 — foil R5: the metallic composite — always on the card, energy redistributed, Chey's compositing dials

Chey's ruling (2026-08-05, verbatim in DECISIONS): no onset range at all — the effect is ALWAYS on the card; dial in how it composites onto the ink; metallic and spectral, not brightening.

• Onset machinery removed (R4c/R4d ramp, REST floor, uOnsetRange/uOnsetCurve sliders on both surfaces). No canon/override file ever stored the keys; CardViewer skips unknown uniforms, so stale keys anywhere are inert.
• R5 metalness law (scan path only): scan = albedo; the pattern field splits around PIVOT 0.40 — above = tinted-specular highlight, below = bounded multiplicative depth (mirror-turned-away darks, sqrt-concave). Energy redistributes; contrast + chroma carry the metallic read.
• Soft-knee headroom: the R4b hard clamp plateaued strong fields into a flat wash (the "brightened" read); the budget now compresses monotonically so pattern structure survives. Per-channel no-clip caps stay hard; glyph ink (inkDark) exempt from highlight AND depth — text contrast preserved (Grubbin 86->97 at peak, Machamp flat).
• New canon-stored dials on both surfaces: Metallic 0.6, Sheen strength 0.55, Sheen tint 0.5 (= the R4 chroma law exactly), Depth 0.55, Texture 1. uMetal 0 / uIntensity 0 / uInkGuard 0 all still yield the plain scan.
• Blank canon lab pixel-unchanged: CDP frame-stepped before/after AE 0 at fuzz 1% (mirror/cosmos/horizontal-sheen/starlight, tilted + rest); one Gemini vision pass agrees (all pairs identical). Shots + per-card rest pairs: ~/.pokedex-dev-hub-legacy/foil-shots/r5/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

4a322e1 — 2026-08-07 — foil R5b: metalness is MIRROR ONLY — every other pattern gets its own colour back

Chey: "You applied the metallic treatment to EVERY HOLO PATTERN when I only wanted it for mirror. So now, every single holo pattern looks like mirror and you can barely see the pattern at all, it's like every single one is being applied in a way that it adds no rainbow color or anything to the card."

R5 shipped its metalness dials as GLOBAL_DEFAULTS, so all 43 recipes inherited a law built for one of them. uMetal is now the scan-path LAW SELECTOR and defaults to 0 — nothing can inherit metalness; mirror is the only recipe that opts in, in its own patterns.ts defaults.

• Metalness branch is the R5 code textually unmoved, so mirror's arithmetic is unchanged by construction. Proven: his hand-tuned canon on Grubbin me05-002
  • Tropius me05-001, rest + tilt 0.2/0.35/0.5, 390px and desktop — 16/16 frames AE 0 at fuzz 1%, same floor as the same-build control pair.
• uMetal 0 runs the ADDITIVE law: the pattern's own emission (uSheen gain) screen-combined per channel over a substrate-attenuated body, tinted toward the ink only as far as uSheenTint asks. Three fixes to what R5 broke — the scalar no-clip cap let the artwork's brightest channel throttle the whole flash; the quartic ink budget (shaped to starve mid-dark glyphs) starved ordinary mid-tone art instead; PIVOT ate the field before anything flashed. Cosmos measured 0.005% of art-window pixels changed >8% under R5.
• uDepth decoupled from uDarken: uDarken is the blank-canon substrate, frozen by 30 saved canon files, and 0 on exactly the pale reverse sheets that need one. The substrate follows the pattern's dark half so it buys contrast, not a uniform dim (a flat 0.34 cost mean dL -21.7 and 29% of text contrast).
• Per-recipe families in patterns.ts (PARTICLE / SHEET / WASH / PEARL) plus seven by-eye corrections, each judged on a card the resolver assigns.
• Onset machinery stays deleted. Text guarantee re-verified: glyphs matte and crisp at peak tilt on a reverse whose foil covers the text box.
• Blank-card canon appearance unchanged: 8 patterns AE 0 (6 direct; cracked-ice and tinsel-ii against the pre-R5b code with canon held constant, since Chey re-saved those two mid-wave).
• SKILL.md composite contract rewritten around per-recipe selection.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2c73ef0 — 2026-08-07 — foil R6: composite defaults re-derived from Chey's canons; the guard learns to tell text from artwork

Chey: "i re-adjusted cracked-ice. take a look at it, and then make the changes based on everything you're observing to make all of this better - i want it so that i have to do as little of work personally on all the rest as possible."

WHAT HIS CANONS TAUGHT

uSheenTint 0 on all three non-mirror canons -> 0 for every non-mirror family (was 0.15/0.5/0.85/0.6). Ink-tinting a pattern's own highlight is exactly "it adds no rainbow color or anything to the card". uSheen he raised all three -> families now 1.2-4.2 (were 1.0-2.6). uDepth NOT universal: 0 on cracked-ice, 1 on tinsel-ii and rgs. It darkens the card WHERE THE PATTERN IS NOT, so its cost scales with 1 - duty. Measured every recipe's duty cycle; his three choices line up with it, and it predicted both dials for the only canon that exercises the defaults. uSat 1 + wide spread -> shipped on the 13 recipes with no canon.

Families are renamed for how light LANDS (flash / line / stamp / field / pearl / metal) and stated explicitly as FoilPattern.family. The stamp tier exists because the first pass got it wrong: energy-symbols-ii lights 5.8% of the face, so a field substrate darkened the other 94% and the card read "someone dimmed this".

THE INK GUARD: SPLIT, DON'T WEAKEN

inkDark is a local-contrast measure, so it fired on every dark mark - including every outline and shading edge in the ILLUSTRATION (measured: 86-89% of strong hits inside the art window are artwork, not text). The additive law spends it as a hard coverage multiply, so guard 1 punched the pattern full of holes. That is why he kept pulling it to ~0.51/0.56 on sheet/wash patterns.

Now split by glyphness into inkGlyph (sacred, zero flash, unchanged) and inkDetail (artwork darks keep the pattern on a tightened budget). inkAny is byte-for-byte the old inkDark and every pre-R6 consumer still reads it.

Success test met: at guard 1 under R6, tinsel-ii and rgs look BETTER than his 0.51/0.56 do today - tinsel-ii's artwork is visible through the streaks instead of greyed out. Both canons migrated to 1, reason recorded in each file's note; every other value in them is his, untouched.

SLIDER RANGES (his second ask)

Four dials sat exactly on their cap: uIntensity 2->4, uSheen 3->6, uSat 1->2, uHueSpread 1.5->3. Not extended: uInkGuard (his carve-out), uMetal (a law selector), the 0..1 mix weights, and uDarken/uDepth (already fully spent at 1).

uSheen needed a shader change to be honest: the additive budget's ceiling was reached at uSheen 2.0, so the top third of the slider bought almost nothing - very likely why he parked two canons at the ceiling. It now resumes climbing above 3, leaving every stored value <= 3 bit-for-bit identical.

The 19 duplicated lines across the two surfaces are one CORE_SLIDERS table; uGrain is dimmed and labelled "metal law only" where it does nothing.

NEW: "Apply composite -> family" copies only the on-card dials to siblings, never pattern shape.

PROVEN, NOT ASSUMED (AE vs a base build compiled from the previous commit)

• blank canon room: AE 0 across all 44 implemented recipes x 4 tones x 3 tilts
• mirror: AE 0 on the card path too - pixel-frozen
• text at flash peak: +0.1 to -1.1 points of glyph/paper contrast on every reverse whose foil covers the text box; the library's two largest costs (rainbow-glitter, water-web) toned down per-recipe and re-checked by eye
• every implemented recipe eyeballed on 1-2 real assigned cards at rest and two tilts; shots in ~/.pokedex-dev-hub-legacy/foil-shots/r6/

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2ca1aed — 2026-08-07 — Mask provenance: sidecar v3, the AI learning loop, and the collapse safeguard

Chey (2026-08-07): mark AI-generated masks, let an AI learn from his hand-done ones, then let him correct the AI's mask so it "can observe the diff and continue to improve without me having to hand-paint all the masks."

derivation_method existed in v2 but the API HARDCODED it to 'hand' — harmless while only a Pencil touched the corpus, actively wrong once mask-refine's Flatten started writing machine geometry through the same save path.

Taxonomy (5 values, 4 cases that must never blur): layout-flatten machine geometry, unpainted weight 0 hand human, from a layout prior weight 1 hand-refined human, on an existing mask weight 1 ai generator, UNREVIEWED weight 0 ai-corrected AI proposed, human edited weight 0.6

The label is never taken from the client. The client reports only what the canvas was SEEDED with; writeMaskRecord derives the method by diffing saved pixels against what that seed rasterizes to. A machine label needs a full GeneratorIdentity, which HTTP callers cannot supply. One write path serves the route and generators both.

• Anti-feedback-collapse enforced in selectExemplars(), not by convention: unreviewed ai is weight 0 at any corpus size. Codify now routes through it.
• Corrections keep the parent PNG + change map + metrics (incl. a 4x4 grid of where corrections land) — the supervised pairs a generator lane consumes via the self-describing tuples manifest.
• Seam rule, measured not guessed: canvas roundRect vs the server SDF disagree on 389/330260 px, ALL in the 1px AA band. Without tolerating it every unpainted Flatten would stamp hand.
• Guard: if the file at the target path is an unreviewed ai mask it IS the parent, whatever the client claims — AI ancestry cannot be laundered. That guard caught a real bug during this build (ActionBtn passed React's MouseEvent into saveMask's optional arg, blanking derivation); derivation is now required on the PUT so it fails loudly instead.
• v1/v2 load forever (normalize-on-read); Chey's two masks migrated additively and still read HAND-PAINTED, agreement 0.641, pixels untouched.
• Trial batch wotc-window-trial-1: 6 WOTC Base holo rares from window-artgate@1, all ai/unreviewed, reversible. Leave-one-out on the human corpus beat the rect baseline (0.588 -> 0.688) — real, but n=2 same-set, so explicitly not validation.

Surfaces: provenance badge + detail in the workbench, Mask corpus panel, GET /foil-lab/masks/corpus (+?tuples/?exemplars), artifact routes, and corpus.ts report|exemplars|tuples|migrate. Pure tests added to CI (test:foil). Browser-verified at 390px and desktop. See DECISIONS.md 2026-08-07 and the mask-pipeline SKILL.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

f1e6a83 — 2026-08-08 — foil mask-straighten: read the INTENT out of Chey's hand mask, and make the undo real

Chey, 2026-08-08: "it's impossible to get the lines really straight so I'm hoping you can get computer vision on the mask and card art in tandem to really see my intent there, and make the mask nice and crisp and straight on the lines I was trying to draw along. Will you try that? And then I can either approve or revert."

Smoothing a wobbly boundary gives a smoothly wobbly boundary. The premise here is that a hand-drawn foil boundary is an ATTEMPT TO TRACE A PRINTED EDGE — frame, art box, species strip, stage tag — and those are dead straight in the scan. So: contour the mask, cut each loop into near-axis runs (PCA orientation, wobble gaps bridged), robust-fit each, hunt the scan's directional gradient with a local Hough over (angle, offset) plus a proximity prior, replace, intersect adjacent lines into real corners, rasterize matching his own AA character.

The refusals are the feature, and each is a recorded param: weak evidence, or an ambiguous band of comparable ridges, may nudge but never relocate; with no artwork edge it straightens only to HIS own fit; short and curved runs pass through untouched. A first cut "closed" the Tropius top-left corner — which turned out to be him deliberately tracing around the BASIC stage tag. The stricter rules keep it.

• apps/api/src/foil/line-snap.ts the geometry, pure, no I/O
• generator.ts line-snap@1, a REFINER (requiresSource)
• generate-masks.ts run --refine, archives, restoring revert
• provenance.ts supersedes + verbatim sha256 undo archive

A generator write onto an existing mask now THROWS without explicit supersede { runId }; with it, every artifact the mask had is archived verbatim under superseded/./ with a self-describing manifest, and revert --run-id verifies the whole archive BEFORE deleting anything live, then restores byte-for-byte. supersedes is deliberately NOT correction: one is a human editing a machine (training signal), the other a machine replacing a human with nobody's agreement. --refine also refuses a source with exemplar weight 0, so a refiner can never eat its own unreviewed output.

Tropius me05-001/37184 (run straighten-tropius-1): 8 runs snapped to printed edges, 1 self-straightened, 13 left exactly as drawn, 8 corners closed, 2 specks dropped. Max correction 3.74px, 1.06% of the face changed, Jaccard 0.9798, confidence 0.557. Honest residual: the card's bottom border was too faint to accept, so that run self-straightened and the report says so.

test:foil 25/25 (9 geometry + 5 supersede/restore, incl. byte-for-byte round trip, a deliberate curve surviving, an ambiguous band refused, and a corrupt archive aborting before deletion). Revert exercised twice end to end, sha256sum clean. Shots: ~/.pokedex-dev-hub-legacy/foil-shots/mask-straighten/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id straighten-tropius-1

6de27ac — 2026-08-08 — foil R7: his eye overturns two banked mechanisms; "no catalog cards" gets a reason

Chey's 15 canon-lab comments (2026-08-07), all resolved in issues/foil/.

Eleven shader verdicts. Two overturn mechanisms built from earlier corpus — energy-symbols' parity CHECKERBOARD (R3-GLYPH, from his own words) and prism's solid-facet grid (R3-MISC, from a Gemini delta pass). Having seen both rendered he rejected them. A description is a hypothesis about the render, not a spec.

cosmos the 0.085 UV field slide he read as parallax is now uP4 "Surface drift" at 0.012; zero at rest, canon unchanged striped-v-sheen grey lines: the unconditional silver term is gone AND the second-order lobe cut 0.4 -> 0.16 (it drew a whole second, near-achromatic line population on the far side) rainbow-mirror hue axis is card X, blue -> green -> red, on BOTH the sheen and the spotlight; core whitening 0.85 -> 0.32 energy-symbols checkerboard gone: per-symbol response AXIS + PHASE + transition WIDTH over a low-freq cluster field energy-symbols-ii same treatment (it shared the two-bank mechanism) pokeball-hologram flat SDF icon -> shaded 3D sphere (hemisphere normal, tilt-following light, belt, hub, two-tone shell) tinsel the "blinds" were a literal full-card striation floor; demoted to uP2, default 0 (his canon stores 0) tinsel-ii line-work gated to the colour band, silver 0.60 -> 0.10, band width exposed as uP4 prism pinwheel's spoke fan WITHOUT the per-wedge phase that makes a pinwheel spin; density 33 -> 18 radiant hologram step 0.5 cell -> uP4 0.17, drive rescaled so total travel per unit tilt is unchanged rainbow-glitter invisible on card: bigger white-lifted flakes AND the family fixed (flash -> field; no substrate, no contrast)

Canon policy held: shader.ts / canon.ts / CardViewer.tsx / data/foil-canon have a ZERO diff, sheenGlsl() is byte-identical (all four sheen canons frozen), and every new behaviour went into a NEW uniform old canon files inherit. No stored value migrated. One recommendation left for him, not applied: tinsel-ii's starkness is substantially his own canon (uDepth 1 at ~135 lines).

"Why are there no catalog cards?" — asked four times, three answers. The baked inversion is now v2 and emits a machine-readable diagnosis the lab renders: outranked diagonal-sheen-left — 1,818 SM reverses ARE cited for it, but energy-symbols wins them (the emblem layer vs the sheet rotation; both true). A capped SECONDARY pool is now baked via a new citedFoilPatterns() export and served as via:'cited', so the preview works. Never flip a winner to fill a pool. class-absent pinwheel, pokeball-hologram — the catalog has reverse variants for ex1-ex5 ONLY; the whole late-EX reverse era is missing upstream. The UI now says so instead of "no catalog cards". fixed vertical-sheen-rainbow 0 -> 12 cards. known_residuals recorded twice that the 6 Holo Rare Energies of ex13/ex16 use Sheen but could not be targeted without a card list; the catalog names exactly six per set. Two cardId rows added citing the same Collexy quotes, both residuals closed.

Typecheck (all workspaces) + web build + foil tests green; 14 recipes compile clean; before/after shots on blank + assigned card at 390px and desktop in ~/.pokedex-dev-hub-legacy/foil-shots/r7/. SKILL R7 field notes + DECISIONS appended.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2523581 — 2026-08-08 — foil edge-trace: the artwork holds the geometry, his mask says which of it to follow

Chey, after reviewing line-snap: "Really just 'straighten' isn't quite enough, really just detecting edges and actually tracing accurately around them is the real move." So the premise moves. line-snap treats his strokes as the geometry and nudges near-axis runs onto printed lines. edge-trace treats his mask as a STATEMENT OF INTENT and derives the boundary from the card's own edges.

apps/api/src/foil/edge-trace.ts (pure, no I/O): Di Zenzo colour structure tensor (Scharr per channel) -> NMS -> hysteresis; colour because a green field against a silver border reads 0.13 luminance/px and 44.3 tensor/px at the same edge. Then livewire: a corridor around his boundary (a hard bound, not a suggestion), anchors at fixed spacing PLUS every hard turn in his own line, Dijkstra on ridge + linked + proximity^2 + direction, sub-pixel refinement onto the |dI/dn| peak, and only THEN a straight fit where the traced path really is straight.

His own correction said the same thing before he did: 114 of his 627 correction px sit in one 40x50 box around the species strip's swept tail, which line-snap had squared off by intersecting two lines across it.

Two traps, both now locked by tests:

• the MAD-trim trap: a robust fit discards a small feature as an outlier, reports "straight to 0.2px RMS", and crisping then flattens what it ignored. straightMaxDevPx requires every point in a run to be close.
• the half-pixel frame: mask space centres pixels at x+0.5, a gradient array at integer x. All artwork lookups convert in one place.

Measured, not asserted. generate-masks.ts adherence scores any set of masks by distance to the NEAREST printed edge maximum, with line-snap's own luminance Sobel as the default probe so the incumbent gets the home field: his hand mask 50.7% within 1px / 0.847px mean line-snap@1 74.8% / 0.545px his correction 75.4% / 0.548px edge-trace@1 77.1% / 0.199px (colour probe: 98.4% / 0.082px)

Applied to Tropius me05-001/37184 as run edgetrace-tropius-1, ai/unreviewed, his ai-corrected mask preserved as parent + verbatim supersede archive. Revert proven byte-identical twice before shipping.

test:foil 35/35 (10 new). Typecheck all workspaces, api+mcp+web builds green. Visuals in ~/.pokedex-dev-hub-legacy/foil-shots/mask-edgetrace/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id edgetrace-tropius-1

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

3f72b07 — 2026-08-08 — foil: rule-seeding measured and refused — the era rect is the bug, not the tracer

Chey: "Use it on the next 5 reverse holofoils in the same set."

The five: me05-002/37186, -003/37188, -005/37191, -006/37193, -007/37195 — every me05 card after 001 with a REVERSE variant, in collector order. me05-004 (and 008, 016) skipped: Double rare ex, holo-only, no reverse variant.

STEP 3 CONTROL, on Tropius where his hand mask is ground truth: a rule seed scores IoU 0.7084 against his intent, and edge-trace takes it to 0.7080 — it closed ZERO of the gap. corridorPx 5/12/24/48 give byte-identical output (the corridor is not the knob; anchorSnapPx is, and widening it to 20 makes the boundary hop between the card's two parallel border edges and run off the card). 99.7% of the 70,520px gap is REGIONS the rule over-claims — the silver border ring, the window-top band, the species strip, the BASIC tag — and edge-trace crisps boundaries, never regions.

THE REAL BUG: modern-sv.artWindow.h = 0.315 puts the art window's top edge at y=106. The printed edge is at y=65 on all six me05 cards measured. Chey's hand adjustment (0.3749) lands within 1.1px on all six. Not changed here — era-wide blast radius, his call. Every SV/Mega holo and reverse currently renders 41px of foil error.

The batch ran, all five failed identically (traced window top y=106 vs printed y=65), and was REVERTED. Nothing shipped; the corpus is byte-identical to before.

Also found: adherence measures precision, not correctness — a rule line inside an illustration finds edges to sit on and scores 38.9→84.7% while being 41px wrong. Two tests now lock that. And modern-sv/sheet has ZERO eligible exemplars while the edgetrace-tropius-1 proposal sits on his ai-corrected mask.

Kept (the capability is sound, the seeding is not): --seed rule era rectangle as a refiner's source for unmasked cards; records exemplars: [], halves confidence, notes open "RULE-SEEDED" --param k=v per-run generator overrides, merged into the RECORDED params --dump

dry-run proposal + seed PNGs + report.json, for real review shots seed-neutral report wording in edge-trace ("the source boundary", "as given") revert bugfix it left empty card dirs behind, so undo was not byte-for-byte

typecheck all workspaces clean, api+web builds green, test:foil 39/39 (4 new). Visuals: ~/.pokedex-dev-hub-legacy/foil-shots/mask-edgetrace-batch/

---

foil/mask-learn

Tip d1dd25f85276516af9923f41ceda82ac3fdfc497 · 57 commits · 2026-08-01 → 2026-08-09 · 9,267 words · 32 reproduced below.

bf4674b — 2026-08-01 — foil/main: workbench v1 — three.js viewer, 5-pattern starter library, era layout masks

Quarantined foil workbench at /pokedex/foil-lab (URL-only, linked from nowhere, chrome-free): real owned-card scans from the image cache, tilt via pointer/gyro(iOS permission)/manual with reduced-motion respected, uniform sliders, pattern + mask-scope overrides, mask overlay toggle, copy-recipe JSON. Pattern library v1 tuned against owned eras (Base/WOTC, SV, Mega Evolution): Starlight, Cosmos, SV default holo, SV reverse sheet, Cracked Ice — uniform contract (core + uP0..3) designed for the full 15-20 set. Era layout spec as data (era-layouts.json, rects measured on real scans); resolver v1 (series, rarity, variant kind) -> pattern + mask scope. SKILL.md x2 (foil-effects, mask-pipeline stub). main.tsx gains a dev-only double-createRoot guard (Vite HMR entry-invalidation footgun, see DECISIONS).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

4391fd5 — 2026-08-01 — foil/main v2: Pencil mask editing, comment queue, iPad two-column, branch api

Apple-Pencil hand-mask drawing on the workbench (MaskEditor: card-aligned canvas overlay, pen+mouse default w/ finger toggle, pressure brush, eraser, 12-step undo, touch-action none, starts from layout prior); masks persist as committed ground-truth artifacts data/foil-masks//.png+json (derivation_method hand) via the env-gated branch api (POKEDEX_FOIL_LAB=1, port 3712, PGPOOL_MAX=1 — inert in prod), auto-load and beat the layout tier (uMaskTex contract; CanvasTexture flipY=false, exactly one V flip). Comment button captures card/variant/pattern/scope/mask state + all slider values to issues/foil// (report.md + context.json), out of the fix-issues sweep. Era-grouped owned-card picker (WOTC / SWSH / SV+ME headings); two-column layout (viewer | controls) from 700px — verified at 390px, 744x1133, 1133x744, 1440x900 with simulated pen strokes. SKILL.md x2 updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

38db198 — 2026-08-02 — foil: sidecar v2 (prior+diff), artwork-keyed masks, comment linkage, codify ritual

• Every hand-mask save records its starting prior (era/scope/resolver rect + RESOLVER_VERSION) and the api renders .prior.png + .diff.png (green = human added, red = removed) + Jaccard agreement, server-side via a pure-JS PNG codec (no native addons). Saves without a parsable prior 400.
• Machamp mask backfilled (deterministic prior wotc/window/v1): agreement 0.6409 — his correction is the subject silhouette, the rule's first lesson.
• Artwork keying: variants of one cardId provably share one scan, so mask GETs alias to a sibling variant's mask when prior.scope matches (X-Foil-Mask- Alias-Of). Cross-card reprints unprovable from the catalog -> per-card only.
• Comments auto-link the exact saved mask state (maskFile/maskSavedAt/ maskAliasOf/maskHasPriorDiff).
• mask-pipeline SKILL: sidecar v2 storage, identity rule + limits, and the codify ritual; first worked example data/foil-masks/codified/wotc.md (n=1: WOTC window = art-window rect minus subject silhouette; noted in era-layouts.json wotc.notes).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

bbc5874 — 2026-08-02 — foil: any-card picker — full-catalog browse/search, Owned-only as a toggle

Chey: pull up ANY card in the catalog and tilt it, owned or not. Picker now browses the whole catalog by era -> set (paged strip, promo sets run 300+), adds full-catalog name search, and keeps Owned-only as a filter toggle (default on = prior behavior). Zero new endpoints: existing /series, /series/:slug, /sets/:setId (paged), /search, /cards/:cardId serve it all; unowned cards list their catalog variants (quantity 0) and resolve the same base-guess pattern path. Selection is never clobbered by filters; unmapped series get an honest "Other eras" bucket. Verified on unowned neo2-13 Umbreon (Auto - Starlight WOTC) at 390px + both iPad viewports.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

7164243 — 2026-08-02 — foil: taxonomy integration — all 39 pattern types in the app, cited resolver

Mislabel corrections from the video-reference research, applied to apps/ (research docs untouched):

• sv-holo renamed vertical-sheen (it always rendered the HGSS-XY vertical bands, not SV's). PATTERN_ALIASES keeps old slugs resolving forever — sidecars, comment context.json, Copy-recipe JSON never orphan.
• Sheen family = ONE generator at four rotations + stripe option: five honest implemented slugs (vertical / horizontal=TRUE SV default "Mirage" / diagonal right "/" / diagonal left "" / striped "Line").
• Cosmos label drops "/ Galaxy" (Galaxy is a starlight synonym; synonym moved to the starlight entry). reverse-sheet description notes it ≈ pokeball-masterball stamped sheet.
• Vintage narrowed: starlight = Base/Jungle/Fossil only; Base Set 2 → Call of Legends holos = cosmos.

Pattern dropdown carries the FULL 39-slug vocabulary, grouped: implemented recipes first, then every gap type rendering via its nearest recipe and labeled "approx via ..." (amber note in the UI). Unimplemented != hidden.

Resolver v2: base guess reads the cited usage table (usage-index.json, derived from research/foil-pattern-usage.json by tools/foil/build-usage-index.mjs; set-name match -> era/series token match -> era-default heuristics). The winning row's confidence + citation hosts surface in the workbench ("Guess: set-level citation, high confidence (bulbapedia..., collexy...)"). Reverse rows citing 'mirror' are deprioritized per the research known-gap (ink-design evidence, not foil-pattern evidence). RESOLVER_VERSION -> 2 (scope semantics unchanged from v1).

Verified: typecheck + build green; Playwright at 390/744x1133/1133x744 (~/.pokedex-dev-hub-legacy/foil-shots/w2/) — Machamp auto=starlight set-level-high with his hand mask still resolving, SV holo auto=horizontal-sheen, SV reverse auto=energy-symbols-ii, fireworks approx labeling.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

fefd125 — 2026-08-02 — foil: fix diffuse darkening — scan sampled sRGB-decoded but never re-encoded (issue ls9u0y)

The face texture was uploaded SRGB8_ALPHA8 (hardware sRGB->linear decode at sample time) while the foil ShaderMaterial writes gl_FragColor raw — three.js only appends the linear->sRGB colorspace_fragment chunk to built-in materials — so every scan rendered in linear values displayed as sRGB (measured flat 184->123 / 199->147, exactly the sRGB->linear curve). Sample the scan undecoded (NoColorSpace): the compositing model is authored in display space and pattern=none at rest is now pixel-comparable to the flat view (mean abs luma diff 44.55 -> 2.16, mapping identity +/-2). Also zero pattern-none's default specular so the baseline is truly the plain scan.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

18ea484 — 2026-08-02 — foil: workbench split — canon pattern lab (A) | card adjustment surface (B)

Chey's comment issues/foil/2026-08-02_12-59-52-368_4aq756 (resolved here):

• /pokedex/foil-lab/canon (CanonLab.tsx): bare holofoil pattern on a blank card (no ink; black/dark/silver/white flat-tone base via the normal texture path) beside the REAL reference tilt clip + 8 keyframes streamed from research/foil-video-reference/ by new env-gated dev routes (byte-range via res.sendFile for iOS Safari). Full 39-slug picker, sliders, tilt, and Save canon -> data/foil-canon/.json (full uniform snapshot that replaces recipe code defaults as the baseline).
• /pokedex/foil-lab (FoilLab.tsx reframed): per-card masks/comments unchanged, plus per-card uniform overrides saved as a SPARSE diff vs the canon baseline -> data/foil-overrides//.json (untouched uniforms keep tracking canon). Off-canon sliders get a dot; saving with sliders matching canon deletes the override.
• Layering model in foil/canon.ts (code defaults < canon < override < live); shared atoms + surface tabs in foil/ui.tsx; routes wired in main.tsx; AppShell chrome-free check widened to /foil-lab/*; both data dirs re-included in .gitignore; ports 5184/3713 recorded in ORCHESTRATION.md.

Verified in-browser at 390x844 / 744x1133 / 1133x744 (clip playing beside the bare cosmos render; Machamp base1-8 hand mask still resolves) — shots in ~/.pokedex-dev-hub-legacy/foil-shots/canon-lab/. Typecheck + build green (web, api). patterns.ts untouched (R0 re-tunes concurrently).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

dd84216 — 2026-08-02 — foil R0: re-tune implemented recipes per Gemini verification notes

cosmos: full activation-model rewrite — dark field, smaller/denser orbs mostly near-invisible, cluster pops via low-freq noise over cell ids, pinprick 4-point twinkles (was: dense wall of large saturated orbs). starlight (+ starlight-ii via shared GLSL): parallax architecture untouched; tighter visibility lobe (pow 5->9, floor .18->.08), sharper glyph cores + narrower flares, smaller blobs, saturated star color, galaxy wash halved; starlight-ii art gate lowered for the mid-orange Evolutions field. cracked-ice: authored intra-shard grain removed per Chey's accuracy ruling — facets flash as solid smooth planes; whitening at peak reduced, uSat up. sheen family: generator grows sharp/beam/barcode options. vertical-sheen -> new barcode variant (thin spectral lines of varying width, CD-groove slide) with tamed beam + art gate. diagonals: sharp 3.0, beam 0.55, specular 0.35 (center blow-out), diagonal-sheen-right uP0 2->7 (same-sheet consistency). striped-vertical-sheen: stripes finer + more blended (90->130, floor .40). horizontal-sheen: untouched (20/20).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

8958cfc — 2026-08-02 — foil R0: rounds 2-3 re-tunes + verification report, field notes, decisions

starlight: longer thin-armed 4-point glyphs + 8-point-burst subset, tighter pop lobe (pow 11, floor .04), full-saturation star color, parallax default 2.4, gain 3.6 — best of 3 rounds 11/20 (still-frame judging cannot see the parallax; recorded as honest residual with a video-judge recommendation). starlight-ii: 20/20 (uP2 pinned so base-default drift can't invalidate it). cracked-ice: facet gain 1.1 -> 0.55 (art visible through flashes), density 10, less whitening — 15/20 yay; shard-geometry uniformity queued for R1. Verification report gains the R0 section (before->after per pattern, the diag-left judge-noise event, diffuse-fix supersession note); foil-effects SKILL.md gains the R0 field notes; DECISIONS entry for root cause + wave.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

f83e807 — 2026-08-02 — foil: cited per-card pattern assignments across the whole catalog (resolver v3)

Chey's mlmwmp comment: the library was funneling to era defaults ("cosmos"). Full-catalog enumeration (20,964 cards / 37,627 variants, via the branch read api) + an 18-lane Ringer research swarm (foil-card-assignments, GLM-5.2, executed selector-integrity checks) produced:

• research/foil-card-assignments.json — 111 set/rarity/kind/card-level rows, 19 catalog foil-facet mappings (holo-foil-cosmos, reverse-foil-masterball, normal-foil-galaxy...), 55 honest residuals; per-lane provenance + method.
• tools/foil/build-assignments-index.mjs -> apps/web/src/foil/assignments-index.json (trimmed for the bundle, duplicate-facet guard).
• resolver v3 (RESOLVER_VERSION=3): assignment tier above the v2 usage table; specificity cardIds > variantKinds > facet > rarities > set > v2 > heuristic; normal-kind variants with a declared foil facet now render window foil (previously scope none); FoilLab passes cardId.
• era-layouts: wotc seriesSlugs 'ecard' -> 'e-card', + 'legendary-collection' (both had fallen through to modern-sv rects and heuristics).

Coverage (foil-bearing variants): 79.5% set-cited / 20.4% era-token / 0.2% heuristic -> 93.4% set / 4.3% facet-or-card / 2.4% era-token / 0 heuristic. 13 in-browser spot-checks in ~/.pokedex-dev-hub-legacy/foil-shots/assignments/. Issue mlmwmp resolved. Weak spots + curation ledger in DECISIONS.md.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

98ba9ca — 2026-08-02 — foil R1: dedicated recipes for the twelve owned-era gap patterns (10/12 verified)

Chey's comment dml369 ("develop out the rest of the missing holofoil patterns"): dedicated GLSL for fireworks, ace-spec, energy-symbols-ii, rainbow-glitter-sheen, ex-starfoil, prismatic-pokeball, tinsel-ii, cosmos-iii-smooth, pokeball-masterball, radiant, rainbow-glitter, confetti — implemented flags flipped, approx labels dropped, slugs stable. Gemini-verified 10/12 match over 3 rounds (verdict table + honest residuals in research/foil-verification.md R1 section). Honest nays: tinsel-ii (procedural static plateaus at judge score 2) and prismatic-pokeball (screen-only blend cannot darken a white reverse body — structural, Chey's call).

R0 residuals cleared in the same pass: cracked-ice anisotropic shattered-glass metric (re-judged yay 14/20); starlight milky wash + per-layer visibility curves with starlight-ii re-banked clean 20/20 through both shared-GLSL changes; starlight parallax judged from a new 16-frame fine sweep and still scored flat — verified real by eye, recorded as judge-blindness.

Field notes appended to foil-effects SKILL.md; dated DECISIONS entry; issue dml369 resolved.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

773b5a0 — 2026-08-02 — foil R2: uDarken blend-model term — prismatic-pokeball 17/20 yay, tinsel-ii 16/20 yay

One opt-in substrate-darkening term in the shared composite (main()): body = scan * (1 - uDarken * mask * gate) before the foil screen-blends. Default 0 is bit-identical to the old render; absent canon/override keys seed 0. Full 21-pattern regression sweep on provably identical renders: no true regressions (and a measurement of Gemini single-roll noise, ±3-6; diag-right + pokeball-masterball double-nays recorded as unstable verdicts on geometry proof). prismatic-pokeball rebuilt on the term (dark-mirror base, facet-quantized flash lobe, ball watermark as ink-overprint suppression): structural nay -> 17/20 yay. tinsel-ii one-line opt-in (uDarken 0.4): static plateau -> 16/20 yay. 20 of 21 implemented types now match; sole nay starlight (judge parallax blindness).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

0e3b4a1 — 2026-08-02 — foil R2 stage 3: thirteen dedicated recipes (gap patterns + RC dots) + subset-commons resolver tier

• patterns.ts: real recipes for mirror, rainbow-mirror, vertical-sheen-rainbow (dark-mirror family on uDarken), pokeball-hologram, prism, water-web, energy-symbols, pinwheel, ex-emerald, tinsel, cosmos-ii-pixel, crosshatch, radiant-collection-dots — implemented: true, approx labels dropped, slugs stable. Unused plain SHEEN_V generator removed (last fallback users gone).
• resolver v4: card-level 'normal'-class assignment rows consulted before the scope-none early return — subset cards the catalog declares plain but which physically carry foil (RC commons dot overprint) resolve their subset's pattern with scope full.
• foil-card-assignments.json: two card-level rows (bw11-RC1..20, g1 25 non-ultra RC ids) with video citation (non-holo RC Teddiursa keeps the dot overprint); builder enforces cardIds on 'normal' rows; index regenerated.

Verified in-browser: g1-RC1 + bw11-RC1 render the RC dot treatment via Auto (guess card|medium). Gemini verdicts in research/foil-verification.md (R2 recipe-wave section, next commit).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b3a573e — 2026-08-02 — foil R2b: four §40–43 vocabulary recipes (all match) + resolver v5 scope overrides

The last lane of the wave — dedicated recipes for the vocab-extension types: gold-secret (warm-locked goldRamp field + chromatic glitter pops + SWSH burst rays — 19/20 yay), vstar-pearl (near-white pearl on uDarken 0.3, warm-led diagonal wash, frame-line streaks — 13/20 yay), shiny-vault (silvery field + band-keyed sparkle-glyph amplifiers, uDarken 0.15 — 11 nay → 16/20 yay r2), detective-pikachu (diagonal sheen × photo-luminance via the documented uFace contract exception — 20/20 yay). 38/43 taxonomy types real, 34 hold match.

Resolver v5: assignment rows carry an optional per-row scope override (window/ full/sheet), honored by the set-row tier and the cls-'normal' tier — baby shinies render window despite the 'Shiny rare' full-foil rarity, VSTARs full despite plain holo kinds, det1 window despite 'ultra rare'.

Assignment research: +2 facet rows (gold, gold-jumbo → gold-secret; retires both facet residuals, 1-card metal collateral flagged) and +10 cited rows — sma 51 baby/35 GX/8 gold splits, swsh4.5sv 104 baby + 18 full-face (incl. the black shiny Eternatus pair, name-verified not gold), sv04.5 rarity split, VSTAR via the dedicated 'Holo Rare VSTAR' rarity, det1 ultra-rare + cls-normal rows (all 18 cards window). Residuals annotated with resolutions.

build-usage-index.mjs: tolerate era-wide rows with no scope.sets (the vocab gold rows crashed the build — the 122-row usage index never compiled until now). Verdicts/frames under ~/.pokedex-dev-hub-legacy/foil-verify/ (verify-

-r2b*); shots in foil-shots/r2b/. Docs: foil-verification.md R2b section, SKILL R2b field notes, DECISIONS entry.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

c48fbad — 2026-08-03 — foil R3: sheen-family streak-field rework — all 8 of Chey's canon-lab comments resolved, 7/7 Gemini 20/20

sheenGlsl is now a streak field (Chey's critique = the spec): sparse irregular streaks, per-streak lean following card tilt (crisscross; converging pairs terminate where they meet), stretched-ellipse taper, hue running ALONG each strip as well as across. striped-vertical-sheen rebuilt from a fresh Gemini corpus re-spec run with Chey's description embedded (grouped reveal via wide sweeping windows over a subtle bottom-converging fan; gemini-spec-r3.md); ex-starfoil rebased on the reworked diagonal base + fine CD-line layer; vstar-pearl rebuilt from the horizontal-sheen streak field, pearl kept.

Canon migrations (full-snapshot rule, DECISIONS): striped + horizontal uDarken 0 -> 0.32 (bright-substrate legibility, third confirmation). All other canon values carry over; uniform semantics unchanged. validate_spec.py frame regex single-digit fix (undercounted two-digit citations, failed an honest worker).

Verdicts (Chey's words folded into every judge prompt as acceptance criteria): vertical/right/left/vstar-pearl 20/20 first roll; striped + ex-starfoil 20/20 round 2; horizontal 20/20 round 3. Table + residuals in research/foil-verification.md R3.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

e71f83f — 2026-08-03 — foil R3-MOTION: axis-split starlight, per-dot cosmos, hologram radiant, the rainbow-glitter-sheen delta

Chey's six motion-model canon-lab comments resolved (his eye = ground truth):

• starlight (+ starlight-ii via shared GLSL): top-to-bottom hue BANDING (soft-quantized bands; his canon 0.62/0.6 lands the reference order blue->green->red on the R->B->G cosine ramp) + AXIS-SPLIT tilt: vertical slides the whole field (~10% card height), horizontal drives the per-star random fade; his 3-layer uP1 parallax rides on top untouched. Judge: starlight 17/20 yay round 2 — the FIRST starlight yay (standing parallax nay broken); starlight-ii 20/20.
• cosmos: whole-field slide + per-dot random response (independent random tilt axes for brightness and hue per orb — no shared sweep axis); at tilt 0 his canon appearance is pixel-identical. Judge: 20/20.
• radiant: discrete hologram steps (half-cell up/down, 60% hold / 40% crossfade — one position fades out as the next fades in), uP1 wired as Hologram travel (canon migrated 0 -> 2.2; was an unused placeholder), grid lines ~45% thicker. Judge nay 13/20 across 3 rounds on a pixel-refuted "slides continuously" claim — recorded as still-frame motion blindness (same class as starlight parallax); adjacent-frame crops prove discrete positions + crossfade. Chey's live tilt is the tiebreak.
• rainbow-glitter-sheen: dedicated Gemini delta-articulation pass first (he couldn't name the difference), pixel-verified, then fixed: narrow laser-saturated striped chevron (sigma 0.010->0.0035, hue traversal 5x->9x, pow-deepened at low gain), fainter +-0.8 repeats, chevron angle 1.3->1.9, denser colored always-visible glitter, uDarken 0->0.4 (4th legibility-physics data point). Judge: 19/20 yay.

Docs: research/foil-verification.md R3-MOTION section, DECISIONS entry, foil-effects SKILL field notes. Frames/verdicts in ~/.pokedex-dev-hub-legacy/foil-verify (jobs -r3m, manifest-r3m{,2,3}, delta-rainbow-glitter-sheen); gate shots in foil-shots/r3-motion.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9c57c84 — 2026-08-03 — foil mask-refine: window handles → flatten → hand refine

Chey's requested workflow on the Card-adjust surface: draggable corner/edge handles adjust the layout window rect per card (WindowEditor.tsx, 44px touch targets, finger-drag allowed), the adjusted geometry persists as artwork-keyed scope-agnostic sidecars (data/foil-windows//.json, /foil-lab/windows routes), and Flatten bakes it through the STANDARD hand-mask save path (sidecar prior stays the era rule; new optional prior.window records the adjustment) then opens the existing paint tooling.

• rasterizeWindowRect shared by Flatten + loadLayoutRect (identical bakes)
• saveMask prior = layoutMask (rule), never the effective mask
• windows GET aliases across variants of the same scan, newest savedAt
• verified in-browser desktop + 390px: sma-SV1/sma-SV2 (SM baby shiny, borrowed modern-sv rect) and me01-034 (zero-coverage gap); base1-8 + base1-5 hand-mask regression clean; test artifacts deleted (corpus rule)
• ports: vite 5186 / api 3714 recorded in ORCHESTRATION.md

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

dfd8f56 — 2026-08-03 — foil R3-GLYPH: glyph-pattern reworks per Chey's 7 comments + the drop-in glyph slot for his SVGs

Behavior (his notes verbatim = acceptance criteria): reverse-sheet rainbow only on grainy glyphs over a neutral sheet; energy-symbols checkerboard banks that swap with tilt; energy-symbols-ii sporadic random banks (~half faint at any time); ace-spec diamonds grow/shrink with tilt + a touch of blur; prismatic-pokeball ball never darkens (white-mixed pale response + coherent plane flash, suppression deleted); radiant-collection-dots shapes catch a traveling rainbow band (16/20 YAY — standing R2 nay broken).

Glyph slot: research/foil-glyphs//glyph[-N].svg -> lab-gated api routes -> atlas rasterizer (foil/glyphs.ts) -> uGlyphTex contract + glyphTex() helper; ~2.5s auto-pickup on save, procedural fallback when absent, README tells Chey exactly where his four promised files go.

Canon migration: reverse-sheet uP1 0 -> 0.6 (dead placeholder now drives his requested glyph grain). Verdicts + pixel-refutations recorded in research/foil-verification.md (R3-GLYPH); lessons in foil-effects SKILL.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9d7a70d — 2026-08-03 — foil R3-MISC: the uTint ink-metallic blend term + Chey's twelve remaining canon-lab comments (12/12 yay)

The reverse-holo fix (his chat report): screen-blending achromatic foil light over colored art compresses chroma — "dull and grayish, rather than making the color look metallic." New core uniform uTint (default 0 = bit-exact legacy) multiplies the foil layer + in-mask specular by the luminance-normalized scan chroma SQUARED (the double ink pass of a real mirror reflection); neutral on blank cards, so no canon-lab appearance moved. Reverse-family recipes opt in; verified before/after at 5 tilt angles on Victini ball-reverse, Crystal Energy, Pineco. uP4/uP5 param slots added in the same contract change.

The twelve comments, each mechanism-mapped in its issue frontmatter and the verification doc: cracked-ice triangulated + half-off gate; fireworks grid lattice + center-emanating radial hue; rainbow-mirror rebuilt as mirror + hue-banded spotlight; tinsel-ii silver lines + vertical rainbow band (band lives only on the line-work); prism rebuilt as pinwheel-kin from the corpus-vs-corpus delta pass; cosmos-ii pixel-quantized stamp edges; crosshatch thinned; gold-secret adjustable burst origin (uP4/uP5, center default, per-card via overrides) + holographic grain; dedicated recipes for sequin (glyph-family glints), tcg-classic (flat starlight + rainbow glitter), acid-wash (water-web kin), disco (perfect disc lattice, galaxy ignition) — zero approxVia fallbacks remain.

Canon migrations (appearance-preserving unless he asked otherwise): fireworks uP0 3→4.5 (dead two-octave structure), gold-secret uP4/uP5 0.5 added, sequin uP0-uP3 re-keyed to the new recipe. Verdicts: research/foil-verification.md "R3-MISC" — 12/12 Gemini yay incl. the tcg-classic ink-scope re-roll.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

932cc3a — 2026-08-03 — foil R4-COMPOSITE: the ink-density invariant — foil adds pop, never washes text or mutes ink

Chey's two me05 comments (7rtnzx mirror reverse "blows out the darks/text", 19mo4l holo "darkens/muddies the colors") ruled a CORE COMPOSITE INVARIANT, engineered in shader.ts main(): ink density estimated from the scan (inkDark = darker-than-local-8-tap-field average, exactly 0 on flat blank bases; inkColor = chroma above the canon-lab-tone floor) now gates every layer — dark ink blocks the flash + shields the specular (text crisp at every tilt), ALL ink is exempt from uDarken substrate attenuation (printed color never muted; the mirror stays dark only where the mirror is visible), colored ink auto-tints its flash (max(uTint, inkColor)) and uInkPop pumps chroma under the flash so colors read MORE saturated and metallic.

New core uniforms uInkGuard (default 1) / uInkPop (default 0.5), both 0 = bit-exact legacy composite; sliders in both labs; canon files untouched (absent keys inherit the new defaults — catalog-wide fix, zero migrations).

Proof: canon-lab blank-card pairs (mirror dark+silver, cosmos, horizontal-sheen) pixel-identical (compare -metric AE 0) under a frame-stepped rAF harness with a same-settings AE-0 control, plus a Gemini sanity pass (all pairs "identical"). On-card before/after pairs across the 8-card sample (both issue cards, Victini, Pineco, Darkrai, Cetitan, Machamp WOTC, Hydreigon SIR; 3 tilts, 390px + desktop) in ~/.pokedex-dev-hub-legacy/foil-shots/r4-composite/. Both issues resolved with mechanism-mapped notes; composite contract documented in the foil-effects SKILL; DECISIONS + research/foil-verification.md updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2292524 — 2026-08-04 — foil R4b: scan-additive composite (rest parity by construction) + canon-lab card preview

Chey's Grubbin (me05-002, Pitch Black) review refuted R4's ink heuristics: the reverse still muddied the card at rest (measured ΔC +32, green→yellow), darkened the header under tilt (−22 L), and blew out text (glyph contrast −26%). New law on real scans (uScanBase 1): the scan is a photograph of the card at rest — foil is additive dynamic light only, tilt-gated to ZERO at rest, luminance-headroom + per-channel clamped (text sacred at every angle, uArtGate opens the clamp where dark scan pixels ARE the foil), chroma- preserving with a pastel-safe tint ramp and an L²-gated chroma pump. Canon lab blank bases run the classic composite textually unchanged: AE-0 proven (CDP frame-stepped harness, control + knob pairs; Gemini pass agreed). uInkGuard 0 still reproduces the pre-R4 composite exactly.

Canon-lab card preview (his request verbatim): blank/on-card/↻-another chips render the live sliders on a random catalog card the resolver assigns the pattern to — server-sampled from the baked inversion (tools/foil/ build-pattern-cards.mts → data/foil-pattern-cards.json, gitignored; GET /foil-lab/pattern-cards/:patternId), honoring hand masks + adjusted windows; empty-pool patterns say "no catalog cards". Also commits Chey's live mirror-canon tweak from the Grubbin session. Full metrics + verdicts: DECISIONS 2026-08-04.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

46c872d — 2026-08-04 — foil R4c: onset ease — the sheen leans in instead of switching on

Chey on R4b's rest gate: "it's like it isn't there at all until I've tilted the card a tiny bit and then it's like it just suddenly appears." Measured: the headroom clamp saturated R4b's smoothstep(.02,.28) gate, so ALL added light arrived inside tilt 0.03-0.13 (0 -> 16.1 dL mean) then froze flat to tilt 1.0 — a step smeared over 4% of the tilt range.

Now: one wide ease ramp = smoothstep(0, .45, |tilt|) (full at 0.45 — every R4b tilt-0.5 metric unchanged), flash gated by ramp^1.5 (cubic-slow start), specular by ramp^2.5 (the broad gloss band now trails the flash instead of leading it), plus a REST=0.006 sub-JND living rest sheen blended into both gates (rest +0.43 dL card-mean, max 2/255 — visually still the scan). After-curve: smooth S over 0->0.25, no dead zone, no knee; text clamp untouched (glyph contrast 33.6 -> 33.6 at flash peak); blank-card canon lab bit-level unchanged (fuzz-1% AE 0 vs control floor).

Harness + curves + filmstrips: ~/.pokedex-dev-hub-legacy/foil-shots/r4c/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d1e1988 — 2026-08-04 — foil R4d: motion-first onset — tiny gestures move the sheen; glow arrives late; Chey gets the dials

Chey rejected R4c too ('It still lights up pretty noticeably when I tilt the card a tiny bit. I don't like it.') — measured the gesture→tilt mapping first this time: pointer is 0.0051 tilt/px at 390px and gyro is 0.036 tilt/deg, so a 30px thumb drag (tilt 0.154) was already getting HALF the full glow under R4c. The curve was never the felt variable.

• shader.ts: onset gates now uniform-driven — uOnsetRange (glow full at |tilt|, default 0.5) and uOnsetCurve (flash exponent, default 3.5; specular trails at +1). (0.45, 1.5) reproduces R4c exactly (verified to the third decimal); defaults are strictly softer at every tilt and equal from 0.5 up, so all R4b/R4c peak metrics are unchanged (Grubbin 17.976 ΔL / text 34.02, Machamp 18.035 / text 45.6 — exact).
• FoilLab + CanonLab: 'Onset range' / 'Onset curve' sliders, canon- stored like every other uniform; absent canon keys inherit defaults.
• patterns.ts: CoreDefaults admits the two keys per recipe.

Pointer-path proof (real pointermove, 390px, 30px drag, first 300ms): before mean ΔL climbs to 6.6 and rising; after stays within ±0.08 while per-pixel movement reaches 6.1 mean / 18.9 p95 — moves, doesn't light up. Blank canon lab at fuzz-1% AE 0 vs control. Artifacts in ~/.pokedex-dev-hub-legacy/foil-shots/r4d/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

845821f — 2026-08-05 — foil R5: the metallic composite — always on the card, energy redistributed, Chey's compositing dials

Chey's ruling (2026-08-05, verbatim in DECISIONS): no onset range at all — the effect is ALWAYS on the card; dial in how it composites onto the ink; metallic and spectral, not brightening.

• Onset machinery removed (R4c/R4d ramp, REST floor, uOnsetRange/uOnsetCurve sliders on both surfaces). No canon/override file ever stored the keys; CardViewer skips unknown uniforms, so stale keys anywhere are inert.
• R5 metalness law (scan path only): scan = albedo; the pattern field splits around PIVOT 0.40 — above = tinted-specular highlight, below = bounded multiplicative depth (mirror-turned-away darks, sqrt-concave). Energy redistributes; contrast + chroma carry the metallic read.
• Soft-knee headroom: the R4b hard clamp plateaued strong fields into a flat wash (the "brightened" read); the budget now compresses monotonically so pattern structure survives. Per-channel no-clip caps stay hard; glyph ink (inkDark) exempt from highlight AND depth — text contrast preserved (Grubbin 86->97 at peak, Machamp flat).
• New canon-stored dials on both surfaces: Metallic 0.6, Sheen strength 0.55, Sheen tint 0.5 (= the R4 chroma law exactly), Depth 0.55, Texture 1. uMetal 0 / uIntensity 0 / uInkGuard 0 all still yield the plain scan.
• Blank canon lab pixel-unchanged: CDP frame-stepped before/after AE 0 at fuzz 1% (mirror/cosmos/horizontal-sheen/starlight, tilted + rest); one Gemini vision pass agrees (all pairs identical). Shots + per-card rest pairs: ~/.pokedex-dev-hub-legacy/foil-shots/r5/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

33a744c — 2026-08-07 — foil R5b: metalness is MIRROR ONLY — every other pattern gets its own colour back

Chey: "You applied the metallic treatment to EVERY HOLO PATTERN when I only wanted it for mirror. So now, every single holo pattern looks like mirror and you can barely see the pattern at all, it's like every single one is being applied in a way that it adds no rainbow color or anything to the card."

R5 shipped its metalness dials as GLOBAL_DEFAULTS, so all 43 recipes inherited a law built for one of them. uMetal is now the scan-path LAW SELECTOR and defaults to 0 — nothing can inherit metalness; mirror is the only recipe that opts in, in its own patterns.ts defaults.

• Metalness branch is the R5 code textually unmoved, so mirror's arithmetic is unchanged by construction. Proven: his hand-tuned canon on Grubbin me05-002
  • Tropius me05-001, rest + tilt 0.2/0.35/0.5, 390px and desktop — 16/16 frames AE 0 at fuzz 1%, same floor as the same-build control pair.
• uMetal 0 runs the ADDITIVE law: the pattern's own emission (uSheen gain) screen-combined per channel over a substrate-attenuated body, tinted toward the ink only as far as uSheenTint asks. Three fixes to what R5 broke — the scalar no-clip cap let the artwork's brightest channel throttle the whole flash; the quartic ink budget (shaped to starve mid-dark glyphs) starved ordinary mid-tone art instead; PIVOT ate the field before anything flashed. Cosmos measured 0.005% of art-window pixels changed >8% under R5.
• uDepth decoupled from uDarken: uDarken is the blank-canon substrate, frozen by 30 saved canon files, and 0 on exactly the pale reverse sheets that need one. The substrate follows the pattern's dark half so it buys contrast, not a uniform dim (a flat 0.34 cost mean dL -21.7 and 29% of text contrast).
• Per-recipe families in patterns.ts (PARTICLE / SHEET / WASH / PEARL) plus seven by-eye corrections, each judged on a card the resolver assigns.
• Onset machinery stays deleted. Text guarantee re-verified: glyphs matte and crisp at peak tilt on a reverse whose foil covers the text box.
• Blank-card canon appearance unchanged: 8 patterns AE 0 (6 direct; cracked-ice and tinsel-ii against the pre-R5b code with canon held constant, since Chey re-saved those two mid-wave).
• SKILL.md composite contract rewritten around per-recipe selection.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

92770db — 2026-08-07 — foil R6: composite defaults re-derived from Chey's canons; the guard learns to tell text from artwork

Chey: "i re-adjusted cracked-ice. take a look at it, and then make the changes based on everything you're observing to make all of this better - i want it so that i have to do as little of work personally on all the rest as possible."

WHAT HIS CANONS TAUGHT

uSheenTint 0 on all three non-mirror canons -> 0 for every non-mirror family (was 0.15/0.5/0.85/0.6). Ink-tinting a pattern's own highlight is exactly "it adds no rainbow color or anything to the card". uSheen he raised all three -> families now 1.2-4.2 (were 1.0-2.6). uDepth NOT universal: 0 on cracked-ice, 1 on tinsel-ii and rgs. It darkens the card WHERE THE PATTERN IS NOT, so its cost scales with 1 - duty. Measured every recipe's duty cycle; his three choices line up with it, and it predicted both dials for the only canon that exercises the defaults. uSat 1 + wide spread -> shipped on the 13 recipes with no canon.

Families are renamed for how light LANDS (flash / line / stamp / field / pearl / metal) and stated explicitly as FoilPattern.family. The stamp tier exists because the first pass got it wrong: energy-symbols-ii lights 5.8% of the face, so a field substrate darkened the other 94% and the card read "someone dimmed this".

THE INK GUARD: SPLIT, DON'T WEAKEN

inkDark is a local-contrast measure, so it fired on every dark mark - including every outline and shading edge in the ILLUSTRATION (measured: 86-89% of strong hits inside the art window are artwork, not text). The additive law spends it as a hard coverage multiply, so guard 1 punched the pattern full of holes. That is why he kept pulling it to ~0.51/0.56 on sheet/wash patterns.

Now split by glyphness into inkGlyph (sacred, zero flash, unchanged) and inkDetail (artwork darks keep the pattern on a tightened budget). inkAny is byte-for-byte the old inkDark and every pre-R6 consumer still reads it.

Success test met: at guard 1 under R6, tinsel-ii and rgs look BETTER than his 0.51/0.56 do today - tinsel-ii's artwork is visible through the streaks instead of greyed out. Both canons migrated to 1, reason recorded in each file's note; every other value in them is his, untouched.

SLIDER RANGES (his second ask)

Four dials sat exactly on their cap: uIntensity 2->4, uSheen 3->6, uSat 1->2, uHueSpread 1.5->3. Not extended: uInkGuard (his carve-out), uMetal (a law selector), the 0..1 mix weights, and uDarken/uDepth (already fully spent at 1).

uSheen needed a shader change to be honest: the additive budget's ceiling was reached at uSheen 2.0, so the top third of the slider bought almost nothing - very likely why he parked two canons at the ceiling. It now resumes climbing above 3, leaving every stored value <= 3 bit-for-bit identical.

The 19 duplicated lines across the two surfaces are one CORE_SLIDERS table; uGrain is dimmed and labelled "metal law only" where it does nothing.

NEW: "Apply composite -> family" copies only the on-card dials to siblings, never pattern shape.

PROVEN, NOT ASSUMED (AE vs a base build compiled from the previous commit)

• blank canon room: AE 0 across all 44 implemented recipes x 4 tones x 3 tilts
• mirror: AE 0 on the card path too - pixel-frozen
• text at flash peak: +0.1 to -1.1 points of glyph/paper contrast on every reverse whose foil covers the text box; the library's two largest costs (rainbow-glitter, water-web) toned down per-recipe and re-checked by eye
• every implemented recipe eyeballed on 1-2 real assigned cards at rest and two tilts; shots in ~/.pokedex-dev-hub-legacy/foil-shots/r6/

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

f2b312f — 2026-08-07 — Mask provenance: sidecar v3, the AI learning loop, and the collapse safeguard

Chey (2026-08-07): mark AI-generated masks, let an AI learn from his hand-done ones, then let him correct the AI's mask so it "can observe the diff and continue to improve without me having to hand-paint all the masks."

derivation_method existed in v2 but the API HARDCODED it to 'hand' — harmless while only a Pencil touched the corpus, actively wrong once mask-refine's Flatten started writing machine geometry through the same save path.

Taxonomy (5 values, 4 cases that must never blur): layout-flatten machine geometry, unpainted weight 0 hand human, from a layout prior weight 1 hand-refined human, on an existing mask weight 1 ai generator, UNREVIEWED weight 0 ai-corrected AI proposed, human edited weight 0.6

The label is never taken from the client. The client reports only what the canvas was SEEDED with; writeMaskRecord derives the method by diffing saved pixels against what that seed rasterizes to. A machine label needs a full GeneratorIdentity, which HTTP callers cannot supply. One write path serves the route and generators both.

• Anti-feedback-collapse enforced in selectExemplars(), not by convention: unreviewed ai is weight 0 at any corpus size. Codify now routes through it.
• Corrections keep the parent PNG + change map + metrics (incl. a 4x4 grid of where corrections land) — the supervised pairs a generator lane consumes via the self-describing tuples manifest.
• Seam rule, measured not guessed: canvas roundRect vs the server SDF disagree on 389/330260 px, ALL in the 1px AA band. Without tolerating it every unpainted Flatten would stamp hand.
• Guard: if the file at the target path is an unreviewed ai mask it IS the parent, whatever the client claims — AI ancestry cannot be laundered. That guard caught a real bug during this build (ActionBtn passed React's MouseEvent into saveMask's optional arg, blanking derivation); derivation is now required on the PUT so it fails loudly instead.
• v1/v2 load forever (normalize-on-read); Chey's two masks migrated additively and still read HAND-PAINTED, agreement 0.641, pixels untouched.
• Trial batch wotc-window-trial-1: 6 WOTC Base holo rares from window-artgate@1, all ai/unreviewed, reversible. Leave-one-out on the human corpus beat the rect baseline (0.588 -> 0.688) — real, but n=2 same-set, so explicitly not validation.

Surfaces: provenance badge + detail in the workbench, Mask corpus panel, GET /foil-lab/masks/corpus (+?tuples/?exemplars), artifact routes, and corpus.ts report|exemplars|tuples|migrate. Pure tests added to CI (test:foil). Browser-verified at 390px and desktop. See DECISIONS.md 2026-08-07 and the mask-pipeline SKILL.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

36f9eef — 2026-08-08 — foil mask-straighten: read the INTENT out of Chey's hand mask, and make the undo real

Chey, 2026-08-08: "it's impossible to get the lines really straight so I'm hoping you can get computer vision on the mask and card art in tandem to really see my intent there, and make the mask nice and crisp and straight on the lines I was trying to draw along. Will you try that? And then I can either approve or revert."

Smoothing a wobbly boundary gives a smoothly wobbly boundary. The premise here is that a hand-drawn foil boundary is an ATTEMPT TO TRACE A PRINTED EDGE — frame, art box, species strip, stage tag — and those are dead straight in the scan. So: contour the mask, cut each loop into near-axis runs (PCA orientation, wobble gaps bridged), robust-fit each, hunt the scan's directional gradient with a local Hough over (angle, offset) plus a proximity prior, replace, intersect adjacent lines into real corners, rasterize matching his own AA character.

The refusals are the feature, and each is a recorded param: weak evidence, or an ambiguous band of comparable ridges, may nudge but never relocate; with no artwork edge it straightens only to HIS own fit; short and curved runs pass through untouched. A first cut "closed" the Tropius top-left corner — which turned out to be him deliberately tracing around the BASIC stage tag. The stricter rules keep it.

• apps/api/src/foil/line-snap.ts the geometry, pure, no I/O
• generator.ts line-snap@1, a REFINER (requiresSource)
• generate-masks.ts run --refine, archives, restoring revert
• provenance.ts supersedes + verbatim sha256 undo archive

A generator write onto an existing mask now THROWS without explicit supersede { runId }; with it, every artifact the mask had is archived verbatim under superseded/./ with a self-describing manifest, and revert --run-id verifies the whole archive BEFORE deleting anything live, then restores byte-for-byte. supersedes is deliberately NOT correction: one is a human editing a machine (training signal), the other a machine replacing a human with nobody's agreement. --refine also refuses a source with exemplar weight 0, so a refiner can never eat its own unreviewed output.

Tropius me05-001/37184 (run straighten-tropius-1): 8 runs snapped to printed edges, 1 self-straightened, 13 left exactly as drawn, 8 corners closed, 2 specks dropped. Max correction 3.74px, 1.06% of the face changed, Jaccard 0.9798, confidence 0.557. Honest residual: the card's bottom border was too faint to accept, so that run self-straightened and the report says so.

test:foil 25/25 (9 geometry + 5 supersede/restore, incl. byte-for-byte round trip, a deliberate curve surviving, an ambiguous band refused, and a corrupt archive aborting before deletion). Revert exercised twice end to end, sha256sum clean. Shots: ~/.pokedex-dev-hub-legacy/foil-shots/mask-straighten/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id straighten-tropius-1

63f1b95 — 2026-08-08 — foil R7: his eye overturns two banked mechanisms; "no catalog cards" gets a reason

Chey's 15 canon-lab comments (2026-08-07), all resolved in issues/foil/.

Eleven shader verdicts. Two overturn mechanisms built from earlier corpus — energy-symbols' parity CHECKERBOARD (R3-GLYPH, from his own words) and prism's solid-facet grid (R3-MISC, from a Gemini delta pass). Having seen both rendered he rejected them. A description is a hypothesis about the render, not a spec.

cosmos the 0.085 UV field slide he read as parallax is now uP4 "Surface drift" at 0.012; zero at rest, canon unchanged striped-v-sheen grey lines: the unconditional silver term is gone AND the second-order lobe cut 0.4 -> 0.16 (it drew a whole second, near-achromatic line population on the far side) rainbow-mirror hue axis is card X, blue -> green -> red, on BOTH the sheen and the spotlight; core whitening 0.85 -> 0.32 energy-symbols checkerboard gone: per-symbol response AXIS + PHASE + transition WIDTH over a low-freq cluster field energy-symbols-ii same treatment (it shared the two-bank mechanism) pokeball-hologram flat SDF icon -> shaded 3D sphere (hemisphere normal, tilt-following light, belt, hub, two-tone shell) tinsel the "blinds" were a literal full-card striation floor; demoted to uP2, default 0 (his canon stores 0) tinsel-ii line-work gated to the colour band, silver 0.60 -> 0.10, band width exposed as uP4 prism pinwheel's spoke fan WITHOUT the per-wedge phase that makes a pinwheel spin; density 33 -> 18 radiant hologram step 0.5 cell -> uP4 0.17, drive rescaled so total travel per unit tilt is unchanged rainbow-glitter invisible on card: bigger white-lifted flakes AND the family fixed (flash -> field; no substrate, no contrast)

Canon policy held: shader.ts / canon.ts / CardViewer.tsx / data/foil-canon have a ZERO diff, sheenGlsl() is byte-identical (all four sheen canons frozen), and every new behaviour went into a NEW uniform old canon files inherit. No stored value migrated. One recommendation left for him, not applied: tinsel-ii's starkness is substantially his own canon (uDepth 1 at ~135 lines).

"Why are there no catalog cards?" — asked four times, three answers. The baked inversion is now v2 and emits a machine-readable diagnosis the lab renders: outranked diagonal-sheen-left — 1,818 SM reverses ARE cited for it, but energy-symbols wins them (the emblem layer vs the sheet rotation; both true). A capped SECONDARY pool is now baked via a new citedFoilPatterns() export and served as via:'cited', so the preview works. Never flip a winner to fill a pool. class-absent pinwheel, pokeball-hologram — the catalog has reverse variants for ex1-ex5 ONLY; the whole late-EX reverse era is missing upstream. The UI now says so instead of "no catalog cards". fixed vertical-sheen-rainbow 0 -> 12 cards. known_residuals recorded twice that the 6 Holo Rare Energies of ex13/ex16 use Sheen but could not be targeted without a card list; the catalog names exactly six per set. Two cardId rows added citing the same Collexy quotes, both residuals closed.

Typecheck (all workspaces) + web build + foil tests green; 14 recipes compile clean; before/after shots on blank + assigned card at 390px and desktop in ~/.pokedex-dev-hub-legacy/foil-shots/r7/. SKILL R7 field notes + DECISIONS appended.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

8f476f7 — 2026-08-08 — foil edge-trace: the artwork holds the geometry, his mask says which of it to follow

Chey, after reviewing line-snap: "Really just 'straighten' isn't quite enough, really just detecting edges and actually tracing accurately around them is the real move." So the premise moves. line-snap treats his strokes as the geometry and nudges near-axis runs onto printed lines. edge-trace treats his mask as a STATEMENT OF INTENT and derives the boundary from the card's own edges.

apps/api/src/foil/edge-trace.ts (pure, no I/O): Di Zenzo colour structure tensor (Scharr per channel) -> NMS -> hysteresis; colour because a green field against a silver border reads 0.13 luminance/px and 44.3 tensor/px at the same edge. Then livewire: a corridor around his boundary (a hard bound, not a suggestion), anchors at fixed spacing PLUS every hard turn in his own line, Dijkstra on ridge + linked + proximity^2 + direction, sub-pixel refinement onto the |dI/dn| peak, and only THEN a straight fit where the traced path really is straight.

His own correction said the same thing before he did: 114 of his 627 correction px sit in one 40x50 box around the species strip's swept tail, which line-snap had squared off by intersecting two lines across it.

Two traps, both now locked by tests:

• the MAD-trim trap: a robust fit discards a small feature as an outlier, reports "straight to 0.2px RMS", and crisping then flattens what it ignored. straightMaxDevPx requires every point in a run to be close.
• the half-pixel frame: mask space centres pixels at x+0.5, a gradient array at integer x. All artwork lookups convert in one place.

Measured, not asserted. generate-masks.ts adherence scores any set of masks by distance to the NEAREST printed edge maximum, with line-snap's own luminance Sobel as the default probe so the incumbent gets the home field: his hand mask 50.7% within 1px / 0.847px mean line-snap@1 74.8% / 0.545px his correction 75.4% / 0.548px edge-trace@1 77.1% / 0.199px (colour probe: 98.4% / 0.082px)

Applied to Tropius me05-001/37184 as run edgetrace-tropius-1, ai/unreviewed, his ai-corrected mask preserved as parent + verbatim supersede archive. Revert proven byte-identical twice before shipping.

test:foil 35/35 (10 new). Typecheck all workspaces, api+mcp+web builds green. Visuals in ~/.pokedex-dev-hub-legacy/foil-shots/mask-edgetrace/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id edgetrace-tropius-1

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

f5f20a7 — 2026-08-08 — Mask editor: pan + pinch-zoom that keeps the strokes honest, and no more "selected" card

Chey, before a batch of hand masks: "something that would be nice is to be able to pan and pinch to zoom while editing masks. In addition, sometimes the card image gets 'selected' when I'm trying to draw which is annoying."

Zoom is a camera view offset, not a CSS scale. camera.setViewOffset(Wz, Hz, x, y, W, H) re-rasterizes the card at the zoomed size, so 4x zoom is 4x real detail — the point of zooming is tracing a printed edge, and a magnified 390px framebuffer would not help. The overlays get the same transform on one wrapper div, written imperatively (no re-render per frame). Because that is a pure crop of a linearly-scaled render, u lands at u*z - offset in both places, so getBoundingClientRect already reports the truth and the pointer->texel math is unchanged. Verified numerically: a 3.375x corner stroke lands at texels [84,180,108,196] vs a predicted [85,180,109,196]; touch at 3x lands within one texel at 390 and 744.

Brush is constant in SCREEN px (lineWidth = brushSize / zoom): zooming in must buy finer control, not a fatter-looking stroke. Two fingers pinch/pan; one finger pans unless it is a drawing finger or on a window handle; a pen owns the surface (palm rejection); a second finger mid-stroke aborts AND rolls the stroke back. Desktop gets wheel/trackpad zoom at the cursor, middle- or Space-drag pan, and +/-/0. HUD shows the zoom and fits back in one tap. Pan is clamped so the card cannot be lost. Zoom is live only while editing, resets on exit, and never reaches the artifact — a mask saved at 3.375x round-trips as 490x674 with an identical bbox. WindowEditor divides drag deltas by zoom and counter-scales its handles, so they stay 44px on screen (44.0 -> 44.0 measured at 2.25x).

Selection fix: user-select/touch-callout/user-drag none, draggable=false, and preventDefault on dragstart/selectstart and the paint pointerdown across both canvases and the viewer column; touch-action stays none so the two fixes do not fight. iOS also needs the gesturestart/change/end veto — touch-action alone does not stop Safari page zoom.

Cost the most time: window listeners registered in the controller factory were killed by StrictMode's simulated unmount, so the first pinch started and never ended and every later gesture and stroke was dead. They are now a symmetric attach()/detach() pair driven by an effect. See DECISIONS.md.

44/44 browser checks at 1280 / 744 / 390 (touch via CDP), his committed masks still load, scratch card artifacts removed. Shots in ~/.pokedex-dev-hub-legacy/foil-shots/mask-ux/.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

cbba612 — 2026-08-08 — foil region-learn: the REGIONS come from his masks, the geometry from the artwork

The last lane measured why rule-seeded edge-trace failed on every me05 reverse: 99.7% of the gap between the layout rule and Chey's intent is REGION decisions, and a tracer can only move a boundary, never add or remove a region. This is the missing half.

region-learn@1 partitions a card face into five structural classes computed from its OWN printing (border / furniture / frameBody / windowBackground / windowSubject), reads off his masks WHICH of them carry foil, applies that to a new card, and only then hands the boundary to edge-trace. Nothing about Pokemon is hard-coded — the same code reads OPPOSITE policies from the two classes in the corpus, and a test asserts the votes, not just the verdict.

The policy, in his pixels: modern-sv/sheet foil = the COLOURED FRAME BODY, nothing else. Species strip (tails included), stage tag, evolution medallion + its sprite, "evolves from" bar, copyright footer: all excluded. And THE SILVER BORDER RING IS NOT FOIL — 0.2/0.2/0.4% across his three masks. That was flagged as "his call, not a measurement"; it is now a measurement, and unanimous. wotc/window foil = the illustration's own background, minus the subject silhouette, minus the stage box where it overlaps. 4/4.

Bar written down before any number existed: mean IoU >= 0.90, no held-out card below 0.85. Leave-one-out: modern-sv/sheet 0.7566 -> 0.9691 (worst 0.9519) PASS -> batch of 8 shipped wotc/window 0.6146 -> 0.8971 (worst 0.8599) FAIL by 0.0029 -> no batch

The two residuals differ in kind, and that is the finding: modern-sv's is a 1-2px registration hairline; wotc's is 8,343px of Machamp's body read as background because his blue-grey sits inside the colour distance of the teal holo field. Every frame-level decision on WOTC is right. The subject silhouette is not.

Also found and NOT changed (era-wide, his call): wotc.artWindow is 11px too high at the top and 6px too wide at the right — his four masks and each card's own printed edge agree, and base1-8 is the card the era was "measured on". Same pattern as me04-007 for modern-sv.

Also: the bevel side. An illustration box has TWO parallel printed edges; which one the foil stops at is which side the foil is on, not which is strongest. Picking the strongest put base1-6's window top 8px wrong and scored it 0.0001; picking the foil side took it to 0.8789.

Batch regionlearn-me05-1 (8 cards, 5 frame colours, 4 with medallions, all ai / unreviewed) applied and proven reversible byte-for-byte — including a revert bug fixed here: it left the empty card directory behind, so a reverted card still looked masked to anything that walks data/foil-masks.

Recommend reverting wotc-window-trial-1: those five base1 masks were made by window-artgate@1, which scores 0.7694 mean / 0.7330 worst on this same test.

test:foil 46/46 (11 new, including the anti-collapse rule locked three ways).

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

---

foil/mask-straighten

Tip 9db4e93eebd4bef87896f1b15ae1636ce9a37a66 · 51 commits · 2026-08-01 → 2026-08-09 · 7,803 words · 29 reproduced below.

0b9dbdd — 2026-08-01 — foil/main: workbench v1 — three.js viewer, 5-pattern starter library, era layout masks

Quarantined foil workbench at /pokedex/foil-lab (URL-only, linked from nowhere, chrome-free): real owned-card scans from the image cache, tilt via pointer/gyro(iOS permission)/manual with reduced-motion respected, uniform sliders, pattern + mask-scope overrides, mask overlay toggle, copy-recipe JSON. Pattern library v1 tuned against owned eras (Base/WOTC, SV, Mega Evolution): Starlight, Cosmos, SV default holo, SV reverse sheet, Cracked Ice — uniform contract (core + uP0..3) designed for the full 15-20 set. Era layout spec as data (era-layouts.json, rects measured on real scans); resolver v1 (series, rarity, variant kind) -> pattern + mask scope. SKILL.md x2 (foil-effects, mask-pipeline stub). main.tsx gains a dev-only double-createRoot guard (Vite HMR entry-invalidation footgun, see DECISIONS).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

f0bc78e — 2026-08-01 — foil/main v2: Pencil mask editing, comment queue, iPad two-column, branch api

Apple-Pencil hand-mask drawing on the workbench (MaskEditor: card-aligned canvas overlay, pen+mouse default w/ finger toggle, pressure brush, eraser, 12-step undo, touch-action none, starts from layout prior); masks persist as committed ground-truth artifacts data/foil-masks//.png+json (derivation_method hand) via the env-gated branch api (POKEDEX_FOIL_LAB=1, port 3712, PGPOOL_MAX=1 — inert in prod), auto-load and beat the layout tier (uMaskTex contract; CanvasTexture flipY=false, exactly one V flip). Comment button captures card/variant/pattern/scope/mask state + all slider values to issues/foil// (report.md + context.json), out of the fix-issues sweep. Era-grouped owned-card picker (WOTC / SWSH / SV+ME headings); two-column layout (viewer | controls) from 700px — verified at 390px, 744x1133, 1133x744, 1440x900 with simulated pen strokes. SKILL.md x2 updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

d725642 — 2026-08-02 — foil: sidecar v2 (prior+diff), artwork-keyed masks, comment linkage, codify ritual

• Every hand-mask save records its starting prior (era/scope/resolver rect + RESOLVER_VERSION) and the api renders .prior.png + .diff.png (green = human added, red = removed) + Jaccard agreement, server-side via a pure-JS PNG codec (no native addons). Saves without a parsable prior 400.
• Machamp mask backfilled (deterministic prior wotc/window/v1): agreement 0.6409 — his correction is the subject silhouette, the rule's first lesson.
• Artwork keying: variants of one cardId provably share one scan, so mask GETs alias to a sibling variant's mask when prior.scope matches (X-Foil-Mask- Alias-Of). Cross-card reprints unprovable from the catalog -> per-card only.
• Comments auto-link the exact saved mask state (maskFile/maskSavedAt/ maskAliasOf/maskHasPriorDiff).
• mask-pipeline SKILL: sidecar v2 storage, identity rule + limits, and the codify ritual; first worked example data/foil-masks/codified/wotc.md (n=1: WOTC window = art-window rect minus subject silhouette; noted in era-layouts.json wotc.notes).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

efada09 — 2026-08-02 — foil: any-card picker — full-catalog browse/search, Owned-only as a toggle

Chey: pull up ANY card in the catalog and tilt it, owned or not. Picker now browses the whole catalog by era -> set (paged strip, promo sets run 300+), adds full-catalog name search, and keeps Owned-only as a filter toggle (default on = prior behavior). Zero new endpoints: existing /series, /series/:slug, /sets/:setId (paged), /search, /cards/:cardId serve it all; unowned cards list their catalog variants (quantity 0) and resolve the same base-guess pattern path. Selection is never clobbered by filters; unmapped series get an honest "Other eras" bucket. Verified on unowned neo2-13 Umbreon (Auto - Starlight WOTC) at 390px + both iPad viewports.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

eeb6031 — 2026-08-02 — foil: taxonomy integration — all 39 pattern types in the app, cited resolver

Mislabel corrections from the video-reference research, applied to apps/ (research docs untouched):

• sv-holo renamed vertical-sheen (it always rendered the HGSS-XY vertical bands, not SV's). PATTERN_ALIASES keeps old slugs resolving forever — sidecars, comment context.json, Copy-recipe JSON never orphan.
• Sheen family = ONE generator at four rotations + stripe option: five honest implemented slugs (vertical / horizontal=TRUE SV default "Mirage" / diagonal right "/" / diagonal left "" / striped "Line").
• Cosmos label drops "/ Galaxy" (Galaxy is a starlight synonym; synonym moved to the starlight entry). reverse-sheet description notes it ≈ pokeball-masterball stamped sheet.
• Vintage narrowed: starlight = Base/Jungle/Fossil only; Base Set 2 → Call of Legends holos = cosmos.

Pattern dropdown carries the FULL 39-slug vocabulary, grouped: implemented recipes first, then every gap type rendering via its nearest recipe and labeled "approx via ..." (amber note in the UI). Unimplemented != hidden.

Resolver v2: base guess reads the cited usage table (usage-index.json, derived from research/foil-pattern-usage.json by tools/foil/build-usage-index.mjs; set-name match -> era/series token match -> era-default heuristics). The winning row's confidence + citation hosts surface in the workbench ("Guess: set-level citation, high confidence (bulbapedia..., collexy...)"). Reverse rows citing 'mirror' are deprioritized per the research known-gap (ink-design evidence, not foil-pattern evidence). RESOLVER_VERSION -> 2 (scope semantics unchanged from v1).

Verified: typecheck + build green; Playwright at 390/744x1133/1133x744 (~/.pokedex-dev-hub-legacy/foil-shots/w2/) — Machamp auto=starlight set-level-high with his hand mask still resolving, SV holo auto=horizontal-sheen, SV reverse auto=energy-symbols-ii, fireworks approx labeling.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

839a58f — 2026-08-02 — foil: fix diffuse darkening — scan sampled sRGB-decoded but never re-encoded (issue ls9u0y)

The face texture was uploaded SRGB8_ALPHA8 (hardware sRGB->linear decode at sample time) while the foil ShaderMaterial writes gl_FragColor raw — three.js only appends the linear->sRGB colorspace_fragment chunk to built-in materials — so every scan rendered in linear values displayed as sRGB (measured flat 184->123 / 199->147, exactly the sRGB->linear curve). Sample the scan undecoded (NoColorSpace): the compositing model is authored in display space and pattern=none at rest is now pixel-comparable to the flat view (mean abs luma diff 44.55 -> 2.16, mapping identity +/-2). Also zero pattern-none's default specular so the baseline is truly the plain scan.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

12a0bba — 2026-08-02 — foil: workbench split — canon pattern lab (A) | card adjustment surface (B)

Chey's comment issues/foil/2026-08-02_12-59-52-368_4aq756 (resolved here):

• /pokedex/foil-lab/canon (CanonLab.tsx): bare holofoil pattern on a blank card (no ink; black/dark/silver/white flat-tone base via the normal texture path) beside the REAL reference tilt clip + 8 keyframes streamed from research/foil-video-reference/ by new env-gated dev routes (byte-range via res.sendFile for iOS Safari). Full 39-slug picker, sliders, tilt, and Save canon -> data/foil-canon/.json (full uniform snapshot that replaces recipe code defaults as the baseline).
• /pokedex/foil-lab (FoilLab.tsx reframed): per-card masks/comments unchanged, plus per-card uniform overrides saved as a SPARSE diff vs the canon baseline -> data/foil-overrides//.json (untouched uniforms keep tracking canon). Off-canon sliders get a dot; saving with sliders matching canon deletes the override.
• Layering model in foil/canon.ts (code defaults < canon < override < live); shared atoms + surface tabs in foil/ui.tsx; routes wired in main.tsx; AppShell chrome-free check widened to /foil-lab/*; both data dirs re-included in .gitignore; ports 5184/3713 recorded in ORCHESTRATION.md.

Verified in-browser at 390x844 / 744x1133 / 1133x744 (clip playing beside the bare cosmos render; Machamp base1-8 hand mask still resolves) — shots in ~/.pokedex-dev-hub-legacy/foil-shots/canon-lab/. Typecheck + build green (web, api). patterns.ts untouched (R0 re-tunes concurrently).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d2a2dcd — 2026-08-02 — foil R0: re-tune implemented recipes per Gemini verification notes

cosmos: full activation-model rewrite — dark field, smaller/denser orbs mostly near-invisible, cluster pops via low-freq noise over cell ids, pinprick 4-point twinkles (was: dense wall of large saturated orbs). starlight (+ starlight-ii via shared GLSL): parallax architecture untouched; tighter visibility lobe (pow 5->9, floor .18->.08), sharper glyph cores + narrower flares, smaller blobs, saturated star color, galaxy wash halved; starlight-ii art gate lowered for the mid-orange Evolutions field. cracked-ice: authored intra-shard grain removed per Chey's accuracy ruling — facets flash as solid smooth planes; whitening at peak reduced, uSat up. sheen family: generator grows sharp/beam/barcode options. vertical-sheen -> new barcode variant (thin spectral lines of varying width, CD-groove slide) with tamed beam + art gate. diagonals: sharp 3.0, beam 0.55, specular 0.35 (center blow-out), diagonal-sheen-right uP0 2->7 (same-sheet consistency). striped-vertical-sheen: stripes finer + more blended (90->130, floor .40). horizontal-sheen: untouched (20/20).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

632c59b — 2026-08-02 — foil R0: rounds 2-3 re-tunes + verification report, field notes, decisions

starlight: longer thin-armed 4-point glyphs + 8-point-burst subset, tighter pop lobe (pow 11, floor .04), full-saturation star color, parallax default 2.4, gain 3.6 — best of 3 rounds 11/20 (still-frame judging cannot see the parallax; recorded as honest residual with a video-judge recommendation). starlight-ii: 20/20 (uP2 pinned so base-default drift can't invalidate it). cracked-ice: facet gain 1.1 -> 0.55 (art visible through flashes), density 10, less whitening — 15/20 yay; shard-geometry uniformity queued for R1. Verification report gains the R0 section (before->after per pattern, the diag-left judge-noise event, diffuse-fix supersession note); foil-effects SKILL.md gains the R0 field notes; DECISIONS entry for root cause + wave.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d99cc07 — 2026-08-02 — foil: cited per-card pattern assignments across the whole catalog (resolver v3)

Chey's mlmwmp comment: the library was funneling to era defaults ("cosmos"). Full-catalog enumeration (20,964 cards / 37,627 variants, via the branch read api) + an 18-lane Ringer research swarm (foil-card-assignments, GLM-5.2, executed selector-integrity checks) produced:

• research/foil-card-assignments.json — 111 set/rarity/kind/card-level rows, 19 catalog foil-facet mappings (holo-foil-cosmos, reverse-foil-masterball, normal-foil-galaxy...), 55 honest residuals; per-lane provenance + method.
• tools/foil/build-assignments-index.mjs -> apps/web/src/foil/assignments-index.json (trimmed for the bundle, duplicate-facet guard).
• resolver v3 (RESOLVER_VERSION=3): assignment tier above the v2 usage table; specificity cardIds > variantKinds > facet > rarities > set > v2 > heuristic; normal-kind variants with a declared foil facet now render window foil (previously scope none); FoilLab passes cardId.
• era-layouts: wotc seriesSlugs 'ecard' -> 'e-card', + 'legendary-collection' (both had fallen through to modern-sv rects and heuristics).

Coverage (foil-bearing variants): 79.5% set-cited / 20.4% era-token / 0.2% heuristic -> 93.4% set / 4.3% facet-or-card / 2.4% era-token / 0 heuristic. 13 in-browser spot-checks in ~/.pokedex-dev-hub-legacy/foil-shots/assignments/. Issue mlmwmp resolved. Weak spots + curation ledger in DECISIONS.md.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

4c8e8bf — 2026-08-02 — foil R1: dedicated recipes for the twelve owned-era gap patterns (10/12 verified)

Chey's comment dml369 ("develop out the rest of the missing holofoil patterns"): dedicated GLSL for fireworks, ace-spec, energy-symbols-ii, rainbow-glitter-sheen, ex-starfoil, prismatic-pokeball, tinsel-ii, cosmos-iii-smooth, pokeball-masterball, radiant, rainbow-glitter, confetti — implemented flags flipped, approx labels dropped, slugs stable. Gemini-verified 10/12 match over 3 rounds (verdict table + honest residuals in research/foil-verification.md R1 section). Honest nays: tinsel-ii (procedural static plateaus at judge score 2) and prismatic-pokeball (screen-only blend cannot darken a white reverse body — structural, Chey's call).

R0 residuals cleared in the same pass: cracked-ice anisotropic shattered-glass metric (re-judged yay 14/20); starlight milky wash + per-layer visibility curves with starlight-ii re-banked clean 20/20 through both shared-GLSL changes; starlight parallax judged from a new 16-frame fine sweep and still scored flat — verified real by eye, recorded as judge-blindness.

Field notes appended to foil-effects SKILL.md; dated DECISIONS entry; issue dml369 resolved.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ee7260c — 2026-08-02 — foil R2: uDarken blend-model term — prismatic-pokeball 17/20 yay, tinsel-ii 16/20 yay

One opt-in substrate-darkening term in the shared composite (main()): body = scan * (1 - uDarken * mask * gate) before the foil screen-blends. Default 0 is bit-identical to the old render; absent canon/override keys seed 0. Full 21-pattern regression sweep on provably identical renders: no true regressions (and a measurement of Gemini single-roll noise, ±3-6; diag-right + pokeball-masterball double-nays recorded as unstable verdicts on geometry proof). prismatic-pokeball rebuilt on the term (dark-mirror base, facet-quantized flash lobe, ball watermark as ink-overprint suppression): structural nay -> 17/20 yay. tinsel-ii one-line opt-in (uDarken 0.4): static plateau -> 16/20 yay. 20 of 21 implemented types now match; sole nay starlight (judge parallax blindness).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

6cd8803 — 2026-08-02 — foil R2 stage 3: thirteen dedicated recipes (gap patterns + RC dots) + subset-commons resolver tier

• patterns.ts: real recipes for mirror, rainbow-mirror, vertical-sheen-rainbow (dark-mirror family on uDarken), pokeball-hologram, prism, water-web, energy-symbols, pinwheel, ex-emerald, tinsel, cosmos-ii-pixel, crosshatch, radiant-collection-dots — implemented: true, approx labels dropped, slugs stable. Unused plain SHEEN_V generator removed (last fallback users gone).
• resolver v4: card-level 'normal'-class assignment rows consulted before the scope-none early return — subset cards the catalog declares plain but which physically carry foil (RC commons dot overprint) resolve their subset's pattern with scope full.
• foil-card-assignments.json: two card-level rows (bw11-RC1..20, g1 25 non-ultra RC ids) with video citation (non-holo RC Teddiursa keeps the dot overprint); builder enforces cardIds on 'normal' rows; index regenerated.

Verified in-browser: g1-RC1 + bw11-RC1 render the RC dot treatment via Auto (guess card|medium). Gemini verdicts in research/foil-verification.md (R2 recipe-wave section, next commit).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

f1a797e — 2026-08-02 — foil R2b: four §40–43 vocabulary recipes (all match) + resolver v5 scope overrides

The last lane of the wave — dedicated recipes for the vocab-extension types: gold-secret (warm-locked goldRamp field + chromatic glitter pops + SWSH burst rays — 19/20 yay), vstar-pearl (near-white pearl on uDarken 0.3, warm-led diagonal wash, frame-line streaks — 13/20 yay), shiny-vault (silvery field + band-keyed sparkle-glyph amplifiers, uDarken 0.15 — 11 nay → 16/20 yay r2), detective-pikachu (diagonal sheen × photo-luminance via the documented uFace contract exception — 20/20 yay). 38/43 taxonomy types real, 34 hold match.

Resolver v5: assignment rows carry an optional per-row scope override (window/ full/sheet), honored by the set-row tier and the cls-'normal' tier — baby shinies render window despite the 'Shiny rare' full-foil rarity, VSTARs full despite plain holo kinds, det1 window despite 'ultra rare'.

Assignment research: +2 facet rows (gold, gold-jumbo → gold-secret; retires both facet residuals, 1-card metal collateral flagged) and +10 cited rows — sma 51 baby/35 GX/8 gold splits, swsh4.5sv 104 baby + 18 full-face (incl. the black shiny Eternatus pair, name-verified not gold), sv04.5 rarity split, VSTAR via the dedicated 'Holo Rare VSTAR' rarity, det1 ultra-rare + cls-normal rows (all 18 cards window). Residuals annotated with resolutions.

build-usage-index.mjs: tolerate era-wide rows with no scope.sets (the vocab gold rows crashed the build — the 122-row usage index never compiled until now). Verdicts/frames under ~/.pokedex-dev-hub-legacy/foil-verify/ (verify-

-r2b*); shots in foil-shots/r2b/. Docs: foil-verification.md R2b section, SKILL R2b field notes, DECISIONS entry.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

7a1451b — 2026-08-03 — foil R3: sheen-family streak-field rework — all 8 of Chey's canon-lab comments resolved, 7/7 Gemini 20/20

sheenGlsl is now a streak field (Chey's critique = the spec): sparse irregular streaks, per-streak lean following card tilt (crisscross; converging pairs terminate where they meet), stretched-ellipse taper, hue running ALONG each strip as well as across. striped-vertical-sheen rebuilt from a fresh Gemini corpus re-spec run with Chey's description embedded (grouped reveal via wide sweeping windows over a subtle bottom-converging fan; gemini-spec-r3.md); ex-starfoil rebased on the reworked diagonal base + fine CD-line layer; vstar-pearl rebuilt from the horizontal-sheen streak field, pearl kept.

Canon migrations (full-snapshot rule, DECISIONS): striped + horizontal uDarken 0 -> 0.32 (bright-substrate legibility, third confirmation). All other canon values carry over; uniform semantics unchanged. validate_spec.py frame regex single-digit fix (undercounted two-digit citations, failed an honest worker).

Verdicts (Chey's words folded into every judge prompt as acceptance criteria): vertical/right/left/vstar-pearl 20/20 first roll; striped + ex-starfoil 20/20 round 2; horizontal 20/20 round 3. Table + residuals in research/foil-verification.md R3.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b3ccdc6 — 2026-08-03 — foil R3-MOTION: axis-split starlight, per-dot cosmos, hologram radiant, the rainbow-glitter-sheen delta

Chey's six motion-model canon-lab comments resolved (his eye = ground truth):

• starlight (+ starlight-ii via shared GLSL): top-to-bottom hue BANDING (soft-quantized bands; his canon 0.62/0.6 lands the reference order blue->green->red on the R->B->G cosine ramp) + AXIS-SPLIT tilt: vertical slides the whole field (~10% card height), horizontal drives the per-star random fade; his 3-layer uP1 parallax rides on top untouched. Judge: starlight 17/20 yay round 2 — the FIRST starlight yay (standing parallax nay broken); starlight-ii 20/20.
• cosmos: whole-field slide + per-dot random response (independent random tilt axes for brightness and hue per orb — no shared sweep axis); at tilt 0 his canon appearance is pixel-identical. Judge: 20/20.
• radiant: discrete hologram steps (half-cell up/down, 60% hold / 40% crossfade — one position fades out as the next fades in), uP1 wired as Hologram travel (canon migrated 0 -> 2.2; was an unused placeholder), grid lines ~45% thicker. Judge nay 13/20 across 3 rounds on a pixel-refuted "slides continuously" claim — recorded as still-frame motion blindness (same class as starlight parallax); adjacent-frame crops prove discrete positions + crossfade. Chey's live tilt is the tiebreak.
• rainbow-glitter-sheen: dedicated Gemini delta-articulation pass first (he couldn't name the difference), pixel-verified, then fixed: narrow laser-saturated striped chevron (sigma 0.010->0.0035, hue traversal 5x->9x, pow-deepened at low gain), fainter +-0.8 repeats, chevron angle 1.3->1.9, denser colored always-visible glitter, uDarken 0->0.4 (4th legibility-physics data point). Judge: 19/20 yay.

Docs: research/foil-verification.md R3-MOTION section, DECISIONS entry, foil-effects SKILL field notes. Frames/verdicts in ~/.pokedex-dev-hub-legacy/foil-verify (jobs -r3m, manifest-r3m{,2,3}, delta-rainbow-glitter-sheen); gate shots in foil-shots/r3-motion.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

60d304d — 2026-08-03 — foil mask-refine: window handles → flatten → hand refine

Chey's requested workflow on the Card-adjust surface: draggable corner/edge handles adjust the layout window rect per card (WindowEditor.tsx, 44px touch targets, finger-drag allowed), the adjusted geometry persists as artwork-keyed scope-agnostic sidecars (data/foil-windows//.json, /foil-lab/windows routes), and Flatten bakes it through the STANDARD hand-mask save path (sidecar prior stays the era rule; new optional prior.window records the adjustment) then opens the existing paint tooling.

• rasterizeWindowRect shared by Flatten + loadLayoutRect (identical bakes)
• saveMask prior = layoutMask (rule), never the effective mask
• windows GET aliases across variants of the same scan, newest savedAt
• verified in-browser desktop + 390px: sma-SV1/sma-SV2 (SM baby shiny, borrowed modern-sv rect) and me01-034 (zero-coverage gap); base1-8 + base1-5 hand-mask regression clean; test artifacts deleted (corpus rule)
• ports: vite 5186 / api 3714 recorded in ORCHESTRATION.md

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

57f28b5 — 2026-08-03 — foil R3-GLYPH: glyph-pattern reworks per Chey's 7 comments + the drop-in glyph slot for his SVGs

Behavior (his notes verbatim = acceptance criteria): reverse-sheet rainbow only on grainy glyphs over a neutral sheet; energy-symbols checkerboard banks that swap with tilt; energy-symbols-ii sporadic random banks (~half faint at any time); ace-spec diamonds grow/shrink with tilt + a touch of blur; prismatic-pokeball ball never darkens (white-mixed pale response + coherent plane flash, suppression deleted); radiant-collection-dots shapes catch a traveling rainbow band (16/20 YAY — standing R2 nay broken).

Glyph slot: research/foil-glyphs//glyph[-N].svg -> lab-gated api routes -> atlas rasterizer (foil/glyphs.ts) -> uGlyphTex contract + glyphTex() helper; ~2.5s auto-pickup on save, procedural fallback when absent, README tells Chey exactly where his four promised files go.

Canon migration: reverse-sheet uP1 0 -> 0.6 (dead placeholder now drives his requested glyph grain). Verdicts + pixel-refutations recorded in research/foil-verification.md (R3-GLYPH); lessons in foil-effects SKILL.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

6fadaab — 2026-08-03 — foil R3-MISC: the uTint ink-metallic blend term + Chey's twelve remaining canon-lab comments (12/12 yay)

The reverse-holo fix (his chat report): screen-blending achromatic foil light over colored art compresses chroma — "dull and grayish, rather than making the color look metallic." New core uniform uTint (default 0 = bit-exact legacy) multiplies the foil layer + in-mask specular by the luminance-normalized scan chroma SQUARED (the double ink pass of a real mirror reflection); neutral on blank cards, so no canon-lab appearance moved. Reverse-family recipes opt in; verified before/after at 5 tilt angles on Victini ball-reverse, Crystal Energy, Pineco. uP4/uP5 param slots added in the same contract change.

The twelve comments, each mechanism-mapped in its issue frontmatter and the verification doc: cracked-ice triangulated + half-off gate; fireworks grid lattice + center-emanating radial hue; rainbow-mirror rebuilt as mirror + hue-banded spotlight; tinsel-ii silver lines + vertical rainbow band (band lives only on the line-work); prism rebuilt as pinwheel-kin from the corpus-vs-corpus delta pass; cosmos-ii pixel-quantized stamp edges; crosshatch thinned; gold-secret adjustable burst origin (uP4/uP5, center default, per-card via overrides) + holographic grain; dedicated recipes for sequin (glyph-family glints), tcg-classic (flat starlight + rainbow glitter), acid-wash (water-web kin), disco (perfect disc lattice, galaxy ignition) — zero approxVia fallbacks remain.

Canon migrations (appearance-preserving unless he asked otherwise): fireworks uP0 3→4.5 (dead two-octave structure), gold-secret uP4/uP5 0.5 added, sequin uP0-uP3 re-keyed to the new recipe. Verdicts: research/foil-verification.md "R3-MISC" — 12/12 Gemini yay incl. the tcg-classic ink-scope re-roll.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

872e2e2 — 2026-08-03 — foil R4-COMPOSITE: the ink-density invariant — foil adds pop, never washes text or mutes ink

Chey's two me05 comments (7rtnzx mirror reverse "blows out the darks/text", 19mo4l holo "darkens/muddies the colors") ruled a CORE COMPOSITE INVARIANT, engineered in shader.ts main(): ink density estimated from the scan (inkDark = darker-than-local-8-tap-field average, exactly 0 on flat blank bases; inkColor = chroma above the canon-lab-tone floor) now gates every layer — dark ink blocks the flash + shields the specular (text crisp at every tilt), ALL ink is exempt from uDarken substrate attenuation (printed color never muted; the mirror stays dark only where the mirror is visible), colored ink auto-tints its flash (max(uTint, inkColor)) and uInkPop pumps chroma under the flash so colors read MORE saturated and metallic.

New core uniforms uInkGuard (default 1) / uInkPop (default 0.5), both 0 = bit-exact legacy composite; sliders in both labs; canon files untouched (absent keys inherit the new defaults — catalog-wide fix, zero migrations).

Proof: canon-lab blank-card pairs (mirror dark+silver, cosmos, horizontal-sheen) pixel-identical (compare -metric AE 0) under a frame-stepped rAF harness with a same-settings AE-0 control, plus a Gemini sanity pass (all pairs "identical"). On-card before/after pairs across the 8-card sample (both issue cards, Victini, Pineco, Darkrai, Cetitan, Machamp WOTC, Hydreigon SIR; 3 tilts, 390px + desktop) in ~/.pokedex-dev-hub-legacy/foil-shots/r4-composite/. Both issues resolved with mechanism-mapped notes; composite contract documented in the foil-effects SKILL; DECISIONS + research/foil-verification.md updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2c23fef — 2026-08-04 — foil R4b: scan-additive composite (rest parity by construction) + canon-lab card preview

Chey's Grubbin (me05-002, Pitch Black) review refuted R4's ink heuristics: the reverse still muddied the card at rest (measured ΔC +32, green→yellow), darkened the header under tilt (−22 L), and blew out text (glyph contrast −26%). New law on real scans (uScanBase 1): the scan is a photograph of the card at rest — foil is additive dynamic light only, tilt-gated to ZERO at rest, luminance-headroom + per-channel clamped (text sacred at every angle, uArtGate opens the clamp where dark scan pixels ARE the foil), chroma- preserving with a pastel-safe tint ramp and an L²-gated chroma pump. Canon lab blank bases run the classic composite textually unchanged: AE-0 proven (CDP frame-stepped harness, control + knob pairs; Gemini pass agreed). uInkGuard 0 still reproduces the pre-R4 composite exactly.

Canon-lab card preview (his request verbatim): blank/on-card/↻-another chips render the live sliders on a random catalog card the resolver assigns the pattern to — server-sampled from the baked inversion (tools/foil/ build-pattern-cards.mts → data/foil-pattern-cards.json, gitignored; GET /foil-lab/pattern-cards/:patternId), honoring hand masks + adjusted windows; empty-pool patterns say "no catalog cards". Also commits Chey's live mirror-canon tweak from the Grubbin session. Full metrics + verdicts: DECISIONS 2026-08-04.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ba09108 — 2026-08-04 — foil R4c: onset ease — the sheen leans in instead of switching on

Chey on R4b's rest gate: "it's like it isn't there at all until I've tilted the card a tiny bit and then it's like it just suddenly appears." Measured: the headroom clamp saturated R4b's smoothstep(.02,.28) gate, so ALL added light arrived inside tilt 0.03-0.13 (0 -> 16.1 dL mean) then froze flat to tilt 1.0 — a step smeared over 4% of the tilt range.

Now: one wide ease ramp = smoothstep(0, .45, |tilt|) (full at 0.45 — every R4b tilt-0.5 metric unchanged), flash gated by ramp^1.5 (cubic-slow start), specular by ramp^2.5 (the broad gloss band now trails the flash instead of leading it), plus a REST=0.006 sub-JND living rest sheen blended into both gates (rest +0.43 dL card-mean, max 2/255 — visually still the scan). After-curve: smooth S over 0->0.25, no dead zone, no knee; text clamp untouched (glyph contrast 33.6 -> 33.6 at flash peak); blank-card canon lab bit-level unchanged (fuzz-1% AE 0 vs control floor).

Harness + curves + filmstrips: ~/.pokedex-dev-hub-legacy/foil-shots/r4c/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

eeed482 — 2026-08-04 — foil R4d: motion-first onset — tiny gestures move the sheen; glow arrives late; Chey gets the dials

Chey rejected R4c too ('It still lights up pretty noticeably when I tilt the card a tiny bit. I don't like it.') — measured the gesture→tilt mapping first this time: pointer is 0.0051 tilt/px at 390px and gyro is 0.036 tilt/deg, so a 30px thumb drag (tilt 0.154) was already getting HALF the full glow under R4c. The curve was never the felt variable.

• shader.ts: onset gates now uniform-driven — uOnsetRange (glow full at |tilt|, default 0.5) and uOnsetCurve (flash exponent, default 3.5; specular trails at +1). (0.45, 1.5) reproduces R4c exactly (verified to the third decimal); defaults are strictly softer at every tilt and equal from 0.5 up, so all R4b/R4c peak metrics are unchanged (Grubbin 17.976 ΔL / text 34.02, Machamp 18.035 / text 45.6 — exact).
• FoilLab + CanonLab: 'Onset range' / 'Onset curve' sliders, canon- stored like every other uniform; absent canon keys inherit defaults.
• patterns.ts: CoreDefaults admits the two keys per recipe.

Pointer-path proof (real pointermove, 390px, 30px drag, first 300ms): before mean ΔL climbs to 6.6 and rising; after stays within ±0.08 while per-pixel movement reaches 6.1 mean / 18.9 p95 — moves, doesn't light up. Blank canon lab at fuzz-1% AE 0 vs control. Artifacts in ~/.pokedex-dev-hub-legacy/foil-shots/r4d/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

c8079bc — 2026-08-05 — foil R5: the metallic composite — always on the card, energy redistributed, Chey's compositing dials

Chey's ruling (2026-08-05, verbatim in DECISIONS): no onset range at all — the effect is ALWAYS on the card; dial in how it composites onto the ink; metallic and spectral, not brightening.

• Onset machinery removed (R4c/R4d ramp, REST floor, uOnsetRange/uOnsetCurve sliders on both surfaces). No canon/override file ever stored the keys; CardViewer skips unknown uniforms, so stale keys anywhere are inert.
• R5 metalness law (scan path only): scan = albedo; the pattern field splits around PIVOT 0.40 — above = tinted-specular highlight, below = bounded multiplicative depth (mirror-turned-away darks, sqrt-concave). Energy redistributes; contrast + chroma carry the metallic read.
• Soft-knee headroom: the R4b hard clamp plateaued strong fields into a flat wash (the "brightened" read); the budget now compresses monotonically so pattern structure survives. Per-channel no-clip caps stay hard; glyph ink (inkDark) exempt from highlight AND depth — text contrast preserved (Grubbin 86->97 at peak, Machamp flat).
• New canon-stored dials on both surfaces: Metallic 0.6, Sheen strength 0.55, Sheen tint 0.5 (= the R4 chroma law exactly), Depth 0.55, Texture 1. uMetal 0 / uIntensity 0 / uInkGuard 0 all still yield the plain scan.
• Blank canon lab pixel-unchanged: CDP frame-stepped before/after AE 0 at fuzz 1% (mirror/cosmos/horizontal-sheen/starlight, tilted + rest); one Gemini vision pass agrees (all pairs identical). Shots + per-card rest pairs: ~/.pokedex-dev-hub-legacy/foil-shots/r5/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ddac585 — 2026-08-07 — foil R5b: metalness is MIRROR ONLY — every other pattern gets its own colour back

Chey: "You applied the metallic treatment to EVERY HOLO PATTERN when I only wanted it for mirror. So now, every single holo pattern looks like mirror and you can barely see the pattern at all, it's like every single one is being applied in a way that it adds no rainbow color or anything to the card."

R5 shipped its metalness dials as GLOBAL_DEFAULTS, so all 43 recipes inherited a law built for one of them. uMetal is now the scan-path LAW SELECTOR and defaults to 0 — nothing can inherit metalness; mirror is the only recipe that opts in, in its own patterns.ts defaults.

• Metalness branch is the R5 code textually unmoved, so mirror's arithmetic is unchanged by construction. Proven: his hand-tuned canon on Grubbin me05-002
  • Tropius me05-001, rest + tilt 0.2/0.35/0.5, 390px and desktop — 16/16 frames AE 0 at fuzz 1%, same floor as the same-build control pair.
• uMetal 0 runs the ADDITIVE law: the pattern's own emission (uSheen gain) screen-combined per channel over a substrate-attenuated body, tinted toward the ink only as far as uSheenTint asks. Three fixes to what R5 broke — the scalar no-clip cap let the artwork's brightest channel throttle the whole flash; the quartic ink budget (shaped to starve mid-dark glyphs) starved ordinary mid-tone art instead; PIVOT ate the field before anything flashed. Cosmos measured 0.005% of art-window pixels changed >8% under R5.
• uDepth decoupled from uDarken: uDarken is the blank-canon substrate, frozen by 30 saved canon files, and 0 on exactly the pale reverse sheets that need one. The substrate follows the pattern's dark half so it buys contrast, not a uniform dim (a flat 0.34 cost mean dL -21.7 and 29% of text contrast).
• Per-recipe families in patterns.ts (PARTICLE / SHEET / WASH / PEARL) plus seven by-eye corrections, each judged on a card the resolver assigns.
• Onset machinery stays deleted. Text guarantee re-verified: glyphs matte and crisp at peak tilt on a reverse whose foil covers the text box.
• Blank-card canon appearance unchanged: 8 patterns AE 0 (6 direct; cracked-ice and tinsel-ii against the pre-R5b code with canon held constant, since Chey re-saved those two mid-wave).
• SKILL.md composite contract rewritten around per-recipe selection.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2db0685 — 2026-08-07 — foil R6: composite defaults re-derived from Chey's canons; the guard learns to tell text from artwork

Chey: "i re-adjusted cracked-ice. take a look at it, and then make the changes based on everything you're observing to make all of this better - i want it so that i have to do as little of work personally on all the rest as possible."

WHAT HIS CANONS TAUGHT

uSheenTint 0 on all three non-mirror canons -> 0 for every non-mirror family (was 0.15/0.5/0.85/0.6). Ink-tinting a pattern's own highlight is exactly "it adds no rainbow color or anything to the card". uSheen he raised all three -> families now 1.2-4.2 (were 1.0-2.6). uDepth NOT universal: 0 on cracked-ice, 1 on tinsel-ii and rgs. It darkens the card WHERE THE PATTERN IS NOT, so its cost scales with 1 - duty. Measured every recipe's duty cycle; his three choices line up with it, and it predicted both dials for the only canon that exercises the defaults. uSat 1 + wide spread -> shipped on the 13 recipes with no canon.

Families are renamed for how light LANDS (flash / line / stamp / field / pearl / metal) and stated explicitly as FoilPattern.family. The stamp tier exists because the first pass got it wrong: energy-symbols-ii lights 5.8% of the face, so a field substrate darkened the other 94% and the card read "someone dimmed this".

THE INK GUARD: SPLIT, DON'T WEAKEN

inkDark is a local-contrast measure, so it fired on every dark mark - including every outline and shading edge in the ILLUSTRATION (measured: 86-89% of strong hits inside the art window are artwork, not text). The additive law spends it as a hard coverage multiply, so guard 1 punched the pattern full of holes. That is why he kept pulling it to ~0.51/0.56 on sheet/wash patterns.

Now split by glyphness into inkGlyph (sacred, zero flash, unchanged) and inkDetail (artwork darks keep the pattern on a tightened budget). inkAny is byte-for-byte the old inkDark and every pre-R6 consumer still reads it.

Success test met: at guard 1 under R6, tinsel-ii and rgs look BETTER than his 0.51/0.56 do today - tinsel-ii's artwork is visible through the streaks instead of greyed out. Both canons migrated to 1, reason recorded in each file's note; every other value in them is his, untouched.

SLIDER RANGES (his second ask)

Four dials sat exactly on their cap: uIntensity 2->4, uSheen 3->6, uSat 1->2, uHueSpread 1.5->3. Not extended: uInkGuard (his carve-out), uMetal (a law selector), the 0..1 mix weights, and uDarken/uDepth (already fully spent at 1).

uSheen needed a shader change to be honest: the additive budget's ceiling was reached at uSheen 2.0, so the top third of the slider bought almost nothing - very likely why he parked two canons at the ceiling. It now resumes climbing above 3, leaving every stored value <= 3 bit-for-bit identical.

The 19 duplicated lines across the two surfaces are one CORE_SLIDERS table; uGrain is dimmed and labelled "metal law only" where it does nothing.

NEW: "Apply composite -> family" copies only the on-card dials to siblings, never pattern shape.

PROVEN, NOT ASSUMED (AE vs a base build compiled from the previous commit)

• blank canon room: AE 0 across all 44 implemented recipes x 4 tones x 3 tilts
• mirror: AE 0 on the card path too - pixel-frozen
• text at flash peak: +0.1 to -1.1 points of glyph/paper contrast on every reverse whose foil covers the text box; the library's two largest costs (rainbow-glitter, water-web) toned down per-recipe and re-checked by eye
• every implemented recipe eyeballed on 1-2 real assigned cards at rest and two tilts; shots in ~/.pokedex-dev-hub-legacy/foil-shots/r6/

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

a4050a8 — 2026-08-07 — Mask provenance: sidecar v3, the AI learning loop, and the collapse safeguard

Chey (2026-08-07): mark AI-generated masks, let an AI learn from his hand-done ones, then let him correct the AI's mask so it "can observe the diff and continue to improve without me having to hand-paint all the masks."

derivation_method existed in v2 but the API HARDCODED it to 'hand' — harmless while only a Pencil touched the corpus, actively wrong once mask-refine's Flatten started writing machine geometry through the same save path.

Taxonomy (5 values, 4 cases that must never blur): layout-flatten machine geometry, unpainted weight 0 hand human, from a layout prior weight 1 hand-refined human, on an existing mask weight 1 ai generator, UNREVIEWED weight 0 ai-corrected AI proposed, human edited weight 0.6

The label is never taken from the client. The client reports only what the canvas was SEEDED with; writeMaskRecord derives the method by diffing saved pixels against what that seed rasterizes to. A machine label needs a full GeneratorIdentity, which HTTP callers cannot supply. One write path serves the route and generators both.

• Anti-feedback-collapse enforced in selectExemplars(), not by convention: unreviewed ai is weight 0 at any corpus size. Codify now routes through it.
• Corrections keep the parent PNG + change map + metrics (incl. a 4x4 grid of where corrections land) — the supervised pairs a generator lane consumes via the self-describing tuples manifest.
• Seam rule, measured not guessed: canvas roundRect vs the server SDF disagree on 389/330260 px, ALL in the 1px AA band. Without tolerating it every unpainted Flatten would stamp hand.
• Guard: if the file at the target path is an unreviewed ai mask it IS the parent, whatever the client claims — AI ancestry cannot be laundered. That guard caught a real bug during this build (ActionBtn passed React's MouseEvent into saveMask's optional arg, blanking derivation); derivation is now required on the PUT so it fails loudly instead.
• v1/v2 load forever (normalize-on-read); Chey's two masks migrated additively and still read HAND-PAINTED, agreement 0.641, pixels untouched.
• Trial batch wotc-window-trial-1: 6 WOTC Base holo rares from window-artgate@1, all ai/unreviewed, reversible. Leave-one-out on the human corpus beat the rect baseline (0.588 -> 0.688) — real, but n=2 same-set, so explicitly not validation.

Surfaces: provenance badge + detail in the workbench, Mask corpus panel, GET /foil-lab/masks/corpus (+?tuples/?exemplars), artifact routes, and corpus.ts report|exemplars|tuples|migrate. Pure tests added to CI (test:foil). Browser-verified at 390px and desktop. See DECISIONS.md 2026-08-07 and the mask-pipeline SKILL.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

b9c96f5 — 2026-08-08 — foil mask-straighten: read the INTENT out of Chey's hand mask, and make the undo real

Chey, 2026-08-08: "it's impossible to get the lines really straight so I'm hoping you can get computer vision on the mask and card art in tandem to really see my intent there, and make the mask nice and crisp and straight on the lines I was trying to draw along. Will you try that? And then I can either approve or revert."

Smoothing a wobbly boundary gives a smoothly wobbly boundary. The premise here is that a hand-drawn foil boundary is an ATTEMPT TO TRACE A PRINTED EDGE — frame, art box, species strip, stage tag — and those are dead straight in the scan. So: contour the mask, cut each loop into near-axis runs (PCA orientation, wobble gaps bridged), robust-fit each, hunt the scan's directional gradient with a local Hough over (angle, offset) plus a proximity prior, replace, intersect adjacent lines into real corners, rasterize matching his own AA character.

The refusals are the feature, and each is a recorded param: weak evidence, or an ambiguous band of comparable ridges, may nudge but never relocate; with no artwork edge it straightens only to HIS own fit; short and curved runs pass through untouched. A first cut "closed" the Tropius top-left corner — which turned out to be him deliberately tracing around the BASIC stage tag. The stricter rules keep it.

• apps/api/src/foil/line-snap.ts the geometry, pure, no I/O
• generator.ts line-snap@1, a REFINER (requiresSource)
• generate-masks.ts run --refine, archives, restoring revert
• provenance.ts supersedes + verbatim sha256 undo archive

A generator write onto an existing mask now THROWS without explicit supersede { runId }; with it, every artifact the mask had is archived verbatim under superseded/./ with a self-describing manifest, and revert --run-id verifies the whole archive BEFORE deleting anything live, then restores byte-for-byte. supersedes is deliberately NOT correction: one is a human editing a machine (training signal), the other a machine replacing a human with nobody's agreement. --refine also refuses a source with exemplar weight 0, so a refiner can never eat its own unreviewed output.

Tropius me05-001/37184 (run straighten-tropius-1): 8 runs snapped to printed edges, 1 self-straightened, 13 left exactly as drawn, 8 corners closed, 2 specks dropped. Max correction 3.74px, 1.06% of the face changed, Jaccard 0.9798, confidence 0.557. Honest residual: the card's bottom border was too faint to accept, so that run self-straightened and the report says so.

test:foil 25/25 (9 geometry + 5 supersede/restore, incl. byte-for-byte round trip, a deliberate curve surviving, an ambiguous band refused, and a corrupt archive aborting before deletion). Revert exercised twice end to end, sha256sum clean. Shots: ~/.pokedex-dev-hub-legacy/foil-shots/mask-straighten/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id straighten-tropius-1

b028e83 — 2026-08-08 — foil R7: his eye overturns two banked mechanisms; "no catalog cards" gets a reason

Chey's 15 canon-lab comments (2026-08-07), all resolved in issues/foil/.

Eleven shader verdicts. Two overturn mechanisms built from earlier corpus — energy-symbols' parity CHECKERBOARD (R3-GLYPH, from his own words) and prism's solid-facet grid (R3-MISC, from a Gemini delta pass). Having seen both rendered he rejected them. A description is a hypothesis about the render, not a spec.

cosmos the 0.085 UV field slide he read as parallax is now uP4 "Surface drift" at 0.012; zero at rest, canon unchanged striped-v-sheen grey lines: the unconditional silver term is gone AND the second-order lobe cut 0.4 -> 0.16 (it drew a whole second, near-achromatic line population on the far side) rainbow-mirror hue axis is card X, blue -> green -> red, on BOTH the sheen and the spotlight; core whitening 0.85 -> 0.32 energy-symbols checkerboard gone: per-symbol response AXIS + PHASE + transition WIDTH over a low-freq cluster field energy-symbols-ii same treatment (it shared the two-bank mechanism) pokeball-hologram flat SDF icon -> shaded 3D sphere (hemisphere normal, tilt-following light, belt, hub, two-tone shell) tinsel the "blinds" were a literal full-card striation floor; demoted to uP2, default 0 (his canon stores 0) tinsel-ii line-work gated to the colour band, silver 0.60 -> 0.10, band width exposed as uP4 prism pinwheel's spoke fan WITHOUT the per-wedge phase that makes a pinwheel spin; density 33 -> 18 radiant hologram step 0.5 cell -> uP4 0.17, drive rescaled so total travel per unit tilt is unchanged rainbow-glitter invisible on card: bigger white-lifted flakes AND the family fixed (flash -> field; no substrate, no contrast)

Canon policy held: shader.ts / canon.ts / CardViewer.tsx / data/foil-canon have a ZERO diff, sheenGlsl() is byte-identical (all four sheen canons frozen), and every new behaviour went into a NEW uniform old canon files inherit. No stored value migrated. One recommendation left for him, not applied: tinsel-ii's starkness is substantially his own canon (uDepth 1 at ~135 lines).

"Why are there no catalog cards?" — asked four times, three answers. The baked inversion is now v2 and emits a machine-readable diagnosis the lab renders: outranked diagonal-sheen-left — 1,818 SM reverses ARE cited for it, but energy-symbols wins them (the emblem layer vs the sheet rotation; both true). A capped SECONDARY pool is now baked via a new citedFoilPatterns() export and served as via:'cited', so the preview works. Never flip a winner to fill a pool. class-absent pinwheel, pokeball-hologram — the catalog has reverse variants for ex1-ex5 ONLY; the whole late-EX reverse era is missing upstream. The UI now says so instead of "no catalog cards". fixed vertical-sheen-rainbow 0 -> 12 cards. known_residuals recorded twice that the 6 Holo Rare Energies of ex13/ex16 use Sheen but could not be targeted without a card list; the catalog names exactly six per set. Two cardId rows added citing the same Collexy quotes, both residuals closed.

Typecheck (all workspaces) + web build + foil tests green; 14 recipes compile clean; before/after shots on blank + assigned card at 390px and desktop in ~/.pokedex-dev-hub-legacy/foil-shots/r7/. SKILL R7 field notes + DECISIONS appended.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

---

foil/mask-vector

Tip bf1ff3f219aa7a102c6de8ba5083b694c32cadfb · 62 commits · 2026-08-01 → 2026-08-09 · 10,533 words · 35 reproduced below.

350bb60 — 2026-08-01 — foil/main: workbench v1 — three.js viewer, 5-pattern starter library, era layout masks

Quarantined foil workbench at /pokedex/foil-lab (URL-only, linked from nowhere, chrome-free): real owned-card scans from the image cache, tilt via pointer/gyro(iOS permission)/manual with reduced-motion respected, uniform sliders, pattern + mask-scope overrides, mask overlay toggle, copy-recipe JSON. Pattern library v1 tuned against owned eras (Base/WOTC, SV, Mega Evolution): Starlight, Cosmos, SV default holo, SV reverse sheet, Cracked Ice — uniform contract (core + uP0..3) designed for the full 15-20 set. Era layout spec as data (era-layouts.json, rects measured on real scans); resolver v1 (series, rarity, variant kind) -> pattern + mask scope. SKILL.md x2 (foil-effects, mask-pipeline stub). main.tsx gains a dev-only double-createRoot guard (Vite HMR entry-invalidation footgun, see DECISIONS).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

bef8153 — 2026-08-01 — foil/main v2: Pencil mask editing, comment queue, iPad two-column, branch api

Apple-Pencil hand-mask drawing on the workbench (MaskEditor: card-aligned canvas overlay, pen+mouse default w/ finger toggle, pressure brush, eraser, 12-step undo, touch-action none, starts from layout prior); masks persist as committed ground-truth artifacts data/foil-masks//.png+json (derivation_method hand) via the env-gated branch api (POKEDEX_FOIL_LAB=1, port 3712, PGPOOL_MAX=1 — inert in prod), auto-load and beat the layout tier (uMaskTex contract; CanvasTexture flipY=false, exactly one V flip). Comment button captures card/variant/pattern/scope/mask state + all slider values to issues/foil// (report.md + context.json), out of the fix-issues sweep. Era-grouped owned-card picker (WOTC / SWSH / SV+ME headings); two-column layout (viewer | controls) from 700px — verified at 390px, 744x1133, 1133x744, 1440x900 with simulated pen strokes. SKILL.md x2 updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

3707066 — 2026-08-02 — foil: sidecar v2 (prior+diff), artwork-keyed masks, comment linkage, codify ritual

• Every hand-mask save records its starting prior (era/scope/resolver rect + RESOLVER_VERSION) and the api renders .prior.png + .diff.png (green = human added, red = removed) + Jaccard agreement, server-side via a pure-JS PNG codec (no native addons). Saves without a parsable prior 400.
• Machamp mask backfilled (deterministic prior wotc/window/v1): agreement 0.6409 — his correction is the subject silhouette, the rule's first lesson.
• Artwork keying: variants of one cardId provably share one scan, so mask GETs alias to a sibling variant's mask when prior.scope matches (X-Foil-Mask- Alias-Of). Cross-card reprints unprovable from the catalog -> per-card only.
• Comments auto-link the exact saved mask state (maskFile/maskSavedAt/ maskAliasOf/maskHasPriorDiff).
• mask-pipeline SKILL: sidecar v2 storage, identity rule + limits, and the codify ritual; first worked example data/foil-masks/codified/wotc.md (n=1: WOTC window = art-window rect minus subject silhouette; noted in era-layouts.json wotc.notes).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

b8fc7c7 — 2026-08-02 — foil: any-card picker — full-catalog browse/search, Owned-only as a toggle

Chey: pull up ANY card in the catalog and tilt it, owned or not. Picker now browses the whole catalog by era -> set (paged strip, promo sets run 300+), adds full-catalog name search, and keeps Owned-only as a filter toggle (default on = prior behavior). Zero new endpoints: existing /series, /series/:slug, /sets/:setId (paged), /search, /cards/:cardId serve it all; unowned cards list their catalog variants (quantity 0) and resolve the same base-guess pattern path. Selection is never clobbered by filters; unmapped series get an honest "Other eras" bucket. Verified on unowned neo2-13 Umbreon (Auto - Starlight WOTC) at 390px + both iPad viewports.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

1cfc897 — 2026-08-02 — foil: taxonomy integration — all 39 pattern types in the app, cited resolver

Mislabel corrections from the video-reference research, applied to apps/ (research docs untouched):

• sv-holo renamed vertical-sheen (it always rendered the HGSS-XY vertical bands, not SV's). PATTERN_ALIASES keeps old slugs resolving forever — sidecars, comment context.json, Copy-recipe JSON never orphan.
• Sheen family = ONE generator at four rotations + stripe option: five honest implemented slugs (vertical / horizontal=TRUE SV default "Mirage" / diagonal right "/" / diagonal left "" / striped "Line").
• Cosmos label drops "/ Galaxy" (Galaxy is a starlight synonym; synonym moved to the starlight entry). reverse-sheet description notes it ≈ pokeball-masterball stamped sheet.
• Vintage narrowed: starlight = Base/Jungle/Fossil only; Base Set 2 → Call of Legends holos = cosmos.

Pattern dropdown carries the FULL 39-slug vocabulary, grouped: implemented recipes first, then every gap type rendering via its nearest recipe and labeled "approx via ..." (amber note in the UI). Unimplemented != hidden.

Resolver v2: base guess reads the cited usage table (usage-index.json, derived from research/foil-pattern-usage.json by tools/foil/build-usage-index.mjs; set-name match -> era/series token match -> era-default heuristics). The winning row's confidence + citation hosts surface in the workbench ("Guess: set-level citation, high confidence (bulbapedia..., collexy...)"). Reverse rows citing 'mirror' are deprioritized per the research known-gap (ink-design evidence, not foil-pattern evidence). RESOLVER_VERSION -> 2 (scope semantics unchanged from v1).

Verified: typecheck + build green; Playwright at 390/744x1133/1133x744 (~/.pokedex-dev-hub-legacy/foil-shots/w2/) — Machamp auto=starlight set-level-high with his hand mask still resolving, SV holo auto=horizontal-sheen, SV reverse auto=energy-symbols-ii, fireworks approx labeling.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

fbf6a4c — 2026-08-02 — foil: fix diffuse darkening — scan sampled sRGB-decoded but never re-encoded (issue ls9u0y)

The face texture was uploaded SRGB8_ALPHA8 (hardware sRGB->linear decode at sample time) while the foil ShaderMaterial writes gl_FragColor raw — three.js only appends the linear->sRGB colorspace_fragment chunk to built-in materials — so every scan rendered in linear values displayed as sRGB (measured flat 184->123 / 199->147, exactly the sRGB->linear curve). Sample the scan undecoded (NoColorSpace): the compositing model is authored in display space and pattern=none at rest is now pixel-comparable to the flat view (mean abs luma diff 44.55 -> 2.16, mapping identity +/-2). Also zero pattern-none's default specular so the baseline is truly the plain scan.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

3960827 — 2026-08-02 — foil: workbench split — canon pattern lab (A) | card adjustment surface (B)

Chey's comment issues/foil/2026-08-02_12-59-52-368_4aq756 (resolved here):

• /pokedex/foil-lab/canon (CanonLab.tsx): bare holofoil pattern on a blank card (no ink; black/dark/silver/white flat-tone base via the normal texture path) beside the REAL reference tilt clip + 8 keyframes streamed from research/foil-video-reference/ by new env-gated dev routes (byte-range via res.sendFile for iOS Safari). Full 39-slug picker, sliders, tilt, and Save canon -> data/foil-canon/.json (full uniform snapshot that replaces recipe code defaults as the baseline).
• /pokedex/foil-lab (FoilLab.tsx reframed): per-card masks/comments unchanged, plus per-card uniform overrides saved as a SPARSE diff vs the canon baseline -> data/foil-overrides//.json (untouched uniforms keep tracking canon). Off-canon sliders get a dot; saving with sliders matching canon deletes the override.
• Layering model in foil/canon.ts (code defaults < canon < override < live); shared atoms + surface tabs in foil/ui.tsx; routes wired in main.tsx; AppShell chrome-free check widened to /foil-lab/*; both data dirs re-included in .gitignore; ports 5184/3713 recorded in ORCHESTRATION.md.

Verified in-browser at 390x844 / 744x1133 / 1133x744 (clip playing beside the bare cosmos render; Machamp base1-8 hand mask still resolves) — shots in ~/.pokedex-dev-hub-legacy/foil-shots/canon-lab/. Typecheck + build green (web, api). patterns.ts untouched (R0 re-tunes concurrently).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

49c4db6 — 2026-08-02 — foil R0: re-tune implemented recipes per Gemini verification notes

cosmos: full activation-model rewrite — dark field, smaller/denser orbs mostly near-invisible, cluster pops via low-freq noise over cell ids, pinprick 4-point twinkles (was: dense wall of large saturated orbs). starlight (+ starlight-ii via shared GLSL): parallax architecture untouched; tighter visibility lobe (pow 5->9, floor .18->.08), sharper glyph cores + narrower flares, smaller blobs, saturated star color, galaxy wash halved; starlight-ii art gate lowered for the mid-orange Evolutions field. cracked-ice: authored intra-shard grain removed per Chey's accuracy ruling — facets flash as solid smooth planes; whitening at peak reduced, uSat up. sheen family: generator grows sharp/beam/barcode options. vertical-sheen -> new barcode variant (thin spectral lines of varying width, CD-groove slide) with tamed beam + art gate. diagonals: sharp 3.0, beam 0.55, specular 0.35 (center blow-out), diagonal-sheen-right uP0 2->7 (same-sheet consistency). striped-vertical-sheen: stripes finer + more blended (90->130, floor .40). horizontal-sheen: untouched (20/20).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

0597b19 — 2026-08-02 — foil R0: rounds 2-3 re-tunes + verification report, field notes, decisions

starlight: longer thin-armed 4-point glyphs + 8-point-burst subset, tighter pop lobe (pow 11, floor .04), full-saturation star color, parallax default 2.4, gain 3.6 — best of 3 rounds 11/20 (still-frame judging cannot see the parallax; recorded as honest residual with a video-judge recommendation). starlight-ii: 20/20 (uP2 pinned so base-default drift can't invalidate it). cracked-ice: facet gain 1.1 -> 0.55 (art visible through flashes), density 10, less whitening — 15/20 yay; shard-geometry uniformity queued for R1. Verification report gains the R0 section (before->after per pattern, the diag-left judge-noise event, diffuse-fix supersession note); foil-effects SKILL.md gains the R0 field notes; DECISIONS entry for root cause + wave.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2419ed9 — 2026-08-02 — foil: cited per-card pattern assignments across the whole catalog (resolver v3)

Chey's mlmwmp comment: the library was funneling to era defaults ("cosmos"). Full-catalog enumeration (20,964 cards / 37,627 variants, via the branch read api) + an 18-lane Ringer research swarm (foil-card-assignments, GLM-5.2, executed selector-integrity checks) produced:

• research/foil-card-assignments.json — 111 set/rarity/kind/card-level rows, 19 catalog foil-facet mappings (holo-foil-cosmos, reverse-foil-masterball, normal-foil-galaxy...), 55 honest residuals; per-lane provenance + method.
• tools/foil/build-assignments-index.mjs -> apps/web/src/foil/assignments-index.json (trimmed for the bundle, duplicate-facet guard).
• resolver v3 (RESOLVER_VERSION=3): assignment tier above the v2 usage table; specificity cardIds > variantKinds > facet > rarities > set > v2 > heuristic; normal-kind variants with a declared foil facet now render window foil (previously scope none); FoilLab passes cardId.
• era-layouts: wotc seriesSlugs 'ecard' -> 'e-card', + 'legendary-collection' (both had fallen through to modern-sv rects and heuristics).

Coverage (foil-bearing variants): 79.5% set-cited / 20.4% era-token / 0.2% heuristic -> 93.4% set / 4.3% facet-or-card / 2.4% era-token / 0 heuristic. 13 in-browser spot-checks in ~/.pokedex-dev-hub-legacy/foil-shots/assignments/. Issue mlmwmp resolved. Weak spots + curation ledger in DECISIONS.md.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

780368f — 2026-08-02 — foil R1: dedicated recipes for the twelve owned-era gap patterns (10/12 verified)

Chey's comment dml369 ("develop out the rest of the missing holofoil patterns"): dedicated GLSL for fireworks, ace-spec, energy-symbols-ii, rainbow-glitter-sheen, ex-starfoil, prismatic-pokeball, tinsel-ii, cosmos-iii-smooth, pokeball-masterball, radiant, rainbow-glitter, confetti — implemented flags flipped, approx labels dropped, slugs stable. Gemini-verified 10/12 match over 3 rounds (verdict table + honest residuals in research/foil-verification.md R1 section). Honest nays: tinsel-ii (procedural static plateaus at judge score 2) and prismatic-pokeball (screen-only blend cannot darken a white reverse body — structural, Chey's call).

R0 residuals cleared in the same pass: cracked-ice anisotropic shattered-glass metric (re-judged yay 14/20); starlight milky wash + per-layer visibility curves with starlight-ii re-banked clean 20/20 through both shared-GLSL changes; starlight parallax judged from a new 16-frame fine sweep and still scored flat — verified real by eye, recorded as judge-blindness.

Field notes appended to foil-effects SKILL.md; dated DECISIONS entry; issue dml369 resolved.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

1764745 — 2026-08-02 — foil R2: uDarken blend-model term — prismatic-pokeball 17/20 yay, tinsel-ii 16/20 yay

One opt-in substrate-darkening term in the shared composite (main()): body = scan * (1 - uDarken * mask * gate) before the foil screen-blends. Default 0 is bit-identical to the old render; absent canon/override keys seed 0. Full 21-pattern regression sweep on provably identical renders: no true regressions (and a measurement of Gemini single-roll noise, ±3-6; diag-right + pokeball-masterball double-nays recorded as unstable verdicts on geometry proof). prismatic-pokeball rebuilt on the term (dark-mirror base, facet-quantized flash lobe, ball watermark as ink-overprint suppression): structural nay -> 17/20 yay. tinsel-ii one-line opt-in (uDarken 0.4): static plateau -> 16/20 yay. 20 of 21 implemented types now match; sole nay starlight (judge parallax blindness).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

efd1402 — 2026-08-02 — foil R2 stage 3: thirteen dedicated recipes (gap patterns + RC dots) + subset-commons resolver tier

• patterns.ts: real recipes for mirror, rainbow-mirror, vertical-sheen-rainbow (dark-mirror family on uDarken), pokeball-hologram, prism, water-web, energy-symbols, pinwheel, ex-emerald, tinsel, cosmos-ii-pixel, crosshatch, radiant-collection-dots — implemented: true, approx labels dropped, slugs stable. Unused plain SHEEN_V generator removed (last fallback users gone).
• resolver v4: card-level 'normal'-class assignment rows consulted before the scope-none early return — subset cards the catalog declares plain but which physically carry foil (RC commons dot overprint) resolve their subset's pattern with scope full.
• foil-card-assignments.json: two card-level rows (bw11-RC1..20, g1 25 non-ultra RC ids) with video citation (non-holo RC Teddiursa keeps the dot overprint); builder enforces cardIds on 'normal' rows; index regenerated.

Verified in-browser: g1-RC1 + bw11-RC1 render the RC dot treatment via Auto (guess card|medium). Gemini verdicts in research/foil-verification.md (R2 recipe-wave section, next commit).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9574a4a — 2026-08-02 — foil R2b: four §40–43 vocabulary recipes (all match) + resolver v5 scope overrides

The last lane of the wave — dedicated recipes for the vocab-extension types: gold-secret (warm-locked goldRamp field + chromatic glitter pops + SWSH burst rays — 19/20 yay), vstar-pearl (near-white pearl on uDarken 0.3, warm-led diagonal wash, frame-line streaks — 13/20 yay), shiny-vault (silvery field + band-keyed sparkle-glyph amplifiers, uDarken 0.15 — 11 nay → 16/20 yay r2), detective-pikachu (diagonal sheen × photo-luminance via the documented uFace contract exception — 20/20 yay). 38/43 taxonomy types real, 34 hold match.

Resolver v5: assignment rows carry an optional per-row scope override (window/ full/sheet), honored by the set-row tier and the cls-'normal' tier — baby shinies render window despite the 'Shiny rare' full-foil rarity, VSTARs full despite plain holo kinds, det1 window despite 'ultra rare'.

Assignment research: +2 facet rows (gold, gold-jumbo → gold-secret; retires both facet residuals, 1-card metal collateral flagged) and +10 cited rows — sma 51 baby/35 GX/8 gold splits, swsh4.5sv 104 baby + 18 full-face (incl. the black shiny Eternatus pair, name-verified not gold), sv04.5 rarity split, VSTAR via the dedicated 'Holo Rare VSTAR' rarity, det1 ultra-rare + cls-normal rows (all 18 cards window). Residuals annotated with resolutions.

build-usage-index.mjs: tolerate era-wide rows with no scope.sets (the vocab gold rows crashed the build — the 122-row usage index never compiled until now). Verdicts/frames under ~/.pokedex-dev-hub-legacy/foil-verify/ (verify-

-r2b*); shots in foil-shots/r2b/. Docs: foil-verification.md R2b section, SKILL R2b field notes, DECISIONS entry.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

17f114d — 2026-08-03 — foil R3: sheen-family streak-field rework — all 8 of Chey's canon-lab comments resolved, 7/7 Gemini 20/20

sheenGlsl is now a streak field (Chey's critique = the spec): sparse irregular streaks, per-streak lean following card tilt (crisscross; converging pairs terminate where they meet), stretched-ellipse taper, hue running ALONG each strip as well as across. striped-vertical-sheen rebuilt from a fresh Gemini corpus re-spec run with Chey's description embedded (grouped reveal via wide sweeping windows over a subtle bottom-converging fan; gemini-spec-r3.md); ex-starfoil rebased on the reworked diagonal base + fine CD-line layer; vstar-pearl rebuilt from the horizontal-sheen streak field, pearl kept.

Canon migrations (full-snapshot rule, DECISIONS): striped + horizontal uDarken 0 -> 0.32 (bright-substrate legibility, third confirmation). All other canon values carry over; uniform semantics unchanged. validate_spec.py frame regex single-digit fix (undercounted two-digit citations, failed an honest worker).

Verdicts (Chey's words folded into every judge prompt as acceptance criteria): vertical/right/left/vstar-pearl 20/20 first roll; striped + ex-starfoil 20/20 round 2; horizontal 20/20 round 3. Table + residuals in research/foil-verification.md R3.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

712702f — 2026-08-03 — foil R3-MOTION: axis-split starlight, per-dot cosmos, hologram radiant, the rainbow-glitter-sheen delta

Chey's six motion-model canon-lab comments resolved (his eye = ground truth):

• starlight (+ starlight-ii via shared GLSL): top-to-bottom hue BANDING (soft-quantized bands; his canon 0.62/0.6 lands the reference order blue->green->red on the R->B->G cosine ramp) + AXIS-SPLIT tilt: vertical slides the whole field (~10% card height), horizontal drives the per-star random fade; his 3-layer uP1 parallax rides on top untouched. Judge: starlight 17/20 yay round 2 — the FIRST starlight yay (standing parallax nay broken); starlight-ii 20/20.
• cosmos: whole-field slide + per-dot random response (independent random tilt axes for brightness and hue per orb — no shared sweep axis); at tilt 0 his canon appearance is pixel-identical. Judge: 20/20.
• radiant: discrete hologram steps (half-cell up/down, 60% hold / 40% crossfade — one position fades out as the next fades in), uP1 wired as Hologram travel (canon migrated 0 -> 2.2; was an unused placeholder), grid lines ~45% thicker. Judge nay 13/20 across 3 rounds on a pixel-refuted "slides continuously" claim — recorded as still-frame motion blindness (same class as starlight parallax); adjacent-frame crops prove discrete positions + crossfade. Chey's live tilt is the tiebreak.
• rainbow-glitter-sheen: dedicated Gemini delta-articulation pass first (he couldn't name the difference), pixel-verified, then fixed: narrow laser-saturated striped chevron (sigma 0.010->0.0035, hue traversal 5x->9x, pow-deepened at low gain), fainter +-0.8 repeats, chevron angle 1.3->1.9, denser colored always-visible glitter, uDarken 0->0.4 (4th legibility-physics data point). Judge: 19/20 yay.

Docs: research/foil-verification.md R3-MOTION section, DECISIONS entry, foil-effects SKILL field notes. Frames/verdicts in ~/.pokedex-dev-hub-legacy/foil-verify (jobs -r3m, manifest-r3m{,2,3}, delta-rainbow-glitter-sheen); gate shots in foil-shots/r3-motion.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

c961b1a — 2026-08-03 — foil mask-refine: window handles → flatten → hand refine

Chey's requested workflow on the Card-adjust surface: draggable corner/edge handles adjust the layout window rect per card (WindowEditor.tsx, 44px touch targets, finger-drag allowed), the adjusted geometry persists as artwork-keyed scope-agnostic sidecars (data/foil-windows//.json, /foil-lab/windows routes), and Flatten bakes it through the STANDARD hand-mask save path (sidecar prior stays the era rule; new optional prior.window records the adjustment) then opens the existing paint tooling.

• rasterizeWindowRect shared by Flatten + loadLayoutRect (identical bakes)
• saveMask prior = layoutMask (rule), never the effective mask
• windows GET aliases across variants of the same scan, newest savedAt
• verified in-browser desktop + 390px: sma-SV1/sma-SV2 (SM baby shiny, borrowed modern-sv rect) and me01-034 (zero-coverage gap); base1-8 + base1-5 hand-mask regression clean; test artifacts deleted (corpus rule)
• ports: vite 5186 / api 3714 recorded in ORCHESTRATION.md

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

0a72545 — 2026-08-03 — foil R3-GLYPH: glyph-pattern reworks per Chey's 7 comments + the drop-in glyph slot for his SVGs

Behavior (his notes verbatim = acceptance criteria): reverse-sheet rainbow only on grainy glyphs over a neutral sheet; energy-symbols checkerboard banks that swap with tilt; energy-symbols-ii sporadic random banks (~half faint at any time); ace-spec diamonds grow/shrink with tilt + a touch of blur; prismatic-pokeball ball never darkens (white-mixed pale response + coherent plane flash, suppression deleted); radiant-collection-dots shapes catch a traveling rainbow band (16/20 YAY — standing R2 nay broken).

Glyph slot: research/foil-glyphs//glyph[-N].svg -> lab-gated api routes -> atlas rasterizer (foil/glyphs.ts) -> uGlyphTex contract + glyphTex() helper; ~2.5s auto-pickup on save, procedural fallback when absent, README tells Chey exactly where his four promised files go.

Canon migration: reverse-sheet uP1 0 -> 0.6 (dead placeholder now drives his requested glyph grain). Verdicts + pixel-refutations recorded in research/foil-verification.md (R3-GLYPH); lessons in foil-effects SKILL.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

82a4f5a — 2026-08-03 — foil R3-MISC: the uTint ink-metallic blend term + Chey's twelve remaining canon-lab comments (12/12 yay)

The reverse-holo fix (his chat report): screen-blending achromatic foil light over colored art compresses chroma — "dull and grayish, rather than making the color look metallic." New core uniform uTint (default 0 = bit-exact legacy) multiplies the foil layer + in-mask specular by the luminance-normalized scan chroma SQUARED (the double ink pass of a real mirror reflection); neutral on blank cards, so no canon-lab appearance moved. Reverse-family recipes opt in; verified before/after at 5 tilt angles on Victini ball-reverse, Crystal Energy, Pineco. uP4/uP5 param slots added in the same contract change.

The twelve comments, each mechanism-mapped in its issue frontmatter and the verification doc: cracked-ice triangulated + half-off gate; fireworks grid lattice + center-emanating radial hue; rainbow-mirror rebuilt as mirror + hue-banded spotlight; tinsel-ii silver lines + vertical rainbow band (band lives only on the line-work); prism rebuilt as pinwheel-kin from the corpus-vs-corpus delta pass; cosmos-ii pixel-quantized stamp edges; crosshatch thinned; gold-secret adjustable burst origin (uP4/uP5, center default, per-card via overrides) + holographic grain; dedicated recipes for sequin (glyph-family glints), tcg-classic (flat starlight + rainbow glitter), acid-wash (water-web kin), disco (perfect disc lattice, galaxy ignition) — zero approxVia fallbacks remain.

Canon migrations (appearance-preserving unless he asked otherwise): fireworks uP0 3→4.5 (dead two-octave structure), gold-secret uP4/uP5 0.5 added, sequin uP0-uP3 re-keyed to the new recipe. Verdicts: research/foil-verification.md "R3-MISC" — 12/12 Gemini yay incl. the tcg-classic ink-scope re-roll.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

32f8bf3 — 2026-08-03 — foil R4-COMPOSITE: the ink-density invariant — foil adds pop, never washes text or mutes ink

Chey's two me05 comments (7rtnzx mirror reverse "blows out the darks/text", 19mo4l holo "darkens/muddies the colors") ruled a CORE COMPOSITE INVARIANT, engineered in shader.ts main(): ink density estimated from the scan (inkDark = darker-than-local-8-tap-field average, exactly 0 on flat blank bases; inkColor = chroma above the canon-lab-tone floor) now gates every layer — dark ink blocks the flash + shields the specular (text crisp at every tilt), ALL ink is exempt from uDarken substrate attenuation (printed color never muted; the mirror stays dark only where the mirror is visible), colored ink auto-tints its flash (max(uTint, inkColor)) and uInkPop pumps chroma under the flash so colors read MORE saturated and metallic.

New core uniforms uInkGuard (default 1) / uInkPop (default 0.5), both 0 = bit-exact legacy composite; sliders in both labs; canon files untouched (absent keys inherit the new defaults — catalog-wide fix, zero migrations).

Proof: canon-lab blank-card pairs (mirror dark+silver, cosmos, horizontal-sheen) pixel-identical (compare -metric AE 0) under a frame-stepped rAF harness with a same-settings AE-0 control, plus a Gemini sanity pass (all pairs "identical"). On-card before/after pairs across the 8-card sample (both issue cards, Victini, Pineco, Darkrai, Cetitan, Machamp WOTC, Hydreigon SIR; 3 tilts, 390px + desktop) in ~/.pokedex-dev-hub-legacy/foil-shots/r4-composite/. Both issues resolved with mechanism-mapped notes; composite contract documented in the foil-effects SKILL; DECISIONS + research/foil-verification.md updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

7de053a — 2026-08-04 — foil R4b: scan-additive composite (rest parity by construction) + canon-lab card preview

Chey's Grubbin (me05-002, Pitch Black) review refuted R4's ink heuristics: the reverse still muddied the card at rest (measured ΔC +32, green→yellow), darkened the header under tilt (−22 L), and blew out text (glyph contrast −26%). New law on real scans (uScanBase 1): the scan is a photograph of the card at rest — foil is additive dynamic light only, tilt-gated to ZERO at rest, luminance-headroom + per-channel clamped (text sacred at every angle, uArtGate opens the clamp where dark scan pixels ARE the foil), chroma- preserving with a pastel-safe tint ramp and an L²-gated chroma pump. Canon lab blank bases run the classic composite textually unchanged: AE-0 proven (CDP frame-stepped harness, control + knob pairs; Gemini pass agreed). uInkGuard 0 still reproduces the pre-R4 composite exactly.

Canon-lab card preview (his request verbatim): blank/on-card/↻-another chips render the live sliders on a random catalog card the resolver assigns the pattern to — server-sampled from the baked inversion (tools/foil/ build-pattern-cards.mts → data/foil-pattern-cards.json, gitignored; GET /foil-lab/pattern-cards/:patternId), honoring hand masks + adjusted windows; empty-pool patterns say "no catalog cards". Also commits Chey's live mirror-canon tweak from the Grubbin session. Full metrics + verdicts: DECISIONS 2026-08-04.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ca7fa20 — 2026-08-04 — foil R4c: onset ease — the sheen leans in instead of switching on

Chey on R4b's rest gate: "it's like it isn't there at all until I've tilted the card a tiny bit and then it's like it just suddenly appears." Measured: the headroom clamp saturated R4b's smoothstep(.02,.28) gate, so ALL added light arrived inside tilt 0.03-0.13 (0 -> 16.1 dL mean) then froze flat to tilt 1.0 — a step smeared over 4% of the tilt range.

Now: one wide ease ramp = smoothstep(0, .45, |tilt|) (full at 0.45 — every R4b tilt-0.5 metric unchanged), flash gated by ramp^1.5 (cubic-slow start), specular by ramp^2.5 (the broad gloss band now trails the flash instead of leading it), plus a REST=0.006 sub-JND living rest sheen blended into both gates (rest +0.43 dL card-mean, max 2/255 — visually still the scan). After-curve: smooth S over 0->0.25, no dead zone, no knee; text clamp untouched (glyph contrast 33.6 -> 33.6 at flash peak); blank-card canon lab bit-level unchanged (fuzz-1% AE 0 vs control floor).

Harness + curves + filmstrips: ~/.pokedex-dev-hub-legacy/foil-shots/r4c/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

cb24f25 — 2026-08-04 — foil R4d: motion-first onset — tiny gestures move the sheen; glow arrives late; Chey gets the dials

Chey rejected R4c too ('It still lights up pretty noticeably when I tilt the card a tiny bit. I don't like it.') — measured the gesture→tilt mapping first this time: pointer is 0.0051 tilt/px at 390px and gyro is 0.036 tilt/deg, so a 30px thumb drag (tilt 0.154) was already getting HALF the full glow under R4c. The curve was never the felt variable.

• shader.ts: onset gates now uniform-driven — uOnsetRange (glow full at |tilt|, default 0.5) and uOnsetCurve (flash exponent, default 3.5; specular trails at +1). (0.45, 1.5) reproduces R4c exactly (verified to the third decimal); defaults are strictly softer at every tilt and equal from 0.5 up, so all R4b/R4c peak metrics are unchanged (Grubbin 17.976 ΔL / text 34.02, Machamp 18.035 / text 45.6 — exact).
• FoilLab + CanonLab: 'Onset range' / 'Onset curve' sliders, canon- stored like every other uniform; absent canon keys inherit defaults.
• patterns.ts: CoreDefaults admits the two keys per recipe.

Pointer-path proof (real pointermove, 390px, 30px drag, first 300ms): before mean ΔL climbs to 6.6 and rising; after stays within ±0.08 while per-pixel movement reaches 6.1 mean / 18.9 p95 — moves, doesn't light up. Blank canon lab at fuzz-1% AE 0 vs control. Artifacts in ~/.pokedex-dev-hub-legacy/foil-shots/r4d/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2746168 — 2026-08-05 — foil R5: the metallic composite — always on the card, energy redistributed, Chey's compositing dials

Chey's ruling (2026-08-05, verbatim in DECISIONS): no onset range at all — the effect is ALWAYS on the card; dial in how it composites onto the ink; metallic and spectral, not brightening.

• Onset machinery removed (R4c/R4d ramp, REST floor, uOnsetRange/uOnsetCurve sliders on both surfaces). No canon/override file ever stored the keys; CardViewer skips unknown uniforms, so stale keys anywhere are inert.
• R5 metalness law (scan path only): scan = albedo; the pattern field splits around PIVOT 0.40 — above = tinted-specular highlight, below = bounded multiplicative depth (mirror-turned-away darks, sqrt-concave). Energy redistributes; contrast + chroma carry the metallic read.
• Soft-knee headroom: the R4b hard clamp plateaued strong fields into a flat wash (the "brightened" read); the budget now compresses monotonically so pattern structure survives. Per-channel no-clip caps stay hard; glyph ink (inkDark) exempt from highlight AND depth — text contrast preserved (Grubbin 86->97 at peak, Machamp flat).
• New canon-stored dials on both surfaces: Metallic 0.6, Sheen strength 0.55, Sheen tint 0.5 (= the R4 chroma law exactly), Depth 0.55, Texture 1. uMetal 0 / uIntensity 0 / uInkGuard 0 all still yield the plain scan.
• Blank canon lab pixel-unchanged: CDP frame-stepped before/after AE 0 at fuzz 1% (mirror/cosmos/horizontal-sheen/starlight, tilted + rest); one Gemini vision pass agrees (all pairs identical). Shots + per-card rest pairs: ~/.pokedex-dev-hub-legacy/foil-shots/r5/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

23e225b — 2026-08-07 — foil R5b: metalness is MIRROR ONLY — every other pattern gets its own colour back

Chey: "You applied the metallic treatment to EVERY HOLO PATTERN when I only wanted it for mirror. So now, every single holo pattern looks like mirror and you can barely see the pattern at all, it's like every single one is being applied in a way that it adds no rainbow color or anything to the card."

R5 shipped its metalness dials as GLOBAL_DEFAULTS, so all 43 recipes inherited a law built for one of them. uMetal is now the scan-path LAW SELECTOR and defaults to 0 — nothing can inherit metalness; mirror is the only recipe that opts in, in its own patterns.ts defaults.

• Metalness branch is the R5 code textually unmoved, so mirror's arithmetic is unchanged by construction. Proven: his hand-tuned canon on Grubbin me05-002
  • Tropius me05-001, rest + tilt 0.2/0.35/0.5, 390px and desktop — 16/16 frames AE 0 at fuzz 1%, same floor as the same-build control pair.
• uMetal 0 runs the ADDITIVE law: the pattern's own emission (uSheen gain) screen-combined per channel over a substrate-attenuated body, tinted toward the ink only as far as uSheenTint asks. Three fixes to what R5 broke — the scalar no-clip cap let the artwork's brightest channel throttle the whole flash; the quartic ink budget (shaped to starve mid-dark glyphs) starved ordinary mid-tone art instead; PIVOT ate the field before anything flashed. Cosmos measured 0.005% of art-window pixels changed >8% under R5.
• uDepth decoupled from uDarken: uDarken is the blank-canon substrate, frozen by 30 saved canon files, and 0 on exactly the pale reverse sheets that need one. The substrate follows the pattern's dark half so it buys contrast, not a uniform dim (a flat 0.34 cost mean dL -21.7 and 29% of text contrast).
• Per-recipe families in patterns.ts (PARTICLE / SHEET / WASH / PEARL) plus seven by-eye corrections, each judged on a card the resolver assigns.
• Onset machinery stays deleted. Text guarantee re-verified: glyphs matte and crisp at peak tilt on a reverse whose foil covers the text box.
• Blank-card canon appearance unchanged: 8 patterns AE 0 (6 direct; cracked-ice and tinsel-ii against the pre-R5b code with canon held constant, since Chey re-saved those two mid-wave).
• SKILL.md composite contract rewritten around per-recipe selection.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

1705a39 — 2026-08-07 — foil R6: composite defaults re-derived from Chey's canons; the guard learns to tell text from artwork

Chey: "i re-adjusted cracked-ice. take a look at it, and then make the changes based on everything you're observing to make all of this better - i want it so that i have to do as little of work personally on all the rest as possible."

WHAT HIS CANONS TAUGHT

uSheenTint 0 on all three non-mirror canons -> 0 for every non-mirror family (was 0.15/0.5/0.85/0.6). Ink-tinting a pattern's own highlight is exactly "it adds no rainbow color or anything to the card". uSheen he raised all three -> families now 1.2-4.2 (were 1.0-2.6). uDepth NOT universal: 0 on cracked-ice, 1 on tinsel-ii and rgs. It darkens the card WHERE THE PATTERN IS NOT, so its cost scales with 1 - duty. Measured every recipe's duty cycle; his three choices line up with it, and it predicted both dials for the only canon that exercises the defaults. uSat 1 + wide spread -> shipped on the 13 recipes with no canon.

Families are renamed for how light LANDS (flash / line / stamp / field / pearl / metal) and stated explicitly as FoilPattern.family. The stamp tier exists because the first pass got it wrong: energy-symbols-ii lights 5.8% of the face, so a field substrate darkened the other 94% and the card read "someone dimmed this".

THE INK GUARD: SPLIT, DON'T WEAKEN

inkDark is a local-contrast measure, so it fired on every dark mark - including every outline and shading edge in the ILLUSTRATION (measured: 86-89% of strong hits inside the art window are artwork, not text). The additive law spends it as a hard coverage multiply, so guard 1 punched the pattern full of holes. That is why he kept pulling it to ~0.51/0.56 on sheet/wash patterns.

Now split by glyphness into inkGlyph (sacred, zero flash, unchanged) and inkDetail (artwork darks keep the pattern on a tightened budget). inkAny is byte-for-byte the old inkDark and every pre-R6 consumer still reads it.

Success test met: at guard 1 under R6, tinsel-ii and rgs look BETTER than his 0.51/0.56 do today - tinsel-ii's artwork is visible through the streaks instead of greyed out. Both canons migrated to 1, reason recorded in each file's note; every other value in them is his, untouched.

SLIDER RANGES (his second ask)

Four dials sat exactly on their cap: uIntensity 2->4, uSheen 3->6, uSat 1->2, uHueSpread 1.5->3. Not extended: uInkGuard (his carve-out), uMetal (a law selector), the 0..1 mix weights, and uDarken/uDepth (already fully spent at 1).

uSheen needed a shader change to be honest: the additive budget's ceiling was reached at uSheen 2.0, so the top third of the slider bought almost nothing - very likely why he parked two canons at the ceiling. It now resumes climbing above 3, leaving every stored value <= 3 bit-for-bit identical.

The 19 duplicated lines across the two surfaces are one CORE_SLIDERS table; uGrain is dimmed and labelled "metal law only" where it does nothing.

NEW: "Apply composite -> family" copies only the on-card dials to siblings, never pattern shape.

PROVEN, NOT ASSUMED (AE vs a base build compiled from the previous commit)

• blank canon room: AE 0 across all 44 implemented recipes x 4 tones x 3 tilts
• mirror: AE 0 on the card path too - pixel-frozen
• text at flash peak: +0.1 to -1.1 points of glyph/paper contrast on every reverse whose foil covers the text box; the library's two largest costs (rainbow-glitter, water-web) toned down per-recipe and re-checked by eye
• every implemented recipe eyeballed on 1-2 real assigned cards at rest and two tilts; shots in ~/.pokedex-dev-hub-legacy/foil-shots/r6/

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

be4e428 — 2026-08-07 — Mask provenance: sidecar v3, the AI learning loop, and the collapse safeguard

Chey (2026-08-07): mark AI-generated masks, let an AI learn from his hand-done ones, then let him correct the AI's mask so it "can observe the diff and continue to improve without me having to hand-paint all the masks."

derivation_method existed in v2 but the API HARDCODED it to 'hand' — harmless while only a Pencil touched the corpus, actively wrong once mask-refine's Flatten started writing machine geometry through the same save path.

Taxonomy (5 values, 4 cases that must never blur): layout-flatten machine geometry, unpainted weight 0 hand human, from a layout prior weight 1 hand-refined human, on an existing mask weight 1 ai generator, UNREVIEWED weight 0 ai-corrected AI proposed, human edited weight 0.6

The label is never taken from the client. The client reports only what the canvas was SEEDED with; writeMaskRecord derives the method by diffing saved pixels against what that seed rasterizes to. A machine label needs a full GeneratorIdentity, which HTTP callers cannot supply. One write path serves the route and generators both.

• Anti-feedback-collapse enforced in selectExemplars(), not by convention: unreviewed ai is weight 0 at any corpus size. Codify now routes through it.
• Corrections keep the parent PNG + change map + metrics (incl. a 4x4 grid of where corrections land) — the supervised pairs a generator lane consumes via the self-describing tuples manifest.
• Seam rule, measured not guessed: canvas roundRect vs the server SDF disagree on 389/330260 px, ALL in the 1px AA band. Without tolerating it every unpainted Flatten would stamp hand.
• Guard: if the file at the target path is an unreviewed ai mask it IS the parent, whatever the client claims — AI ancestry cannot be laundered. That guard caught a real bug during this build (ActionBtn passed React's MouseEvent into saveMask's optional arg, blanking derivation); derivation is now required on the PUT so it fails loudly instead.
• v1/v2 load forever (normalize-on-read); Chey's two masks migrated additively and still read HAND-PAINTED, agreement 0.641, pixels untouched.
• Trial batch wotc-window-trial-1: 6 WOTC Base holo rares from window-artgate@1, all ai/unreviewed, reversible. Leave-one-out on the human corpus beat the rect baseline (0.588 -> 0.688) — real, but n=2 same-set, so explicitly not validation.

Surfaces: provenance badge + detail in the workbench, Mask corpus panel, GET /foil-lab/masks/corpus (+?tuples/?exemplars), artifact routes, and corpus.ts report|exemplars|tuples|migrate. Pure tests added to CI (test:foil). Browser-verified at 390px and desktop. See DECISIONS.md 2026-08-07 and the mask-pipeline SKILL.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

16cf67a — 2026-08-08 — foil mask-straighten: read the INTENT out of Chey's hand mask, and make the undo real

Chey, 2026-08-08: "it's impossible to get the lines really straight so I'm hoping you can get computer vision on the mask and card art in tandem to really see my intent there, and make the mask nice and crisp and straight on the lines I was trying to draw along. Will you try that? And then I can either approve or revert."

Smoothing a wobbly boundary gives a smoothly wobbly boundary. The premise here is that a hand-drawn foil boundary is an ATTEMPT TO TRACE A PRINTED EDGE — frame, art box, species strip, stage tag — and those are dead straight in the scan. So: contour the mask, cut each loop into near-axis runs (PCA orientation, wobble gaps bridged), robust-fit each, hunt the scan's directional gradient with a local Hough over (angle, offset) plus a proximity prior, replace, intersect adjacent lines into real corners, rasterize matching his own AA character.

The refusals are the feature, and each is a recorded param: weak evidence, or an ambiguous band of comparable ridges, may nudge but never relocate; with no artwork edge it straightens only to HIS own fit; short and curved runs pass through untouched. A first cut "closed" the Tropius top-left corner — which turned out to be him deliberately tracing around the BASIC stage tag. The stricter rules keep it.

• apps/api/src/foil/line-snap.ts the geometry, pure, no I/O
• generator.ts line-snap@1, a REFINER (requiresSource)
• generate-masks.ts run --refine, archives, restoring revert
• provenance.ts supersedes + verbatim sha256 undo archive

A generator write onto an existing mask now THROWS without explicit supersede { runId }; with it, every artifact the mask had is archived verbatim under superseded/./ with a self-describing manifest, and revert --run-id verifies the whole archive BEFORE deleting anything live, then restores byte-for-byte. supersedes is deliberately NOT correction: one is a human editing a machine (training signal), the other a machine replacing a human with nobody's agreement. --refine also refuses a source with exemplar weight 0, so a refiner can never eat its own unreviewed output.

Tropius me05-001/37184 (run straighten-tropius-1): 8 runs snapped to printed edges, 1 self-straightened, 13 left exactly as drawn, 8 corners closed, 2 specks dropped. Max correction 3.74px, 1.06% of the face changed, Jaccard 0.9798, confidence 0.557. Honest residual: the card's bottom border was too faint to accept, so that run self-straightened and the report says so.

test:foil 25/25 (9 geometry + 5 supersede/restore, incl. byte-for-byte round trip, a deliberate curve surviving, an ambiguous band refused, and a corrupt archive aborting before deletion). Revert exercised twice end to end, sha256sum clean. Shots: ~/.pokedex-dev-hub-legacy/foil-shots/mask-straighten/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id straighten-tropius-1

24a13b8 — 2026-08-08 — foil R7: his eye overturns two banked mechanisms; "no catalog cards" gets a reason

Chey's 15 canon-lab comments (2026-08-07), all resolved in issues/foil/.

Eleven shader verdicts. Two overturn mechanisms built from earlier corpus — energy-symbols' parity CHECKERBOARD (R3-GLYPH, from his own words) and prism's solid-facet grid (R3-MISC, from a Gemini delta pass). Having seen both rendered he rejected them. A description is a hypothesis about the render, not a spec.

cosmos the 0.085 UV field slide he read as parallax is now uP4 "Surface drift" at 0.012; zero at rest, canon unchanged striped-v-sheen grey lines: the unconditional silver term is gone AND the second-order lobe cut 0.4 -> 0.16 (it drew a whole second, near-achromatic line population on the far side) rainbow-mirror hue axis is card X, blue -> green -> red, on BOTH the sheen and the spotlight; core whitening 0.85 -> 0.32 energy-symbols checkerboard gone: per-symbol response AXIS + PHASE + transition WIDTH over a low-freq cluster field energy-symbols-ii same treatment (it shared the two-bank mechanism) pokeball-hologram flat SDF icon -> shaded 3D sphere (hemisphere normal, tilt-following light, belt, hub, two-tone shell) tinsel the "blinds" were a literal full-card striation floor; demoted to uP2, default 0 (his canon stores 0) tinsel-ii line-work gated to the colour band, silver 0.60 -> 0.10, band width exposed as uP4 prism pinwheel's spoke fan WITHOUT the per-wedge phase that makes a pinwheel spin; density 33 -> 18 radiant hologram step 0.5 cell -> uP4 0.17, drive rescaled so total travel per unit tilt is unchanged rainbow-glitter invisible on card: bigger white-lifted flakes AND the family fixed (flash -> field; no substrate, no contrast)

Canon policy held: shader.ts / canon.ts / CardViewer.tsx / data/foil-canon have a ZERO diff, sheenGlsl() is byte-identical (all four sheen canons frozen), and every new behaviour went into a NEW uniform old canon files inherit. No stored value migrated. One recommendation left for him, not applied: tinsel-ii's starkness is substantially his own canon (uDepth 1 at ~135 lines).

"Why are there no catalog cards?" — asked four times, three answers. The baked inversion is now v2 and emits a machine-readable diagnosis the lab renders: outranked diagonal-sheen-left — 1,818 SM reverses ARE cited for it, but energy-symbols wins them (the emblem layer vs the sheet rotation; both true). A capped SECONDARY pool is now baked via a new citedFoilPatterns() export and served as via:'cited', so the preview works. Never flip a winner to fill a pool. class-absent pinwheel, pokeball-hologram — the catalog has reverse variants for ex1-ex5 ONLY; the whole late-EX reverse era is missing upstream. The UI now says so instead of "no catalog cards". fixed vertical-sheen-rainbow 0 -> 12 cards. known_residuals recorded twice that the 6 Holo Rare Energies of ex13/ex16 use Sheen but could not be targeted without a card list; the catalog names exactly six per set. Two cardId rows added citing the same Collexy quotes, both residuals closed.

Typecheck (all workspaces) + web build + foil tests green; 14 recipes compile clean; before/after shots on blank + assigned card at 390px and desktop in ~/.pokedex-dev-hub-legacy/foil-shots/r7/. SKILL R7 field notes + DECISIONS appended.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

77a0b60 — 2026-08-08 — foil edge-trace: the artwork holds the geometry, his mask says which of it to follow

Chey, after reviewing line-snap: "Really just 'straighten' isn't quite enough, really just detecting edges and actually tracing accurately around them is the real move." So the premise moves. line-snap treats his strokes as the geometry and nudges near-axis runs onto printed lines. edge-trace treats his mask as a STATEMENT OF INTENT and derives the boundary from the card's own edges.

apps/api/src/foil/edge-trace.ts (pure, no I/O): Di Zenzo colour structure tensor (Scharr per channel) -> NMS -> hysteresis; colour because a green field against a silver border reads 0.13 luminance/px and 44.3 tensor/px at the same edge. Then livewire: a corridor around his boundary (a hard bound, not a suggestion), anchors at fixed spacing PLUS every hard turn in his own line, Dijkstra on ridge + linked + proximity^2 + direction, sub-pixel refinement onto the |dI/dn| peak, and only THEN a straight fit where the traced path really is straight.

His own correction said the same thing before he did: 114 of his 627 correction px sit in one 40x50 box around the species strip's swept tail, which line-snap had squared off by intersecting two lines across it.

Two traps, both now locked by tests:

• the MAD-trim trap: a robust fit discards a small feature as an outlier, reports "straight to 0.2px RMS", and crisping then flattens what it ignored. straightMaxDevPx requires every point in a run to be close.
• the half-pixel frame: mask space centres pixels at x+0.5, a gradient array at integer x. All artwork lookups convert in one place.

Measured, not asserted. generate-masks.ts adherence scores any set of masks by distance to the NEAREST printed edge maximum, with line-snap's own luminance Sobel as the default probe so the incumbent gets the home field: his hand mask 50.7% within 1px / 0.847px mean line-snap@1 74.8% / 0.545px his correction 75.4% / 0.548px edge-trace@1 77.1% / 0.199px (colour probe: 98.4% / 0.082px)

Applied to Tropius me05-001/37184 as run edgetrace-tropius-1, ai/unreviewed, his ai-corrected mask preserved as parent + verbatim supersede archive. Revert proven byte-identical twice before shipping.

test:foil 35/35 (10 new). Typecheck all workspaces, api+mcp+web builds green. Visuals in ~/.pokedex-dev-hub-legacy/foil-shots/mask-edgetrace/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id edgetrace-tropius-1

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

dd8a02d — 2026-08-08 — Mask editor: pan + pinch-zoom that keeps the strokes honest, and no more "selected" card

Chey, before a batch of hand masks: "something that would be nice is to be able to pan and pinch to zoom while editing masks. In addition, sometimes the card image gets 'selected' when I'm trying to draw which is annoying."

Zoom is a camera view offset, not a CSS scale. camera.setViewOffset(Wz, Hz, x, y, W, H) re-rasterizes the card at the zoomed size, so 4x zoom is 4x real detail — the point of zooming is tracing a printed edge, and a magnified 390px framebuffer would not help. The overlays get the same transform on one wrapper div, written imperatively (no re-render per frame). Because that is a pure crop of a linearly-scaled render, u lands at u*z - offset in both places, so getBoundingClientRect already reports the truth and the pointer->texel math is unchanged. Verified numerically: a 3.375x corner stroke lands at texels [84,180,108,196] vs a predicted [85,180,109,196]; touch at 3x lands within one texel at 390 and 744.

Brush is constant in SCREEN px (lineWidth = brushSize / zoom): zooming in must buy finer control, not a fatter-looking stroke. Two fingers pinch/pan; one finger pans unless it is a drawing finger or on a window handle; a pen owns the surface (palm rejection); a second finger mid-stroke aborts AND rolls the stroke back. Desktop gets wheel/trackpad zoom at the cursor, middle- or Space-drag pan, and +/-/0. HUD shows the zoom and fits back in one tap. Pan is clamped so the card cannot be lost. Zoom is live only while editing, resets on exit, and never reaches the artifact — a mask saved at 3.375x round-trips as 490x674 with an identical bbox. WindowEditor divides drag deltas by zoom and counter-scales its handles, so they stay 44px on screen (44.0 -> 44.0 measured at 2.25x).

Selection fix: user-select/touch-callout/user-drag none, draggable=false, and preventDefault on dragstart/selectstart and the paint pointerdown across both canvases and the viewer column; touch-action stays none so the two fixes do not fight. iOS also needs the gesturestart/change/end veto — touch-action alone does not stop Safari page zoom.

Cost the most time: window listeners registered in the controller factory were killed by StrictMode's simulated unmount, so the first pinch started and never ended and every later gesture and stroke was dead. They are now a symmetric attach()/detach() pair driven by an effect. See DECISIONS.md.

44/44 browser checks at 1280 / 744 / 390 (touch via CDP), his committed masks still load, scratch card artifacts removed. Shots in ~/.pokedex-dev-hub-legacy/foil-shots/mask-ux/.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

efc73fe — 2026-08-08 — foil region-learn: the REGIONS come from his masks, the geometry from the artwork

The last lane measured why rule-seeded edge-trace failed on every me05 reverse: 99.7% of the gap between the layout rule and Chey's intent is REGION decisions, and a tracer can only move a boundary, never add or remove a region. This is the missing half.

region-learn@1 partitions a card face into five structural classes computed from its OWN printing (border / furniture / frameBody / windowBackground / windowSubject), reads off his masks WHICH of them carry foil, applies that to a new card, and only then hands the boundary to edge-trace. Nothing about Pokemon is hard-coded — the same code reads OPPOSITE policies from the two classes in the corpus, and a test asserts the votes, not just the verdict.

The policy, in his pixels: modern-sv/sheet foil = the COLOURED FRAME BODY, nothing else. Species strip (tails included), stage tag, evolution medallion + its sprite, "evolves from" bar, copyright footer: all excluded. And THE SILVER BORDER RING IS NOT FOIL — 0.2/0.2/0.4% across his three masks. That was flagged as "his call, not a measurement"; it is now a measurement, and unanimous. wotc/window foil = the illustration's own background, minus the subject silhouette, minus the stage box where it overlaps. 4/4.

Bar written down before any number existed: mean IoU >= 0.90, no held-out card below 0.85. Leave-one-out: modern-sv/sheet 0.7566 -> 0.9691 (worst 0.9519) PASS -> batch of 8 shipped wotc/window 0.6146 -> 0.8971 (worst 0.8599) FAIL by 0.0029 -> no batch

The two residuals differ in kind, and that is the finding: modern-sv's is a 1-2px registration hairline; wotc's is 8,343px of Machamp's body read as background because his blue-grey sits inside the colour distance of the teal holo field. Every frame-level decision on WOTC is right. The subject silhouette is not.

Also found and NOT changed (era-wide, his call): wotc.artWindow is 11px too high at the top and 6px too wide at the right — his four masks and each card's own printed edge agree, and base1-8 is the card the era was "measured on". Same pattern as me04-007 for modern-sv.

Also: the bevel side. An illustration box has TWO parallel printed edges; which one the foil stops at is which side the foil is on, not which is strongest. Picking the strongest put base1-6's window top 8px wrong and scored it 0.0001; picking the foil side took it to 0.8789.

Batch regionlearn-me05-1 (8 cards, 5 frame colours, 4 with medallions, all ai / unreviewed) applied and proven reversible byte-for-byte — including a revert bug fixed here: it left the empty card directory behind, so a reverted card still looked masked to anything that walks data/foil-masks.

Recommend reverting wotc-window-trial-1: those five base1 masks were made by window-artgate@1, which scores 0.7694 mean / 0.7330 worst on this same test.

test:foil 46/46 (11 new, including the anti-collapse rule locked three ways).

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

879a2c0 — 2026-08-08 — foil vector-template@1: the layout is the artifact, and it is analytic geometry

Chey: "what I draw is intent... generated masks should feel like they're derived from clean vectors" — and "we don't need 3,454 vector masks, all of these share the same 2 layouts really."

Mined his 8 correction diffs first. They say the regions were nearly right (0.12% to 2.23% changed) and that the errors were structural, not stylistic:

• the coloured sliver pinched off between the evolution medallion and the border ring IS foil (he added it on all 4 Stage-1 cards). This was the open question flagged in codified/modern-sv.md as "Chey has never ruled on that sliver". It is ruled on now.
• achromatic marks printed ON the coloured field carry foil: the colourless energy symbols, the retreat symbols, the regulation-mark box, the name text. region-learn@1 carved holes around all of them; he filled every one.
• the left/right window edge sat on the bevel's INNER line; he moved it out.
• and, decisively for this lane: where the print is straight, HIS line is straighter than region-learn@1's traced one. The machine was the thing that wobbled, so a smoother tracer was never the fix.

Geometry, measured over his 11 masks: 50.05% of the face is foil on ALL of them, 47.18% bare on all, 2.77% contested — and the Basic-vs-Stage1 difference is a SINGLE 70x85 blob. So it is not 2 layouts, it is one layout plus one optional element, which the fitter discovers from the corpus itself (bimodal split, 80.5pp separation) rather than being told a medallion exists.

vector-template@1 fits lines + circular arcs + exact corners to the consensus of his masks and rasterises that. 58 primitives, 13.2 KB of committed JSON, applied at render time — instead of 3,454 rasters.

Leave-one-out vs the bar committed in a14a163 BEFORE any number existed (mean >= 0.94, none below 0.90): mean 0.9904, worst 0.9838. PASS. region-learn@1 on the same 11: 0.9757 / 0.9519. Boundary p95 1px on 10 of 11 (was 2-72px).

Vector-ness at 0.35px tolerance on SUB-PIXEL contours (measuring on traceLoops is a trap — crack following returns a rectilinear staircase and rates a hand blob as perfectly axis-aligned): primitives/kpx region-learn@1 33.7 HIS 21.3 vector-template@1 10.7 This is the finding: his hand masks are ALREADY more vector-like than the generator that was supposed to be learning from them.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d0bb143 — 2026-08-08 — foil vector-template: make the emitted path actually analytic (two real bugs)

Writing the round-trip test found two things the IoU table was hiding.

1. JOINS WERE RAW CONTOUR POINTS. vectorizeLoop fitted lines, axis-snapped them, and then took each run's endpoint straight off the sampled contour, only substituting an intersection for line/line corners. So the emitted "vertical" left edge of a rounded rect ran from x=25.00 to x=25.66 — a polyline through contour samples with arcs bridging them, wearing a vector's clothes. The straightness this lane exists to deliver was measured but not emitted. Joins are now the actual intersection of the two primitives that meet there (line/line, line/arc, arc/arc all solved), with joinMaxMovePx rejecting a join that lands far from where the contour really turned.

2. ARC SWEEP WAS INVERTED. cross < 0 ? 1 : 0 — with y down, a positive cross product is a clockwise turn on screen and sweep=1 is clockwise, so every fitted arc bulged the wrong way. Round-trip IoU on a rounded rect: 0.9482 before, 0.9994 after. It only cost ~0.5% on the real template because that template is 56 lines and 2 arcs, which is exactly why a synthetic test found it and the corpus numbers did not.

Template refitted. Nothing moved much, because the bugs were small in this geometry — but the artifact now says what it means: in-sample mean IoU 0.9908 -> 0.9911 vector-ness prim/kpx region-learn@1 33.72 | HIS 21.35 | mine 10.34 long-primitive fraction 0.9047 | 0.9679 | 0.9861 axis-aligned fraction 0.7875 | 0.8554 | 0.9321 correction margin: closer to what he drew than to what he rejected on all 10.

test:foil 56/56.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

d10eca8 — 2026-08-08 — foil: serve the template as a mask, and write down what his corrections ruled

API: GET /foil-lab/masks/:cardId/:variantId?era=&scope=&evolves=1 now falls back to the vector template when no hand mask resolves. A hand mask ALWAYS answers first, so this cannot shadow Chey's work — it only fills what used to 404 — and it is labelled X-Foil-Mask-Review-Status: unreviewed with the generator identity, run id and exemplar count.

The cache is TWO PNGs, not 3,454. A template's rasterisation depends only on (era, scope, hasOptionalElement, w, h) and not on the card, so every Basic reverse gets byte-identical pixels. ~17 KB each under IMAGE_CACHE_ROOT/foil-templates/, mirroring the image-cache contract: disk-only, outside git, safe to delete, rebuilds in milliseconds. Committing per-card masks would have been ~100-350 MB against the template's 13.3 KB.

evolves is an INPUT, not something the route detects, for two structural reasons: these routes have no DB and the server never shells out, so the process can neither read card.stage nor decode card art. The caller already knows. This turned out to be the right call for a second reason — the artwork probe agrees with the catalog on 89.3% of Pokemon but misfires on pale frames (it called sv09-120 Dunsparce, a Basic, "medallion present" at 0.0% chroma).

codified/modern-sv.md pass 2 (n=11) records the three rulings his corrections make, with numbers:

• THE SLIVER IS FOIL. Pass 1 flagged it as "Chey has never ruled on that sliver". region-learn@1 scored 0.0% there on all four Stage-1 cards; he added it on all four. "Largest coloured component" was the wrong rule; the right one is every coloured region that reaches the border ring. The medallion's own sprite stays out because it is fully enclosed — topology, not size.
• Achromatic ink printed ON the coloured field carries foil: colourless energy symbols, retreat symbols, the regulation box, the name text. Furniture is distinguished topologically, not by chroma.
• The window edge sits on the bevel's OUTER line.

And what is NOT covered, said out loud: Trainer (431 variants) and Energy (104) are a different layout with zero exemplars — visibly wrong, probe agrees with the catalog on 0/49 of them. Pale frames cannot self-verify (45.5% of Lightning and 41.0% of Colorless flagged vs 0.0% of Fire/Psychic/Darkness) — a contrast effect, calibrated against his own hand masks scoring 0.720 on the same probe.

Revert proven, not asserted: a 2-card run reverts to a byte-identical tree (git sees nothing), and a supersede round-trip restores the earlier mask with matching sha256 on both the PNG and the sidecar.

DECISIONS also records an operational gotcha I caused: pkill -f on "apps/api/dist/index.js" matches the deployed pm2 pokedex-api too. pm2 healed it in seconds (prod 200, unstable restarts 0), but kill dev instances by PID.

typecheck clean across all workspaces; test:foil 56/56; api + web builds green.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

---

foil/tooling-research

Tip 85af395c28d2c4f3b9bf43d0fc6293c9951272d4 · 63 commits · 2026-08-01 → 2026-08-09 · 10,650 words · 35 reproduced below.

c627662 — 2026-08-01 — foil/main: workbench v1 — three.js viewer, 5-pattern starter library, era layout masks

Quarantined foil workbench at /pokedex/foil-lab (URL-only, linked from nowhere, chrome-free): real owned-card scans from the image cache, tilt via pointer/gyro(iOS permission)/manual with reduced-motion respected, uniform sliders, pattern + mask-scope overrides, mask overlay toggle, copy-recipe JSON. Pattern library v1 tuned against owned eras (Base/WOTC, SV, Mega Evolution): Starlight, Cosmos, SV default holo, SV reverse sheet, Cracked Ice — uniform contract (core + uP0..3) designed for the full 15-20 set. Era layout spec as data (era-layouts.json, rects measured on real scans); resolver v1 (series, rarity, variant kind) -> pattern + mask scope. SKILL.md x2 (foil-effects, mask-pipeline stub). main.tsx gains a dev-only double-createRoot guard (Vite HMR entry-invalidation footgun, see DECISIONS).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

0573a82 — 2026-08-01 — foil/main v2: Pencil mask editing, comment queue, iPad two-column, branch api

Apple-Pencil hand-mask drawing on the workbench (MaskEditor: card-aligned canvas overlay, pen+mouse default w/ finger toggle, pressure brush, eraser, 12-step undo, touch-action none, starts from layout prior); masks persist as committed ground-truth artifacts data/foil-masks//.png+json (derivation_method hand) via the env-gated branch api (POKEDEX_FOIL_LAB=1, port 3712, PGPOOL_MAX=1 — inert in prod), auto-load and beat the layout tier (uMaskTex contract; CanvasTexture flipY=false, exactly one V flip). Comment button captures card/variant/pattern/scope/mask state + all slider values to issues/foil// (report.md + context.json), out of the fix-issues sweep. Era-grouped owned-card picker (WOTC / SWSH / SV+ME headings); two-column layout (viewer | controls) from 700px — verified at 390px, 744x1133, 1133x744, 1440x900 with simulated pen strokes. SKILL.md x2 updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

2b038d6 — 2026-08-02 — foil: sidecar v2 (prior+diff), artwork-keyed masks, comment linkage, codify ritual

• Every hand-mask save records its starting prior (era/scope/resolver rect + RESOLVER_VERSION) and the api renders .prior.png + .diff.png (green = human added, red = removed) + Jaccard agreement, server-side via a pure-JS PNG codec (no native addons). Saves without a parsable prior 400.
• Machamp mask backfilled (deterministic prior wotc/window/v1): agreement 0.6409 — his correction is the subject silhouette, the rule's first lesson.
• Artwork keying: variants of one cardId provably share one scan, so mask GETs alias to a sibling variant's mask when prior.scope matches (X-Foil-Mask- Alias-Of). Cross-card reprints unprovable from the catalog -> per-card only.
• Comments auto-link the exact saved mask state (maskFile/maskSavedAt/ maskAliasOf/maskHasPriorDiff).
• mask-pipeline SKILL: sidecar v2 storage, identity rule + limits, and the codify ritual; first worked example data/foil-masks/codified/wotc.md (n=1: WOTC window = art-window rect minus subject silhouette; noted in era-layouts.json wotc.notes).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

1ac3bf4 — 2026-08-02 — foil: any-card picker — full-catalog browse/search, Owned-only as a toggle

Chey: pull up ANY card in the catalog and tilt it, owned or not. Picker now browses the whole catalog by era -> set (paged strip, promo sets run 300+), adds full-catalog name search, and keeps Owned-only as a filter toggle (default on = prior behavior). Zero new endpoints: existing /series, /series/:slug, /sets/:setId (paged), /search, /cards/:cardId serve it all; unowned cards list their catalog variants (quantity 0) and resolve the same base-guess pattern path. Selection is never clobbered by filters; unmapped series get an honest "Other eras" bucket. Verified on unowned neo2-13 Umbreon (Auto - Starlight WOTC) at 390px + both iPad viewports.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

bc971e1 — 2026-08-02 — foil: taxonomy integration — all 39 pattern types in the app, cited resolver

Mislabel corrections from the video-reference research, applied to apps/ (research docs untouched):

• sv-holo renamed vertical-sheen (it always rendered the HGSS-XY vertical bands, not SV's). PATTERN_ALIASES keeps old slugs resolving forever — sidecars, comment context.json, Copy-recipe JSON never orphan.
• Sheen family = ONE generator at four rotations + stripe option: five honest implemented slugs (vertical / horizontal=TRUE SV default "Mirage" / diagonal right "/" / diagonal left "" / striped "Line").
• Cosmos label drops "/ Galaxy" (Galaxy is a starlight synonym; synonym moved to the starlight entry). reverse-sheet description notes it ≈ pokeball-masterball stamped sheet.
• Vintage narrowed: starlight = Base/Jungle/Fossil only; Base Set 2 → Call of Legends holos = cosmos.

Pattern dropdown carries the FULL 39-slug vocabulary, grouped: implemented recipes first, then every gap type rendering via its nearest recipe and labeled "approx via ..." (amber note in the UI). Unimplemented != hidden.

Resolver v2: base guess reads the cited usage table (usage-index.json, derived from research/foil-pattern-usage.json by tools/foil/build-usage-index.mjs; set-name match -> era/series token match -> era-default heuristics). The winning row's confidence + citation hosts surface in the workbench ("Guess: set-level citation, high confidence (bulbapedia..., collexy...)"). Reverse rows citing 'mirror' are deprioritized per the research known-gap (ink-design evidence, not foil-pattern evidence). RESOLVER_VERSION -> 2 (scope semantics unchanged from v1).

Verified: typecheck + build green; Playwright at 390/744x1133/1133x744 (~/.pokedex-dev-hub-legacy/foil-shots/w2/) — Machamp auto=starlight set-level-high with his hand mask still resolving, SV holo auto=horizontal-sheen, SV reverse auto=energy-symbols-ii, fireworks approx labeling.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01MY2totBB6u2Z3sjFdPy8bn

6a68c14 — 2026-08-02 — foil: fix diffuse darkening — scan sampled sRGB-decoded but never re-encoded (issue ls9u0y)

The face texture was uploaded SRGB8_ALPHA8 (hardware sRGB->linear decode at sample time) while the foil ShaderMaterial writes gl_FragColor raw — three.js only appends the linear->sRGB colorspace_fragment chunk to built-in materials — so every scan rendered in linear values displayed as sRGB (measured flat 184->123 / 199->147, exactly the sRGB->linear curve). Sample the scan undecoded (NoColorSpace): the compositing model is authored in display space and pattern=none at rest is now pixel-comparable to the flat view (mean abs luma diff 44.55 -> 2.16, mapping identity +/-2). Also zero pattern-none's default specular so the baseline is truly the plain scan.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

21ea5e1 — 2026-08-02 — foil: workbench split — canon pattern lab (A) | card adjustment surface (B)

Chey's comment issues/foil/2026-08-02_12-59-52-368_4aq756 (resolved here):

• /pokedex/foil-lab/canon (CanonLab.tsx): bare holofoil pattern on a blank card (no ink; black/dark/silver/white flat-tone base via the normal texture path) beside the REAL reference tilt clip + 8 keyframes streamed from research/foil-video-reference/ by new env-gated dev routes (byte-range via res.sendFile for iOS Safari). Full 39-slug picker, sliders, tilt, and Save canon -> data/foil-canon/.json (full uniform snapshot that replaces recipe code defaults as the baseline).
• /pokedex/foil-lab (FoilLab.tsx reframed): per-card masks/comments unchanged, plus per-card uniform overrides saved as a SPARSE diff vs the canon baseline -> data/foil-overrides//.json (untouched uniforms keep tracking canon). Off-canon sliders get a dot; saving with sliders matching canon deletes the override.
• Layering model in foil/canon.ts (code defaults < canon < override < live); shared atoms + surface tabs in foil/ui.tsx; routes wired in main.tsx; AppShell chrome-free check widened to /foil-lab/*; both data dirs re-included in .gitignore; ports 5184/3713 recorded in ORCHESTRATION.md.

Verified in-browser at 390x844 / 744x1133 / 1133x744 (clip playing beside the bare cosmos render; Machamp base1-8 hand mask still resolves) — shots in ~/.pokedex-dev-hub-legacy/foil-shots/canon-lab/. Typecheck + build green (web, api). patterns.ts untouched (R0 re-tunes concurrently).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ed3a26c — 2026-08-02 — foil R0: re-tune implemented recipes per Gemini verification notes

cosmos: full activation-model rewrite — dark field, smaller/denser orbs mostly near-invisible, cluster pops via low-freq noise over cell ids, pinprick 4-point twinkles (was: dense wall of large saturated orbs). starlight (+ starlight-ii via shared GLSL): parallax architecture untouched; tighter visibility lobe (pow 5->9, floor .18->.08), sharper glyph cores + narrower flares, smaller blobs, saturated star color, galaxy wash halved; starlight-ii art gate lowered for the mid-orange Evolutions field. cracked-ice: authored intra-shard grain removed per Chey's accuracy ruling — facets flash as solid smooth planes; whitening at peak reduced, uSat up. sheen family: generator grows sharp/beam/barcode options. vertical-sheen -> new barcode variant (thin spectral lines of varying width, CD-groove slide) with tamed beam + art gate. diagonals: sharp 3.0, beam 0.55, specular 0.35 (center blow-out), diagonal-sheen-right uP0 2->7 (same-sheet consistency). striped-vertical-sheen: stripes finer + more blended (90->130, floor .40). horizontal-sheen: untouched (20/20).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2c83027 — 2026-08-02 — foil R0: rounds 2-3 re-tunes + verification report, field notes, decisions

starlight: longer thin-armed 4-point glyphs + 8-point-burst subset, tighter pop lobe (pow 11, floor .04), full-saturation star color, parallax default 2.4, gain 3.6 — best of 3 rounds 11/20 (still-frame judging cannot see the parallax; recorded as honest residual with a video-judge recommendation). starlight-ii: 20/20 (uP2 pinned so base-default drift can't invalidate it). cracked-ice: facet gain 1.1 -> 0.55 (art visible through flashes), density 10, less whitening — 15/20 yay; shard-geometry uniformity queued for R1. Verification report gains the R0 section (before->after per pattern, the diag-left judge-noise event, diffuse-fix supersession note); foil-effects SKILL.md gains the R0 field notes; DECISIONS entry for root cause + wave.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ddffdcc — 2026-08-02 — foil: cited per-card pattern assignments across the whole catalog (resolver v3)

Chey's mlmwmp comment: the library was funneling to era defaults ("cosmos"). Full-catalog enumeration (20,964 cards / 37,627 variants, via the branch read api) + an 18-lane Ringer research swarm (foil-card-assignments, GLM-5.2, executed selector-integrity checks) produced:

• research/foil-card-assignments.json — 111 set/rarity/kind/card-level rows, 19 catalog foil-facet mappings (holo-foil-cosmos, reverse-foil-masterball, normal-foil-galaxy...), 55 honest residuals; per-lane provenance + method.
• tools/foil/build-assignments-index.mjs -> apps/web/src/foil/assignments-index.json (trimmed for the bundle, duplicate-facet guard).
• resolver v3 (RESOLVER_VERSION=3): assignment tier above the v2 usage table; specificity cardIds > variantKinds > facet > rarities > set > v2 > heuristic; normal-kind variants with a declared foil facet now render window foil (previously scope none); FoilLab passes cardId.
• era-layouts: wotc seriesSlugs 'ecard' -> 'e-card', + 'legendary-collection' (both had fallen through to modern-sv rects and heuristics).

Coverage (foil-bearing variants): 79.5% set-cited / 20.4% era-token / 0.2% heuristic -> 93.4% set / 4.3% facet-or-card / 2.4% era-token / 0 heuristic. 13 in-browser spot-checks in ~/.pokedex-dev-hub-legacy/foil-shots/assignments/. Issue mlmwmp resolved. Weak spots + curation ledger in DECISIONS.md.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

e2bf3b8 — 2026-08-02 — foil R1: dedicated recipes for the twelve owned-era gap patterns (10/12 verified)

Chey's comment dml369 ("develop out the rest of the missing holofoil patterns"): dedicated GLSL for fireworks, ace-spec, energy-symbols-ii, rainbow-glitter-sheen, ex-starfoil, prismatic-pokeball, tinsel-ii, cosmos-iii-smooth, pokeball-masterball, radiant, rainbow-glitter, confetti — implemented flags flipped, approx labels dropped, slugs stable. Gemini-verified 10/12 match over 3 rounds (verdict table + honest residuals in research/foil-verification.md R1 section). Honest nays: tinsel-ii (procedural static plateaus at judge score 2) and prismatic-pokeball (screen-only blend cannot darken a white reverse body — structural, Chey's call).

R0 residuals cleared in the same pass: cracked-ice anisotropic shattered-glass metric (re-judged yay 14/20); starlight milky wash + per-layer visibility curves with starlight-ii re-banked clean 20/20 through both shared-GLSL changes; starlight parallax judged from a new 16-frame fine sweep and still scored flat — verified real by eye, recorded as judge-blindness.

Field notes appended to foil-effects SKILL.md; dated DECISIONS entry; issue dml369 resolved.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

a54f040 — 2026-08-02 — foil R2: uDarken blend-model term — prismatic-pokeball 17/20 yay, tinsel-ii 16/20 yay

One opt-in substrate-darkening term in the shared composite (main()): body = scan * (1 - uDarken * mask * gate) before the foil screen-blends. Default 0 is bit-identical to the old render; absent canon/override keys seed 0. Full 21-pattern regression sweep on provably identical renders: no true regressions (and a measurement of Gemini single-roll noise, ±3-6; diag-right + pokeball-masterball double-nays recorded as unstable verdicts on geometry proof). prismatic-pokeball rebuilt on the term (dark-mirror base, facet-quantized flash lobe, ball watermark as ink-overprint suppression): structural nay -> 17/20 yay. tinsel-ii one-line opt-in (uDarken 0.4): static plateau -> 16/20 yay. 20 of 21 implemented types now match; sole nay starlight (judge parallax blindness).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

cef7ed5 — 2026-08-02 — foil R2 stage 3: thirteen dedicated recipes (gap patterns + RC dots) + subset-commons resolver tier

• patterns.ts: real recipes for mirror, rainbow-mirror, vertical-sheen-rainbow (dark-mirror family on uDarken), pokeball-hologram, prism, water-web, energy-symbols, pinwheel, ex-emerald, tinsel, cosmos-ii-pixel, crosshatch, radiant-collection-dots — implemented: true, approx labels dropped, slugs stable. Unused plain SHEEN_V generator removed (last fallback users gone).
• resolver v4: card-level 'normal'-class assignment rows consulted before the scope-none early return — subset cards the catalog declares plain but which physically carry foil (RC commons dot overprint) resolve their subset's pattern with scope full.
• foil-card-assignments.json: two card-level rows (bw11-RC1..20, g1 25 non-ultra RC ids) with video citation (non-holo RC Teddiursa keeps the dot overprint); builder enforces cardIds on 'normal' rows; index regenerated.

Verified in-browser: g1-RC1 + bw11-RC1 render the RC dot treatment via Auto (guess card|medium). Gemini verdicts in research/foil-verification.md (R2 recipe-wave section, next commit).

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ef8f3ef — 2026-08-02 — foil R2b: four §40–43 vocabulary recipes (all match) + resolver v5 scope overrides

The last lane of the wave — dedicated recipes for the vocab-extension types: gold-secret (warm-locked goldRamp field + chromatic glitter pops + SWSH burst rays — 19/20 yay), vstar-pearl (near-white pearl on uDarken 0.3, warm-led diagonal wash, frame-line streaks — 13/20 yay), shiny-vault (silvery field + band-keyed sparkle-glyph amplifiers, uDarken 0.15 — 11 nay → 16/20 yay r2), detective-pikachu (diagonal sheen × photo-luminance via the documented uFace contract exception — 20/20 yay). 38/43 taxonomy types real, 34 hold match.

Resolver v5: assignment rows carry an optional per-row scope override (window/ full/sheet), honored by the set-row tier and the cls-'normal' tier — baby shinies render window despite the 'Shiny rare' full-foil rarity, VSTARs full despite plain holo kinds, det1 window despite 'ultra rare'.

Assignment research: +2 facet rows (gold, gold-jumbo → gold-secret; retires both facet residuals, 1-card metal collateral flagged) and +10 cited rows — sma 51 baby/35 GX/8 gold splits, swsh4.5sv 104 baby + 18 full-face (incl. the black shiny Eternatus pair, name-verified not gold), sv04.5 rarity split, VSTAR via the dedicated 'Holo Rare VSTAR' rarity, det1 ultra-rare + cls-normal rows (all 18 cards window). Residuals annotated with resolutions.

build-usage-index.mjs: tolerate era-wide rows with no scope.sets (the vocab gold rows crashed the build — the 122-row usage index never compiled until now). Verdicts/frames under ~/.pokedex-dev-hub-legacy/foil-verify/ (verify-

-r2b*); shots in foil-shots/r2b/. Docs: foil-verification.md R2b section, SKILL R2b field notes, DECISIONS entry.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

bfdeee4 — 2026-08-03 — foil R3: sheen-family streak-field rework — all 8 of Chey's canon-lab comments resolved, 7/7 Gemini 20/20

sheenGlsl is now a streak field (Chey's critique = the spec): sparse irregular streaks, per-streak lean following card tilt (crisscross; converging pairs terminate where they meet), stretched-ellipse taper, hue running ALONG each strip as well as across. striped-vertical-sheen rebuilt from a fresh Gemini corpus re-spec run with Chey's description embedded (grouped reveal via wide sweeping windows over a subtle bottom-converging fan; gemini-spec-r3.md); ex-starfoil rebased on the reworked diagonal base + fine CD-line layer; vstar-pearl rebuilt from the horizontal-sheen streak field, pearl kept.

Canon migrations (full-snapshot rule, DECISIONS): striped + horizontal uDarken 0 -> 0.32 (bright-substrate legibility, third confirmation). All other canon values carry over; uniform semantics unchanged. validate_spec.py frame regex single-digit fix (undercounted two-digit citations, failed an honest worker).

Verdicts (Chey's words folded into every judge prompt as acceptance criteria): vertical/right/left/vstar-pearl 20/20 first roll; striped + ex-starfoil 20/20 round 2; horizontal 20/20 round 3. Table + residuals in research/foil-verification.md R3.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

a6bb407 — 2026-08-03 — foil R3-MOTION: axis-split starlight, per-dot cosmos, hologram radiant, the rainbow-glitter-sheen delta

Chey's six motion-model canon-lab comments resolved (his eye = ground truth):

• starlight (+ starlight-ii via shared GLSL): top-to-bottom hue BANDING (soft-quantized bands; his canon 0.62/0.6 lands the reference order blue->green->red on the R->B->G cosine ramp) + AXIS-SPLIT tilt: vertical slides the whole field (~10% card height), horizontal drives the per-star random fade; his 3-layer uP1 parallax rides on top untouched. Judge: starlight 17/20 yay round 2 — the FIRST starlight yay (standing parallax nay broken); starlight-ii 20/20.
• cosmos: whole-field slide + per-dot random response (independent random tilt axes for brightness and hue per orb — no shared sweep axis); at tilt 0 his canon appearance is pixel-identical. Judge: 20/20.
• radiant: discrete hologram steps (half-cell up/down, 60% hold / 40% crossfade — one position fades out as the next fades in), uP1 wired as Hologram travel (canon migrated 0 -> 2.2; was an unused placeholder), grid lines ~45% thicker. Judge nay 13/20 across 3 rounds on a pixel-refuted "slides continuously" claim — recorded as still-frame motion blindness (same class as starlight parallax); adjacent-frame crops prove discrete positions + crossfade. Chey's live tilt is the tiebreak.
• rainbow-glitter-sheen: dedicated Gemini delta-articulation pass first (he couldn't name the difference), pixel-verified, then fixed: narrow laser-saturated striped chevron (sigma 0.010->0.0035, hue traversal 5x->9x, pow-deepened at low gain), fainter +-0.8 repeats, chevron angle 1.3->1.9, denser colored always-visible glitter, uDarken 0->0.4 (4th legibility-physics data point). Judge: 19/20 yay.

Docs: research/foil-verification.md R3-MOTION section, DECISIONS entry, foil-effects SKILL field notes. Frames/verdicts in ~/.pokedex-dev-hub-legacy/foil-verify (jobs -r3m, manifest-r3m{,2,3}, delta-rainbow-glitter-sheen); gate shots in foil-shots/r3-motion.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

44213af — 2026-08-03 — foil mask-refine: window handles → flatten → hand refine

Chey's requested workflow on the Card-adjust surface: draggable corner/edge handles adjust the layout window rect per card (WindowEditor.tsx, 44px touch targets, finger-drag allowed), the adjusted geometry persists as artwork-keyed scope-agnostic sidecars (data/foil-windows//.json, /foil-lab/windows routes), and Flatten bakes it through the STANDARD hand-mask save path (sidecar prior stays the era rule; new optional prior.window records the adjustment) then opens the existing paint tooling.

• rasterizeWindowRect shared by Flatten + loadLayoutRect (identical bakes)
• saveMask prior = layoutMask (rule), never the effective mask
• windows GET aliases across variants of the same scan, newest savedAt
• verified in-browser desktop + 390px: sma-SV1/sma-SV2 (SM baby shiny, borrowed modern-sv rect) and me01-034 (zero-coverage gap); base1-8 + base1-5 hand-mask regression clean; test artifacts deleted (corpus rule)
• ports: vite 5186 / api 3714 recorded in ORCHESTRATION.md

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9299c9f — 2026-08-03 — foil R3-GLYPH: glyph-pattern reworks per Chey's 7 comments + the drop-in glyph slot for his SVGs

Behavior (his notes verbatim = acceptance criteria): reverse-sheet rainbow only on grainy glyphs over a neutral sheet; energy-symbols checkerboard banks that swap with tilt; energy-symbols-ii sporadic random banks (~half faint at any time); ace-spec diamonds grow/shrink with tilt + a touch of blur; prismatic-pokeball ball never darkens (white-mixed pale response + coherent plane flash, suppression deleted); radiant-collection-dots shapes catch a traveling rainbow band (16/20 YAY — standing R2 nay broken).

Glyph slot: research/foil-glyphs//glyph[-N].svg -> lab-gated api routes -> atlas rasterizer (foil/glyphs.ts) -> uGlyphTex contract + glyphTex() helper; ~2.5s auto-pickup on save, procedural fallback when absent, README tells Chey exactly where his four promised files go.

Canon migration: reverse-sheet uP1 0 -> 0.6 (dead placeholder now drives his requested glyph grain). Verdicts + pixel-refutations recorded in research/foil-verification.md (R3-GLYPH); lessons in foil-effects SKILL.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

98e3b47 — 2026-08-03 — foil R3-MISC: the uTint ink-metallic blend term + Chey's twelve remaining canon-lab comments (12/12 yay)

The reverse-holo fix (his chat report): screen-blending achromatic foil light over colored art compresses chroma — "dull and grayish, rather than making the color look metallic." New core uniform uTint (default 0 = bit-exact legacy) multiplies the foil layer + in-mask specular by the luminance-normalized scan chroma SQUARED (the double ink pass of a real mirror reflection); neutral on blank cards, so no canon-lab appearance moved. Reverse-family recipes opt in; verified before/after at 5 tilt angles on Victini ball-reverse, Crystal Energy, Pineco. uP4/uP5 param slots added in the same contract change.

The twelve comments, each mechanism-mapped in its issue frontmatter and the verification doc: cracked-ice triangulated + half-off gate; fireworks grid lattice + center-emanating radial hue; rainbow-mirror rebuilt as mirror + hue-banded spotlight; tinsel-ii silver lines + vertical rainbow band (band lives only on the line-work); prism rebuilt as pinwheel-kin from the corpus-vs-corpus delta pass; cosmos-ii pixel-quantized stamp edges; crosshatch thinned; gold-secret adjustable burst origin (uP4/uP5, center default, per-card via overrides) + holographic grain; dedicated recipes for sequin (glyph-family glints), tcg-classic (flat starlight + rainbow glitter), acid-wash (water-web kin), disco (perfect disc lattice, galaxy ignition) — zero approxVia fallbacks remain.

Canon migrations (appearance-preserving unless he asked otherwise): fireworks uP0 3→4.5 (dead two-octave structure), gold-secret uP4/uP5 0.5 added, sequin uP0-uP3 re-keyed to the new recipe. Verdicts: research/foil-verification.md "R3-MISC" — 12/12 Gemini yay incl. the tcg-classic ink-scope re-roll.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2022159 — 2026-08-03 — foil R4-COMPOSITE: the ink-density invariant — foil adds pop, never washes text or mutes ink

Chey's two me05 comments (7rtnzx mirror reverse "blows out the darks/text", 19mo4l holo "darkens/muddies the colors") ruled a CORE COMPOSITE INVARIANT, engineered in shader.ts main(): ink density estimated from the scan (inkDark = darker-than-local-8-tap-field average, exactly 0 on flat blank bases; inkColor = chroma above the canon-lab-tone floor) now gates every layer — dark ink blocks the flash + shields the specular (text crisp at every tilt), ALL ink is exempt from uDarken substrate attenuation (printed color never muted; the mirror stays dark only where the mirror is visible), colored ink auto-tints its flash (max(uTint, inkColor)) and uInkPop pumps chroma under the flash so colors read MORE saturated and metallic.

New core uniforms uInkGuard (default 1) / uInkPop (default 0.5), both 0 = bit-exact legacy composite; sliders in both labs; canon files untouched (absent keys inherit the new defaults — catalog-wide fix, zero migrations).

Proof: canon-lab blank-card pairs (mirror dark+silver, cosmos, horizontal-sheen) pixel-identical (compare -metric AE 0) under a frame-stepped rAF harness with a same-settings AE-0 control, plus a Gemini sanity pass (all pairs "identical"). On-card before/after pairs across the 8-card sample (both issue cards, Victini, Pineco, Darkrai, Cetitan, Machamp WOTC, Hydreigon SIR; 3 tilts, 390px + desktop) in ~/.pokedex-dev-hub-legacy/foil-shots/r4-composite/. Both issues resolved with mechanism-mapped notes; composite contract documented in the foil-effects SKILL; DECISIONS + research/foil-verification.md updated.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

2fdbce0 — 2026-08-04 — foil R4b: scan-additive composite (rest parity by construction) + canon-lab card preview

Chey's Grubbin (me05-002, Pitch Black) review refuted R4's ink heuristics: the reverse still muddied the card at rest (measured ΔC +32, green→yellow), darkened the header under tilt (−22 L), and blew out text (glyph contrast −26%). New law on real scans (uScanBase 1): the scan is a photograph of the card at rest — foil is additive dynamic light only, tilt-gated to ZERO at rest, luminance-headroom + per-channel clamped (text sacred at every angle, uArtGate opens the clamp where dark scan pixels ARE the foil), chroma- preserving with a pastel-safe tint ramp and an L²-gated chroma pump. Canon lab blank bases run the classic composite textually unchanged: AE-0 proven (CDP frame-stepped harness, control + knob pairs; Gemini pass agreed). uInkGuard 0 still reproduces the pre-R4 composite exactly.

Canon-lab card preview (his request verbatim): blank/on-card/↻-another chips render the live sliders on a random catalog card the resolver assigns the pattern to — server-sampled from the baked inversion (tools/foil/ build-pattern-cards.mts → data/foil-pattern-cards.json, gitignored; GET /foil-lab/pattern-cards/:patternId), honoring hand masks + adjusted windows; empty-pool patterns say "no catalog cards". Also commits Chey's live mirror-canon tweak from the Grubbin session. Full metrics + verdicts: DECISIONS 2026-08-04.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

bb51f4e — 2026-08-04 — foil R4c: onset ease — the sheen leans in instead of switching on

Chey on R4b's rest gate: "it's like it isn't there at all until I've tilted the card a tiny bit and then it's like it just suddenly appears." Measured: the headroom clamp saturated R4b's smoothstep(.02,.28) gate, so ALL added light arrived inside tilt 0.03-0.13 (0 -> 16.1 dL mean) then froze flat to tilt 1.0 — a step smeared over 4% of the tilt range.

Now: one wide ease ramp = smoothstep(0, .45, |tilt|) (full at 0.45 — every R4b tilt-0.5 metric unchanged), flash gated by ramp^1.5 (cubic-slow start), specular by ramp^2.5 (the broad gloss band now trails the flash instead of leading it), plus a REST=0.006 sub-JND living rest sheen blended into both gates (rest +0.43 dL card-mean, max 2/255 — visually still the scan). After-curve: smooth S over 0->0.25, no dead zone, no knee; text clamp untouched (glyph contrast 33.6 -> 33.6 at flash peak); blank-card canon lab bit-level unchanged (fuzz-1% AE 0 vs control floor).

Harness + curves + filmstrips: ~/.pokedex-dev-hub-legacy/foil-shots/r4c/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

7d27447 — 2026-08-04 — foil R4d: motion-first onset — tiny gestures move the sheen; glow arrives late; Chey gets the dials

Chey rejected R4c too ('It still lights up pretty noticeably when I tilt the card a tiny bit. I don't like it.') — measured the gesture→tilt mapping first this time: pointer is 0.0051 tilt/px at 390px and gyro is 0.036 tilt/deg, so a 30px thumb drag (tilt 0.154) was already getting HALF the full glow under R4c. The curve was never the felt variable.

• shader.ts: onset gates now uniform-driven — uOnsetRange (glow full at |tilt|, default 0.5) and uOnsetCurve (flash exponent, default 3.5; specular trails at +1). (0.45, 1.5) reproduces R4c exactly (verified to the third decimal); defaults are strictly softer at every tilt and equal from 0.5 up, so all R4b/R4c peak metrics are unchanged (Grubbin 17.976 ΔL / text 34.02, Machamp 18.035 / text 45.6 — exact).
• FoilLab + CanonLab: 'Onset range' / 'Onset curve' sliders, canon- stored like every other uniform; absent canon keys inherit defaults.
• patterns.ts: CoreDefaults admits the two keys per recipe.

Pointer-path proof (real pointermove, 390px, 30px drag, first 300ms): before mean ΔL climbs to 6.6 and rising; after stays within ±0.08 while per-pixel movement reaches 6.1 mean / 18.9 p95 — moves, doesn't light up. Blank canon lab at fuzz-1% AE 0 vs control. Artifacts in ~/.pokedex-dev-hub-legacy/foil-shots/r4d/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

3f696cf — 2026-08-05 — foil R5: the metallic composite — always on the card, energy redistributed, Chey's compositing dials

Chey's ruling (2026-08-05, verbatim in DECISIONS): no onset range at all — the effect is ALWAYS on the card; dial in how it composites onto the ink; metallic and spectral, not brightening.

• Onset machinery removed (R4c/R4d ramp, REST floor, uOnsetRange/uOnsetCurve sliders on both surfaces). No canon/override file ever stored the keys; CardViewer skips unknown uniforms, so stale keys anywhere are inert.
• R5 metalness law (scan path only): scan = albedo; the pattern field splits around PIVOT 0.40 — above = tinted-specular highlight, below = bounded multiplicative depth (mirror-turned-away darks, sqrt-concave). Energy redistributes; contrast + chroma carry the metallic read.
• Soft-knee headroom: the R4b hard clamp plateaued strong fields into a flat wash (the "brightened" read); the budget now compresses monotonically so pattern structure survives. Per-channel no-clip caps stay hard; glyph ink (inkDark) exempt from highlight AND depth — text contrast preserved (Grubbin 86->97 at peak, Machamp flat).
• New canon-stored dials on both surfaces: Metallic 0.6, Sheen strength 0.55, Sheen tint 0.5 (= the R4 chroma law exactly), Depth 0.55, Texture 1. uMetal 0 / uIntensity 0 / uInkGuard 0 all still yield the plain scan.
• Blank canon lab pixel-unchanged: CDP frame-stepped before/after AE 0 at fuzz 1% (mirror/cosmos/horizontal-sheen/starlight, tilted + rest); one Gemini vision pass agrees (all pairs identical). Shots + per-card rest pairs: ~/.pokedex-dev-hub-legacy/foil-shots/r5/.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

f235040 — 2026-08-07 — foil R5b: metalness is MIRROR ONLY — every other pattern gets its own colour back

Chey: "You applied the metallic treatment to EVERY HOLO PATTERN when I only wanted it for mirror. So now, every single holo pattern looks like mirror and you can barely see the pattern at all, it's like every single one is being applied in a way that it adds no rainbow color or anything to the card."

R5 shipped its metalness dials as GLOBAL_DEFAULTS, so all 43 recipes inherited a law built for one of them. uMetal is now the scan-path LAW SELECTOR and defaults to 0 — nothing can inherit metalness; mirror is the only recipe that opts in, in its own patterns.ts defaults.

• Metalness branch is the R5 code textually unmoved, so mirror's arithmetic is unchanged by construction. Proven: his hand-tuned canon on Grubbin me05-002
  • Tropius me05-001, rest + tilt 0.2/0.35/0.5, 390px and desktop — 16/16 frames AE 0 at fuzz 1%, same floor as the same-build control pair.
• uMetal 0 runs the ADDITIVE law: the pattern's own emission (uSheen gain) screen-combined per channel over a substrate-attenuated body, tinted toward the ink only as far as uSheenTint asks. Three fixes to what R5 broke — the scalar no-clip cap let the artwork's brightest channel throttle the whole flash; the quartic ink budget (shaped to starve mid-dark glyphs) starved ordinary mid-tone art instead; PIVOT ate the field before anything flashed. Cosmos measured 0.005% of art-window pixels changed >8% under R5.
• uDepth decoupled from uDarken: uDarken is the blank-canon substrate, frozen by 30 saved canon files, and 0 on exactly the pale reverse sheets that need one. The substrate follows the pattern's dark half so it buys contrast, not a uniform dim (a flat 0.34 cost mean dL -21.7 and 29% of text contrast).
• Per-recipe families in patterns.ts (PARTICLE / SHEET / WASH / PEARL) plus seven by-eye corrections, each judged on a card the resolver assigns.
• Onset machinery stays deleted. Text guarantee re-verified: glyphs matte and crisp at peak tilt on a reverse whose foil covers the text box.
• Blank-card canon appearance unchanged: 8 patterns AE 0 (6 direct; cracked-ice and tinsel-ii against the pre-R5b code with canon held constant, since Chey re-saved those two mid-wave).
• SKILL.md composite contract rewritten around per-recipe selection.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

42c195a — 2026-08-07 — foil R6: composite defaults re-derived from Chey's canons; the guard learns to tell text from artwork

Chey: "i re-adjusted cracked-ice. take a look at it, and then make the changes based on everything you're observing to make all of this better - i want it so that i have to do as little of work personally on all the rest as possible."

WHAT HIS CANONS TAUGHT

uSheenTint 0 on all three non-mirror canons -> 0 for every non-mirror family (was 0.15/0.5/0.85/0.6). Ink-tinting a pattern's own highlight is exactly "it adds no rainbow color or anything to the card". uSheen he raised all three -> families now 1.2-4.2 (were 1.0-2.6). uDepth NOT universal: 0 on cracked-ice, 1 on tinsel-ii and rgs. It darkens the card WHERE THE PATTERN IS NOT, so its cost scales with 1 - duty. Measured every recipe's duty cycle; his three choices line up with it, and it predicted both dials for the only canon that exercises the defaults. uSat 1 + wide spread -> shipped on the 13 recipes with no canon.

Families are renamed for how light LANDS (flash / line / stamp / field / pearl / metal) and stated explicitly as FoilPattern.family. The stamp tier exists because the first pass got it wrong: energy-symbols-ii lights 5.8% of the face, so a field substrate darkened the other 94% and the card read "someone dimmed this".

THE INK GUARD: SPLIT, DON'T WEAKEN

inkDark is a local-contrast measure, so it fired on every dark mark - including every outline and shading edge in the ILLUSTRATION (measured: 86-89% of strong hits inside the art window are artwork, not text). The additive law spends it as a hard coverage multiply, so guard 1 punched the pattern full of holes. That is why he kept pulling it to ~0.51/0.56 on sheet/wash patterns.

Now split by glyphness into inkGlyph (sacred, zero flash, unchanged) and inkDetail (artwork darks keep the pattern on a tightened budget). inkAny is byte-for-byte the old inkDark and every pre-R6 consumer still reads it.

Success test met: at guard 1 under R6, tinsel-ii and rgs look BETTER than his 0.51/0.56 do today - tinsel-ii's artwork is visible through the streaks instead of greyed out. Both canons migrated to 1, reason recorded in each file's note; every other value in them is his, untouched.

SLIDER RANGES (his second ask)

Four dials sat exactly on their cap: uIntensity 2->4, uSheen 3->6, uSat 1->2, uHueSpread 1.5->3. Not extended: uInkGuard (his carve-out), uMetal (a law selector), the 0..1 mix weights, and uDarken/uDepth (already fully spent at 1).

uSheen needed a shader change to be honest: the additive budget's ceiling was reached at uSheen 2.0, so the top third of the slider bought almost nothing - very likely why he parked two canons at the ceiling. It now resumes climbing above 3, leaving every stored value <= 3 bit-for-bit identical.

The 19 duplicated lines across the two surfaces are one CORE_SLIDERS table; uGrain is dimmed and labelled "metal law only" where it does nothing.

NEW: "Apply composite -> family" copies only the on-card dials to siblings, never pattern shape.

PROVEN, NOT ASSUMED (AE vs a base build compiled from the previous commit)

• blank canon room: AE 0 across all 44 implemented recipes x 4 tones x 3 tilts
• mirror: AE 0 on the card path too - pixel-frozen
• text at flash peak: +0.1 to -1.1 points of glyph/paper contrast on every reverse whose foil covers the text box; the library's two largest costs (rainbow-glitter, water-web) toned down per-recipe and re-checked by eye
• every implemented recipe eyeballed on 1-2 real assigned cards at rest and two tilts; shots in ~/.pokedex-dev-hub-legacy/foil-shots/r6/

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

ea98da1 — 2026-08-07 — Mask provenance: sidecar v3, the AI learning loop, and the collapse safeguard

Chey (2026-08-07): mark AI-generated masks, let an AI learn from his hand-done ones, then let him correct the AI's mask so it "can observe the diff and continue to improve without me having to hand-paint all the masks."

derivation_method existed in v2 but the API HARDCODED it to 'hand' — harmless while only a Pencil touched the corpus, actively wrong once mask-refine's Flatten started writing machine geometry through the same save path.

Taxonomy (5 values, 4 cases that must never blur): layout-flatten machine geometry, unpainted weight 0 hand human, from a layout prior weight 1 hand-refined human, on an existing mask weight 1 ai generator, UNREVIEWED weight 0 ai-corrected AI proposed, human edited weight 0.6

The label is never taken from the client. The client reports only what the canvas was SEEDED with; writeMaskRecord derives the method by diffing saved pixels against what that seed rasterizes to. A machine label needs a full GeneratorIdentity, which HTTP callers cannot supply. One write path serves the route and generators both.

• Anti-feedback-collapse enforced in selectExemplars(), not by convention: unreviewed ai is weight 0 at any corpus size. Codify now routes through it.
• Corrections keep the parent PNG + change map + metrics (incl. a 4x4 grid of where corrections land) — the supervised pairs a generator lane consumes via the self-describing tuples manifest.
• Seam rule, measured not guessed: canvas roundRect vs the server SDF disagree on 389/330260 px, ALL in the 1px AA band. Without tolerating it every unpainted Flatten would stamp hand.
• Guard: if the file at the target path is an unreviewed ai mask it IS the parent, whatever the client claims — AI ancestry cannot be laundered. That guard caught a real bug during this build (ActionBtn passed React's MouseEvent into saveMask's optional arg, blanking derivation); derivation is now required on the PUT so it fails loudly instead.
• v1/v2 load forever (normalize-on-read); Chey's two masks migrated additively and still read HAND-PAINTED, agreement 0.641, pixels untouched.
• Trial batch wotc-window-trial-1: 6 WOTC Base holo rares from window-artgate@1, all ai/unreviewed, reversible. Leave-one-out on the human corpus beat the rect baseline (0.588 -> 0.688) — real, but n=2 same-set, so explicitly not validation.

Surfaces: provenance badge + detail in the workbench, Mask corpus panel, GET /foil-lab/masks/corpus (+?tuples/?exemplars), artifact routes, and corpus.ts report|exemplars|tuples|migrate. Pure tests added to CI (test:foil). Browser-verified at 390px and desktop. See DECISIONS.md 2026-08-07 and the mask-pipeline SKILL.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

57e39db — 2026-08-08 — foil mask-straighten: read the INTENT out of Chey's hand mask, and make the undo real

Chey, 2026-08-08: "it's impossible to get the lines really straight so I'm hoping you can get computer vision on the mask and card art in tandem to really see my intent there, and make the mask nice and crisp and straight on the lines I was trying to draw along. Will you try that? And then I can either approve or revert."

Smoothing a wobbly boundary gives a smoothly wobbly boundary. The premise here is that a hand-drawn foil boundary is an ATTEMPT TO TRACE A PRINTED EDGE — frame, art box, species strip, stage tag — and those are dead straight in the scan. So: contour the mask, cut each loop into near-axis runs (PCA orientation, wobble gaps bridged), robust-fit each, hunt the scan's directional gradient with a local Hough over (angle, offset) plus a proximity prior, replace, intersect adjacent lines into real corners, rasterize matching his own AA character.

The refusals are the feature, and each is a recorded param: weak evidence, or an ambiguous band of comparable ridges, may nudge but never relocate; with no artwork edge it straightens only to HIS own fit; short and curved runs pass through untouched. A first cut "closed" the Tropius top-left corner — which turned out to be him deliberately tracing around the BASIC stage tag. The stricter rules keep it.

• apps/api/src/foil/line-snap.ts the geometry, pure, no I/O
• generator.ts line-snap@1, a REFINER (requiresSource)
• generate-masks.ts run --refine, archives, restoring revert
• provenance.ts supersedes + verbatim sha256 undo archive

A generator write onto an existing mask now THROWS without explicit supersede { runId }; with it, every artifact the mask had is archived verbatim under superseded/./ with a self-describing manifest, and revert --run-id verifies the whole archive BEFORE deleting anything live, then restores byte-for-byte. supersedes is deliberately NOT correction: one is a human editing a machine (training signal), the other a machine replacing a human with nobody's agreement. --refine also refuses a source with exemplar weight 0, so a refiner can never eat its own unreviewed output.

Tropius me05-001/37184 (run straighten-tropius-1): 8 runs snapped to printed edges, 1 self-straightened, 13 left exactly as drawn, 8 corners closed, 2 specks dropped. Max correction 3.74px, 1.06% of the face changed, Jaccard 0.9798, confidence 0.557. Honest residual: the card's bottom border was too faint to accept, so that run self-straightened and the report says so.

test:foil 25/25 (9 geometry + 5 supersede/restore, incl. byte-for-byte round trip, a deliberate curve surviving, an ambiguous band refused, and a corrupt archive aborting before deletion). Revert exercised twice end to end, sha256sum clean. Shots: ~/.pokedex-dev-hub-legacy/foil-shots/mask-straighten/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id straighten-tropius-1

68687de — 2026-08-08 — foil R7: his eye overturns two banked mechanisms; "no catalog cards" gets a reason

Chey's 15 canon-lab comments (2026-08-07), all resolved in issues/foil/.

Eleven shader verdicts. Two overturn mechanisms built from earlier corpus — energy-symbols' parity CHECKERBOARD (R3-GLYPH, from his own words) and prism's solid-facet grid (R3-MISC, from a Gemini delta pass). Having seen both rendered he rejected them. A description is a hypothesis about the render, not a spec.

cosmos the 0.085 UV field slide he read as parallax is now uP4 "Surface drift" at 0.012; zero at rest, canon unchanged striped-v-sheen grey lines: the unconditional silver term is gone AND the second-order lobe cut 0.4 -> 0.16 (it drew a whole second, near-achromatic line population on the far side) rainbow-mirror hue axis is card X, blue -> green -> red, on BOTH the sheen and the spotlight; core whitening 0.85 -> 0.32 energy-symbols checkerboard gone: per-symbol response AXIS + PHASE + transition WIDTH over a low-freq cluster field energy-symbols-ii same treatment (it shared the two-bank mechanism) pokeball-hologram flat SDF icon -> shaded 3D sphere (hemisphere normal, tilt-following light, belt, hub, two-tone shell) tinsel the "blinds" were a literal full-card striation floor; demoted to uP2, default 0 (his canon stores 0) tinsel-ii line-work gated to the colour band, silver 0.60 -> 0.10, band width exposed as uP4 prism pinwheel's spoke fan WITHOUT the per-wedge phase that makes a pinwheel spin; density 33 -> 18 radiant hologram step 0.5 cell -> uP4 0.17, drive rescaled so total travel per unit tilt is unchanged rainbow-glitter invisible on card: bigger white-lifted flakes AND the family fixed (flash -> field; no substrate, no contrast)

Canon policy held: shader.ts / canon.ts / CardViewer.tsx / data/foil-canon have a ZERO diff, sheenGlsl() is byte-identical (all four sheen canons frozen), and every new behaviour went into a NEW uniform old canon files inherit. No stored value migrated. One recommendation left for him, not applied: tinsel-ii's starkness is substantially his own canon (uDepth 1 at ~135 lines).

"Why are there no catalog cards?" — asked four times, three answers. The baked inversion is now v2 and emits a machine-readable diagnosis the lab renders: outranked diagonal-sheen-left — 1,818 SM reverses ARE cited for it, but energy-symbols wins them (the emblem layer vs the sheet rotation; both true). A capped SECONDARY pool is now baked via a new citedFoilPatterns() export and served as via:'cited', so the preview works. Never flip a winner to fill a pool. class-absent pinwheel, pokeball-hologram — the catalog has reverse variants for ex1-ex5 ONLY; the whole late-EX reverse era is missing upstream. The UI now says so instead of "no catalog cards". fixed vertical-sheen-rainbow 0 -> 12 cards. known_residuals recorded twice that the 6 Holo Rare Energies of ex13/ex16 use Sheen but could not be targeted without a card list; the catalog names exactly six per set. Two cardId rows added citing the same Collexy quotes, both residuals closed.

Typecheck (all workspaces) + web build + foil tests green; 14 recipes compile clean; before/after shots on blank + assigned card at 390px and desktop in ~/.pokedex-dev-hub-legacy/foil-shots/r7/. SKILL R7 field notes + DECISIONS appended.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

fa965be — 2026-08-08 — foil edge-trace: the artwork holds the geometry, his mask says which of it to follow

Chey, after reviewing line-snap: "Really just 'straighten' isn't quite enough, really just detecting edges and actually tracing accurately around them is the real move." So the premise moves. line-snap treats his strokes as the geometry and nudges near-axis runs onto printed lines. edge-trace treats his mask as a STATEMENT OF INTENT and derives the boundary from the card's own edges.

apps/api/src/foil/edge-trace.ts (pure, no I/O): Di Zenzo colour structure tensor (Scharr per channel) -> NMS -> hysteresis; colour because a green field against a silver border reads 0.13 luminance/px and 44.3 tensor/px at the same edge. Then livewire: a corridor around his boundary (a hard bound, not a suggestion), anchors at fixed spacing PLUS every hard turn in his own line, Dijkstra on ridge + linked + proximity^2 + direction, sub-pixel refinement onto the |dI/dn| peak, and only THEN a straight fit where the traced path really is straight.

His own correction said the same thing before he did: 114 of his 627 correction px sit in one 40x50 box around the species strip's swept tail, which line-snap had squared off by intersecting two lines across it.

Two traps, both now locked by tests:

• the MAD-trim trap: a robust fit discards a small feature as an outlier, reports "straight to 0.2px RMS", and crisping then flattens what it ignored. straightMaxDevPx requires every point in a run to be close.
• the half-pixel frame: mask space centres pixels at x+0.5, a gradient array at integer x. All artwork lookups convert in one place.

Measured, not asserted. generate-masks.ts adherence scores any set of masks by distance to the NEAREST printed edge maximum, with line-snap's own luminance Sobel as the default probe so the incumbent gets the home field: his hand mask 50.7% within 1px / 0.847px mean line-snap@1 74.8% / 0.545px his correction 75.4% / 0.548px edge-trace@1 77.1% / 0.199px (colour probe: 98.4% / 0.082px)

Applied to Tropius me05-001/37184 as run edgetrace-tropius-1, ai/unreviewed, his ai-corrected mask preserved as parent + verbatim supersede archive. Revert proven byte-identical twice before shipping.

test:foil 35/35 (10 new). Typecheck all workspaces, api+mcp+web builds green. Visuals in ~/.pokedex-dev-hub-legacy/foil-shots/mask-edgetrace/.

Undo: pnpm --filter pokedex-api exec tsx src/foil/generate-masks.ts revert --run-id edgetrace-tropius-1

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

9c2c43d — 2026-08-08 — Mask editor: pan + pinch-zoom that keeps the strokes honest, and no more "selected" card

Chey, before a batch of hand masks: "something that would be nice is to be able to pan and pinch to zoom while editing masks. In addition, sometimes the card image gets 'selected' when I'm trying to draw which is annoying."

Zoom is a camera view offset, not a CSS scale. camera.setViewOffset(Wz, Hz, x, y, W, H) re-rasterizes the card at the zoomed size, so 4x zoom is 4x real detail — the point of zooming is tracing a printed edge, and a magnified 390px framebuffer would not help. The overlays get the same transform on one wrapper div, written imperatively (no re-render per frame). Because that is a pure crop of a linearly-scaled render, u lands at u*z - offset in both places, so getBoundingClientRect already reports the truth and the pointer->texel math is unchanged. Verified numerically: a 3.375x corner stroke lands at texels [84,180,108,196] vs a predicted [85,180,109,196]; touch at 3x lands within one texel at 390 and 744.

Brush is constant in SCREEN px (lineWidth = brushSize / zoom): zooming in must buy finer control, not a fatter-looking stroke. Two fingers pinch/pan; one finger pans unless it is a drawing finger or on a window handle; a pen owns the surface (palm rejection); a second finger mid-stroke aborts AND rolls the stroke back. Desktop gets wheel/trackpad zoom at the cursor, middle- or Space-drag pan, and +/-/0. HUD shows the zoom and fits back in one tap. Pan is clamped so the card cannot be lost. Zoom is live only while editing, resets on exit, and never reaches the artifact — a mask saved at 3.375x round-trips as 490x674 with an identical bbox. WindowEditor divides drag deltas by zoom and counter-scales its handles, so they stay 44px on screen (44.0 -> 44.0 measured at 2.25x).

Selection fix: user-select/touch-callout/user-drag none, draggable=false, and preventDefault on dragstart/selectstart and the paint pointerdown across both canvases and the viewer column; touch-action stays none so the two fixes do not fight. iOS also needs the gesturestart/change/end veto — touch-action alone does not stop Safari page zoom.

Cost the most time: window listeners registered in the controller factory were killed by StrictMode's simulated unmount, so the first pinch started and never ended and every later gesture and stroke was dead. They are now a symmetric attach()/detach() pair driven by an effect. See DECISIONS.md.

44/44 browser checks at 1280 / 744 / 390 (touch via CDP), his committed masks still load, scratch card artifacts removed. Shots in ~/.pokedex-dev-hub-legacy/foil-shots/mask-ux/.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

f135964 — 2026-08-08 — foil region-learn: the REGIONS come from his masks, the geometry from the artwork

The last lane measured why rule-seeded edge-trace failed on every me05 reverse: 99.7% of the gap between the layout rule and Chey's intent is REGION decisions, and a tracer can only move a boundary, never add or remove a region. This is the missing half.

region-learn@1 partitions a card face into five structural classes computed from its OWN printing (border / furniture / frameBody / windowBackground / windowSubject), reads off his masks WHICH of them carry foil, applies that to a new card, and only then hands the boundary to edge-trace. Nothing about Pokemon is hard-coded — the same code reads OPPOSITE policies from the two classes in the corpus, and a test asserts the votes, not just the verdict.

The policy, in his pixels: modern-sv/sheet foil = the COLOURED FRAME BODY, nothing else. Species strip (tails included), stage tag, evolution medallion + its sprite, "evolves from" bar, copyright footer: all excluded. And THE SILVER BORDER RING IS NOT FOIL — 0.2/0.2/0.4% across his three masks. That was flagged as "his call, not a measurement"; it is now a measurement, and unanimous. wotc/window foil = the illustration's own background, minus the subject silhouette, minus the stage box where it overlaps. 4/4.

Bar written down before any number existed: mean IoU >= 0.90, no held-out card below 0.85. Leave-one-out: modern-sv/sheet 0.7566 -> 0.9691 (worst 0.9519) PASS -> batch of 8 shipped wotc/window 0.6146 -> 0.8971 (worst 0.8599) FAIL by 0.0029 -> no batch

The two residuals differ in kind, and that is the finding: modern-sv's is a 1-2px registration hairline; wotc's is 8,343px of Machamp's body read as background because his blue-grey sits inside the colour distance of the teal holo field. Every frame-level decision on WOTC is right. The subject silhouette is not.

Also found and NOT changed (era-wide, his call): wotc.artWindow is 11px too high at the top and 6px too wide at the right — his four masks and each card's own printed edge agree, and base1-8 is the card the era was "measured on". Same pattern as me04-007 for modern-sv.

Also: the bevel side. An illustration box has TWO parallel printed edges; which one the foil stops at is which side the foil is on, not which is strongest. Picking the strongest put base1-6's window top 8px wrong and scored it 0.0001; picking the foil side took it to 0.8789.

Batch regionlearn-me05-1 (8 cards, 5 frame colours, 4 with medallions, all ai / unreviewed) applied and proven reversible byte-for-byte — including a revert bug fixed here: it left the empty card directory behind, so a reverted card still looked masked to anything that walks data/foil-masks.

Recommend reverting wotc-window-trial-1: those five base1 masks were made by window-artgate@1, which scores 0.7694 mean / 0.7330 worst on this same test.

test:foil 46/46 (11 new, including the anti-collapse rule locked three ways).

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

11363c4 — 2026-08-08 — foil vector-template@1: the layout is the artifact, and it is analytic geometry

Chey: "what I draw is intent... generated masks should feel like they're derived from clean vectors" — and "we don't need 3,454 vector masks, all of these share the same 2 layouts really."

Mined his 8 correction diffs first. They say the regions were nearly right (0.12% to 2.23% changed) and that the errors were structural, not stylistic:

• the coloured sliver pinched off between the evolution medallion and the border ring IS foil (he added it on all 4 Stage-1 cards). This was the open question flagged in codified/modern-sv.md as "Chey has never ruled on that sliver". It is ruled on now.
• achromatic marks printed ON the coloured field carry foil: the colourless energy symbols, the retreat symbols, the regulation-mark box, the name text. region-learn@1 carved holes around all of them; he filled every one.
• the left/right window edge sat on the bevel's INNER line; he moved it out.
• and, decisively for this lane: where the print is straight, HIS line is straighter than region-learn@1's traced one. The machine was the thing that wobbled, so a smoother tracer was never the fix.

Geometry, measured over his 11 masks: 50.05% of the face is foil on ALL of them, 47.18% bare on all, 2.77% contested — and the Basic-vs-Stage1 difference is a SINGLE 70x85 blob. So it is not 2 layouts, it is one layout plus one optional element, which the fitter discovers from the corpus itself (bimodal split, 80.5pp separation) rather than being told a medallion exists.

vector-template@1 fits lines + circular arcs + exact corners to the consensus of his masks and rasterises that. 58 primitives, 13.2 KB of committed JSON, applied at render time — instead of 3,454 rasters.

Leave-one-out vs the bar committed in a14a163 BEFORE any number existed (mean >= 0.94, none below 0.90): mean 0.9904, worst 0.9838. PASS. region-learn@1 on the same 11: 0.9757 / 0.9519. Boundary p95 1px on 10 of 11 (was 2-72px).

Vector-ness at 0.35px tolerance on SUB-PIXEL contours (measuring on traceLoops is a trap — crack following returns a rectilinear staircase and rates a hand blob as perfectly axis-aligned): primitives/kpx region-learn@1 33.7 HIS 21.3 vector-template@1 10.7 This is the finding: his hand masks are ALREADY more vector-like than the generator that was supposed to be learning from them.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

8edb243 — 2026-08-08 — foil vector-template: make the emitted path actually analytic (two real bugs)

Writing the round-trip test found two things the IoU table was hiding.

1. JOINS WERE RAW CONTOUR POINTS. vectorizeLoop fitted lines, axis-snapped them, and then took each run's endpoint straight off the sampled contour, only substituting an intersection for line/line corners. So the emitted "vertical" left edge of a rounded rect ran from x=25.00 to x=25.66 — a polyline through contour samples with arcs bridging them, wearing a vector's clothes. The straightness this lane exists to deliver was measured but not emitted. Joins are now the actual intersection of the two primitives that meet there (line/line, line/arc, arc/arc all solved), with joinMaxMovePx rejecting a join that lands far from where the contour really turned.

2. ARC SWEEP WAS INVERTED. cross < 0 ? 1 : 0 — with y down, a positive cross product is a clockwise turn on screen and sweep=1 is clockwise, so every fitted arc bulged the wrong way. Round-trip IoU on a rounded rect: 0.9482 before, 0.9994 after. It only cost ~0.5% on the real template because that template is 56 lines and 2 arcs, which is exactly why a synthetic test found it and the corpus numbers did not.

Template refitted. Nothing moved much, because the bugs were small in this geometry — but the artifact now says what it means: in-sample mean IoU 0.9908 -> 0.9911 vector-ness prim/kpx region-learn@1 33.72 | HIS 21.35 | mine 10.34 long-primitive fraction 0.9047 | 0.9679 | 0.9861 axis-aligned fraction 0.7875 | 0.8554 | 0.9321 correction margin: closer to what he drew than to what he rejected on all 10.

test:foil 56/56.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

0b63ef5 — 2026-08-08 — foil: serve the template as a mask, and write down what his corrections ruled

API: GET /foil-lab/masks/:cardId/:variantId?era=&scope=&evolves=1 now falls back to the vector template when no hand mask resolves. A hand mask ALWAYS answers first, so this cannot shadow Chey's work — it only fills what used to 404 — and it is labelled X-Foil-Mask-Review-Status: unreviewed with the generator identity, run id and exemplar count.

The cache is TWO PNGs, not 3,454. A template's rasterisation depends only on (era, scope, hasOptionalElement, w, h) and not on the card, so every Basic reverse gets byte-identical pixels. ~17 KB each under IMAGE_CACHE_ROOT/foil-templates/, mirroring the image-cache contract: disk-only, outside git, safe to delete, rebuilds in milliseconds. Committing per-card masks would have been ~100-350 MB against the template's 13.3 KB.

evolves is an INPUT, not something the route detects, for two structural reasons: these routes have no DB and the server never shells out, so the process can neither read card.stage nor decode card art. The caller already knows. This turned out to be the right call for a second reason — the artwork probe agrees with the catalog on 89.3% of Pokemon but misfires on pale frames (it called sv09-120 Dunsparce, a Basic, "medallion present" at 0.0% chroma).

codified/modern-sv.md pass 2 (n=11) records the three rulings his corrections make, with numbers:

• THE SLIVER IS FOIL. Pass 1 flagged it as "Chey has never ruled on that sliver". region-learn@1 scored 0.0% there on all four Stage-1 cards; he added it on all four. "Largest coloured component" was the wrong rule; the right one is every coloured region that reaches the border ring. The medallion's own sprite stays out because it is fully enclosed — topology, not size.
• Achromatic ink printed ON the coloured field carries foil: colourless energy symbols, retreat symbols, the regulation box, the name text. Furniture is distinguished topologically, not by chroma.
• The window edge sits on the bevel's OUTER line.

And what is NOT covered, said out loud: Trainer (431 variants) and Energy (104) are a different layout with zero exemplars — visibly wrong, probe agrees with the catalog on 0/49 of them. Pale frames cannot self-verify (45.5% of Lightning and 41.0% of Colorless flagged vs 0.0% of Fire/Psychic/Darkness) — a contrast effect, calibrated against his own hand masks scoring 0.720 on the same probe.

Revert proven, not asserted: a 2-card run reverts to a byte-identical tree (git sees nothing), and a supersede round-trip restores the earlier mask with matching sha256 on both the PNG and the sidecar.

DECISIONS also records an operational gotcha I caused: pkill -f on "apps/api/dist/index.js" matches the deployed pm2 pokedex-api too. pm2 healed it in seconds (prod 200, unstable restarts 0), but kill dev instances by PID.

typecheck clean across all workspaces; test:foil 56/56; api + web builds green.

Co-Authored-By: Claude Opus 5 (1M context) noreply@anthropic.com Claude-Session: https://claude.ai/code/session_01HE3n5mBRZUCv4oH1dWqPi6

---

The remaining commits

The 203 commits whose message body is at or under ~100 words — rebases, chore commits, one-line fixes and short notes. Grouped by branch, chronological within branch. Short SHAs resolve against the mirror.

foil/analysis-source

Commit	Date	Subject
31810c4	2026-08-01	foil corpus: commit Chey's first hand mask + workbench comment, untouched
e84fcc4	2026-08-02	foil: Starlight parallax rework from Chey's critique; resolve his comment
5a7c90d	2026-08-02	video reference corpus, pre-synthesis checkpoint
d1ab2d8	2026-08-02	research/foil-patterns.md: canonical 39-pattern spec (video + Bulbapedia + vision pass reconciled)
21b6365	2026-08-02	research/foil-pattern-usage.json: cited era->pattern map (7 Ringer lanes + editorial conflict rows, 113 rows)
8c55380	2026-08-02	foil: Gemini verification of all 33 catalog patterns + report + diagonal-sheen-left fix
29a5f91	2026-08-02	foil/vocab: harvest checkpoint — keyframes+clips for the four vocabulary-extension patterns
d625d15	2026-08-02	foil/vocab: four vocabulary-extension pattern types — gold-secret, vstar-pearl, shiny-vault, detective-pikachu
afcb885	2026-08-02	foil R2 stage 3 docs: verdict table (10/13 match), field notes, DECISIONS entry
8d1b9d8	2026-08-02	foil: cosmos canon snapshot found saved via canon lab (uDarken 0.2, warm hue)
836018a	2026-08-02	foil corpus: hand mask for base1-5 v19 (sidecar v2, saved via workbench)
cc04c62	2026-08-02	foil corpus: Chey's canon pass — 29 pattern canons saved via canon lab
08cbeac	2026-08-02	foil R3: diagonal-sheen slope swap — right is "", left is "/" (Chey + frames; Gemini was right all along)
f068ef4	2026-08-03	foil CLOSER: merge closeout — CHEY-REVIEW punch list + DECISIONS entry
8a62af3	2026-08-07	foil canon: Chey's hand-tuned mirror canon (his words: "basically perfect")
7c8e3be	2026-08-07	foil canon: Chey's hand-tuned cracked-ice canon (saved 2026-08-07 04:35)
723d63b	2026-08-07	foil canon: Chey's tinsel-ii + rainbow-glitter-sheen, hand-tuned on the R5b additive law
145359c	2026-08-07	foil canon: Chey's re-adjusted cracked-ice (2026-08-07 12:11)
c4c622f	2026-08-08	foil corpus: Chey's Tropius me05-001 sheet mask (hand-refined), pokeball-hologram canon, 7 canon re-tunes, first window adjustment
74e4949	2026-08-08	foil corpus: Chey's correction of the line-snap Tropius mask (ai-corrected)
db4eb50	2026-08-08	foil: fix modern-sv art window top edge (106px -> 66px)
2dca1be	2026-08-08	foil corpus: Chey's mask batch — 2 new modern-sv sheet masks, base1-7 new, base1-5/6/8 refined
49f4a86	2026-08-08	foil corpus: Chey's corrections of all 8 regionlearn-me05-1 masks (modern-sv/sheet now 11 human-authored)
8e4d7e2	2026-08-08	foil/mask-vector: state the bar before the generator exists
699e345	2026-08-09	privacy: scrub personal paths and rename pokedex->deckscout in foil code
602155a	2026-08-09	rebase: re-apply DECISIONS.md entries after main reword
65a683a	2026-08-09	strip: remove legacy deployment-era references

foil/main

Commit	Date	Subject
d9ad551	2026-08-01	foil corpus: commit Chey's first hand mask + workbench comment, untouched
6895176	2026-08-02	foil: Starlight parallax rework from Chey's critique; resolve his comment
bf8078c	2026-08-02	video reference corpus, pre-synthesis checkpoint
a80c40e	2026-08-02	research/foil-patterns.md: canonical 39-pattern spec (video + Bulbapedia + vision pass reconciled)
7dccec7	2026-08-02	research/foil-pattern-usage.json: cited era->pattern map (7 Ringer lanes + editorial conflict rows, 113 rows)
7c016f4	2026-08-02	foil: Gemini verification of all 33 catalog patterns + report + diagonal-sheen-left fix
356da5c	2026-08-02	foil/vocab: harvest checkpoint — keyframes+clips for the four vocabulary-extension patterns
b54c9ea	2026-08-02	foil/vocab: four vocabulary-extension pattern types — gold-secret, vstar-pearl, shiny-vault, detective-pikachu
f118a32	2026-08-02	foil R2 stage 3 docs: verdict table (10/13 match), field notes, DECISIONS entry
d021785	2026-08-02	foil: cosmos canon snapshot found saved via canon lab (uDarken 0.2, warm hue)
6fbfe98	2026-08-02	foil corpus: hand mask for base1-5 v19 (sidecar v2, saved via workbench)
0109828	2026-08-02	foil corpus: Chey's canon pass — 29 pattern canons saved via canon lab
daad1a8	2026-08-02	foil R3: diagonal-sheen slope swap — right is "", left is "/" (Chey + frames; Gemini was right all along)
0fa761c	2026-08-03	foil CLOSER: merge closeout — CHEY-REVIEW punch list + DECISIONS entry
20eb358	2026-08-07	foil canon: Chey's hand-tuned mirror canon (his words: "basically perfect")
f7dde7e	2026-08-07	foil canon: Chey's hand-tuned cracked-ice canon (saved 2026-08-07 04:35)
b8b5fe1	2026-08-07	foil canon: Chey's tinsel-ii + rainbow-glitter-sheen, hand-tuned on the R5b additive law
44307aa	2026-08-07	foil canon: Chey's re-adjusted cracked-ice (2026-08-07 12:11)
1611ab9	2026-08-08	foil corpus: Chey's Tropius me05-001 sheet mask (hand-refined), pokeball-hologram canon, 7 canon re-tunes, first window adjustment
9bd3c21	2026-08-08	foil corpus: Chey's correction of the line-snap Tropius mask (ai-corrected)
67ed307	2026-08-08	foil: fix modern-sv art window top edge (106px -> 66px)
b5d7910	2026-08-08	foil corpus: Chey's mask batch — 2 new modern-sv sheet masks, base1-7 new, base1-5/6/8 refined
8c17091	2026-08-08	foil corpus: Chey's corrections of all 8 regionlearn-me05-1 masks (modern-sv/sheet now 11 human-authored)
95d1031	2026-08-08	foil/mask-vector: state the bar before the generator exists
d34c3b5	2026-08-08	foil analysis-source: honest provenance for the pixels a generator fits to
b50ae1a	2026-08-09	privacy: scrub personal paths, rename pokedex->deckscout in foil code
84bc671	2026-08-09	rebase: re-apply DECISIONS.md entries after main reword
ceb0ab1	2026-08-09	strip: remove legacy deployment-era references

foil/mask-edgetrace

Commit	Date	Subject
f1109ca	2026-08-01	foil corpus: commit Chey's first hand mask + workbench comment, untouched
02c953c	2026-08-02	foil: Starlight parallax rework from Chey's critique; resolve his comment
082b57d	2026-08-02	video reference corpus, pre-synthesis checkpoint
67969cd	2026-08-02	research/foil-patterns.md: canonical 39-pattern spec (video + Bulbapedia + vision pass reconciled)
abb7fa3	2026-08-02	research/foil-pattern-usage.json: cited era->pattern map (7 Ringer lanes + editorial conflict rows, 113 rows)
d44df07	2026-08-02	foil: Gemini verification of all 33 catalog patterns + report + diagonal-sheen-left fix
66f8652	2026-08-02	foil/vocab: harvest checkpoint — keyframes+clips for the four vocabulary-extension patterns
c919cd9	2026-08-02	foil/vocab: four vocabulary-extension pattern types — gold-secret, vstar-pearl, shiny-vault, detective-pikachu
427e0b8	2026-08-02	foil R2 stage 3 docs: verdict table (10/13 match), field notes, DECISIONS entry
156fb73	2026-08-02	foil: cosmos canon snapshot found saved via canon lab (uDarken 0.2, warm hue)
e762077	2026-08-02	foil corpus: hand mask for base1-5 v19 (sidecar v2, saved via workbench)
78682a3	2026-08-02	foil corpus: Chey's canon pass — 29 pattern canons saved via canon lab
a3d1297	2026-08-02	foil R3: diagonal-sheen slope swap — right is "", left is "/" (Chey + frames; Gemini was right all along)
9ac7e58	2026-08-03	foil CLOSER: merge closeout — CHEY-REVIEW punch list + DECISIONS entry
017eb76	2026-08-07	foil canon: Chey's hand-tuned mirror canon (his words: "basically perfect")
97e45d6	2026-08-07	foil canon: Chey's hand-tuned cracked-ice canon (saved 2026-08-07 04:35)
0888844	2026-08-07	foil canon: Chey's tinsel-ii + rainbow-glitter-sheen, hand-tuned on the R5b additive law
2e50c10	2026-08-07	foil canon: Chey's re-adjusted cracked-ice (2026-08-07 12:11)
339f726	2026-08-08	foil corpus: Chey's Tropius me05-001 sheet mask (hand-refined), pokeball-hologram canon, 7 canon re-tunes, first window adjustment
45fe8e1	2026-08-08	foil corpus: Chey's correction of the line-snap Tropius mask (ai-corrected)
7313449	2026-08-09	privacy: scrub personal paths and rename pokedex->deckscout in foil code
9d1b22a	2026-08-09	rebase: re-apply DECISIONS.md entries after main reword
9161536	2026-08-09	strip: remove legacy deployment-era references

foil/mask-edgetrace-batch

Commit	Date	Subject
f2f9500	2026-08-01	foil corpus: commit Chey's first hand mask + workbench comment, untouched
37357d1	2026-08-02	foil: Starlight parallax rework from Chey's critique; resolve his comment
c154b7f	2026-08-02	video reference corpus, pre-synthesis checkpoint
02de13d	2026-08-02	research/foil-patterns.md: canonical 39-pattern spec (video + Bulbapedia + vision pass reconciled)
d172df7	2026-08-02	research/foil-pattern-usage.json: cited era->pattern map (7 Ringer lanes + editorial conflict rows, 113 rows)
16098b2	2026-08-02	foil: Gemini verification of all 33 catalog patterns + report + diagonal-sheen-left fix
512c322	2026-08-02	foil/vocab: harvest checkpoint — keyframes+clips for the four vocabulary-extension patterns
ff4de8b	2026-08-02	foil/vocab: four vocabulary-extension pattern types — gold-secret, vstar-pearl, shiny-vault, detective-pikachu
385b298	2026-08-02	foil R2 stage 3 docs: verdict table (10/13 match), field notes, DECISIONS entry
8295042	2026-08-02	foil: cosmos canon snapshot found saved via canon lab (uDarken 0.2, warm hue)
1ae1531	2026-08-02	foil corpus: hand mask for base1-5 v19 (sidecar v2, saved via workbench)
7c1b1ce	2026-08-02	foil corpus: Chey's canon pass — 29 pattern canons saved via canon lab
7444633	2026-08-02	foil R3: diagonal-sheen slope swap — right is "", left is "/" (Chey + frames; Gemini was right all along)
8b7b8b9	2026-08-03	foil CLOSER: merge closeout — CHEY-REVIEW punch list + DECISIONS entry
599add5	2026-08-07	foil canon: Chey's hand-tuned mirror canon (his words: "basically perfect")
8bddb48	2026-08-07	foil canon: Chey's hand-tuned cracked-ice canon (saved 2026-08-07 04:35)
0f3bab2	2026-08-07	foil canon: Chey's tinsel-ii + rainbow-glitter-sheen, hand-tuned on the R5b additive law
bbbeca9	2026-08-07	foil canon: Chey's re-adjusted cracked-ice (2026-08-07 12:11)
b528e26	2026-08-08	foil corpus: Chey's Tropius me05-001 sheet mask (hand-refined), pokeball-hologram canon, 7 canon re-tunes, first window adjustment
f7e8948	2026-08-08	foil corpus: Chey's correction of the line-snap Tropius mask (ai-corrected)
5575f60	2026-08-09	privacy: scrub personal paths and rename pokedex->deckscout in foil code
71c975b	2026-08-09	rebase: re-apply DECISIONS.md entries after main reword
c611e3f	2026-08-09	strip: remove legacy deployment-era references

foil/mask-learn

Commit	Date	Subject
93a311b	2026-08-01	foil corpus: commit Chey's first hand mask + workbench comment, untouched
0f1a137	2026-08-02	foil: Starlight parallax rework from Chey's critique; resolve his comment
29888e1	2026-08-02	video reference corpus, pre-synthesis checkpoint
a916891	2026-08-02	research/foil-patterns.md: canonical 39-pattern spec (video + Bulbapedia + vision pass reconciled)
1f202b5	2026-08-02	research/foil-pattern-usage.json: cited era->pattern map (7 Ringer lanes + editorial conflict rows, 113 rows)
bd9a463	2026-08-02	foil: Gemini verification of all 33 catalog patterns + report + diagonal-sheen-left fix
1baeb0a	2026-08-02	foil/vocab: harvest checkpoint — keyframes+clips for the four vocabulary-extension patterns
1941d58	2026-08-02	foil/vocab: four vocabulary-extension pattern types — gold-secret, vstar-pearl, shiny-vault, detective-pikachu
a0331b2	2026-08-02	foil R2 stage 3 docs: verdict table (10/13 match), field notes, DECISIONS entry
95bb63c	2026-08-02	foil: cosmos canon snapshot found saved via canon lab (uDarken 0.2, warm hue)
90baae9	2026-08-02	foil corpus: hand mask for base1-5 v19 (sidecar v2, saved via workbench)
f79ce9c	2026-08-02	foil corpus: Chey's canon pass — 29 pattern canons saved via canon lab
187bc4e	2026-08-02	foil R3: diagonal-sheen slope swap — right is "", left is "/" (Chey + frames; Gemini was right all along)
636952d	2026-08-03	foil CLOSER: merge closeout — CHEY-REVIEW punch list + DECISIONS entry
9996e80	2026-08-07	foil canon: Chey's hand-tuned mirror canon (his words: "basically perfect")
b25190a	2026-08-07	foil canon: Chey's hand-tuned cracked-ice canon (saved 2026-08-07 04:35)
e0738a8	2026-08-07	foil canon: Chey's tinsel-ii + rainbow-glitter-sheen, hand-tuned on the R5b additive law
6cf0b43	2026-08-07	foil canon: Chey's re-adjusted cracked-ice (2026-08-07 12:11)
7090006	2026-08-08	foil corpus: Chey's Tropius me05-001 sheet mask (hand-refined), pokeball-hologram canon, 7 canon re-tunes, first window adjustment
4ab13eb	2026-08-08	foil corpus: Chey's correction of the line-snap Tropius mask (ai-corrected)
bb71a9d	2026-08-08	foil: fix modern-sv art window top edge (106px -> 66px)
11ac399	2026-08-08	foil corpus: Chey's mask batch — 2 new modern-sv sheet masks, base1-7 new, base1-5/6/8 refined
9931bc9	2026-08-09	priva