model-v2: anchor — land the v2 engine as the crates/n0 family#5
Merged
Conversation
The full redesign workbench for node geometry/layout/transform, scoped to the Rust engine (crates/grida) and the format spec (format/grida.fbs): - phase 1-2 research corpus: problems, harnesses, peer study, candidate models (anchor / sheet / bake / wire) — anchor decided by survey - the 114-test Rust lab (model-v2/a/lab): four-phase resolver (Taffy flex, oriented envelopes), typed ops incl. cross-zero flip resize (E-A14), canonical text IR, node-arena + SOA storage (11.5x resolver speedup) - the decision register (a/DECISIONS.md) with DEC-0 LOCKED: visual-only rotation (the CSS framing), CSS-pure sizing (a/dec0-visual-only.md) - peer-compat matrix (COMPAT), ship-readiness census (LIMITS), report - demo pages (a/.preview): proof, model walkthrough, edge cases, DEC-0 fork A/B, free editing - E10: the native skia interactive spike (a/spike-canvas) — feel pass + TEXTBOOK.md, the reference for the legacy migration Isolated: nothing outside model-v2/ is touched; no production code paths.
…tests, branch note); fix MODEL.md dec0 link
…ontracts (ENG-0…ENG-5), validity-tagged
…nce the snapshot) The lab's accumulated evolution on the anchor branch since the textbook snapshot, across two threads — the seams anchor-engine consumes and the DEC-0 visual-only feel study — plus a crate-wide rustfmt pass. 121-test suite green. Engine seams (additive; serde stays out of default deps): - ops.rs: typed `Op` + pure-dispatch `apply` (journal/replay/dirty land on it; free fns kept, `apply ≡ free_fn` differential per variant) + `DirtyClass` parameterized on rotation-in-flow [ENG-1.2]. - model.rs: per-slot `generations` column + `gen_of` (semantic PartialEq ignores it — a storage artifact) [ENG-2.3]. - resolve.rs: non-panicking `box_opt`/`local_opt`/`aabb_opt`/`slot_count`. - serde feature (default OFF) — the op-log wire; new tests accessors, generations, op_apply. Visual-only exploration: rotation scenes added to the edge/fork probes (grow × rotation, rotated text/frame/group) and expanded edge_census + visual_only corpora anchoring DEC-0.
…spike re-hosted anchor-engine: document → resolve → drawlist → paint, plus the read tier (query), time-as-data (journal/replay), and the sockets every future optimization plugs into (damage, ident, oracle). It consumes anchor-lab as a library — the relationship the crates/grida migration will have with the model crate. Every ENG-0…ENG-5 contract has a code socket and a guarding test; skia is confined to paint.rs so a later core/paint split is mechanical. The spike is re-hosted onto it (proof, not just scaffolding): the scene painter is deleted → drawlist::build + paint::execute (4 goldens byte-identical, incl. the rotated and cross-zero-flip scenes); pick/hover via query; gesture ops via apply, recorded in the journal (undo stays document snapshots — ENG-5.5); `--record` writes .replay corpus. The scene-cache compositor (ENG-2) rasters once into a backend-matched offscreen and blits camera pans O(1), re-rastering only on doc-change / zoom / pan-beyond-margin. Live per-stage frame instrumentation (a glFinish GPU-vs-vsync split) plus render-on-demand (draw only on real input) land the felt-latency fix. gate (ENG-0.2 oracle law): shots 4/4 IDENTICAL · replays deterministic · scene-cache differential MATCHES fresh (integer pan) · bench within budgets. Docs: ENGINE.md SETUP ground contracts (S-1…S-7), engine README (contract → module → guarding-test map), model-v2 + a/README phase-4 status.
Performance is not one question, so it is not one tool — it splits on a hard line between what a machine can measure alone and what needs a human. This is the doctrine that stops us measuring the wrong axis (as a session's detour did: render *work* was green while the felt lag lived in the *present path*). MEASURE.md — the four axes and which tool owns each: - work + correctness are automated & headless (probe / probe_gpu / gate) — the default loop, no human. - feel & pacing need a real window: the opt-in ANCHOR_FRAMELOG channel (human runs + interacts + kills; the reviewer reads the per-frame gap/stage log). An on-demand bridge, not the iteration loop. - input→photon latency is NOT software-measurable — the lag accrues in the OS compositor/display after swap_buffers returns; measuring it needs a photodiode rig or a human glance. Named so a green gate is never mistaken for good feel. The automated work axis it documents: - bin/probe — deterministic CPU raster, per-stage distributions, JSON baselines, panicking audit-guards. The always-on regression detector. - bin/probe_gpu — headless surfaceless-GL, real flush_and_submit timing, for the wins whose value is a persistent GPU texture (CPU raster reports the wrong sign there). Worked example folded in: render-on-demand (idle frames 466 → 2, throughput 125fps median — the symptom was pacing/latency, not work).
…L doctrine Two gate-verified data-layout moves aligned to browser prior art, plus the decisions doc that grounds them: - Color: fill `Option<String>` → `model::Color(u32)` (0xAARRGGBB), read straight by the drawlist. No per-node heap String, no per-build hex parse; text-IR/SVG convert at the boundary via from_hex/to_hex. Parse byte-identical to the old resolve_color (pixels unchanged, gate 4/4 IDENTICAL); build −31% at 100k nodes. - Resolve caches: HashMap<NodeId,_> → Vec<Option<_>> (union_cache, ops_cache). NodeId is a dense arena index — hashing it is the cc anti-pattern. Correctness preserved (gate PASS, 121 lab tests). - spike-canvas: `pages` stress scene (realistic nested flex cards, ANCHOR_SCENE selected) + --bench rows; goldens still use starter(). - engine/DATA-MODEL.md: six data-layout details decided against cc/Blink/Stylo/ Skia prior art, validity-tagged (ALIGNED/ADOPTED/SOCKET), with honesty corrections to the record.
Package renames (anchor-engine -> n0, anchor-lab -> n0-model, anchor-spike -> n0_dev) with mechanical use-path renames across the three crates; root workspace members += n0-model, n0_dev; path deps rewired; compile-checked include fixes (lab tests' engine/rig includes -> ../../n0/rig, spike font fixture depth); gate.rs re-pointed at the n0_dev binary and ../n0_dev/shots goldens. Patrol ledger (patrol-before-drop) — deliberately dropped as obsolete: - the three [workspace] breakouts and "standalone on purpose" rationale comments (workspace membership is what promotion IS), - [profile.release] debug=true x3 (ignored in a workspace; profiling uses CARGO_PROFILE_RELEASE_DEBUG=true), - both .cargo/config.toml target-dir shims and every "shared target dir reuse" comment clause (automatic in-workspace), - the three standalone Cargo.locks (the root lock takes over; its diff is additions-only — no existing dep version moved), - the engine's phase-4 description (replaced by the stub's "A 2D graphics engine." + categories/keywords/readme metadata). Kept/rewritten per the same ledger: the serde op-log-wire rationale, the trace zero-cost-observability rationale, the native-gl-context GL-free-default rationale, the skia-confined-to-paint.rs invariant, and the shell-dep version-parity constraint (reason updated from target-dir reuse to workspace type unification). Crate-doc (//!) and README framing updated from phase-4/standalone future tense to the landed topology; archived-workbench references repointed to model-v2/ paths; 15 forbidden grida.co/docs/wg links converted to same-repo relative per the links discipline. Part of landing #5; tracking #9.
git rm src/bin/{e1,e3,edge,fork}.rs — the archive-coupled experiment
harnesses whose outputs are frozen under model-v2/a/
(e1-rotation-in-flow/, e3-text-ir/, edge-cases/, dec0-fork/). The
arg-driven e4 (resolver benchmarks) and e5scan (SVG corpus scan) stay.
Patrol ledger (patrol-before-drop) — essence captured into the archive
BEFORE deletion, in this commit:
- e3's truth protocol emits a world-AABB line only for rotated
NON-GROUP nodes — the group exclusion was undocumented (inert on the
frozen corpus); now stated in model-v2/a/e3-text-ir/README.md.
- edge's flipzero scene: only the flip arm is real resize_drag output;
wall and slide are analytic stand-ins — now stated in
model-v2/a/DECISIONS.md DEC-9 (was recoverable from nowhere else).
- fork's demo-meter definitions (escape = max sibling world-AABB
overrun past the container's ORIENTED painted rect, 0 = contained;
overlap = max pairwise sibling world-AABB intersection area) — now
stated alongside rule V-3 in model-v2/a/dec0-visual-only.md; escape
was defined nowhere else.
- e1 carries no at-risk essence (fully covered by
model-v2/a/e1-rotation-in-flow/ README + verdict).
Deliberately dropped: edge.rs's stale s_flipzero doc comment
(v=80/anchor-95 parameterization) — it contradicts the shipped
code and frozen frames.json (v=60/anchor-75); wrong, not essence.
Part of landing #5; tracking #9.
The anchor_lab -> n0_model / anchor_engine -> n0 renames shortened import paths past rustfmt wrap thresholds. No semantic change. Part of landing #5.
Opt n0 and n0-model into [lints] workspace = true — the workspace table's too_many_arguments allow covers all 9 findings (the documented house pattern: geometry/cache passes many scalar params). Fence e5scan's ASCII decision table as a text block so its '+' rows stop parsing as markdown list items. n0_dev's opt-in is deferred (manifest comment): unreachable_pub flags 46 internal pub items in the binary crate; demoting them to pub(crate) is follow-up hygiene, not landing scope. cargo clippy --no-deps --workspace --exclude grida-canvas-wasm -- -D warnings and cargo fmt --all --check both green. Part of landing #5.
model-v2/README.md and a/README.md become the archive's navigational surfaces: archive banner with the promotion map (a/lab -> crates/ n0-model, engine/ -> crates/n0, a/spike-canvas -> crates/n0_dev), the declared convention that frozen papers' relative paths refer to the pre-promotion layout, the stale gridaco/grida#957 tracking note fixed to #9, run commands updated to workspace form, and the bare "PR #437" reference fixed to gridaco/grida#437 per cross-repo link discipline. Both RFDs' v2 evidence citations flip from cited-by-branch-in-prose to real in-tree links, as their own footnotes promised at landing: paint-model.md -> crates/n0-model/src/model.rs, display-list-contract.md -> crates/n0/src/drawlist.rs (the branch-links footnotes retire). AGENTS.md and README.md: crates/n0 is no longer "reserved" — the project-structure table and workspace list gain the n0 family and the model-v2 archive rows; check/test commands extended with the new crates. Part of landing #5; tracking #9.
… code) All four shots (default, crosszero, rot45, ungroup) diffed against their goldens by 11-50 bytes out of 4.9M pixels with max channel delta 1 — sub-LSB antialiasing drift on edge pixels. Etiology, established by A/B experiment: the PRE-promotion standalone build (worktree at the merge commit 80c60ad, original manifests and standalone Cargo.locks) produces byte-identical output to the promoted workspace build in today's environment — all four shot SHA1s equal across the two builds. The promotion is therefore pixel-exact; the goldens were stale against this machine's current rendering stack (recorded in an earlier environment). A stale always-red byte gate catches nothing; re-blessed from the verified-identical output. The bench budgets (rig/baselines.json) are deliberately NOT re-blessed: the same A/B showed today's shell measures both builds uniformly ~1.7-2.5x over the recorded baselines (the old build even slower than the new), i.e. environmental timing variance, not a promotion regression. Loosening budgets from a slow context would blind the gate. CI-hosted gate baselines remain the filed follow-up. Part of landing #5.
softmarshmallow
changed the title
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Important Review skippedToo many files! This PR contains 244 files, which is 144 over the limit of 100. To get a review, narrow the scope: Upgrade to a paid plan to raise the limit. ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: ⛔ Files ignored due to path filters (27)
📒 Files selected for processing (244)
You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
CI's first-ever linux run of the v2 suites exposed a pre-existing
platform sensitivity: paint::read_pixels requested native-N32
(RGBA on Apple Silicon, BGRA on x86 linux), so byte-order assertions
in animation_profile5_paints failed on ubuntu. Pin the readback to
explicit RGBA8888 premul — the same rule tests/grida_xml.rs already
applies ("so the assertions do not depend on native N32"). Byte
output on Apple Silicon is unchanged: n0 suites and the gate's four
shot goldens re-verified IDENTICAL after the fix.
oxfmt (node.yml fmt:check): the frozen model-v2/ archive joins the
formatter's ignore list (kept verbatim per its tombstone); the live
crates/n0 docs are formatted.
Part of landing #5.
softmarshmallow
marked this pull request as ready for review
July 19, 2026 09:45
The experiment-ledger codes (E1/E3/E4/E5/E10) were plan-sequence names, not names of things; live artifacts now carry names per their end-state role: - bins: e4 -> resolve_bench, e5scan -> svg_transform_scan - the "historical E3 TextIr" dialect -> "the historical TextIr dialect" (doc comments, error strings, and the test asserting on them, in lockstep) - window/HUD titles and bin headers drop "E10"/"E4" for descriptive names; crate docs describe the experiments instead of citing their ledger ids (the ledger id survives only as a parenthetical archive citation where provenance matters) Deliberately KEPT: registry citations (R-E1, G-E2, L-E1, D-E1, E-A*, DEC-*, V-*, ENG-*) — these cite the archived conformance edge registries, spec-delta register, and decision/contract catalogues; scrubbing them would sever rule-to-registry traceability. SVG animation Profile 0-6 identifiers are spec profile ids, not plan codes. Part of landing #5.
…n-slots
Renames (63 tracked renames, content byte-identical):
- models/{a,b,c,d}.md -> models/{anchor,sheet,bake,wire}.md — the
files now carry the working identifiers their own README declared
- model-v2/a/ -> model-v2/anchor/ — the winner's workbench named for
what it is
- experiment dirs drop their eN- ledger prefixes: rotation-in-flow/,
format-roundtrip/, text-ir/, resolver-spike/, svg-corpus/,
pluggable-layout/, shape-points-scale/ (dec0-* keeps its decision-
register id)
- crates/n0_dev/SPIKE.md -> model-v2/anchor/spike-readout.md — a
frozen session read-out belongs in the archive, not a live crate
(TEXTBOOK.md stays: living migration doctrine)
Every repo reference rewritten (live crates, archive papers, demo
pages); the E1-E10 ids remain the experiment ledger's register keys in
prose and in anchor/README.md's table, which now also carries the
old-name -> new-name map. The archive convention note is amended:
content verbatim, names canonicalized at landing, frozen papers'
relative paths still mean the pre-promotion layout. One frozen
historical command quote in proof.html deliberately keeps the old bin
name (it records what was run).
Part of landing #5.
The XML source format is v2-native, and v2 is n0 — the format now carries the engine's name (owner directive): - modules/APIs: grida_xml -> n0_xml, grida_xml_source -> n0_xml_source, Dialect::GridaXml -> N0Xml; bin grida_xml_render -> n0_xml_render; all n0/n0-model test files renamed in kind - file extension: .grida.xml -> .n0.xml (rig examples + fixtures renamed; every reference updated) - docs/wg RFD family: format/grida-xml*.md -> format/n0-xml*.md with repo-wide link updates; the main RFD gains a "Formerly" continuity note; frontmatter keeps "grida xml" as a search keyword - the term "Grida XML" -> "n0 XML" repo-wide, including the archive (the archive convention note now records the term rename as one of the two canonicalizations applied at landing) Deliberately unchanged: the Draft-0 wire grammar — the <grida version="0"> root element is NOT renamed by this commit; a grammar rename is a spec decision, flagged as an open item in the RFD rather than smuggled through a term sweep. Legacy crates: zero references, zero touch. Part of landing #5; tracking #9.
…nvention Per the fixtures doctrine (deterministic test inputs live under fixtures/<domain>/), the engine's rig corpus moves out of the crate: - crates/n0/rig/corpus/*.replay -> fixtures/n0-replay/ - crates/n0/rig/examples/*.n0.xml -> fixtures/test-n0-xml/examples/ (incl. social-feed/ with its webp assets) - crates/n0/rig/fixtures/*.n0.xml -> fixtures/test-n0-xml/cases/ (incl. component-program/, slot-program/) - all 14 SVG animation profile sources (showcases + boundaries) -> fixtures/test-svg/animation/ — joining the existing SVG fixture domain, and dissolving the cross-crate include (n0-model tests no longer reach into n0's tree; both crates now include from fixtures/) crates/n0/rig/ keeps only baselines.json + probe-baseline.json — gate state (machine-local budget/probe records), not fixtures. Authored relative resource paths inside the moved .n0.xml documents are rebased to their new location (../../images/...); every code and doc reference updated; the gate re-verified green from the new corpus location (shots IDENTICAL, replays DETERMINISTIC, differential MATCHES). Part of landing #5.
The post-landing program (phases 0-6 with entry/exit gates, first three PRs, the owner decision registry, known unknowns) — synthesized from the three-lens panel + adversarial judge and ratified by the owner's review. Git-persisted so it survives agent sessions (named consolidation.md, not *.plan.md — that suffix is the repo's untracked working-plan convention); the umbrella tracking issue mirrors it. Part of landing #5; tracking #9.
13 tasks
The frozen v2 workbench record moves from the repo root (model-v2/) to archive/model-v2/ — a dedicated home for concluded-program records (owner directive: the workbench should not sit at the repo root on main). archive/README.md declares the convention; future concluded programs land as archive/<program>/. Every live reference updated (crate docs, WG docs, AGENTS/README, issue links follow separately); the oxfmt ignore generalizes from model-v2/ to archive/. The tombstone banner records the relocation; archive-internal relative links are unaffected (the tree moved wholesale). Part of landing #5.
This was referenced Jul 19, 2026
softmarshmallow
added a commit
that referenced
this pull request
Jul 19, 2026
Adopt the complete review-only candidate from GitHub Actions run 29700301719, artifact 8446425713 (zip SHA-256 469e3fbf2a526611f61f1d43220dc966bcc414ba290595dd0e97d6b2f3fed978). The manifest records source ed0f310, ubuntu-24.04 x86_64, Rust 1.92.0, and skia-safe 0.93.1. Captured essence: preserve all four named 1360x900 RGBA states, paired shot/timing ownership, same-host determinism, replay determinism, the ENG-0.2 differential, scene-cache equality, and the generous max(1.5x, +50us) timing rule. The candidate gate passed every portable check before installing and uploading exactly five files. Etiology: the outgoing images came from 7c101b9 on aarch64-macos. Main commit abe0305 then deliberately changed the footer from the plan-code name anchor spike — E10 to n0_dev — anchor spike without updating the shots; that source change accounts for exactly 4,010 pixels in each current macOS render. Comparing current source across macOS and Ubuntu changes 2,627-3,992 pixels per shot (0.215-0.326%); visual diffs are confined to text rasterization, while geometry and colors align. The timing numbers move from incomparable macOS measurements to CI-owned min-of-11 measurements: starter 15.639/1.654us and flat10k 1076.711/269.591us for resolve/build. This is an authority transfer, not a performance claim. Deliberate drop: retire only aarch64-macos baseline authority and the stale E10 footer pixels. Preserve their provenance here and in #5/#44. format/grida.fbs, the published canvas-wasm freeze, probe-baseline.json, renderer behavior, and fixture content remain untouched.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Lands the v2 anchor stack as workspace crates. This PR carries the research branch's full 33-commit history (relocated from gridaco/grida#970 at the engine split) plus the landing commits, and merges
--no-ff— never squash.Tracking: #9 (anchor — the v2 node model). Sequencing per the legacy seam program (#27), which completed first: with the v2 engine in-tree, the extraction rule — a module becomes a workspace crate when its second consumer appears in the workspace — is now armed for every seam M4 reviewed.
The move map
model-v2/a/labcrates/n0-modelanchormodel (node records, ops, resolve, pick, n0_xml, textir)model-v2/enginecrates/n0resolve → drawlist → paint, query, journal/replay, damage — takes the reserved public namemodel-v2/a/spike-canvascrates/n0_devmodel-v2/(rest)Promotion is a
git mv, not a rewrite — the engine's own contract (ENGINE.md S-1, archived). The dependency-direction law survives as cargo boundaries: model (no skia) ← engine (skia inpaint.rsonly) ← host.Nobody absorbs anybody at landing. This PR is topology-only. Absorption runs afterward in the promotion program: n0 consumes crates extracted from
crates/grida(cg vocabulary first — M4 ruled it extraction-ready); grida is decomposed, not absorbed, and remains the executable conformance bar until the v1-vs-n0 scoreboard says otherwise. The freeze contract (@grida/canvas-wasm@0.91.0-canary.22) is untouched.Landing commits (on top of the research history)
git mvof the three trees (the reservedcrates/n0stub retires).anchor-*→n0family), manifests rewritten against a patrol ledger (every caveat comment kept/rewritten, named-obsolete, or re-homed; nothing silently dropped), standalone locks/.cargoshims removed, include paths fixed, 15 forbiddengrida.co/docs/wglinks converted to same-repo relative.e1/e3/edge/forkretired; their unrecorded essence (e3's group-AABB exclusion, edge's flipzero stand-in provenance, fork'sescape/overlapmetric definitions) was captured into the archive docs before deletion.n0_dev's opt-in deferred (46unreachable_pubin a binary crate = follow-up hygiene).n0.e4→resolve_bench,e5scan→svg_transform_scan; "historical E3 TextIr" → "the historical TextIr dialect"; window/HUD titles and crate docs describe things instead of citing ledger ids. The archive canonicalizes too:models/{a,b,c,d}.md→{anchor,sheet,bake,wire}.md,model-v2/a/→model-v2/anchor/, experiment dirs dropeN-prefixes, the spike's frozen read-out moves to the archive. Registry citations (R-E1,E-A*,DEC-*,V-*,ENG-*) are deliberately kept — they cite the archived conformance/decision registers; the E1–E10 ids remain the ledger's register keys, with the old→new map inmodel-v2/anchor/README.md..grida.xml→.n0.xml, modules/APIsgrida_xml→n0_xml, the RFD familydocs/wg/format/grida-xml*.md→n0-xml*.md(with a "Formerly" continuity note), the term swapped repo-wide, and issue anchor — the v2 node model #9's body updated in kind (including its stale pre-splitgridaco/gridalinks → landed in-tree paths). The Draft-0 wire grammar — the<grida version="0">root element — is deliberately NOT renamed; that's flagged as an open RFD item, not smuggled through a sweep. Legacy crates: zero references, zero touch.rig/corpus→fixtures/n0-replay/,rig/examples+rig/fixtures→fixtures/test-n0-xml/{examples,cases}/, and all 14 SVG animation profile sources →fixtures/test-svg/animation/— dissolving the cross-crate include (n0-model tests no longer reach into n0's tree).crates/n0/rig/keeps only the machine-local gate baselines. Authored resource paths inside the moved documents rebased; the gate re-verified green from the new corpus.Gates (all run at the head commit)
cargo test --workspace --exclude grida-canvas-wasm— green (legacy + v2 suites).cargo clippy --no-deps --workspace --exclude grida-canvas-wasm -- -D warningsandcargo fmt --all --check— green (the exact CI commands).git diff main...HEADovercrates/grida*,math2,csscascade,fonts,grida-canvas-wasm,format/,packages/is empty; the rootCargo.lockdiff is additions-only (no existing dep version moved).L0.exact— 0 byte diffs, 65/65.just build(the wasm package) succeeds off the shared lock.Deferred (filed as follow-ups)
WG graduation of the anchor spec (normative rewrite of
models/anchor.md); the extraction cascade (cg first);n0_xmlextraction at its second consumer;math2vsn0-model/mathconvergence; fonts-backed text oracle; the v1-vs-n0 reftest scoreboard; CI-hosted gate with CI-blessed baselines; ENGINE.md contract graduation;n0_devpub-hygiene + workspace-lints opt-in; edition-2024 bump.