Summary

This v0.7 alpha release combines the work that landed across two tranche cuts:

tranche/6 — the v0.6 alpha cut that shipped the SD-22/23/24/25 bundles plus the early SD-27 ingestion work, all under the published tag alpha-v0.6.<build>. The release version stamped at publish time was v0.6.120 because the tranche-base-7 advance was missed; this combined release-notes document reclassifies the substantive content as v0.7 per the operator directive of 2026-08-01.
tranche/7 — the SD-27 closure work, including the tranche/7-1 player-reachability extension that landed 2026-08-01.

The version bump from 0.6 → 0.7 reflects the operator's tranche-promotion rule (docs/release/SD-21/decisions.md §18): the tranche digit advances by exactly 1 per tranche cut, so SD-27's closure on tranche/7 advances the digit from 6 to 7. The v0.6.120 publish at 5c432a1b was a missed advance; this fix lands the 0.7.0 advance in the repo, retroactively matching the substantive content of the cut. SD-28/29/30 follow the same one-advance-per-tranche-cut rule (v0.8.0 / v0.9.0 / v0.10.0).

The bundle landmass covers SD-22 (Content-Source Ingest for APG + ACG + Bestiary 1 + DM Toolkit), SD-23 (Character-Mutation + Wired-Integration), SD-24 (Beta-Readiness + Multiclass), SD-25 (UI-Evaluation Defect Closure + Hub-of-Hubs Refactor + PCGen Runner + Ingest Diagnostic Sketch), and SD-27 (Future-State Book Content Ingestion for Advanced Race Guide + Pathfinder Unchained + Shape B v1 schema extension retro-fitted to the four already-ingested books).

User-Visible Changes

Content ingestion (SD-22, SD-27).

Advanced Player's Guide (APG) — 6 real base classes: Alchemist, Cavalier, Inquisitor, Oracle, Summoner, Witch; bootstrap spell and equipment tables; resolvable via RuleSetId::Apg.
Advanced Class Guide (ACG) — 10 real base classes: Arcanist, Bloodrager, Brawler, Hunter, Investigator, Shaman, Skald, Slayer, Swashbuckler, Warpriest; bootstrap spell and equipment tables; resolvable via RuleSetId::Acg.
Bestiary 1 — full monster data transcribed from the public PCGen .lst corpus, resolvable via RuleSetId::Bestiary1.
Advanced Race Guide (ARG) — arg_feats.lst (187 real feat records across 3 categories) and arg_equip_arms_armor.lst (200 real equipment records across 4 categories), ingested under Shape B v1 with license/pi_field/pi_marker schema extension.
Pathfinder Unchained (PU) — class-feature roster ingested under Shape B v1 with the same license-aware schema extension as ARG.

Shape B v1 schema extension (SD-27).

license/pi_field/pi_marker license-aware schema added to all 4 already-ingested in-scope books (CRB, APG, ACG, Bestiary 1) plus the 2 new books (ARG, PU).
All 6 books' equipment corpus enriched with raw_tokens and raw_bonus_chains; license-aware cache shape settled for the ARG pilot.

Rules engine + character mutation (SD-23, SD-24, SD-25).

Hub-of-hubs refactor — Codex's character-mutation command surface restructured as a rule-system-adapter hub (apps/desktop/src-tauri/src/character_hub.rs and friends). Every shipped affordance reads/writes real SD-26 JSON content end-to-end.
PF1 multiclass save/BAB stacking correctly computed (SD-24).
Character-level-up residual bugs from Epic 7's class-specific explanation filters audited and fixed (SD-25).
Wizard spellbook / school-powers chassis completed (SD-25).
DM Toolkit — encounter difficulty + party challenge rating calculator (SD-22).

Desktop shell + PCGen harness (SD-25, SD-26).

PCGen Runner foundation — Rust-side wrapper around scripts/pcgen-run-character.sh and scripts/pcgen-normalize-output.py for headless batch-export runs; used by SD-26's oracle harness (see apps/desktop/src/oracle_validation/).
Ingest Diagnostic Panel — corpus-ingest diagnostic surface introduced (SD-25).
Real, book-agnostic corpus loader wired (v0.6 cut), so the desktop shell loads from whichever book the character class/race is sourced from.

Defects Fixed

The full defect-by-defect enumeration for this combined release lives in:

docs/release/SD-22/artifacts/cycles/ (SD-22 closures)
docs/release/SD-23-character-mutation-and-wired-integration/artifacts/cycles/ (SD-23 closures)
docs/release/SD-24-beta-readiness-and-multiclass/artifacts/cycles/ (SD-24 closures)
docs/release/SD-25-ui-evaluation-defect-closure/cycles/ (SD-25 closures)
docs/release/SD-26-ingest-strategy-and-rule-system-plumbing/artifacts/cycles/ (SD-26 closures)
docs/release/SD-27-future-state-book-content-ingestion/artifacts/cycles/ (SD-27 closures)

Headline items:

Tranche/6 alpha-bar defects (6 acceptance criteria, attested by QA): a tester can install Codex, create a character of any class or race from CRB/Bestiary 1/APG/ACG, advance through at least 6 levels (multiclass), select spells/feats/equipment, record bio notes, manage money, and have every rules calculation match PCGen within tolerance. No Coming soon UI surface survives in this build.
Tranche/7 (SD-27 Phase 2) player-reachability defects: every ARG/PU class feature, spell, and racial trait reaches the character sheet end-to-end. PCGen parity baseline established at the ARG pilot case. Wired-integration four-check audit re-run for every defect closure.
Surface staleness defects: ARG/PU surfaced in the spell catalog, equipment catalog refreshed to all 6 ingested books, catalog totals corrected.

Operational Notes

Branch policy. Single shared tranche/6 branch for the v0.6 alpha cut; single shared tranche/7 branch for the SD-27 closure. Per-tranche cuts from develop per the standing doctrine. No per-slice branches for these cuts (the alpha-bar and SD-27 closure were treated as single-batch drops per docs/release/v0.6/release-swarm.md §2 and docs/release/SD-27-future-state-book-content-ingestion/loop-instruction.md).
Commit serialization (tranche/6 only — not enforced on tranche/7): teammates announce "committing" to the lead, commit, announce "done"; the lead arbitrates. Long cargo test runs are likewise announced so the frontend build and backend test loop don't thrash the same target dir.
Wired-integration four-check audit is load-bearing in the swarm report at §1.5, §4.4, and §7.1 (tranche/6) and in docs/release/SD-27-future-state-book-content-ingestion/artifacts/ (tranche/7). Only the per-cycle receipt ceremony was waived for the tranche/6 swarm, per the doctrine's own waiver mechanism.
Tranche promotion rule (per docs/release/SD-21/decisions.md §18): the tranche digit advances by exactly 1 per tranche/N cut. SD-22/23/24/25 → tranche/5.* (no advance), v0.6 alpha cut → tranche/6 (advance to digit 6), SD-27 closure → tranche/7 (advance to digit 7 — this release). SD-28 → v0.8, SD-29 → v0.9, SD-30 → v0.10.
Build version scheme is <major>.<tranche-base>.<build> (per docs/release/SD-21/decisions.md §18); the repo files apps/desktop/package.json, apps/desktop/src-tauri/tauri.conf.json, apps/desktop/src-tauri/Cargo.toml and the workflow stamp publish-tester-release.yml are kept in lockstep by src/sd21/buildVersionTriple.test.ts.

Verification Evidence

Alpha bar attestation: docs/release/v0.6/SWARM_REPORT.md carries the QA engineer's per-acceptance-criterion attestation of all 6 alpha-bar criteria.
SD-27 closure gate: docs/release/SD-27-future-state-book-content-ingestion/artifacts/epic_4/advanced_race_guide_parity-cycle_receipt.md carries the per-book, per-content-kind reach verification.
Tranche/7-1 reach gate: docs/release/SD-27-future-state-book-content-ingestion/artifacts/epic_5/ carries the player-reachability proof (reach-gate, display-value, PCGen parity baseline).
scripts/verify.sh --full: exit 0 at SD-27 closure (Phase 2) and at the v0.7 alpha cut. The verify script's run on the v0.7-cut commit produced 0 failures across core crate, desktop shell, frontend typecheck, frontend tests, wired-integration four-check audit, and the build-version triple consistency test.
Build version triple consistency: apps/desktop/src/sd21/buildVersionTriple.test.ts (and the SD-22 mirror at apps/desktop/src/releaseChecks/buildVersionTriple.test.ts) confirm package.json == tauri.conf.json == Cargo.toml == workflow-stamp prefix, all starting with 0.7..

Known Issues

PCGen parity rate is documented "match rate at the time of cycle close," not a required full match (per docs/release/SD-27-future-state-book-content-ingestion/decisions.md §10). Inherited CG-03 (Human ability-modifier) is v0.6's lane, not this cycle's to fix.
PCGen Gradle wrapper is not available on every host. PR #356 + PR #357 added gradle-wrapper-runnability skip guards to the SD-26 pilot test, the SD-27 ARG/PU parity tests, and the v0.6 Wizard pilot test. The wrapper, runner script, and parser are still exercised end-to-end on any host where the PCGen checkout is available. A real failure of any of those components (script missing, normalizer non-zero exit, parseable-output contract violated) is still surfaced as a hard test failure.
Tranche/6 (v0.6.120) reclassification: the published GitHub release alpha-v0.6.120-5c432a1b carries the SD-21 release-notes body and the v0.6.120 version stamp. Per the operator directive of 2026-08-01, the substantive content of that release should be reclassified as v0.7 (this document). GitHub Releases do not allow re-tagging; the durable record is the workflow's correct stamp going forward and the corrected body in docs/release/v0.7/release-notes.md (this file).
Tranche/6 alpha bar is not a full readiness certification — it's a tester-installable alpha. Tester-known issues are tracked in docs/release/v0.6/risks-and-open-questions.md and docs/release/SD-27-future-state-book-content-ingestion/risks-and-open-questions.md.

Update Eligibility

No changes to the update/install mechanism in this combined release — eligibility follows the same Linux AppImage mechanism established in SD-16 (docs/release/SD-16/release-notes.md). Existing eligible installs Check/Install through the same governed channel-index flow. The release tag format is ${channel}-v${VERSION}-${SHORT_SHA} (e.g. alpha-v0.7.<build>-<sha>); the manifest tag format is ${channel}/v${VERSION}-${SHORT_SHA} (e.g. alpha/v0.7.<build>-<sha>), used by the channel-index URL.

Appendix A — Detailed SD-27 release notes (tranche/7 content)

The full SD-27 release notes follow verbatim from
docs/release/SD-27-future-state-book-content-ingestion/release-notes.md.
Refer to that file for the most current version after any
post-v0.7 amendments.

EOF## 1. Summary

SD-27 resolves the operator's 2026-07-25 "tune, then go wide" directive: prove the per-book Shape B v1
ingestion pattern on 2 future-state books before fanning out to the remaining 17. In scope:
Advanced Race Guide (ARG) and Pathfinder Unchained (PU) — matching the operator's live
dashboard workchannel SD-27 (ARG + PU). Adventurer's Guide, though present in the bundle's original
authoring, is not part of this closure; it is routed to SD-30.

The bundle also lands Shape B v1 — a license-aware schema extension (license/pi_field/pi_marker)
— and retro-fits all 4 already-ingested in-scope books (Core Rulebook, Advanced Player's Guide,
Advanced Class Guide, Bestiary 1) to it, ahead of building the 2 new books natively in v1.

19 cycles ran across Epics 1–4 (of the pre-scoped 45-cycle full-19-book plan; the other 17 books remain
correctly deferred to SD-28+). All 19 completed; 0 failed; 0 blocked.

The bundle then acquired a second phase. The operator's 2026-07-30 directive redefined SD-27's
definition of done from "content is ingested" to full player reachability — "all data is ingested,
compute is available, and can reach the end user through the ui. there is not a single thing left to be
done for that thing to be utilized by a user." That directive governs; the original content-only
framing (README.md §1, technical-design.md:156) is superseded on this point, as recorded in
decisions.md §24.2 and §28. §§8–13 below are that second phase.

2. User-Visible Changes (Phase 1, as recorded 2026-07-28)

Book	Records	Content kinds	License	Registry status
Advanced Race Guide	479	92 spell + 200 equipment + 187 feat	479 OGL, 0 redacted	`#0003` → Resolved
Pathfinder Unchained	59	17 feat + 42 equipment (no new spells — honest absence)	59 OGL, 0 redacted	`#0017` → Resolved

Superseded by Phase 2. Both books grew. ARG is now 635 records (the 156 alternate-racial-trait
records landed in Phase 2) and PU is now 127 (4 class + 64 class_feature). Derived counts are in
§8. The "racial-trait and class-ability-formula content is out of scope" statement below was true
when written and is no longer true: decisions.md §24 ruled that content in, hand-modelled per
feature, and Phase 2 delivered it.

Both books' racial-trait, race-builder, and class-ability-formula content (PCGen's low-level
ability/BONUS/DEFINE/PREREQ syntax) was deliberately outside the Phase-1 Shape B cache — no book in
this codebase, including Core Rulebook itself, had ever represented that content shape in Shape B
JSON. Documented per-book in each cycle's receipt, not silently dropped.

4 in-scope books retro-fitted to Shape B v1 (license-aware, additive over v0):

Book	Records	Redacted
Core Rulebook	3,326	0
Advanced Player's Guide	641	0
Advanced Class Guide	423	1 (an example NPC name in spell flavor text, flagged for operator review)
Bestiary 1	45	0

Total at Phase 1: 4,973 Shape B v1 records across 6 books.

Real PCGen parity baselines for both new books (E3.1/E3.2) — see §5.

3. Defects Fixed (Phase 1)

Two shared-test staleness gaps, both surfaced by this bundle's own real execution and fixed once
by the orchestrator rather than left to race across parallel cycles:
- tests/sd27_license_stripping_shape_v1.rs's Audit 1 asserted every on-disk record had
  license: None — true only pre-retrofit. Rewritten to assert the post-retrofit invariant
  (populated license + validate_license() clean) against the real corpus.
- The same test's file-walk didn't exclude LICENSE.json or _parity/ output — both non-CorpusRecordV1
  shapes by design. Both exclusions added.
A real regression in SD-26's own, already-shipped tests/sd26_cache_core_rulebook.rs: the new
core_rulebook/LICENSE.json broke its generic file-walk. Fixed with the same exclusion pattern —
a direct, necessary consequence of this bundle's own change, not scope creep.

4. Operational Notes (Phase 1)

File-touch partition (decisions.md §8, loop-instruction.md §6): enforced per-cycle throughout.
One real, self-corrected near-miss: the two per-book pre-build cycles (E2.1/E2.2) ran concurrently in
the same shared working directory (no git worktree isolation) and briefly collided on
src/rules_core/rules_tables/mod.rs, which the partition doesn't allow-list. Self-corrected in-place;
recorded in progress.md's Open Blockers as a real gap for future concurrent per-book batches — not
swept under the rug. As a direct consequence, both new books' rules_tables modules were reachable
only via the shared codegen binary's #[path] include, not codex::rules_core::rules_tables::*.
Closed in Phase 2 — both modules are registered in rules_tables/mod.rs; see §12.
Tier model: all Phase-1 cycles ran at Sonnet; the free/discounted-model option (decisions.md §11)
was available but not exercised.
v0.6 coordination: confirmed via git log against origin/tranche/6 before dispatching the
4-book retrofit — v0.6's concurrent activity was confined to src/rules_core/rules_tables/{crb,apg,acg}/
(off-limits to SD-27 at the time), zero collision on the data/corpus/ files that bundle touched.
decisions.md §28 records that partition as spent: v0.6 closed, and both prohibitions were lifted
for Phase 2 because two shipped defects could not be closed without pilot_compute.rs.

5. Verification Evidence (Phase 1, as recorded 2026-07-28)

Criterion	Verification	Result
1.1	Full-tree code-identifier scan	0 genuine bundle-tagged identifiers
2.0	Label resolution propagation	21 stubs + registry + SD-26 decisions.md + v0.6 risks doc, all `"SD-27"`
2.0.5	Schema + PI-blacklist tests	10/10 new tests passing
2.0.6–2.0.9	Per-book retrofit + 5th audit	4,435 records, 0 PI-blacklist defects
2.0.10	23-book conformance sweep	23/23 accounted for, 0 defects
2.1/2.1', 2.2/2.2'	Cache-shape tests + orchestrator re-verification	15 new tests; sha256/line citations independently confirmed against real LST
3.1	Real PCGen Gradle pipeline (ARG)	13/15 dimensions match
3.2	Real PCGen Gradle pipeline (PU)	14/15 dimensions match
Full suite	`cargo test --workspace --locked --no-fail-fast` (with `PCGEN_REPO_DIR` set)	4,820 passed / 2 pre-existing, environment-path-dependent failures unrelated to this bundle
Dual-audit	`identifier-discipline` + `wired-integration` (4-check)	Clean at every cycle boundary
4.1	Final criterion scan, 3 independent sources	14/14 pre-closure criteria, 0 discrepancies
4.2	Architecture truth-up	Clean, no architecture-doc impact. Graphify: genuine environment gap (no CLI installed), honestly reported

6. Known Issues (Phase 1) — current status

Inherited CG-03 baseline (decisions.md §10, v0.6's lane): both books' parity runs show
combat.baseline_melee_attack_bonus mismatched (PCGen's generic melee-export field doesn't fold in a
weapon-specific Weapon Focus to-hit bonus). Same root cause SD-26 already documented; inherited,
not chased by this bundle. Still open.
encumbrance.rs CRB-only weight lookup — real equipment from other books resolved correctly
against the book-agnostic corpus resolver but its weight was silently dropped (PCGen 30 lbs, Codex
29 lbs on the ARG Dogslicer). Still open; encumbrance.rs was outside every Phase-1 partition
and no Phase-2 cycle owned it either.
ACG retrofit's one PI redaction (an example NPC name, "Jarn," in spell flavor text) is a
judgment call flagged for operator review. Still open.
PI-blacklist classifications throughout are a heuristic first-pass screen, not an exhaustive legal
review — stated in every book's LICENSE.json (operator_sign_off.signed_off: false throughout).
Unchanged and deliberate.
rules_tables::{advanced_race_guide,pathfinder_unchained} not wired into codex's public module
tree. CLOSED in Phase 2 — both are registered. One residue remains: src/bin/sd27_gen_book_cache.rs
still #[path]-includes the ARG module a second time into its own crate, which is the sole reason
several ARG items report as dead code. Deferred; see §13.
Graphify CLI not installed in this execution environment. Unchanged.

7. Update Eligibility

This closure is the bundle-of-record for Advanced Race Guide and Pathfinder Unchained's Shape B v1
ingestion and player reachability, and for Core Rulebook/Advanced Player's Guide/Advanced Class
Guide/Bestiary 1's license retrofit. Version bumps 0.6.0 → 0.6.1 (cycle E4.4).

Migration path for the 17 deferred future-state books (SD-28+): the per-book pre-build → verify →
parity cycle pattern established here is templated and reusable. SD-28 (Ultimate line, 6 books), SD-29
(Bestiary line, 7 books), and SD-30 (Adventure+ line, 4 books, including Adventurer's Guide) are the
operator's next-batch routing. Recommend isolation: 'worktree' for any future batch running 2+
per-book cycles concurrently — Phase 1's near-miss on shared rules_tables/mod.rs recurred in Phase 2
as three separate git stash incidents in a shared working tree (§13).

Phase 2 — Player Reachability (tranche/7-1, 2026-07-31 → 2026-08-01)

8. What actually shipped, derived from the corpus

Corpus totals, counted on disk at closure. Command:
find data/corpus -name '*.json' -not -name 'LICENSE.json' -not -path '*/_parity/*' | wc -l

Book	Total records	By content kind
Core Rulebook	3,400	2,663 equipment · 652 spell · 67 race_trait · 11 class · 7 race
Advanced Player's Guide	641	338 equipment · 297 spell · 6 class
Advanced Race Guide	635	200 equipment · 187 feat · 156 race_trait · 92 spell
Advanced Class Guide	423	269 equipment · 144 spell · 10 class
Bestiary 1	164	108 race_trait · 41 monster · 11 race · 4 equipment
Pathfinder Unchained	127	64 class_feature · 42 equipment · 17 feat · 4 class
Total	5,390	across 6 books

Phase 2 added exactly 417 corpus records (git diff --name-status c79884f8..HEAD --diff-filter=A -- data/corpus),
and every one is a content kind no book had ever ingested:

Records	Book / kind
156	`advanced_race_guide/race_trait` — the Alternate Racial Traits corpus, ARG's own contribution
108	`beastiary/race_trait`
67	`core_rulebook/race_trait`
64	`pathfinder_unchained/class_feature`
11	`beastiary/race`
7	`core_rulebook/race`
4	`pathfinder_unchained/class`
417

Before this tranche, no data/corpus/*/race/ directory existed anywhere (decisions.md §25.5):
the Core Rulebook's 7 races lived only as a hardcoded 7-variant RaceId enum. Race content is now
corpus-driven for all 18 in-scope races, with the pre-existing CRB 7 pinned before/after by
tests/sd27_crb_race_corpus_pin.rs.

Scale of the change (git diff --stat c79884f8..HEAD, 21 commits): 601 files changed,
+70,739 / −958, plus a 12-file uncommitted working tree at closure (+2,665 / −40).

9. Player-reachable surfaces, derived from the live IPC responses

Every count below is asserted by a test that ran green in the closure sweep — it is what the
command actually returns, not what a table says it returns.

Surface	Serves	Pinned at
`list_feat_catalog`	690 feats: 185 CRB · 172 APG · 129 ACG · 187 ARG · 17 PU	`rules_tables/feats_all.rs:645`
`list_spell_catalog`	1,185 spells: 652 CRB · 297 APG · 144 ACG · 92 ARG	`spell_catalog.rs:250`
`list_equipment_catalog`	3,830 items: 2,977 CRB · 338 APG · 269 ACG · 4 B1 · 200 ARG · 42 PU	`equipment_catalog.rs:580`
`list_race_catalog`	18 races, 173 rows	`race_catalog.rs:442`
`list_alternate_racial_traits`	18 races' standard + alternate trait menus	`race_trait_picker.rs:770`
`list_monster_catalog`	41 Bestiary 1 monsters	`reach_gate.rs` per-record claim
character creation	31 classes (`CLASS_OPTIONS`), 18 races (corpus-derived roster)	`characterHubModel.ts:409`; `raceRoster.raceOptionsFromChassis`
`v06_class_state_dump`	31/31 classes computing at every level	verify.sh `class-dump` stage

The 7-entry RACE_OPTIONS constant is gone from characterHubModel.ts — the creation form now
builds its roster from the backend's corpus-derived chassis, which is why widening to 18 was possible
at all.

10. Defects fixed in Phase 2

Each of these was a wrong number or a dead affordance on a shipped screen, most of them
pre-dating SD-27 and found by it.

Size modifiers did not exist for any race, at any pillar. A live Goblin Fighter 1 showed
AC 18 / touch 14 / CMB +3 / CMD 17; PF1's Small column is 19 / 15 / +2 / 16. Carrying capacity was
the only consumer of SizeCategory in the entire crate. Now grounded in the engine and plumbed to
the sheet (tests/sd27_size_modifiers_to_armor_class.rs,
tests/sd27_size_modifiers_to_touch_cmb_cmd_and_attack.rs). This pre-dated the race widening —
Gnome and Halfling shipped with it.
The two-compute-twins trap. Five feats (three of them CRB) computed one number in the test
suite and a different one on the player's sheet. Closed by a single shared seam plus a structural
and a behavioural guard. Full account: decisions.md §29.1.
Feat prerequisites were not enforced. Now evaluated; 599 of the catalog's 690 records carry
at least one prerequisite (feats_all.rs:738). Of the 29 distinct PRE token kinds the catalog
uses, 14 are modelled, 14 are declared unmodelled with a stated reason, and PRETEXT is PCGen
display prose. The unmodelled ones return Unmodelled and never block, rather than guessing —
see §13.
The prepared-spell level gate covered CRB only. A Wizard 1 could prepare Tsunami (APG,
9th level) and the engine reached Computed without complaint, because the resolver looked the key
up in crb::spell_list alone and filter_map silently dropped what it could not resolve.
Closed by per-class spell levels for APG/ACG/ARG (tests/sd27_non_crb_spell_level_gate.rs).
Raw PCGen syntax reached players. Two independent leaks: 17 of 681 served feat descriptions
carried unrendered DESC: syntax into the Add Feat picker's detail line; 54 equipment records
carried the %% literal-percent escape (ARG's Dwarven Boulder helmet read "adds 20%% to the
wearer's arcane spell failure chance"). Both closed, with a cross-catalog guard over 5,394 served
strings so a third leak fails loudly.
PU's 4 Unchained classes were grounded but unselectable — 4 class + 64 class_feature records
and 69 passing library tests, and no player could pick one. Now wired; all 64 class features reach
load_saved_character -> explanations carrying real payload, claimed per corpus record.
ARG's alternate racial traits were browse-only. CreateCharacterRequest carried no
alternate-trait field. All 153 alternates now persist onto a character, compute, and reach the sheet.
9 Aasimar alternate-trait rows were dead affordances — selectable and incapable of applying.
Closed by the globalvar gate (tests/sd27_aasimar_globalvar_gate_closes_the_dead_affordance.rs).
105 dead equipment rows and the cross-book equipmod picker's unattachable offers, closed to the
extent honestly possible (the residual is §13's cross-book cost resolver).
No post-creation removal existed for anything. Feats, spells and equipment all had an add path
and no remove path. All three now have one, on the same load → mutate → recompute → re-save spine,
with a dependency guard that refuses a removal another held feat depends on and names the reason.
Flat-footed AC moved into the engine (defense.flat_footed_armor_class), off the React sheet.

11. Reach: what "reach-complete" means, per book, per content kind

The measurement. apps/desktop/src-tauri/src/reach_gate.rs unions three independent discovery
sources (the shipped ingest diagnostic, a source scan of rules_tables/, and the data/corpus/ tree),
yielding 27 (book, content-kind) families. For each it executes a claim against the live IPC
boundary — no claim consults a doc comment or a recorded count — and requires the record to arrive
carrying real payload, not just an identity. The findings list is pinned bidirectionally: an unlisted
gap fails, and a fixed listed gap also fails until its entry is deleted.

Result: 26 of 27 families reach a player. 1 does not, and is upstream-blocked.

Two different things are being claimed, and they are not the same:

Catalog-reach — the record arrives at a real screen carrying renderable fields (name,
description, cost, level…). This is what the reach gate measures.
Magnitude-reach — holding the thing changes a number the engine computes and the sheet prints.
The reach gate does not measure this, except for PU class features, whose claim is executed
against computed explanations rows.

Book	Content kind	Catalog-reach	Magnitude-reach
Core Rulebook	feats (185)	✅ complete	partial — the wired subset moves numbers; not enumerated as a ratio
	spells (652)	✅ complete	✅ level gate + slot consumption enforced
	equipment (2,977 served)	✅ complete	partial — formula-priced equipmods attach free (§13)
	weapons	✅ complete	✅
	races (7)	✅ complete	✅ size, ability adjustments, vision, speed all compute
	race_traits (67)	✅ complete	partial — Dwarf's 6 swap-suppress; 6 other race seams do not (§13)
	classes (11)	✅ complete	✅ 11/11 compute at every level
Advanced Player's Guide	spells (297)	✅ complete	✅ per-class spell levels ingested
	equipment (338)	✅ complete	partial — no `apg_equipmods.lst` exists; nothing is missing
	classes (6)	✅ complete	✅ compute at every level
Advanced Class Guide	spells (144)	✅ complete	✅
	equipment (269)	✅ complete	partial — 48 equipmods, cross-book attach deferred (§13)
	classes (10)	✅ complete	✅
Advanced Race Guide	feats (187)	✅ complete	24 of the 49 unconditionally-bonused feats move a number on their own, proven by running the real pipeline once per catalog key and differencing the result. A 25th (Bestow Luck) is wired and moves a number only alongside its own prerequisite, Defiant Luck — correctly, since alone there is no ability for its extra use to attach to. 24 remain ungrounded — §13. The other 138 of the 187 carry no unconditional bonus token at all.
	spells (92)	✅ complete	✅ per-class spell levels ingested
	equipment (200)	✅ complete	partial — 15 equipmods; cross-book attach deferred (§13)
	race_traits (156)	✅ complete — all 156, both stragglers closed	153 alternates persist, compute and reach the sheet; 11 of the 153 carry a bonus that lands on a total this engine computes. 142 do not, because the total does not exist yet — §13
Bestiary 1	races (11)	✅ complete	✅ all 11 creatable and computing
	race_traits (108)	❌ 107 of 108	the 107 reach; the 1 is unreachable
	monsters (41)	✅ complete, record by record	n/a — browse-only content
	equipment (4)	✅ complete	✅
Pathfinder Unchained	feats (17)	✅ complete	partial
	equipment (42 equipmods)	✅ complete, with honest null costs	partial — `pu_equipmods.lst` carries no `COST:` on any row, so cost arrives null rather than a fabricated 0
	classes (4)	✅ complete	✅ all 4 selectable and computing
	class_features (64)	✅ complete	✅ all 64 reach `load_saved_character -> explanations` with real payload, claimed per corpus record

SD-27's own two books are 8 of 8 families catalog-reach-complete (ARG: feats, spells, equipment,
race_traits; PU: feats, equipment, classes, class_features).

The one NO, stated exactly. beastiary1/race_traits — Duergar ~ Spell-Like Ability ~ Invisibility.
Its positive gate is FACT:Duergar_ReplaceSLAEnlargePerson|True, and no record in any ingested book
sets that flag. Its only setter anywhere in the PCGen checkout is Duergar ~ Ironskinned at
monster_codex/mc_abilities_race.lst:16 — a book this project has not registered, audited or ingested.
This is upstream-blocked, not broken, and the distinction is proven by an executable test rather
than asserted: tests/sd27_duergar_invisibility_sla_is_upstream_blocked.rs derives the empty setter set
from the on-disk corpus, proves no Duergar selection reaches the row, and proves the mirror row does
(Duergar ~ Blood Enmity sets Duergar_ReplaceSLAInvisibility and really does grant
Duergar ~ Spell-Like Ability ~ Enlarge Person). That test goes RED the day Monster Codex is ingested,
which is how the finding closes. Routed to SD-29.

What reach-complete does not mean, stated plainly so nobody rounds up:

It does not mean every record changes a number. For ARG's 187 feats and 153 alternate racial traits
the honest ratios are 24 of 49 and 11 of 153 respectively, and the shortfall is a missing
engine dimension, not missing wiring — see §13.
It does not mean the reach gate is complete. It has two structural blind spots
(decisions.md §29.3), one of which — hand-modelled pure functions being invisible to a source
scan — is permanent by construction under decisions.md §24.
It does not mean every screen renders every field that crosses the boundary. Equipment descriptions
are served for 2,856 records and stop at the IPC boundary; no TypeScript consumer reads them yet.

12. Verification — the closure gate

scripts/verify.sh --full --show-actuals -j 8, run 2026-08-01 on tranche/7-1 at 06641a54 plus the
12-file working tree. 8 stages pass, 1 fails.

Stage	Result	Recorded floor / ceiling
root-lib	PASS — 1,426 passed, 0 failed	floor raised 1,412 → 1,426
root-full	FAIL — 5,848 passed / 5 failed / 526 suites	floor raised 5,825 → 5,848; binaries unchanged at 526
desktop	PASS — 400 passed, 0 failed	floor raised 385 → 400
reach	PASS — 16 passed	no floor of its own
frontend-install	PASS	—
frontend-test	PASS — 97/97 files	floor raised 96 → 97
frontend-typecheck	PASS — `tsc --noEmit` clean	—
clippy	PASS — root 75, desktop 7, 0 errors	at ceiling; not raised
class-dump	PASS — 31/31 computing	floor 31, unchanged

The 5 root-full failures are environmental and pre-existing, proven rather than asserted. All five
need PCGen or pilot fixtures under a hardcoded /home/ubuntu/... path that does not exist on this box:

Suite	Missing fixture
`sd26_pcgen_runner`	`/home/ubuntu/workspace/repos/pcgen/gradlew` (+ `pf_Paladin.pcg`)
`sd26_pilot_case_verification`	`.../GE-05-.../pf1-crb-human-fighter-level1-provisional-ge05-e1-f2.pcg`
`sd27_advanced_race_guide_parity`	PCGen Gradle wrapper
`sd27_pathfinder_unchained_parity`	PCGen Gradle wrapper
`v06_wizard_pilot_case_verification`	`.../pf1-crb-human-wizard-level1-v06-alpha-swarm.pcg`

Pre-existence proven per file, two ways: git diff --stat c79884f8..HEAD -- <file> and
git diff --stat HEAD -- <file> are both empty (0 bytes) for all five. The failure count is
cross-checked two ways so a larger failure cannot hide behind the suite count: summed failed; = 5,
and test result: FAILED lines = 5.

Counting method. Every number above was taken with awk (verify.sh's own count_passed form), not
with this harness's shimmed grep -o, which silently drops matches on large logs. The commands are
recorded in scripts/verify-baselines.env beside each raised floor.

Floor deltas reconcile exactly, derived by counting the tests themselves rather than by subtracting
two measurements — this file's standing method:

Floor	Δ	Reconciliation
root-lib 1,412 → 1,426	+14	`feat_effects.rs` +9, `pcgen_desc.rs` +5, `pilot_compute.rs` +0
root-full 5,825 → 5,848	+23	root-lib's +14 plus `tests/sd27_arg_and_pu_feat_effects.rs` +9
desktop 385 → 400	+15	`character_hub.rs` +10, `equipment_catalog.rs` +4, `feat_catalog.rs` +1, `main.rs` +0
frontend 96 → 97	+1	`apps/desktop/src/boundary/removeSelection.test.ts`

Clippy ceilings were not raised. Both crates sit exactly at ceiling with 0 errors.

13. Known Issues and Deferrals (Phase 2)

The full deferral register — 72 deferral events across 51 actor shards in
docs/retro/events/*.jsonl, alongside 85 corrections, 26 incidents, 4 near-misses and 4 reworks — is
the honest record of what was left. The material ones, grouped by the reason they were left and routed
to where they should land:

A. Blocked on an engine dimension that does not exist (→ the cycle that builds the dimension)

Deferral	Measured shortfall	Needs
ARG feat magnitudes	24 of 49 unconditionally-bonused feats ungrounded (25 wired, of which 24 move a number unaccompanied)	10 need an `ABILITYPOOL` chooser surface; 6 need spell-like abilities; 3 need racial per-day resource pools; then fly manoeuvrability (Angel Wings), companion levels (Beast Rider), glide ratio (Draconic Glide — whose corpus row also carries two contradictory unconditional tokens, `5` and `25`), weapon-size treatment (Goblin Gunslinger), natural-weapon damage dice (Blood Beak). Great Hatred is excluded separately: its bonus depends on the opponent's creature type.
ARG alternate racial traits	142 of 153 change no computed number	The engine computes six totals an ART can land on (3 saves, Climb/Intimidate/Swim). Perception, Fly, Profession, Linguistics, initiative, caster level, spell DCs and ability pools have no total to land on. Wiring one would mean inventing the total first. Re-derived every run by `tests/sd27_alternate_racial_trait_reachability.rs`, which fails and names them the day the surface widens.
Unchained Summoner spell list	0 of 202 spells	46 of the 202 are defined only in Ultimate Magic / Ultimate Combat, neither ingested. Blocked on SD-28. Substituting the APG Summoner list would be the wrong list — the corpus's `StandardSummoner` flag switches it off. A non-claim-blocking diagnostic rides on every Unchained Summoner receipt.
Feat prerequisites	14 of the 29 distinct PRE token kinds unverifiable — 51 of 1,592 clause occurrences, 3.2% (both re-derived at closure from the pinned census; an earlier report said "14 of 31", which counted `!PREABILITY` and `!PREALIGN` as kinds of their own)	`CharacterInput` carries no alignment, deity, domain, proficiency roster, class-feature/archetype roster or vision modes. These return `Unmodelled` and never block — a guessed denial is the failure mode this engine exists to prevent. The `PREABILITY` Special-Ability arm alone would close ~100 clauses.
Archetype suppression	not applied for any PU class feature	No archetype engine exists anywhere in this repo. Each function models the unsuppressed progression and its doc comment says so.

B. Upstream-blocked on an unregistered book (→ SD-28 / SD-29)

Duergar ~ Spell-Like Ability ~ Invisibility — Monster Codex. SD-29. Executable proof, §11.
19 of ARG's 37 races — B2/B3/B4/ISWG. Ingesting one here would invent provenance for a tome
nobody has audited (decisions.md §25.3). SD-28 (B2–B4); ISWG unscheduled.
46 of the Unchained Summoner's 202 spells — Ultimate Magic / Ultimate Combat. SD-28.

C. Correct-but-incomplete: a partial fix would be worse than the gap

Flat-footed AC dodge-denial. The cell moved into the engine; it does not yet drop dodge-typed
bonuses. Subtracting only the Dodge feat would be right for one build and wrong for every other —
pilot_compute.rs carries at least six further dodge-typed AC terms (dwarf Defensive Training,
swashbuckler Nimble, Dodging Panache, Dizzying Defense, Trap Sense, Inspire Heroics). Needs an
engine-side inventory of every dodge-typed contribution first.
Cross-book equipmod attach. 57 non-CRB equipmods the picker offers cannot attach. The
recognition gate and the cost resolver are both CRB-only, so widening the gate alone would attach
a 500gp ARG Whipwood for free — silent mispricing, strictly worse than today's honest refusal.
Two pre-existing pricing defects, pinned rather than fixed: CRB's Holy Symbol (Wooden/Silver)
display 1gp/25gp and attach free (duplicate KEY: across two categories, first match wins); and
formula-priced equipmods (CRB +1..+10, ARG's COST:WT*375 Darkleaf Cloth) attach free because
COST: is a PCGen formula decisions.md §24 forbids evaluating.
Non-caster spell picker. A Fighter is still offered all 1,185 spells. list_class_spell_levels
reports one bit, and that bit conflates "casts, but no list ingested" (Magus, Summoner, Oracle)
with "casts nothing" (Fighter, Barbarian, Rogue). Narrowing on it would hide legal spells from the
first group. Needs a third state.
PF1's minimum-ability rule for learning a spell (Int ≥ 10 + spell level). Landing it alone would
convert a creation-time defect into un-loadable saved characters: pf1_adapter's own wizard fixture
is Int 10. Creation-time validation must come first.
Halfling ~ Adaptable Luck's reduced-bonus magnitude is dropped from the served description
(DESC arg 2 is Halfling_AdaptableLuck_Bonus-1, an expression). Reading a constant off the row
is transcription and allowed; evaluating <Var>-1 is not, under §24. The clause is kept and only the
unread number omitted, because dropping the clause would imply no downside.
Armor of the Pit's Scaled Skin branch grounds no resistance number: the token names all three
of cold/electricity/fire but the rule grants resistance 5 to two, player-chosen. Grounding three
overstates it; grounding a guessed two fabricates the choice. A +0 record names the branch instead.

D. Ran out of scope, not out of reason (→ next cycle owning the file)

Equipment descriptions stop at the IPC boundary. The DTO now carries a rendered description
for 2,856 records; loadEquipmentCatalog.ts, EquipmentCatalogScreen.tsx and itemPickerFilter.ts
do not read it yet. A genuine half-hop, flagged rather than claimed as reach.
The headline HIT POINTS panel contradicts the Defense tab for any Toughness holder — the third
TypeScript twin. decisions.md §29.2.
The sheet's Vision cell ignores a vision-replacing alternate (ARG Halo). The engine is correct;
character_hub::race_creation_chassis resolves vision at race level with no alternate selections.
Standard-trait suppression is gated for Dwarf only. The other 6 CRB race seams emit their
standard traits unconditionally, so an alternate replacing one removes nothing from the sheet's
explanation list. Bounded deliberately to keep the pilot_compute.rs change reviewable under §28's
standing guard; the remaining 6 are the same mechanical edit repeated.
ARG feat magnitudes are not yet read by pilot_compute_corpus.rs for two specific producers
(armor_of_the_pit natural armor; arg_computed_climb_bonus), plus the parity test that would fail
when a term exists on one twin and not the other. Found during verification; a verifier silently
patching the thing it is measuring destroys the measurement.
The Known/spellbook path is ungated. A Wizard 1 can record Antilife Shell (Cleric/Druid-only)
as a Wizard spell. The off-list half is illegal under any reading; the out-of-level half is a
genuine rules question that should not be settled by an agent inventing a rule.
The Add Spell picker offers the whole 1,185-row catalog with no class or level filter. The
backend's refusal is correct PF1 enforcement — the defect is the unfiltered menu.
3 ARG racial-trait records reach no picker surface (Feral ~ Languages,
Scion of Humanity ~ Languages, Saltbeard ~ Dwarf ~ Greed): build_menu filters to
Default|Alternate. All three do reach through resolve_race_alternate_selection, which is why the
family is catalog-reach-complete; being absent from the menu is a product decision.
monster_catalog.rs::race_subtype serves a raw PCGen | to the player — Hell Hound reads
Evil|Extraplanar|Fire|Lawful. A display join, not a DESC: render, so render_pcgen_desc is the
wrong tool.
Post-creation removal of alternate racial traits was not built: there is no post-creation add
path to mirror, and a remove-only path would be a one-way door in the other direction.
16 APG/ACG classes absent from the Class Progression browser (300 rows across 15 of the 31
creatable classes). Pre-existing; PU's rows landing there made the omission more visible, not less.
3 APG spell records still carry %% on disk
(chameleon_stride, fiery_body, ghostbane_dirge). No player is affected — the screen path
renders through render_pcgen_desc. Fix at cache_gen/apg.rs:580 and :780.
One PU class_feature record carries a literal &nl; (unchained_rogue_debilitating_injury.json)
because ingest_pu_classes.rs has its own copy of the desc treatment. Not player-reaching.
src/bin/sd27_gen_book_cache.rs double-includes the ARG module, which is the sole cause of
several ARG dead-code warnings. Deleting the include should drop warnings off the root clippy
baseline.
biosettings height/weight profiles absent for the 11 newly-creatable Bestiary races. No book's
ingest reads <race>_biosettings.lst for any book; the 7 shipped profiles are hand-entered
constants with no corpus source. Those races honestly show "No height/weight profile" and offer no
reroll button, rather than being given invented numbers.

E. Stale data fields, uncorrected because the file was not owned

core_rulebook/LICENSE.json reports 3,326; the corpus holds 3,400. The 74-record gap is exactly
this tranche's 7 race + 67 race_trait additions.
beastiary/LICENSE.json reports 45; the corpus holds 164. The 119-record gap is exactly this
tranche's 11 race + 108 race_trait additions.
tests/sd27_book_license_record_counts.rs's BOOKS covers pathfinder_unchained and
advanced_race_guide only — both of which are correct (127 and 635, verified). Extending it to
the other two is the one-line change that would have caught the above, and is deliberately left as
the only edit needed.
records_processed is written from what one generator binary emitted
(sd27_gen_book_cache.rs), so re-running the generator resets corrected values. A drift test now
fails loudly if the regression returns.
v06_work_inventory.rs::rule_set_for() still maps four books, so ARG and PU corpus units read
as future_state with evidence no_compiled_rule_set_for_book — false since Phase 1.

F. Process incidents worth carrying forward

Three separate git stash incidents in a shared working tree. Agents measuring a "before"
baseline by stashing repeatedly stashed concurrent agents' uncommitted work; one stash was popped
by another process mid-run, producing a junk 13-failure measurement. No work was lost. The method
is unsafe and the remedy is isolation: 'worktree', which Phase 1's release notes already
recommended for a different reason.
Two TDD deviations self-reported rather than glossed: implementation written before the test,
with RED reconstructed afterwards by stubbing the guard and confirming the intended failure.
Two agents wrote outside their declared file scope and said so — feat_catalog.rs (because the
brief's own deliverable, "a test that fails if ANY catalog serves raw PCGen syntax," was
unsatisfiable without it, and pinning 17 known player-visible leaks as expected would be theatre)
and boundary/removeSelection.ts (because the brief required following the add-path pattern
exactly, and that pattern puts the invoke() wrapper in boundary/). Both are recorded as
judgement calls made rather than questions asked.

Navigation Menu

alpha v0.10.125 (09177a93)

Choose a tag to compare

Sorry, something went wrong.