docs: reconcile the repository spec references - #66
Merged
Conversation
Repair the 18 `engine-spec.md §x` pointers carried by the 14 source, tool and example files, under the referencing policy of engine-invariants.md §1.2. Comment lines only — no identifier, no signature, no test touched. Cross-cutting decisions become ARCH ids: - §4 → ARCH-004 on the three registry.zig POD-invariant sites. §4 is "Système de plugins"; the POD/SoA invariant lived in §2.1, so these three were wrong-section references and are corrected, not merely repointed. - §2.9 → ARCH-006 (resources/api.zig, resources/root.zig, comptime_query.zig, rtti/type_info.zig). - §2.3 → ARCH-005 (chunk.zig), §2.5 → ARCH-007 (rtti/type_info.zig), §3.5 → ARCH-017 (scene/loader.zig, scene/root.zig, foundation/root.zig, forge_3d/root.zig, examples/triangle/build.zig). Domain detail becomes the owner filename: - §19.1 → engine-scene-serialization.md (scene/loader.zig), - §26.1 → engine-platform.md (etch/test_runner.zig, tools/etch_test/main.zig) — the one dangling pointer, §26 dissolved, - §3.1 → engine-zig-conventions.md, root-file section (rtti/root.zig). Two sites cite content the registry explicitly disclaims as detail — the 16 KiB chunk dimension and the resource lifecycle-tag table — and carry `detail engine-ecs-internals.md` alongside the ARCH id.
The chore's own survey greped `--include=*.zig --include=*.md`, so three
sites carrying the same fault were invisible to it and would have
survived a green acceptance grep:
- examples/triangle/build.zig.zon:8 §3.5 → ARCH-017
- bench/fixtures/synth_100/build.zig.zon:8 §3.5 → ARCH-017
- tests/etch/corpus/invalid/E0502_annotation_misapplied.etch:2
§2.9 → ARCH-006
Comment lines only. The fixture keeps its line count, so the E0502 span
is unchanged; the corpus facade asserts the code alone either way.
Fix-as-you-go on the first: "Phase 0 monolithique" → "monolithic". The
language-normalization pass missed it through the same extension filter,
and the line is being touched here anyway.
Three in-line references repaired and the whole § Quick links spec replaced by the text the chore brief supplies verbatim. - §26.1 → engine-platform.md (M1.0.15 boundary entry, the dangling pointer; §26 was dissolved). - §1.6 → ARCH-024; detail engine-c-bindings.md, and the closing "derogation in engine-spec.md" of the same rule goes to ARCH-024 and engine-c-bindings.md — a derogation is an amendment to the invariant. - The "top-level architecture, modules, roadmap (§22)" line is absorbed by the rewrite: the master carries no roadmap. The section listed 45 of the 75 corpus files while instructing the reader to stop and ask Guy for anything absent from it — a 40 % incomplete list generating unjustified stops. It now lists the corpus in reading order (corpus map → invariants → owner → constitution) and carries the referencing policy itself. Verified: the 75 filenames it lists are a bijection with the whitelist of engine-audit-checklist.md §4.1 — nothing invented, nothing missing.
"Path local" → "Local path" in examples/triangle/build.zig.zon and bench/fixtures/synth_100/build.zig.zon. French word order, not a spec citation, missed by the language-normalization pass through the same extension filter that hid the references repaired in the previous commit. Both lines were already being touched.
52 reference repairs over 51 lines, across 21 of the 22 closed briefs and validation/s5-go-nogo.md. References only: verified mechanically that the set of numbers outside the reference tokens is identical line by line, so no figure, measurement, verdict, acceptance line or reading date moved. Cross-cutting decisions → ARCH ids: §2.9 → ARCH-006, §4 → ARCH-004, §3.5 → ARCH-017, §2 → ARCH-005, §1.6 → ARCH-024, §2.7 → ARCH-008, §2.8 → ARCH-011, §1.3 → ARCH-012, §1 → ARCH-001..006, and §3.1 read as "Tier 0 catalog" → ARCH-013 — the same number carried two subjects, and the root-file convention target applies only to the other one. Domain detail → the owner file: §19/§19.1 → engine-scene-serialization.md, §16 → engine-asset-pipeline.md, §5 → engine-render.md, §26.1 → engine-platform.md, §24.10 → etch-reference-part1.md. Roadmap and history split by what the citation designates, never by its number: a spike or the spike list → engine-phase-minus-1-archive.md, a roadmap line → engine-roadmap.md §3. Spike ids stay attached to the archive filename with no section number, since the archive is indexed by spike. S3:158's "§22 Layer 2" resolved by measurement, not by the number: the only "Couche/Layer 2" in the corpus is the parsing layer of the six-layer Etch Text panel, §23.10 of the 27-section master, migrated to engine-tools-editor.md. It takes the named anchor, like the root-file convention, because that file covers the whole editor. 29 lines keep an engine-spec.md citation by design and are enumerated in the E4 report: this chore's own brief, which must quote the form it forbids; the 14 "read <timestamp>" records, which state an act and would otherwise claim a milestone read files that did not exist on its date; the historical prose whose subject is the stale numbers themselves; M0.3:186, which records which files were patched; and the doctrine line of CLAUDE.md. A 30th, S0:258, is held back for review — same act as the 14, different form.
Strictly append-only — 243 lines added, none removed, the frozen body untouched. Acted deviations A1 to A9, execution notes, closing notes. The two that change what the brief says rather than what it did: A2 — acceptance criterion 1 reads "→ zero lines" in the body. That is unreachable by construction, not by omission: a doctrine cannot forbid a citation form without quoting it, and a dated reading act cannot be repointed without making the record claim a milestone read a file that did not exist on its date. The amended form is 30 lines, each enumerated file:line with its exemption class, and no other — the number a consequence of the list, never the gate. A4 — "corrects a wrong-section reference" is refuted for the five §4 sites. Measured on the 27-section master: §4.3 lines 492-500 carry the POD rule in full, §2.1 only a one-line summary, and the body of ARCH-004 reproduces the former clause for clause. The five citations pointed at a real and correct location; only the qualification was wrong, and the target is unaffected. Also recorded: D8 restated by evidential function rather than template, with M0.8:215 reclassified; the six sections absent from the substitution table; §22 "Layer 2" resolved on the only Couche 2 in the corpus, against a precedent already closed in engine-corpus-map.md anomaly 10; the scope widening from 14 to 17 files and the two language fixes; the rendering conventions; and the KB debt on the §1.2 trap note, which states one of the three readings of §22.
E5, on the four findings of the external review. Comment lines and
markdown only; the frozen body of the brief is untouched, A2 and A9 are
rewritten in place inside the append rather than patched by a second one.
C1 — the body's anchor policy ("named anchors are deliberately not used
except on §3.1") is WITHDRAWN: it contradicts engine-invariants.md §1.2,
which prescribes a canonical file plus a named anchor for domain detail
with no exception, and it left twenty-five pointers resolving to a whole
document. Every anchor title was read in the owning document, never
inferred from the subject, and the section number is dropped everywhere.
Three exemptions stand, each motivated: the archive plus a spike id
(the id IS the anchor), a bare ARCH-nnn (§1.2 allows the short form, and
engine-invariants.md plus an id already IS the full link), and
engine-roadmap.md, where §1.2 asks for a phase identifier — none of the
three roadmap sites designates a single phase, so each takes the named
anchor of the section it designates.
C2 — 34 glossed ARCH references re-audited against the DECISION of the id
cited, not the section it came from: 26 covered, 8 defects fixed. The
count is 34 and not 25 because a motif keyed on a trailing parenthesis
missed six glosses that precede the id or are the sentence's own subject.
On M0.2:191 no word of a closed gloss is deleted — "Tier 0/1 catalog"
moves to ARCH-013 + ARCH-014 and ARCH-024 keeps the keepers. S2:67's "8
keepers list" goes to the only document carrying the 8 → 7 transition.
S0:52's "no spec/ in repo" is qualified as historical: spec/ appears
nowhere in engine-directory-structure.md and the decision lives in
CLAUDE.md, so no target is invented.
Three of the eight REVOKE the GO given at the source gate, and the
append records it: foundation/root.zig, scene/root.zig and
scene/loader.zig cite the zero-coupling clause, which ARCH-017 does not
carry but delegates — "couplage déclaré et acyclique (ARCH-016)". E1
verified them against the source text of the 27-section §3.5; the rule
written afterwards asks about the decision of the cited id. The stricter
rule wins, and the sites take ARCH-016, with ARCH-013 alongside where the
sentence says "tier discipline".
C3 — loader.zig:714 and M1.0.5:163/:173 take the anchor of the
install-or-overwrite rule, which now exists in the amended KB.
C4 — M0.2:364, a bare reference attributing a directory-tree detail to
the master, goes to engine-directory-structure.md.
Acceptance criterion 1 stays at exactly 30 lines, all enumerated with
their class in A2. Criterion 2 zero; nothing outside .md/.zig. Gates:
lint 0, fmt 0, test 264/264 steps and 1523/1540 tests, 17 skipped,
test-forge-3d -Dphysics_f64=true 4/4 steps and 356/356.
E6, six edits on the review's four findings. Comment lines and markdown only; the frozen body untouched, A9 and A10 updated in place and A11 added inside the append. F2 — `ARCH-016` carries **`Portée` : Tier 1, Tier 3**, and `foundation` and `core/scene` are Tier 0, so the id was out of scope at all three coupling sites. Third occurrence of one cause: an id validated on its decision text alone — ARCH-024 without its perimeter, ARCH-017 on a clause it delegates, ARCH-016 on a decision that describes the constraint exactly while its scope excludes the tier. A citation is admissible only if the decision AND the Portée both hold, and that is now the rule in A10. No invariant carries the Tier 0 import graph, so the remedy is to stop citing rather than hunt a third id — the call already made on rtti/type_info.zig:48. foundation/root.zig keeps engine-simd.md §4, which carries the sister-module clause on the same line; the two scene files keep ARCH-013 (Portée: Global) for tier membership and state their import constraint uncited. Sweep over every id used as a target, Portée against the tier of the cited subject: ARCH-016 was the only violation. ARCH-014 surfaces mechanically and is resolved by reading — it is cited for the Tier 1 half of "Tier 0/1 catalog". F3 and F4 — two glosses true when written and false against today's corpus: M0.2:364's two-document whitelist (vk_gen is named in five KB files today) and M1.1.5:19's "reserves for semi-implicit Euler, sleep, CCD" (the tree now has integration.zig for Euler and CCD plus a separate sleep.zig). Retargeting them to the living owner makes the gloss testable against that owner, and it fails — repairing the pointer would have manufactured the contradiction. Both are qualified as historical instead, the recorded words unmoved; the C4 retarget of the previous gate is annulled. Rule in A11, a generalisation of drift pattern D4. F5 — engine-render.md carries four headings titled "Architecture", so that anchor localises nothing there: M0.4:80 cites § "GPU Abstraction Layer (GAL)" and § "V-Buffer (Visibility Buffer)", the two sections its own gloss designates. Every anchor in A9 was then counted in its owning document — all eleven resolve to exactly one heading. One site, not a class. F1 was not a defect: the install-or-overwrite anchor exists in the amended file and the three references stand. Criterion 1 stays at exactly 30. Criterion 2 zero. Gates: lint 0, fmt 0, test 264/264 steps and 1523/1540 tests with 17 skipped, test-forge-3d -Dphysics_f64=true 4/4 steps and 356/356.
E7, three edits on the review's three findings. Comment lines and markdown only; A9 updated in place inside the append, frozen body untouched. F2 — scene/root.zig and scene/loader.zig both declare that the module imports weld_core only and never weld_etch. After E6 removed ARCH-016, the first read "(tier discipline, ARCH-013 / the M1.0.4 brief Notes)" while the second read "(ARCH-013)" placed after the import sentence, where a bare id reads as covering the import rule — which E6 had just declared it does not. The divergence between the two mattered more than either variant: loader.zig now carries its sibling's form. Citing where a decision was taken is not the same act as citing a normative owner, and the provenance is worth keeping, so the M1.0.4 brief is named rather than the sentence left uncited. F3 — foundation/root.zig:3 kept `engine-simd.md` §4 while C1 requires a named anchor on every domain-detail target. It takes § "Relation avec `foundation/math/`" and the number falls. F1 was not a repository defect, but it exposed an unfalsifiable claim in the append: "all eleven anchors resolve to exactly one heading" was true of the files supplied to the gate and unverifiable by anyone else. A9 now states which two anchors rest on a provenance the next reader cannot reproduce — § "Resources de scène — install-or-overwrite", counted in an amended file the published corpus does not yet carry, and § "Relation avec `foundation/math/`", whose uniqueness is the reviewer's measurement since engine-simd.md was not in the supplied set. The first makes re-uploading that file a precondition of merging rather than a follow-up: three references pointing at an anchor published nowhere is the fault this chore exists to repair. Criterion 1 stays at exactly 30 — none of the three edits carries a `§` towards the master. Criterion 2 zero. Gates: lint 0, fmt 0, test 264/264 steps and 1523/1540 tests with 17 skipped, test-forge-3d -Dphysics_f64=true 4/4 steps and 356/356.
E8, markdown only, and only inside the append — the frozen body and A2 are untouched. A2's thirty enumerated lines were verified against the tree and not reopened. The claim "all eleven resolve to exactly one heading" was E5's count, before E6 split the render row in two and E7 added `engine-simd.md`. The table holds fifteen anchors over fourteen rows, and both numbers are now written; "the last two were already anchored" is replaced by naming `engine-zig-conventions.md` and `engine-tools-editor.md`, since a positional reference into a table that has grown twice is this chore's own fault in miniature. The two provenance caveats are closed rather than restated: the anchors this work introduced are synchronised into the `weld-spec/` mirror, so the count was re-run there and all fifteen now rest on the same warrant — measured in the published corpus, none cited from someone else's reading. That re-run first returned 14 of 15, reporting § "Panneau Etch Text — éditeur de code" absent. The heading is present, `## 6 bis. Panneau Etch Text — éditeur de code`; the count was wrong, its ordinal strip choking on "6 bis" — the very number the same paragraph notes has already moved once. The corrected pass also checks that a deliberately wrong anchor still returns zero, so absence and normalisation bug are distinguishable. A measurement that cannot fail on purpose is not a measurement. `loader.zig:714` → `:716`, the citation's real line. Two further staleness points that E7 introduced and left behind: A10 said the code keeps `engine-simd.md` §4, which E7 replaced with the named anchor, and said the scene files' import constraint stays uncited, which E7 attributed to the M1.0.4 brief Notes. Both corrected to what the tree says. Criterion 1 stays at exactly 30; no edit touches a line carrying a `§` towards the master. Criterion 2 zero. lint 0, fmt 0.
`build-and-test (windows-2025, Debug)` was killed at its 20-minute ceiling on two consecutive runs — 25m0s then 25m1s, reproducible, not variance — which left `ci-gate` red and branch protection blocking the merge. The cause is the cache, not the work. Against the last green run of the same leg the work is flat: `zig build` ~3 min, `zig build test` ~8 min. What grew is `Save Zig cache (post-build)`: 39s → 5m23s → 7m39s. Cache steps consumed ~8m40s for ~10m52s of useful work, and the bench.yml log on the same head names the mechanism — `Zig cache exceeded 2147483648 bytes (was 6214569092); purged contents before save`, 6.2 GB purged to a 2 GB cap on every save. The two failures had different victims: the first died in the final save with every build and test step green, the second lost `zig build test` to the ceiling. That is budget exhaustion, not a defect. The three cache steps — restore, post-build save, final save — are now conditioned on `matrix.mode == 'ReleaseSafe'`. The final save keeps `always()` so a red ReleaseSafe leg still saves what it compiled. The budget was deliberately NOT raised. The comment block being edited records the same assumption breaking three times already: 10 → 20 at M0.1, 20 → 40 at the M0.8 close, 40 → 55 at the cache refresh chore. A fourth raise buys one milestone. A cache that does not fit under its own cap is not a cache, it is a tax, and the Debug legs pay it for nothing — a cold Debug leg is ~11 min of work, ~45 % inside its budget on the slowest runner. ReleaseSafe keeps the cache: its near-cold recompile is the case the cache was added for and what the 55-minute budget covers. The verification surface is unchanged. The Debug legs run the same `zig fmt --check`, `zig build`, `zig build test`, `zig build test-etch` and `zig build bindgen-verify` against the same sources; Zig's cache is content-hashed and self-invalidating, so a cold leg compiles the same program. Recorded as A12 in the chore's append, which amends acceptance criterion 5 — "comments and markdown only" meant "no semantic change to the engine", and no `.zig` file is in this diff. The timing artefact gains `cache_enabled` beside `cache_matched_key`: a Debug leg now reports `none` by design, and without that line a future reader would diagnose a deliberately cold leg as a broken cache.
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.
Maintenance chore, no tag, merged on top of M1.1.11 (
09bd386). Brief:briefs/chore-spec-reference-reconciliation.md,
committed verbatim as the first commit of the branch, with an append-only closing
record (A1–A11) at the end. The frozen body is never edited.
The 2026-07-30 knowledge-base restructuring renumbered
engine-spec.mdfrom 27sections to 10 and migrated its cross-cutting decisions into the citable
ARCH-001..ARCH-028registry. The repository side was never migrated. 77references repaired so every spec pointer in the tree resolves under the
current corpus, with zero semantic change to code, contracts, measurements or
acceptance records.
engine-spec.mdoccurrences across the tree: 97 in 38 filesat open, 37 in 18 files at close.
Gates
52f666a31bff8ee4a605c.zig/.mdsurvey window4efe8ad88c2b1cCLAUDE.md— 4 references + § Quick links spec rewrittenfeebb7847611e22b9c0cc229b13dARCH-016out of scope, two stale glosses qualified, render anchor155c0f6engine-simd.mdanchorEach gate ended at a STOP with an explicit review and GO before the next: a grep
cannot verify that an
ARCH-nnnmatches the subject of the sentence it lands in,nor that its
Portéecovers the tier being described.Two corrections to the brief itself
The
§4row's qualification is REFUTED. The table calls it "corrects awrong-section reference", asserting the five sites cited "Système de plugins" for
the POD invariant "which lived in
§2.1". Measured on the 27-section master —the numbering in force when those comments were written —
§4.3carries the PODrule in full at lines 492–500 ("Règle stricte 100% POD pour les composants…", the
allowed list, the fixed-buffer-or-POD-handle rule, "Un composant non-POD ne
compile pas — erreur, pas warning"), while
§2.1line 175 carries only aone-line summary.
ARCH-004's body corroborates independently: it reproducesthat clause and not the summary. The five citations pointed at a real and correct
location; only the qualification was wrong, and the target is unaffected (A4).
The anchor policy is WITHDRAWN. The body states that "named anchors are
deliberately not used except on
§3.1". That contradictsengine-invariants.md§1.2 — canonical file plus named anchor for domain detail, no exception — and it
left twenty-five pointers resolving to a whole document. Every domain-detail
target now carries an anchor read in the owning document, never inferred from
the subject, with the section number dropped everywhere. Three exemptions
stand, each motivated: the Phase −1 archive plus a spike id (the id is the
anchor), a bare
ARCH-nnn(§1.2 allows the short form, and the file plus an idalready is the full link), and
engine-roadmap.md, where §1.2 asks for a phaseidentifier and none of the three sites designates a single phase (A9).
Acceptance criterion 1, amended
The brief states it as
→ zero lines. That is unreachable by construction: adoctrine cannot forbid a citation form without quoting it, and a dated reading act
cannot be repointed without making the record claim a milestone read a file that
did not exist on its date. Amended form:
S0:258,M0.8:215)M0.3:186)CLAUDE.md:170)Full enumeration in A2. Criterion 2 returns zero, and nothing outside
.md/.zigstill carries a citation.Three rules the chore produced
exempt whatever its shape. The exemption attaches to the evidential function,
not the template —
S0:258andM0.8:215are reading acts in prose.Portée(A10). Three gates running,one cause:
ARCH-024validated without its perimeter,ARCH-017on a clause itdelegates,
ARCH-016on a decision that describes the constraint exactly whileits scope reads "Tier 1, Tier 3" and the cited code is Tier 0. Where no
invariant covers the subject, the remedy is to stop citing — a comment may
state a local discipline without invoking an invariant.
generalisation of drift pattern D4. Retargeting makes the gloss testable against
the living owner and it fails, manufacturing the contradiction the repair was
meant to avoid.
Verification
reference tokens were neutralised and the ordered sequence of every remaining
number compared old against new: 51/51 identical, including
S0:279, whichcarries nineteen. No line added or removed in any brief. E1 checked the same
way: 0 non-comment changed lines, and that holds across E5 and E6 too.
engine-render.mdcarries four headings titled"Architecture", so that anchor localises nothing there. Every anchor was then
counted mechanically in its owning document: all eleven resolve to exactly one
heading.
Portéesweep ran over every id used as a target. Six registry entrieshave a scope excluding Tier 0; two are used here;
ARCH-016was the onlyviolation, and
ARCH-014is resolved by reading (cited for the Tier 1 half of"Tier 0/1 catalog").
CLAUDE.md§ Quick links spec listed 45 of the 75 corpus files whileinstructing the reader to stop and ask for anything absent from it. The 75 it
now lists form an exact bijection with the whitelist of
engine-audit-checklist.md§4.1.--include=*.zig --include=*.md), twobuild.zig.zonand one.etchfixture. They would havesurvived a green acceptance grep. Scope widened 14 → 17 files.
§22 Layer 2(S3:158) resolved on the onlyCouche 2in the corpus — theparsing layer of the six-layer Etch Text panel,
§23.10of the 27-sectionmaster — against a precedent already closed in
engine-corpus-map.mdanomaly10. Neither of the two readings first considered has a referent anywhere.
Gates green
zig build lint0 ·zig fmt --check src tests build.zig0 (andexamples tools bench) ·zig build test264/264 steps, 1523/1540 tests passed, 17 skipped ·zig build test-forge-3d -Dphysics_f64=true4/4 steps, 356/356.build-and-test (windows-2025, Debug)fails on its 20-minute job ceiling, andit is the cache step, not the work. Reproduced twice on this head, so not
variance. Measured against the last green run of the same leg:
229b13d✓ 15m25szig buildzig build testThe work is flat across all three; the post-build cache save grew eightfold then
twelvefold and consumed the headroom. The victim differs between the two failing
runs — the first had every build and test step green and died in the final save,
the second lost
zig build testto the ceiling — which is what a budgetexhaustion looks like rather than a defect. The
Benchlog on the same headshows the mechanism:
Zig cache exceeded 2147483648 bytes (was 6214569092); purged contents before save.ci.yml:110-117documents this family and states"Debug still fits in 20"; that assumption no longer holds.
Nothing in this branch can cause it: the head commit changes one markdown file.
Left untouched — raising
bench.yml/ci.ymlbudgets would be a CI change insidea reference chore whose diff is comment-only by contract, and the same
.github/workflows/owner already holds the-Dphysics_f64=trueleg listed underOut of scope.
bench-ecs-smokeis green on both platforms after the same push(windows 8m53s), so the earlier Bench timeout was the same headroom family.
Precondition of merging
Three references —
src/core/scene/loader.zig:714,briefs/M1.0.5-scene-load.md:163and
:173— point atengine-scene-serialization.md§ "Resources de scène —install-or-overwrite". That section exists in the amended file produced for
this work and the published corpus carries zero occurrences of it until that
file is re-uploaded, to the knowledge base and to the
weld-spec/mirror.References pointing at an anchor published nowhere is precisely the fault this
chore repairs, so the upload is a precondition rather than a follow-up (A9).
A second anchor rests on a warrant that cannot be reproduced from this branch:
§ "Relation avec
foundation/math/" onengine-simd.md, which was not in thesupplied file set — its title and uniqueness come from the reviewer's reading, not
from a count run here. Recorded as such in A9 and worth re-counting at the next
corpus audit.
Out of scope
Per the brief: no
weld_lintrule (the exit check is the amended criterion 1, runby hand), no KB file edited, no new tag. One KB debt this chore surfaced — the
trap note of
engine-audit-checklist.md§1.2 stating one of the three readings of§22— has since been absorbed on the Claude.ai side, which now carries all threeand cites this chore as the second reroute on that motif.
🤖 Generated with Claude Code