Closes GT-668 — the proof that a guard had been seen failing was pinned to a branch, so it expired the day the fix landed

[beyondnetPeru]

The defect, reproduced before it was acted on

GT-666 shipped a self-test whose most important case checks its guard against the real pre-fix artifact rather than a sketch of one. Its comment says why, and the reasoning is right:

"before this fix, read out of git rather than reconstructed … Reconstructed fixtures agree with whatever the author believed was wrong; this one cannot."

It obtained that artifact with git show ${PRE_FIX_REF}:iso-5055-mapping.json, PRE_FIX_REF defaulting to origin/develop. True while the fix lived on a branch. 59d62bae merged, origin/develop began serving the CORRECTED artifact, and a case written to count 64 findings started counting 0. Green on its own branch, red the moment the fix landed — blocking promotion PR #483 in Governance guards (GT-578).

Measured on develop at 59d62bae before touching anything:

ℹ tests 19 · pass 18 · fail 1
✖ the REAL pre-fix artifact, against the REAL packs, is RED — 16 rules, 4 findings each
  AssertionError: each of the 16 rules should fail the class rule, the naming rule and both false-claim rules
  0 !== 64

The worse half is the one nobody would have seen. The case carried a shallow-clone escape — if (before.status !== 0) return void assert.ok(true, 'SKIPPED: …'). Under fetch-depth: 1, a detached history, or a fork without origin/develop, the identical rot passes in silence. That is an escape hatch out of the one case establishing this guard has ever been observed failing rather than merely declared able to — the property 43-validate-guard-negative-fixtures exists to establish.

The fix: freeze the evidence, keep the intent

The intent was right; only the mechanism was wrong. A real pre-fix artifact is not the same thing as fetching one at test time, and the two were conflated.

The artifact is now committed. .harness/fixtures/standards-rule-class/iso-5055-mapping.pre-gt-666.json, taken from 01308346 (59d62bae^, the last commit carrying the defect), blob 6684e8a4. That is the fixtures home this repo already uses for committed CI evidence (.harness/fixtures/<subject>/, alongside rag-eval/) — no new home invented.

Not trimmed. All 412 rows kept, not just the 16 that carry the assertion. Trimming is an editorial act, and an editorial act on a fixture whose entire value is that no author's belief touched it is the thing the comment warns against. 284 KB, the same order as the live artifact already tracked.

One key added, and the claim is checkable rather than asserted. JSON has no comments, so the provenance header is a top-level _fixture string naming what the file is, where it came from, and that it is deliberately frozen. checkClassification reads only rules and summary.byClass, so it is inert. Drop _fixture, re-serialise, and the blob reproduces byte for byte against git cat-file blob 6684e8a4.

Deleted: the git show, the PRE_FIX_REF env var, the skip. No replacement skip. A shallow clone constrains a test that shells out to git, not one that reads a file in its own working tree — so an unreadable fixture is a genuine failure and red is the right outcome.

Nothing weakened; one case added

Assertions are the same two the case always made: 64 findings, 16 of them is classified \governance``, against the frozen fixture and today's live packs. Breakdown measured: 16 class + 16 "no international structural equivalent" + 16 "a governance invariant over Evolith artifacts" + 8/4/4 notes naming no standard.

The added case — the frozen fixture is still the DEFECT, not a refresh of the fixed mapping — closes the failure mode freezing creates: somebody regenerates the fixture from the live artifact and the repo's only observed failure of this guard becomes a copy of the thing it catches. Observed, not argued: the refresh was simulated by writing the live mapping into the fixture, and 2 of 20 cases turn red, the counting case among them.

Measured

Check	Result
node --test .../65-…test.mjs	20 tests, 20 pass, 0 fail, 0 skipped (was 19 / 18 pass / 1 fail)
42-validate-guard-denominators	78 guards classified, 57/57 refuse a zero-element scan
43-validate-guard-negative-fixtures	54/54 exercised guards turned RED
ci-runner.mjs governance	✅ 17/17 scripts
08 / 09 / 46 / 04 / 01 / 49	pass; chain current and at a fixed point
core-domain (clean rebuild, tsbuildinfo removed)	157 suites / 1853 tests green

Coordination

No corpus file, no generator and no derived artifact was touched. build-iso-5055-mapping.mjs and iso-5055-mapping.json are deliberately untouched — a parallel session is changing analyser.adoptable on the four ISO5055-* rows. That work moves neither ruleClass nor the pack rule count, so the frozen 64/16 figures are unaffected by it; the number was not chased.

Registered GT-668, not GT-667. GT-667 is claimed by branch name (guard 50 claims by gt-NNN) on two concurrent branches — gt-667-analyser-adoptable and claude/gt-667-iso5055-adoptable — carrying the analyser.adoptable follow-on that GT-666's own closure record deferred. 49-validate-gap-id-allocation compares against origin/main and could not have seen either, so the id was taken by union of branches.

Not claimed

The frozen fixture pins the pre-fix side; the live side is still read from the repository, so these counts hold while the three packs keep their 16 rules (asserted elsewhere in the same file). This PR does not re-audit the other guard self-tests for the same moving-ref pattern — whether git show against a branch appears anywhere else was not measured.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[github-actions]

📊 Bilingual Coverage Impact

PR Changes

• Paired EN/ES files modified: 4
• New EN files needing ES translation: 0

Repository Coverage

Metric	Value
Total EN files	527
Total ES files	501
Paired files	0
Coverage	0%

✅ Good: All EN changes have ES counterparts.

---

Generated by GitHub Actions