TASK-IMP-145: TRACE citation truth for the skill quality floor

Summary

TASK-SKILL-202 is done and every one of its seven acceptance criteria cites tools/install/tests/test_skill_floor.sh — a file that was never created. The floor it describes shipped as scripts/tests/test_skill_stub_lint.sh and is green. This task repoints the citations at the suite that exists, adds the three small checks needed so no criterion cites nothing, records the one half that stayed deferred, and fixes the same stale paths where benchmark gates G7 and G8 inherited them.

Problem

TASK-SKILL-202's own testing-evidence.md names the divergence under the heading "Cited suite (TRACE-004) — path residual": the implementation put the floor in scripts/tests/test_skill_stub_lint.sh (7 tests, green, auto-registered by run_all.sh's glob), while the spec's new_files: and all seven AC lines kept naming tools/install/tests/test_skill_floor.sh plus a tools/install/check-skill-floor.sh checker whose build wiring was the optional F4 half and never landed. Gate-1 accepted the deviation and asked gate-2 to confirm; gate-2 accepted the task and the citations stayed.

Two other surfaces inherited the same paths: docs/verification/benchmark-gates.md (G7 and G8 Owner + Checked-files rows) and the embedded G7/G8 definitions inside TASK-IMP-140's spec, which is that doc's content contract.

So the corpus carries a shipped task whose acceptance resolves nowhere and two benchmark gates naming a checker nobody can run. verify-goals.mjs refuses exactly this shape for goal predicates ("the acceptance cites a test that no longer exists, which IS the finding"); spec-level citations have no such guard, which is why this one survived a done flip.

Proposed Solution

Repoint each SKILL-202 criterion at the real function that proves it and correct the new_files: list to what shipped. Three clauses have no matching test today — §1.1 (delist + allowlist cleanliness), §1.2 (the certify-nfrs.md degradation notice) and §1.7 (the CHANGELOG record) — so rather than let those ACs cite nothing, add t08_delist_and_allowlist_clean, t09_workflow_degrades_loud and t10_changelog_records_floor to the existing suite: three greps over build.sh, chain-allowlist.txt, certify-nfrs.md and CHANGELOG.md, in the suite's existing counter/exit style. Rewrite SKILL-202 §1.5/§1.6 to describe the as-built mechanism (the detector is the suite's floor_check; the standalone checker and its build wiring stayed deferred) and record that deviation in source_decisions, so the normative text stops promising a file the repo does not have. Finally, correct the G7/G8 Owner and Checked-files rows in benchmark-gates.md and in TASK-IMP-140's embedded definitions, leaving severities and tiers untouched so test_benchmark_gates.sh::t01's spec↔doc cross-check keeps passing.

Alternatives Considered

• Create tools/install/tests/test_skill_floor.sh as a shim that calls the real suite. Rejected: a second registration for one behaviour is the failure mode run_all.sh's glob header argues against, and the shim would make the stale name permanent rather than retired.
• Build the missing check-skill-floor.sh + build wiring so the citations become true as written. Rejected for this task: that is real product work (a build-blocking gate), it was consciously deferred at ship time, and it does not need to gate the correction of a citation. Recording the deferral honestly is the smaller true statement.
• Leave the spec alone; the residual is documented in the evidence file. Rejected: the documentation is one file deep and the corpus surface (spec ACs, new_files:, two benchmark-gate rows) still reads as if the checker exists. A note that contradicts the artefact it annotates loses to the artefact.
• Repoint the citations and skip the three new tests. Rejected: three of the seven ACs would then cite nothing at all, which is worse than citing a wrong path — at least a wrong path is greppable.

Success Metrics

• Primary: zero live references to test_skill_floor.sh / check-skill-floor.sh outside the historical docs/batches/ ship notes, every SKILL-202 AC resolving to a function that exists, and scripts/tests/test_skill_stub_lint.sh green at 10/10.
• Guardrail: scripts/tests/test_benchmark_gates.sh and the full run_all.sh stay green; no vendored skill, no build behaviour and no gate severity/tier changes.

Scope

In scope: the three new suite functions, SKILL-202's frontmatter + §1.5/§1.6 + AC block, the G7/G8 rows in benchmark-gates.md and TASK-IMP-140's spec, CHANGELOG.

Out of scope / Non-Goals

• Building tools/install/check-skill-floor.sh or wiring a floor check into build.sh (the deferred F4 half stays deferred — now recorded, not implied done).
• Re-litigating any SKILL-202 outcome: the delisting, the 20 injection-discipline backports and the SCOPE expansion all shipped and are re-verified by the suite this task cites.
• The historical docs/batches/batch-8b-* notes, which correctly record what was true on the day they were written.

Dependencies

None. TASK-SKILL-202 and TASK-IMP-140 are both done; this task edits their citation surfaces only, changing no clause's meaning.

AI Authorship Disclosure

• Tools used: Cursor agent (Composer) auditing the post-1.2.0 plan's Wave 2 residuals against live main.
• Scope: the audit's own figures were treated as claims to verify, not as inputs. re-derived and CONFIRMED: scripts/tests/test_skill_stub_lint.sh passes 7/7 at HEAD, and its t01–t07 function names are as testing-evidence.md records them. re-derived and CORRECTED: the plan and the batch-8b notes describe the residual as a SKILL-202 spec matter only; a repo-wide grep at HEAD found the same stale paths in docs/verification/benchmark-gates.md and in TASK-IMP-140's embedded G7/G8 definitions, so the scope is four files, not one. measured and ADDED: three SKILL-202 clauses (§1.1, §1.2, §1.7) have no covering test in the as-built suite — measured by mapping each clause to the suite's functions — which is why this task adds t08–t10 rather than only repointing citations. No unscoped attestation is made about clauses outside that mapping.
• Human review: session operator instruction (2026-07-25) to close genuine plan gaps; HITL verdicts recorded with evidence at both gates.

1. Description

• 1.1 scripts/tests/test_skill_stub_lint.sh MUST gain t08_delist_and_allowlist_clean: tools/install/build.sh's VENDORED_SKILLS list contains no nfr- entry and tools/install/chain-allowlist.txt contains no nfr- line.
• 1.2 The suite MUST gain t09_workflow_degrades_loud: certify-nfrs.md (source and vendored copies that exist) states at its routing step that the NFR skills are not yet shipped.
• 1.3 The suite MUST gain t10_changelog_records_floor: CHANGELOG.md records the delisting naming TASK-CUO-209, the backport count, and the parity SCOPE expansion.
• 1.4 Every acceptance criterion in TASK-SKILL-202's spec MUST cite a function that exists in scripts/tests/test_skill_stub_lint.sh, and its new_files: MUST name what shipped rather than paths that were never created.
• 1.5 TASK-SKILL-202 §1.5 and §1.6 MUST describe the as-built mechanism (floor detector inside the suite) and MUST record the deferred standalone-checker/build-wiring half as a named deviation rather than an unmet MUST.
• 1.6 docs/verification/benchmark-gates.md G7 and G8 MUST name scripts/tests/test_skill_stub_lint.sh in their Owner and Checked-files rows, and TASK-IMP-140's embedded G7/G8 definitions MUST match — severities and tiers unchanged.
• 1.7 CHANGELOG.md Unreleased MUST record the citation repair.
• 1.8 The suite MUST gain t11_skill_202_citations_resolve, which parses every test: <path>::<fn> citation out of TASK-SKILL-202's acceptance block and fails when the path is absent or the function is not defined in it — so this class of rot cannot survive another done flip silently.

Acceptance Criteria

• [ ] AC 1 (traces_to: #1.1) — build.sh's vendored list and chain-allowlist.txt carry no nfr- entry, and the check fails on a fixture that re-adds one - test: scripts/tests/test_skill_stub_lint.sh::t08_delist_and_allowlist_clean
• [ ] AC 2 (traces_to: #1.2) — every certify-nfrs.md copy states the NFR skills are not yet shipped at its routing step - test: scripts/tests/test_skill_stub_lint.sh::t09_workflow_degrades_loud
• [ ] AC 3 (traces_to: #1.3, #1.7) — CHANGELOG.md records the delisting naming TASK-CUO-209, the backport count and the SCOPE expansion, plus this task's citation repair - test: scripts/tests/test_skill_stub_lint.sh::t10_changelog_records_floor
• [ ] AC 4 (traces_to: #1.4, #1.5, #1.8) — every SKILL-202 acceptance citation resolves to a function defined in the cited file, and the deferred standalone-checker half is named in the spec rather than left as an unmet MUST - test: scripts/tests/test_skill_stub_lint.sh::t11_skill_202_citations_resolve
• [ ] AC 5 (traces_to: #1.6) — G7/G8 name the real suite in their Owner and Checked-files rows while the spec↔doc severity/tier cross-check stays green - test: scripts/tests/test_benchmark_gates.sh::t01_doc_complete_and_consistent

Test plan

1. bash scripts/tests/test_skill_stub_lint.sh
2. bash scripts/tests/test_benchmark_gates.sh
3. bash scripts/tests/run_all.sh
4. bash .cyberos/cuo/gates/run-gates.sh

3. Edge cases

• certify-nfrs.md exists in more than one place (source + built payload). t09 checks every copy it finds and fails naming the copy that lost the notice, rather than passing on the first hit.
• A future release rewrites the CHANGELOG's Unreleased section. t10 keys on the delisting record anywhere in the file, not on the top entry, so a release rollup does not turn a true statement into a red test.
• The NFR skills are genuinely implemented later and re-vendored. t08 then fails loudly, which is correct: re-vendoring is a reviewed change that must revisit SKILL-202's delisting record in the same commit.
• Security-class: documentation and test-only; the new functions read and pattern-match repo files, execute nothing, and add no network or secret surface.