feat(gate): emit self JUnit + reclassify doc/scaffold test_refs as evidence_refs

spec.yaml

@@ -53,4 +53,4 @@ inventory:
   scenarios: 2
   capabilities: 6
   test_files: 170
-  last_synced: "2026-07-01"
+  last_synced: "2026-07-02"

spec/attestation.yaml

@@ -29,7 +29,7 @@ attested:
   F-021: 8a1a82a59a1c45c7
   F-022: 8f596a1c737f6d42
   F-02343cd1: 77875ee09a7ea3bf
-  F-023: 0c14948e5a91bb0f
+  F-023: 222ad5805d96f197
   F-024: f67a86816b06f8ee
   F-025: 187339b684896b8e
   F-026: bb35faf43ba582cb
@@ -67,7 +67,7 @@ attested:
   F-059: 140dab087ebf5346
   F-060: 96dc261212a0dd3d
   F-061: c16123610e8fe7fc
-  F-062: 0ab83282a7f7b1ef
+  F-062: 6a7566e43056bc20
   F-063: 76a719993cc71fa8
   F-064: e176668e53f61c6e
   F-065: e6ed3ef916201947
@@ -107,10 +107,10 @@ attested:
   F-2de65d: 84ad71574d306c81
   F-315fd7: c3b042c80fa7c187
   F-31eeb8: d88a9880d29ae411
-  F-32b1e0: 9b29e21313d121b1
+  F-32b1e0: 9792422c1c008207
   F-3788c2: af9778dea8687b29
   F-37b4a8: e067655bad681488
-  F-3a5339: b2b2ea8775f99267
+  F-3a5339: 0dfdeec23abfa841
   F-3b3690: 6a36aad282d36f3a
   F-40327b: 8295358f7b813c8a
   F-417ff0: 0cc5eeefc5e08377
@@ -125,7 +125,7 @@ attested:
   F-570a3f: 3f60012b22c9b715
   F-59f093: 26735424fba6308c
   F-5b188856: 92b72281c248eba3
-  F-5b9f9f: 0b972209be8b642f
+  F-5b9f9f: 9d362a585206c5fa
   F-5d3ed2: 9452eac28760fb99
   F-5f6b45: 15323c4f5b619de7
   F-64a5c159: adedb516a257c7ec
@@ -184,7 +184,7 @@ attested:
   F-d2c806: b3d8668905855a6c
   F-d3bde4: 915d13b33258d3fc
   F-d49585: 11e3ac2dce796fc6
-  F-d6b93648: f755a47c66e07635
+  F-d6b93648: 2191277c0e5982ee
   F-d7312b: 000237d094145b6a
   F-d8223c: 0501e9564231899b
   F-d980359c: 8f1559276afc5c03

spec/features/ab-evaluation-4db939.yaml

@@ -51,10 +51,10 @@ acceptance_criteria:
     action: surface the spec-gated-detector limitation honestly — vanilla's "0 errors" is absence of signal, not absence of drift, because detectors need spec to evaluate
     response: 'renderFindings explicitly notes "absence of signal, not absence of drift" in the per-case markdown; docs/ab-evaluation/README.md §Limitations item 2 expands on which detectors are spec-gated (REFERENCE_INTEGRITY, MISSING_IMPLEMENTATION, ARCHITECTURE_FROM_SPEC, CAPABILITIES_FEATURE_MAPPING); docs/ab-evaluation/summary.md verdict on H4 marks it ⚠️ Partially supported with major caveat'
     text: The system shall surface the spec-gated-detector limitation honestly in all 3 documents.
-    test_refs: [docs/ab-evaluation/README.md, docs/ab-evaluation/summary.md]
+    evidence_refs: [docs/ab-evaluation/README.md, docs/ab-evaluation/summary.md]
   - id: AC-007
     ears: ubiquitous
     action: ship the vanilla file sets at senior quality (real zod validation, executable handlers, multiple test cases, proper README) so the comparison stays fair and cladding's value does not depend on vanilla being underwritten
     response: 'tests/scenarios/ab/_vanilla-sim.ts vanilla code is real TypeScript with zod validation, executable handlers, multiple test cases, proper README structure (Install/Usage/API), production-grade package.json with vitest + typescript devDeps; docs/ab-evaluation/README.md §Limitations item 1 explicitly invites readers to inspect the file sets and judge fairness for themselves'
     text: The system shall ship vanilla file sets at senior quality so the comparison stays fair and cladding's value does not depend on vanilla being underwritten.
-    test_refs: [tests/scenarios/ab/_vanilla-sim.ts, docs/ab-evaluation/README.md]
+    evidence_refs: [tests/scenarios/ab/_vanilla-sim.ts, docs/ab-evaluation/README.md]

spec/features/ab-ext-dashboard-ef2fd9.yaml

@@ -19,7 +19,7 @@ acceptance_criteria:
     ears: ubiquitous
     action: 'extract shared React + Vite + TS + Tailwind scaffold into _shared-scaffold.ts so multiple scenarios can reuse package.json / tsconfig / vite.config / tailwind.config / index.html / index.css / main.tsx / README templates without duplication'
     text: The system shall extract shared React + Vite scaffold into a single module reused by all extended scenarios.
-    test_refs: [tests/scenarios/ab-extended/_shared-scaffold.ts]
+    evidence_refs: [tests/scenarios/ab-extended/_shared-scaffold.ts]
   - id: AC-002
     ears: ubiquitous
     action: 'define 30 analytics-dashboard features across 5 categories (layout 8 + cards 8 + charts 5 + data 5 + preferences 4) with deterministic F-hash6 ids and a dashboard-domain capabilities binding'
@@ -44,4 +44,4 @@ acceptance_criteria:
     ears: ubiquitous
     action: 'rewrite docs/ab-evaluation-extended/summary.md as the cross-scenario verdict matrix declaring AB-마무리 (AB-evaluation wrap-up) complete — task-manager + dashboard share identical 3/4 cladding-exclusive drift catch rate proving cladding value generalizes across domains'
     text: The system shall declare AB-evaluation wrap-up complete with a cross-scenario verdict matrix proving domain-independence.
-    test_refs: [docs/ab-evaluation-extended/summary.md]
+    evidence_refs: [docs/ab-evaluation-extended/summary.md]

spec/features/ab-ext-scenarios-emit-f334fa.yaml

@@ -24,9 +24,9 @@ acceptance_criteria:
     ears: ubiquitous
     action: 'with scenario shards in place, AI-query Q5 answers in 1 directory read per scenario — bringing the cladding low-cost answer rate from 2 of 5 to 3 of 5 in both scenarios while vanilla stays at 0 of 5'
     text: The system shall raise the AI-query low-cost answer rate from 2 of 5 to 3 of 5 in both scenarios.
-    test_refs: [docs/ab-evaluation-extended/scenarios/task-manager/report.md, docs/ab-evaluation-extended/scenarios/dashboard/report.md]
+    evidence_refs: [docs/ab-evaluation-extended/scenarios/task-manager/report.md, docs/ab-evaluation-extended/scenarios/dashboard/report.md]
   - id: AC-004
     ears: ubiquitous
     action: 'cross-scenario summary upgrades the H10 verdict from partial to supported (with caveat about Q1 and Q2 linear-scan) so AB-evaluation wrap-up reflects the new measurement'
     text: The system shall upgrade the H10 cross-scenario verdict to supported.
-    test_refs: [docs/ab-evaluation-extended/summary.md]
+    evidence_refs: [docs/ab-evaluation-extended/summary.md]

spec/features/ab-ext-uncommit-demos-9a3b61.yaml

@@ -14,14 +14,14 @@ acceptance_criteria:
     ears: ubiquitous
     action: 'remove the 4 committed React project directories (task-manager + dashboard × cladding + vanilla) from docs/ab-evaluation-extended/scenarios/ — together they were ~160 files / ~10K LoC of regeneratable bloat'
     text: The system shall remove the 4 committed React demo projects under docs/ab-evaluation-extended/scenarios/.
-    test_refs: [.gitignore]
+    evidence_refs: [.gitignore]
   - id: AC-002
     ears: ubiquitous
     action: 'add gitignore rules pattern-excluding scenarios/*/cladding/ + scenarios/*/vanilla/ so the demos stay local-only when regenerated via UPDATE_AB_REPORTS=1'
     text: The system shall pattern-exclude scenarios cladding and vanilla directories from version control.
-    test_refs: [.gitignore]
+    evidence_refs: [.gitignore]
   - id: AC-003
     ears: ubiquitous
     action: 'update README + summary to clarify that demos are regeneratable on demand rather than committed — preserve report.md as the permanent metric record while pointing users at the UPDATE_AB_REPORTS=1 workflow'
     text: The system shall document the regenerate-on-demand workflow in README and summary.
-    test_refs: [docs/ab-evaluation-extended/README.md, docs/ab-evaluation-extended/summary.md]
+    evidence_refs: [docs/ab-evaluation-extended/README.md, docs/ab-evaluation-extended/summary.md]

spec/features/ab-extended-task-manager-0144b9.yaml

@@ -44,4 +44,4 @@ acceptance_criteria:
     ears: ubiquitous
     action: 'verify all measurements support the new hypotheses H9 (linear scale), H10 (AI query cost partial — domain-tuned limitation), H11 (drift catch rate preserved at 3/4), H12 (capture duration bounded under 2s)'
     text: The system shall report H9, H10, H11, H12 verdicts with quantitative anchors.
-    test_refs: [docs/ab-evaluation-extended/summary.md]
+    evidence_refs: [docs/ab-evaluation-extended/summary.md]

spec/features/ab-outcome-quality-ba2e05.yaml

@@ -44,4 +44,4 @@ acceptance_criteria:
     action: surface real cladding bugs that the test had to work around — LLM seed architecture format incompatible with ARCHITECTURE_FROM_SPEC detector schema, and inline F-001 placeholder in init-seeded spec.yaml blocks sharded feature loading
     response: '_drift-injection.ts exports claddingifyForDriftCatch helper that (a) clears spec.yaml inline features so sharded F-4db939 loads via spec.load.ts heuristic, and (b) rewrites spec/architecture.yaml in canonical string[][] + {from,to}[] schema; comment block in the helper documents both as real cladding bugs tracked in docs/ab-evaluation/summary.md Future Work table'
     text: The system shall surface the architecture-schema mismatch and inline-F-001 blocker as known cladding bugs in the future-work table.
-    test_refs: [tests/scenarios/ab/_drift-injection.ts, docs/ab-evaluation/summary.md]
+    evidence_refs: [tests/scenarios/ab/_drift-injection.ts, docs/ab-evaluation/summary.md]

spec/features/ai-hints-preferred-patterns-32b1e0.yaml

@@ -37,4 +37,4 @@ acceptance_criteria:
     ears: ubiquitous
     action: 'dogfood cladding-self spec.yaml with 4 preferred_patterns covering new detector implementation, spec authoring, AI agent lookup, and Iron Law stage structure — each triple uses when plus prefer plus over to be both readable and actionable'
     text: The system shall dogfood preferred_patterns on cladding-self spec.yaml with 4 representative triples.
-    test_refs: [spec.yaml]
+    evidence_refs: [spec.yaml]

spec/features/blind-author-agent-d8223c.yaml

@@ -25,5 +25,6 @@ acceptance_criteria:
     text: "The agent shall be mirrored by the plugin build (agents/ + codex skills) and registered in the glossary like every persona, with the size-budget ratchet extended, so the new definition lives under the same drift discipline as the existing five."
 
     test_refs:
-      - "tests/scenarios/_size-budgets.ts"
-      - "tests/self-consistency.test.ts#every persona file under src/agents/ has a glossary row"
+      - "tests/self-consistency.test.ts#every persona file under src/agents/ has a glossary row"
+    evidence_refs:
+      - "tests/scenarios/_size-budgets.ts"

spec/features/spec-yaml-inventory-and-hints-5b9f9f.yaml

@@ -32,7 +32,7 @@ acceptance_criteria:
     ears: ubiquitous
     action: 'dogfood cladding-self spec.yaml with ai_hints (preferred_persona software-engineer, token_budget_per_session 4000, test_framework vitest, primary_branch develop, forbidden_patterns including eval and innerHTML and dangerouslySetInnerHTML) plus the auto-emitted inventory block'
     text: The system shall dogfood ai_hints + inventory on cladding-self spec.yaml.
-    test_refs: [spec.yaml]
+    evidence_refs: [spec.yaml]
   - id: AC-005
     ears: ubiquitous
     action: 'last_synced field uses YYYY-MM-DD ISO date (not full ISO-8601 with time) so multiple sync runs on the same day produce identical spec.yaml content and do not create commit churn'

spec/features/spec-yaml-metadata-3a5339.yaml

@@ -26,4 +26,4 @@ acceptance_criteria:
     ears: ubiquitous
     action: 'dogfood the new fields in both cladding-self spec.yaml (description, version 0.3.49, repository URL, intent_summary) and the A/B-extended task-manager curator template; regenerate the committed task-manager spec.yaml so reviewers see the enriched front door'
     text: The system shall dogfood the new metadata fields in cladding-self plus the task-manager curator and regenerate committed outputs.
-    test_refs: [spec.yaml, tests/scenarios/ab-extended/_curator.ts]
+    evidence_refs: [spec.yaml, tests/scenarios/ab-extended/_curator.ts]

vitest.config.ts

@@ -9,6 +9,16 @@ import {defineConfig} from 'vitest/config';
 export default defineConfig({
   test: {
     include: ['tests/**/*.test.ts'],
+    // Emit a JUnit report alongside the console output so cladding's own
+    // UNVERIFIED_AC detector can VERIFY (not just existence-check) that each
+    // done AC's test_refs actually ran and passed — closing the self-gate's
+    // AC→test→observed-pass loop (cladding dogfoods its own JUnit feature). The
+    // path is a DEFAULT_REPORT_CANDIDATE the detector auto-discovers; it lives
+    // under .cladding/ (gitignored) so it never pollutes git status. In CI
+    // `npm test` runs before `clad check`, so the gate reads a fresh, complete
+    // report; standalone `clad check` reads the prior run's (the documented
+    // degrade-to-existence baseline applies when no report is present yet).
+    reporters: ['default', ['junit', {outputFile: '.cladding/test-report.junit.xml'}]],
     // The heavy scenario suites (init + observed-path scan over the 8-file
     // fixtures, then detector snapshots) run ~6s locally but several × slower
     // under the 2-core GitHub runner's worker contention. The default 5s