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

[yuyu04]

What

cladding's UNVERIFIED_AC detector verifies each done AC's test_refs actually ran and passed via a JUnit report — but cladding never emitted one for itself, so on its own self-gate the check silently degraded to existence-only: zero self-benefit (the gap flagged while triaging #201/#202).

This wires the emission and fixes what enabling it exposed.

Changes

1. Emit JUnit (vitest.config.ts)

reporters: ['default', ['junit', {outputFile: '.cladding/test-report.junit.xml'}]] — a DEFAULT_REPORT_CANDIDATE the detector auto-discovers, under .cladding/ (gitignored, never pollutes git status). In CI npm test runs before clad check, so the gate reads a fresh, complete report and now genuinely verifies AC↔test↔observed-pass.

2. Reclassify 15 mis-categorized test_refs → evidence_refs (10 features)

Turning emission on first surfaced 21 warnings: 15 test_refs pointed at things that are not executable tests and so have no observed result in any JUnit:

• docs (docs/ab-evaluation*/**.md), .gitignore, spec.yaml (dogfood markers), and underscore-prefixed scaffolds (_vanilla-sim.ts, _shared-scaffold.ts, _drift-injection.ts, _curator.ts, _size-budgets.ts).

MISSING_TESTS is satisfied by either test_refs or evidence_refs, so each AC stays verified — the refs are now recorded as the doc/artifact evidence they actually are (undoing the v0.2.2 "lump everything into test_refs" dishonesty that had recurred). One AC (F-d8223c) kept its real self-consistency.test.ts test_ref and moved only the scaffold.

Note: src/stages/detectors/README.md row | 37 | was not touched — it's the catalog ordinal, not a count (handled separately in #219).

Verification

• UNVERIFIED_AC 0 / UNTESTED_AC 0 / MISSING_TESTS 0; strict gate GREEN with the report present — the self-gate now exercises its own JUnit feature instead of perpetually degrading.
• Full local CI mirror green: build+diff, typecheck, lint, npm test, conformance, clad check --tier=pre-commit --strict.
• attestation re-stamped (the 10 edited shards' features).

Base note

Branched off develop (independent of #219). Both re-stamp spec/attestation.yaml, so whichever of #219/this merges second needs a trivial clad sync + re-attest — no logical conflict.

Closes the ② half of the #202 follow-up (① = detector count, shipped in #219).

🤖 Generated with Claude Code