Publish AV-EXP-001 real-repository shadow benchmark

[sneakocom]

Summary

• preserves the AV-EXP-001 preregistration before comparative outcomes
• benchmarks FULL, native Vitest related, AV_CORE, and AV_WITH_NATIVE_EVIDENCE on pinned Zustand
• publishes the complete full-oracle evidence, skip reasons, Visible Value receipts, and deterministic replay
• remains SHADOW; no CI or trust gate is replaced

Result

AV selected 8/8 oracle-relevant checks with no observed miss in the frozen ten-scenario corpus. The native tests-only arm selected 6/8 and missed one relevant lint check and one relevant typecheck. This is bounded calibration evidence, not general safety.

Verification

• npm run verify
• node benchmark/av-exp-001/verify-results.mjs
• 20 Visible Value receipts validated by the authoritative research validator
• actionlint
• gitleaks git --no-banner --redact .
• same-host clean replay matched baseline, all scenario, summary, and analysis semantic identities