Regenerate retrieval corpus v2 and republish benchmark results

[sarthakagrawal927]

The retrieval harness now rejects v1 corpora because added files cannot be scored against a pre-fix revision. The checked-in retrieval corpora and reports remain archival until rebuilt and rerun.\n\n- [ ] Rebuild every checked-in corpus with codevetter.context-retrieval-corpus.v2\n- [ ] Confirm created_files are excluded from required_files\n- [ ] Rerun required controls and provider arms under the revision-derived tier protocol\n- [ ] Confirm controls lose at 1k, 4k, and 16k within each repository\n- [ ] Republish reports with per-arm case coverage and abandonment status\n- [ ] Record exact commands, revisions, costs, and artifact hashes

[sarthakagrawal927]

Blocked

Blocked on: external accounts, paid API spend, and multi-day compute.

All 7 checked-in corpora (benchmarks/context-retrieval/corpora/corpus-{express,fastify,flask,gin,gin76,got,hugo}.json) still declare schema_version: codevetter.context-retrieval-corpus.v1, while scripts/context-retrieval/build-corpus.mjs and score.mjs require v2. Every checked-in corpus is currently unscoreable, and benchmarks/context-retrieval/results/ holds only full-field-got.

Rebuilding means cloning 7 external upstream repos at pinned revisions, then re-running ~25 provider arms × ~108 cases through third-party MCP retrieval tools (serena, cocoindex-code, codegrok) that each need installation, indexing, and in some cases API keys or paid tokens. HANDOFF.md records one arm at ~90s/case and another burning ~55 minutes in timeouts, and the issue asks for exact costs to be recorded.

Who can unblock: owner — decide whether the benchmark spend is worth it, then provision the tools and keys.