Make TREC CAsT evaluation tractable and publish Gemini baseline

[bigboateng]

Summary

Follow-up to merged PR #7, driven by execution of the pinned TREC CAsT profile.

• Make the combined MS MARCO/TREC CAR judged-passage workflow tractable
• Add durable, paced, retryable Gemini rewrite prefetch
• Use stable gemini-3.1-flash-lite
• Omit the resolved oracle when resolved queries are unavailable
• Publish the completed CAsT result and README headline table
• Fix CI's unavailable build>=1.5.1 requirement

Result

20 judged topics, 173 turns, 21,726 judged passages, MiniLM, 10-neighbor kNN, candidate limit 100, and 1,000 paired bootstrap draws.

• Pointwise follow-up nDCG@3: 0.325
• Map-matched β=1.0: 0.352, delta +0.027, 95% CI [+0.008, +0.049]
• Gemini Flash-Lite rewrite: 0.516, delta +0.191, 95% CI [+0.078, +0.296]
• Claim gate: pass

These are judged-passage micro-corpus results, not full-corpus retrieval evidence.

Runtime improvements

• TREC CAR dependency and one-shot docstore failure handling
• Blockwise NumPy kNN graph construction
• SciPy sparse shortest paths with compact bounded caching
• β=0 graph bypass
• Vectorized full-corpus scoring
• Gemini JSON checkpoints, prefetch, pacing, request timeouts, and transient retries

Test plan

• 88 tests passed, 1 skipped
• Ruff and format checks
• Strict mypy
• Source and wheel build
• Corrected cached CAsT/Gemini profile completed
• Result artifact and README table
• All seven GitHub Actions checks pass