refactor(benchmark): split benchmark fetching and lookup modules (#41)

[RaghavRD]

Refs #41

What

Splits src/whichllm/models/benchmark.py into focused modules:

• benchmark_cache.py — benchmark cache read/write helpers
• benchmark_fetch.py — source fetching and layered merge policy
• benchmark_lineage.py — frozen-score lineage recency demotion
• benchmark_index.py — score indices and model-line interpolation
• benchmark_lookup.py — benchmark evidence lookup and inheritance rules
• benchmark_types.py — shared BenchmarkEvidence dataclass
• benchmark.py — compatibility shim / legacy import surface

Why

Per #41, benchmark.py mixed cache I/O, network source orchestration, merge rules, lineage staleness handling, score indexing, and lookup/evidence heuristics. Splitting these responsibilities makes future benchmark and evidence changes easier to review independently.

Testing

• uv run pytest -q tests/test_benchmark_lookup.py tests/test_cache_encoding.py tests/test_r3_regressions.py — 41 passed
• uv run pytest -q — 466 passed
• .venv/bin/ruff check . — all checks passed
• .venv/bin/ruff format --check . — 98 files already formatted

Notes

Behavior-preserving refactor:

• No ranking behavior changes
• No CLI behavior changes
• No source score formula changes
• Public import surface preserved through whichllm.models.benchmark
• Existing benchmark cache monkey-patch behavior preserved

[Andyyyy64]

Sorry this took me a few days to review.

I checked the split carefully. The important benchmark helpers moved cleanly into focused modules, and whichllm.models.benchmark keeps the old import surface through the shim. I also checked the cache monkeypatch path because our tests patch BENCHMARK_CACHE through the old module, and that still works.

I verified the moved helpers against main and ran local QA around benchmark lookup, cache encoding, ranker regressions, full pytest, lint, format, compileall, and a CLI smoke run with --gpu "RTX 4060" --gpu-only --speed usable --json.

No issues found from my side. Thanks for doing this refactor, and sorry again for the slow review.

[RaghavRD]

Thanks for the careful review and for running the extra QA @Andyyyy64. No worries on timing.

I’m glad the shim and cache monkeypatch path held up. I’ll sync with main and continue with the next #41 slice separately.