perf: cache global BM25 idf per query

[BubbleCal]

this could avoid scanning all partitions for scoring each token, helpful when we have many partitions and long query

[github-actions]

Code Review

Summary: This PR caches the BM25 IDF (query weight) per token during search, avoiding redundant computation across partitions. The optimization is straightforward and correct.

Assessment: ✅ LGTM

The change correctly separates the BM25 score computation into its two parts:

1. Query weight (IDF): Now cached per unique token in idf_cache
2. Doc weight: Computed per document as before

What's good:

• The math is preserved: score = query_weight(token) * doc_weight(freq, doc_length) matches the existing Scorer::score implementation
• The test validates correctness by checking that scores match the expected IDF across partitions
• Using HashMap<String, f32> for the cache is appropriate given query tokens are typically few

Minor observations (not blocking):

• The cache is local to the search call, so no memory concerns across requests
• For queries with many unique tokens, the cache provides linear speedup proportional to partition count

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!