fix: allocate too much memory for block max scores

[BubbleCal]

this could reduce the memory footprint by 128x less for block max scores while indexing

[github-actions]

Code Review

Summary: Good fix for a memory over-allocation bug. The PR correctly reduces pre-allocated capacity from length (number of documents) to num_blocks (number of blocks), which is ~128x smaller since BLOCK_SIZE = 128.

Feedback

No P0/P1 issues found. The changes are correct and well-tested.

Minor observations (non-blocking):

• The idf_scale hoisting is a nice micro-optimization that avoids redundant computation in the hot loop.
• Test correctly validates both len() and capacity() to ensure the fix works as intended.

LGTM ✓

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!