perf: parallelize FTS prewarming

[BubbleCal]

20%+ faster for 2GB index, could be more for larger index

[github-actions]

ACTION NEEDED
Lance follows the Conventional Commits specification for release automation.

The PR title and description are used as the merge commit message. Please update your PR title and description to match the specification.

For details on the error please inspect the "PR Title Check" action.

[github-actions]

PR Review

P1: Memory retention in the modern prewarm path

In the modern (compressed) prewarm branch, each CompressedPostingList is constructed via:

postings.value(idx).as_binary::<i64>().clone()

This is a shallow clone — all cached entries retain Arc references to the full batch data buffer. Unlike the legacy path which calls shrink_to_fit() to compact each slice into independent memory, the modern path skips this. The comment says this is because "DeepSizeOf uses logical byte size," but the concern isn't accounting — it's actual memory behavior.

When cache eviction later removes individual posting lists, no memory is freed until every entry from that batch is evicted. If this index has many tokens, eviction becomes effectively a no-op for memory reclamation. Consider either:

• Calling shrink_to_fit on the sliced binary array (or equivalent deep copy), or
• Documenting that this is intentional because compressed posting lists are expected to always be fully cached

P1: No tests

The PR description says "Not run (not requested)," but repository standards require tests for all changes. The prewarm parallelization and the new modern layout code path both introduce new behavior that should be covered. At minimum, a test verifying that the modern prewarm path correctly populates the cache would catch regressions.

Minor observations

• The list_i32_used_bytes helper assumes the child array element size is always size_of::<i32>(). This is correct for current usage (positions), but the function name doesn't communicate this constraint. Consider renaming or adding a comment.
• PR description mentions "worker error handling" and "monitoring loop for slow posting lists" but the diff only contains prewarm and DeepSizeOf changes. Worth updating the description to match actual changes.

[codecov]

Codecov Report

❌ Patch coverage is 83.33333% with 22 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
rust/lance-index/src/scalar/inverted/index.rs	83.33%	16 Missing and 6 partials ⚠️

📢 Thoughts on this report? Let us know!