fix: deduplicate stale BTree entries during optimize with stable row IDs

[wkalt]

When stable row IDs are used, the BTree index stores stable row IDs (not physical addresses) in its _rowid column. During optimize, the old entries for updated rows would survive fragment-based filtering because stable row IDs don't encode fragment IDs in their upper 32 bits. Both old (stale) and new entries for the same row ID ended up in the merged index, causing FlatIndex::try_new to fail with "RowAddrTreeMap:: from_sorted_iter called with non-sorted input" due to duplicate IDs.

Fix: before merging old and new data in combine_old_new(), collect the new data's row IDs and filter old entries that have matching IDs. This removes stale entries for rows that have been updated.

[wkalt]

This PR was generated automatically from the output of some generative testing

[github-actions]

Review

Solid correctness fix for a real bug. The approach — collect new row IDs, then filter stale entries from the old stream — is straightforward and correct.

P1: Consider RoaringTreemap instead of HashSet<u64>

The codebase convention (per CLAUDE.md) is to prefer RoaringBitmap over HashSet<u32> for memory efficiency. For u64 keys, the equivalent is RoaringTreemap, which is already used in 13+ files across the repo. During a large optimize where many rows were updated, the HashSet<u64> could use significantly more memory than a RoaringTreemap. Since roaring is already a dependency here, the change would be minimal:

use roaring::RoaringTreemap;

let new_row_ids: RoaringTreemap = new_batches
    .iter()
    .flat_map(|batch| { ... })
    .collect();

No other issues found — the test coverage is good and the fix is well-targeted.

[codecov]

Codecov Report

❌ Patch coverage is 92.50000% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
rust/lance-index/src/scalar/btree.rs	92.50%	0 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!

[Xuanwo]

This will collect all our new data, seems bad.