fix: disallow brute force on index unless explicitly requested

[westonpace]

Currently an attempt to search an index with a distance type that does not match the trained distance type will result in a brute force search (with a warning logged). However, this should be an error. An implicit brute force search can be a very costly thing. We already have a way to request a brute force search (use_index=False) and that should be used if this is truly the behavior that is desired.

[github-actions]

Review of PR #6105

The behavioral change (error instead of silent brute force) is well-motivated and the test coverage is good. Two minor issues:

P1: Dead code after early return

After the change, the else branch at the metric mismatch now does return Err(...), which means use_this_index is always true when execution continues. This makes the if use_this_index { Some(...) } else { None } block (and the variable itself) dead code — the else { None } path is unreachable. Consider simplifying: remove the use_this_index variable entirely and just set matching_index directly, since the only two outcomes are now "metrics match → use index" or "metrics don't match → return error".

P1: Test name is stale

test_knn_metric_mismatch_falls_back_to_flat_search no longer describes the test behavior — it now tests that a metric mismatch returns an error. Consider renaming to something like test_knn_metric_mismatch_returns_error.

[codecov]

Codecov Report

❌ Patch coverage is 90.90909% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
rust/lance/src/dataset/scanner.rs	90.90%	0 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[wkalt]

this seems like a significant behavior change, since any clients previously doing the "wrong" thing but still getting results will now get errors

[wjones127]

I think this change is reasonable, but it should definitely be labelled as a breaking change.