test(index): cover KeepFiniteVectors row filtering

[jackylee-ch]

KeepFiniteVectors runs on every IVF build path — five construction sites in vector/ivf.rs — and has three behaviours: drop rows whose vector is null, drop rows containing NaN or an infinity, and error on a column that is not a fixed size list. None of them was tested. The existing tests in this module exercise the is_all_finite free function and the l2 overflow bounds, never the transformer itself; grep KeepFiniteVectors::new returns only the five production sites.

Adds coverage for all three behaviours, plus the all-null and empty-batch cases AGENTS.md asks for in index tests, and the missing-column pass-through. The non-list case also pins the error message, which already names the column and the offending type.

Testing

• cargo test -p lance-index --lib vector::transform — 12 pass
• Confirmed test_keep_finite_vectors_drops_null_and_non_finite_rows fails when the validity check in transform is bypassed, so it is a real regression test
• cargo fmt --all -- --check, cargo clippy -p lance-index --tests -- -D warnings

[lance-gatekeeper]

✅ Gate recommendation: approve.

The transformer-level tests directly pin null and non-finite row filtering, empty and all-null inputs, missing-column pass-through, and contextual type errors. This is the right boundary for protecting the shared filtering behavior used by the IVF construction paths.

[jackylee-ch]

The red checks here are not from this change — main at 018386c (#8024) does not compile. rust/lance/src/index/vector/ivf/v2.rs:660 calls ReaderProjection::from_column_names, which no longer exists, and rust/lance-file/src/reader.rs references LanceFileVersion without importing it.

Same run on main itself: https://github.com/lance-format/lance/actions/runs/30903838621 — clippy, rustdoc, MSRV, cargo-deny, linux-arm, mac-build, windows-build, build-no-lock, query-integration-tests, linux-coverage-build all fail there too, the same 10 jobs that fail here.

cargo-deny is separately red on main for RUSTSEC-2026-0233/0234/0235 (rkyv 0.8.16, fixed in 0.8.17).

This PR only appends to #[cfg(test)] mod tests in rust/lance-index/src/vector/transform.rs and touches nothing else; cargo test -p lance-index --lib vector::transform passes locally against 51da66d. Happy to rebase once main is green.

[jackylee-ch]

Rebased onto d95cb36 to pick up #8261, which fixed the main breakage.

Correcting my earlier comment: the first bad commit was a2d3047 (#8227), not #8024 — a duplicate LanceFileVersion import in the lance-file reader tests, as #8261 describes. My diagnosis that the failures were pre-existing on main was right, the attribution was not.

rust/lance-index/src/vector/transform.rs was untouched by the 16 intervening commits, so the rebase was clean. Re-verified at the new base: cargo test -p lance-index --lib vector::transform 12 pass, cargo fmt --all -- --check, cargo clippy -p lance-index --tests -- -D warnings.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The rebase preserves the previously reviewed transformer-level coverage unchanged, and the focused tests pass on the new base. The direct unit boundary continues to protect the shared filtering behavior across the IVF construction paths.