fix(index): prevent maintenance from undoing index drops

[lance-gatefixer]

Summary

• reject stale index replacement when a concurrent transaction removes the exact source segment
• reject append-style index segments when the same logical index is concurrently removed
• keep pure concurrent drops and disjoint physical-segment changes compatible

Root cause

CreateIndex conflict detection compared only new_indices names. A drop is encoded with an empty new_indices list, so stale optimize or append operations could be rebased after it and publish index metadata again.

Validation

• cargo test -p lance io::commit::conflict_resolver::tests --lib
• cargo fmt --all -- --check
• cargo clippy --all --tests --benches -- -D warnings

Fixes #8982

[lance-gatefixer]

Blocked: PR #8984 remains blocked by an unrelated flaky Rust linux-build check at head 7c77814.

The linux-build check (Actions run 33836370883, job 100909623017) failed in test_create_ivf_hnsw_pq_multivec because recall was 0.4, below the required 0.5. The PR changes only rust/lance/src/io/commit/conflict_resolver.rs; focused runs passed repeatedly before reproducing the identical failure after 36 consecutive passes, and all 65 conflict-resolver tests pass.

Please re-run the failed linux-build job. Alternatively, stabilize the flaky vector-recall test independently and then re-run CI.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The resolver now rejects stale replacement and append-style index maintenance when a concurrent drop invalidates the relevant source identity, while preserving idempotent drops and disjoint segment changes.

[lance-gatefixer]

Blocked: PR #8984 remains blocked at head 7c77814 by the unrelated flaky Rust linux-build check.

The current-head linux-build check (Actions run 33836370883, job 100909623017) failed in test_create_ivf_hnsw_pq_multivec because recall was 0.4, below the required 0.5. Same-head focused runs pass repeatedly, with a stress loop reproducing the identical assertion only after 36 consecutive passes; the repair changes only conflict-resolution logic, all 65 conflict-resolver tests pass, and every other CI check passes.

Please re-run the failed linux-build job. Alternatively, stabilize the flaky vector-recall test independently and then re-run CI.

[majin1102]

LGTM. The conflict semantics and regression coverage look correct. The nested predicates could potentially be simplified into normalized index-change sets in a follow-up, but I don’t consider that blocking.