[wanglun]

C and C++ callers cannot restrict a scan or search to a known set of rows before ranking: a _rowid IN (...) filter is evaluated by a full row-id read plus an IN-list parse, which costs tens of milliseconds even for a few ids and seconds for large lists. Lance's Scanner::with_row_addr_prefilter does this natively; pylance exposes it as row_addr_allowlist / row_addr_blocklist.

Add lance_scanner_set_row_addr_allowlist and
lance_scanner_set_row_addr_blocklist, which take arrays of _rowid values (stable row ids when enabled, otherwise row addresses), and the C++ Scanner::row_addr_allowlist / row_addr_blocklist wrappers. The lists are combined like RowAddrMask::from_serialized_parts (allowed AND NOT blocked) and ANDed with any filter expression. An empty allowlist allows no rows.

A prepared FTS query context rebuilds its Match/Phrase execs around the shared global scorer, and the rebuilt execs did not carry the external mask, so the mask was silently dropped. The rewrite now re-attaches it, so prepared Match/Phrase searches score only the allowed rows with unchanged BM25 scores. Combining a row list with a segment-scoped scalar scan, which installs its own row mask, is rejected.

Tests cover plain scans (allow, block, both, empty, replacement, AND with a SQL filter) on a two-fragment dataset with stable row ids, invalid inputs (NULL with len > 0, setting after the scan starts, scalar_index_segment), plain and prepared full-text search (OR and AND operators, scores equal to the unmasked search), and nearest-neighbor top-k over allowed rows. C and C++ runtime tests exercise the exported symbols and wrappers.

Validation:

• cargo fmt -- --check
• cargo clippy --all-targets -- -D warnings
• cargo test
• cargo test --test compile_and_run_test -- --ignored --test-threads=1

[lance-gatekeeper]

✅ Gate recommendation: approve.

The native row-mask approach preserves selection before ranking and shared BM25 scoring while keeping the C and C++ bindings thin. The new Rust tests, prepared-FTS regressions, additional phrase/top-k and indexed-vector checks, and both C/C++ runtime suites passed.

[wanglun]

@jja725 thanks for the review, it seems I cannot merge it, can you help? thanks!