Why

Part of #1791. Multiword Postgres full-text searches repeatedly fetch and decompress chunk vectors to determine which query words each document contains. This is expensive on large projects even after GIN candidate selection.

What Changed

Collect each operand's parent/chunk matches through GIN, group by project and document identity, and project a document-level operand vector once behind the existing candidate fence. Preserve Boolean semantics, cross-chunk matches, scores, and exact counts. Add semantic and query-plan regression tests plus a reproducible local benchmark.

Implementation Details

The grouped operand matches also provide candidate keys, avoiding a second candidate scan. Pure/optional negation retains the scoped document universe. Single-word and mixed-script queries retain their existing path.

On a generated 32,000-note / 320,000-chunk fixture, compared with main 976a10c89, second-run inclusive buffer touches fell:

These are local SQL measurements, not production latency guarantees. See docs/postgres-fts-large-project-benchmark.md for methodology and reproduction.

Testing

• BASIC_MEMORY_TEST_POSTGRES=1 LOGFIRE_IGNORE_NO_CONFIG=1 uv run pytest -q tests/repository/test_postgres_search_repository.py --no-cov — 40 passed.
• BASIC_MEMORY_TEST_POSTGRES=1 LOGFIRE_IGNORE_NO_CONFIG=1 uv run pytest -q tests/repository/test_postgres_operand_search.py tests/repository/test_search_repository.py tests/repository/test_script_ngrams.py tests/services/test_search_service.py tests/api/v2/test_scoped_search_router.py --no-cov — 211 passed, 19 skipped.
• just fast-check, just doctor, and uv run ty check scripts/benchmark_postgres_fts.py — passed.
• The query-plan regression fails against the baseline compiler and passes with this change.
• Full result identities and score floats matched baseline for six queries on the large fixture.
• Benchmark seed/measure smoke passed with 20 notes and 200 chunks.

Risks / Follow-ups

Broad single-word ranking remains expensive: archive matched most fixture documents and took 30–40 seconds with roughly 3.1 million buffer touches. This PR does not close #1791. Further work should reduce chunk ranking reads without truncating candidates or changing scores. Production/Neon cold-cache and end-to-end API latency remain unverified.