perf(search): stored tsvector column + saner short-term prefixing for ⌘K search

[16francej]

What

Speeds up the ⌘K chat search dramatically by fixing where the query actually spends its time.

1. Stored search_tsv generated column, GIN-indexed. The previous index was an expression index over a plpgsql function that JSON-parses the payload. Prefix tsqueries force a recheck of every lossy bitmap candidate, and each recheck re-ran that parse. On a 1M-row / 878MB benchmark corpus this was ~96% of query time: 102.6s → 4.0s for a representative two-term query, identical result sets. The recheck now reads a stored column instead.
2. Short terms match whole words instead of prefixes. A 1–2 char prefix (de:*) expands to an enormous GIN key set — 14.7s inside the index scan alone on the same corpus. Terms under 3 chars now match exactly; the in-memory store's mirror (matchesSearchTerms) follows the same rule so both stores agree.
3. entry_search_text is marked PARALLEL SAFE (it's pure), unblocking parallel plans.

Migration notes

• The new column is GENERATED ALWAYS … STORED, so the ALTER TABLE rewrites session_entries once at boot. The replacement index builds CONCURRENTLY (same machinery as before), then the old expression index is dropped.
• searchEntries output text still uses entry_search_text(payload) for the ≤40 returned rows — cheap, and avoids storing the text twice.

Benchmarks (1M entries, worst-case dense vocabulary)

Query	Before	After
deploy:* & error:*	102.6 s	4.0 s
de:*	92.9 s	16.8 s (now exact-match, far cheaper still)

Tests

• test/entry-search.test.ts — new cases for short-term exact matching in both tsPrefixQuery and matchesSearchTerms.
• test/session-search.test.ts and test/postgres-store.test.ts (against a real Postgres) pass, exercising searchEntries through the new column and index.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}