perf(search): stored tsvector column + saner short-term prefixing for ⌘K search (#597)

* perf(search): store the search tsvector in a generated column

The chat-search GIN index was built over an expression that calls a
plpgsql JSON parse per row, so every lossy bitmap recheck re-parsed
entry payloads — benchmarked at ~96% of query time on a 1M-row corpus
(102.6s -> 4.0s for a two-term query). Store the tsvector in a
generated column, index that, and query it directly.

Also stop prefix-expanding 1-2 character terms: a short prefix expands
to a huge GIN key set (14.7s inside the index scan alone for 'de:*').
Short terms now match whole words; the in-memory mirror follows suit.

The old expression index is dropped after the replacement index builds.
Note: adding the stored generated column rewrites session_entries once
at boot.

* style: prettier

---------

Co-authored-by: Josh France <josh@ycombinator.com>