Better search: BM25 ranking, snippets, error surfacing

[ezwep]

Summary

Three concrete improvements to the conversation search:

• Relevance ranking via SQLite FTS5 bm25(). Title hits get a 5× weight and a 200pt absolute boost so a query that matches a title outranks one that only matches body text. Old behaviour (date-sort while searching) is preserved as an option.
• Snippets with highlighting. Each result now shows a <mark>-highlighted excerpt of the best matching message (or title/excerpt). Built on FTS5's snippet() function, sanitised in JS so only <mark> survives.
• Inline error reporting. Invalid FTS syntax (e.g. unbalanced quotes) used to silently return zero results; now a status banner under the search input shows the actual SQLite error and the parsed FTS query.

Plus a smarter query parser:

• Multi-word input → implicit-AND with prefix matching on each token (e.g. claude code → claude* code*).
• Quoted phrases, NEAR, NOT, and column filters pass through unchanged so power users keep full FTS5.

Code shape

get_conversations() now returns {conversations, search} instead of a bare list. The JS callsite handles both shapes (back-compat). When search is active each conversation row is annotated with snippet_html, search_score, and match_kind (title, message, or title+message).

Test plan

• Single-word search returns ranked results, snippet visible
• Multi-word search treats words as AND-prefix
• Quoted phrase "hello world" matches verbatim
• Invalid syntax "unclosed surfaces error in banner, no crash
• Empty/null search returns full list with no search_score annotation
• Manual UI smoke test in pywebview window
• BM25 weight tuning if title-boost feels too aggressive