[codex] Fix critical search indexing and performance

[Techie5879]

What changed

This PR fixes the smart session picker indexing path so search no longer rebuilds the whole corpus or reopens/migrates the sidecar database on each search path. It adds a shared sidecar connection, WAL/busy timeout settings, incremental session-level indexing, schema-version cache reset, FTS5-safe rebuild/delete handling, and vector-state preservation across opens.

It also tightens the search corpus and picker behavior:

• indexes user-visible user message content plus session titles, while skipping assistant/tool/subtask/agent internals
• runs keyword and vector search independently so semantic-only matches can rank
• caches fzf and embedding health checks with timeouts
• narrows fzf candidates through indexed keyword hits before invoking fzf
• keeps empty-query picker behavior instant and recency-ordered
• preserves the shadowed session.list command path with higher-priority keymap bindings

The second commit adds deterministic fuzzy-search performance coverage, an opt-in readonly live benchmark, and upstream API verification notes.

Why

The root issue was that the plugin treated index invalidation as a reason to rebuild too much state too often. Search and preview paths could repeatedly open the sidecar, write metadata, rebuild external-content FTS rows, and fetch full message corpora even when only one session changed. That created latency, write contention, and an FTS corruption risk when external-content rows did not mirror the content table exactly.

Impact

Users should see faster picker startup/search refreshes, safer cache recovery, less log noise, and more predictable hybrid/fzf behavior on large workspaces. The sidecar remains plugin-owned derived data, so schema mismatches or corruption are handled by deleting and rebuilding only the cache.

Validation

• bun run test
• bun run test:perf