Problem

Session search (and especially global search across projects) is slow because every search re-parses and
re-processes each JSONL file from scratch.

Per session file searched, the current code:

1. Full file I/O — streams entire JSONL
2. Full JSON parsing — parses every line into ParsedMessage
3. Full chunk building — buildChunks() including message classification, semantic step extraction, tool
   execution tracking, subagent linking
4. Then finally performs the text match

For a project with 100 sessions, that's 100 full parse+build cycles per search. Global search multiplies this
across all projects.

Root Cause

SessionSearcher.searchSessionFile() calls parseJsonlFile() + buildChunks() on every search. The
DataCache exists but is not used for search.

Search only needs user message text and AI output text — it doesn't need full chunk building.

Proposed Fix

1. Lightweight text extraction — skip buildChunks() entirely for search. Extract only searchable text
   (user message strings + assistant text content blocks) directly from parsed messages
2. Search cache — cache extracted searchable text per session, invalidate via mtime check. Extend existing
   DataCache or add a dedicated search cache
3. (Stretch) ripgrep for initial filtering — use rg to search JSONL files as raw text first, then only
   parse matching files for result metadata