Session list performance: disk summary cache, display-name cache, and count alignment

[clean6378-max-it]

Closes #109

Large installs were painful on cold start — every session list request re-parsed every .jsonl, and the landing page reopened files on every load just to get display names and counts. The in-memory LRU from #82 helps within a run but dies on restart.

This adds a SQLite summary cache (~/.claude-code-chat-browser/session_summary_cache.sqlite) keyed on path + mtime + exclusion-rules fingerprint. get_project_sessions skips the full parse when we already have a complete row. get_projects goes through the same cache for peek data instead of calling quick_session_info on every file every time.

Also cached display names in session_path.py — keyed on the project dir and max .jsonl mtime, so repeat landing page loads don't reopen files when nothing changed.

Project card session_count now uses the same titled-session filter as the session list when there are no exclusion rules. If exclusions are on, the card count can still diverge (same trade-off we accepted on cppa).

Partial cache rows from peeks get upgraded to full rows on the first session-list visit. mtime change = cache miss.

Files: utils/session_summary_cache.py (new), api/projects.py, utils/session_path.py, tests in test_session_summary_cache.py + integration/path tests.

To verify: run the test suite; optionally restart the server, hit a project session list twice — second request shouldn't touch get_cached_session for unchanged files. Touch one .jsonl and confirm only that one re-parses.

FTS search / search UX is separate — that's the Monday PR.

Summary by CodeRabbit

• New Features
  • Project and session listings now leverage a disk-backed session summary cache keyed by path, file mtime, and exclusion rules to speed up repeat requests.
  • Added a cache for project display names to reduce repeated JSONL scanning.
• Bug Fixes
  • Improved session counting/filtering by aligning untitled/excluded handling with cached summary completeness.
  • “Last modified” project activity is now derived from cached summary timestamps.
• Tests
  • Added/updated integration and cache-coverage tests, including cache reuse, invalidation, and LRU eviction behavior.