Wynelson94/longhand

Bugfix release: per-project rollup counters were inflated, and sessions could be
filed under the wrong project. No schema-breaking changes; a one-time migration
repairs existing databases automatically on first open.

Fixed

• Project misattribution: sessions no longer drift to the wrong project.
  session.cwd (set by the Stop-hook live tail via MAX(cwd), a meaningless
  lexicographic pick) and project_id (set only by the full analysis pass) were
  written by independent paths and could desync — so work launched from $HOME
  ended up filed under the home catch-all project, invisible to per-project
  recall. On a real 265-session corpus, 35 sessions (13%) were affected. Now
  project attribution and cwd are always derived together from the same events
  (attribute_session_project()), and the live tail derives cwd the same way
  build_session does instead of MAX(cwd). New longhand reattribute command
  re-derives every session's project from the events table (not the
  transcript, which may have rotated away) to repair existing databases.
• projects.session_count and projects.total_edits no longer double-count.
  upsert_project() incremented both columns on every (re-)ingest of a session
  — SessionEnd, the live-tail Stop hook's analysis path, and reconcile
  re-ingests — so they counted ingest events, not distinct sessions. On a real
  corpus the home directory showed 2,068 "sessions" against 264 real ones (and
  53,952 edits against ~7,200 actual). The columns are now derived rollups,
  recomputed authoritatively from the sessions table
  (recompute_project_stats()) after each session is attached. Affected
  longhand projects, the list_projects MCP tool, and recall_project_status;
  raw session/event data, search, and recall were never affected.
• Migration v6 backfills existing databases, recomputing both columns for all
  projects from the sessions table on first open. Idempotent and non-destructive.