v0.12.0 — update channel, honest episodes, attribution overhaul

Feature release: the update channel, cleaner attribution, sharper episodes,
and storage hygiene. One new migration pair (7 and 8) — both O(1) and applied
automatically on first open; existing data is never re-processed without your
say-so. Everything below was dogfooded on a real 331-session corpus before
release, which itself surfaced (and fixed) three long-standing bugs.

Added

• Update channel. longhand --version / -V prints the installed
  version; the interactive CLI checks pypi.org for a newer release at most
  once a day (cache-only hints, printed after your command's output, TTY
  only) and doctor gains a Version row. The check never runs from hooks or
  the MCP server — enforced by a test — never blocks a command, and
  LONGHAND_NO_UPDATE_CHECK=1 disables it entirely. The README privacy
  section discloses exactly what it does. (#46)
• longhand db vacuum (opt-in, never automatic) — reclaim disk space
  with a free-space preflight and the ingest lock held; --prune-aux
  first deletes stored harness-noise events. doctor now shows SQLite +
  vector store sizes. (#51)
• Reconcile sees more. New partially_indexed bucket for sessions whose
  ingest crashed mid-pipeline (repaired by --fix), and skipped_oversize
  for transcripts past the parser's size cap. Pre-existing rows are treated
  as complete — no re-analysis stampede on upgrade. (#48, #52)

Changed

• Attribution overhaul. The project-root walk now prefers the highest
  .git over nearer non-git markers (monorepo packages stop splitting into
  per-package projects); paths are normalized to the filesystem's own casing
  (case-dupe spellings mint one project); and sessions with no cwd or a
  markerless cwd in reserved locations ($HOME root, temp/pytest dirs, plugin
  caches, Claude Code internals) land in one explicit unattributed bucket
  instead of minting junk projects. longhand reattribute is now dry-run
  by default — review the move list, then --fix to apply and prune
  emptied projects. (#49)
• Episode extraction got honest. Verification requires an explicit
  success flag or an error-free result from a command-executing tool —
  Read/Grep output no longer marks episodes "resolved". A new error always
  starts a new episode instead of being swallowed by the previous one.
  Harness plumbing (<task-notification> and friends) is never a problem
  anchor, and the leaked Ask: scaffold is gone from problem descriptions
  (migration 8 cleans existing rows). The find_episodes MCP tool now
  defaults to min_confidence: 0.5 (pass 0 for everything); low-confidence
  episodes are filtered by default, never deleted. Historical labels change
  only when you run longhand analyze --all. (#50)
• analyze --all now reaches your whole corpus and truly replaces.
  Sessions whose transcripts rotated off disk are rebuilt from the events
  table (previously ~75% of a mature corpus was unreachable), and
  re-analysis deletes a session's old episodes/segments before extracting —
  no more stale-boundary duplicates accumulating in SQLite and vector
  search. (#53, #54)
• The prompt hook never embeds inline. The first-recall-after-upgrade
  episode backfill (up to the whole corpus, synchronously, inside the
  UserPromptSubmit hook) is now a detached backfill-episodes worker that
  claims the ingest lock. (#47)

Fixed

• Harness UI-state entries (last-prompt, mode, permission-mode,
  attachment, ai-title, agent-name) are no longer stored as "unknown"
  events with full raw JSON — they were the 3rd-largest event type on a real
  corpus. (#51)
• The hook installer's no-PATH fallback wrote dead -m longhand.cli
  commands (the v0.11.2 background-spawn bug class); hooks installed without
  longhand on PATH now actually run. (#52)
• The MCP server reuses one store across tool calls (it re-ran migrations
  and re-opened ChromaDB on every call) and wraps handler failures in a
  readable JSON error instead of a raw traceback. (#52)
• Session-id prefix lookup (longhand timeline <prefix> etc.) now searches
  the whole corpus in SQL — it previously scanned only the 1,000 most
  recent sessions, so older prefixes silently failed to resolve. (#52)
• Segment ids no longer collide when rebuilt events carry duplicate
  sequence numbers — a collision failed the session's whole embedding
  batch. (#54)
• The credit-card redaction pattern is skipped unless a cheap digit-run
  pre-check fires, cutting redaction CPU on numeric-heavy transcripts;
  masking behavior is unchanged. (#52)

Internal

• First direct tests for session_summary_embedding, segment_search,
  vector_store, and cli/helpers; suite now 421 tests. Third-party
  deprecation noise filtered in pytest config (6,426 → 95 warnings). (#52)