fix(mcp): record_turn chat fallback + most-relevant-thread selection

[momo590]

Summary

Two MCP bugs surfaced by the v0.2.3 + v0.3.0 combo on Claude Desktop chat:

Bug 1 — record_turn failed on every chat call

Two compounding causes:

• _resolve_thread_id_for_write had no fallback once the v0.2.3 migration retired ~/.tinm/current_thread. Claude Desktop chat has no cwd, so every call without an explicit thread_id raised ValueError.
• The success-string formatter dereferenced result['alpha_used'], but the v0.3.0 hot path moved alpha/update_count ownership to the async worker and no longer returns them → KeyError.

Fix:

• Add a per-host chat-thread fallback persisted in ~/.tinm/chat-thread. Auto-created once on first use with origin=user, project_root=None. Disambiguates to claude-desktop-chat-<sha6> if the bare slug already belongs to a real project. Last-resort only — the explicit-slug path and the legacy CURRENT_FILE path are unchanged.
• Make the formatter use .get() everywhere with a — placeholder for missing alpha. Degrades cleanly on both v0.3.0 hot path and legacy sync path.

Bug 2 — Resume loaded the avant-dernier thread

Two root causes in _most_recent_thread:

• 2a: Empty SessionStart auto-creates spawned by cron/autopilot from /root or /tmp had their last_updated bumped to "now" every minute, shadowing the user's real WIP thread.
• 2b: Same-last_updated threads fell back to Path.glob() order — ext4 hash-table order, i.e. non-deterministic across runs.

Fix:

• Replace the scalar timestamp comparison with a four-tuple relevance score (is_substantive, last_updated, n_user_turns, slug), extracted into _thread_relevance_score() so the read-path resolver can adopt the same ranking later.
• Substance gate: is_substantive = 1 iff n_user_turns ≥ 1 AND last_updated != created_at. Empty autocreates score 0 and are only chosen when no substantive thread exists.
• Slug as final tiebreaker eliminates the glob-order randomness.

Test plan

• pytest mvp/tests/test_record_turn.py — 10 new tests, all green (resolver paths, chat-fallback create/persist/collision/recovery, legacy CURRENT_FILE precedence, formatter never crashes on missing alpha).
• pytest mvp/tests/test_thread_selection.py — 12 new tests, all green (pure score helper + integration: substantive beats recent-empty, same-second multi-turn wins, only-autocreates falls back, corrupt/missing dirs handled).
• pytest mvp/tests/ full suite — 420 passed, 1 pre-existing skip. No regressions.

Notes

• Constraint adherence: two bug fixes in two separate commits on this branch.
• Async worker / embedding code untouched.
• ~/.tinm/current_thread was NOT reintroduced as the resolver — the new fallback uses a distinct per-host file (~/.tinm/chat-thread) that the read resolver never consults.

🤖 Generated with Claude Code