fix(daemon): recover interrupt seq from the live watermark - #257

What and why

Follow-up to #255 / #254. Crash recovery assigned turn-interrupted from MAX(persisted seq). Live text-delta / reasoning-delta rows are not stored, so that sequence sits behind envelopes a client already received. Reconnect after=N then missed the terminal event.

A second hole: if turn-complete was persisted and finishTurn did not run, recovery appended turn-interrupted after an existing terminal.

How

• Persist turns.last_seq (schema 7) on every emit, including deltas that skip daemon_events.
• Recovery continues from max(last_seq, persisted max) + 1.
• If the last persisted event is already terminal, finish the turn to match it and do not append another.

Graceful stop() still drains via cancelAll + waitForIdle and ends with turn-complete. Planned SIGTERM already has a terminal event. This PR is crash recovery, not drain-before-update.

Verification

• bun run typecheck
• bun test (running)
• bun run build, and ran the resulting binary
• Negative control: the three new turns tests failed on current dev (after=4 returned no events; leftover turn-complete became interrupted; live deltas left last_seq at 0). After the fix they pass.

Cross-platform impact

• Touches file paths, file I/O, process spawning, signals, or shell invocation
• SQLite schema + integer watermark. No path or spawn changes. Schema 6 is already the schedule backoff columns from feat(scheduler): back off then auto-pause after consecutive failures #256; this is 7.

Notes for the reviewer

• Does not change graceful update restart. That path already terminalizes as turn-complete.
• last_seq is an integer update per live delta, not a JSON daemon_events row. Deltas stay omitted from history.