Sweep overdue ledger rows into orchestrator_inbox

OVERRIDES runtime profile: uses direct Python (no claude CLI) because the sweep is a pure SQL scan over mc.db — no LLM call, no provider routing.

Every 30 minutes:

1. Parses any Expected check-in: line on open tickets and writes the typed tickets.expected_check_in_at column (idempotent backfill).
2. Lists open tickets and active runtime_sessions whose expected_check_in_at is in the past.
3. For each overdue ticket inserts one pending row into orchestrator_inbox with source_type='followup_sweep', deduped so a single overdue episode produces a single inbox row (matches the _has_open_row discipline used by orchestrator_triage).
4. Posts a single [follow-up] N min past expected check-in system comment on the ticket when the inbox row is created.

The orchestrator pickup loop drains orchestrator_inbox into the next Luci digest, so overdue check-ins surface in the standard control-plane channel instead of triggering ad-hoc Telegrams.

Closes gap §5 item 2 in runbooks/luci-delegation-control-plane.md.

Verification: python3 /home/lucienne/workspace/scripts/luci_followup_sweep.py --dry-run --timeout 90 --db-timeout 5 prints a JSON summary without writing.

The command uses an internal 90s wall deadline and a 5s SQLite busy timeout, both below the scheduler's 120s kill timeout. Dry-run opens the DB read-only so manual reproduction cannot run schema migrations/backfills.

Logs to the standard scheduler task_runs.output field — one JSON line per run with counts of backfilled, overdue, inserted and deduped rows. Progress is emitted as JSON lines on stderr and error JSON includes the active phase so a future scheduler failure shows where it stalled.