Scheduler watchdog — observability for scheduled tasks

OVERRIDES runtime profile: uses direct_python (plain Python, no model) because the command chain never invokes the claude CLI or any LLM API — pure-infra task; scheduler provider env injection is a no-op (MC-4942 U12 sweep).

Hourly watchdog that detects silent scheduler failures and creates MC tickets for Luci to investigate.

Four checks, each produces dedup'd MC tickets (stable identifier embedded in title prefix [watchdog:<check>:<key>]):

• A. YAML parse — scans scheduler.log for WARNING: Skipping <file> — bad frontmatter. Catches race conditions where the scheduler trips on a task file while syncs are writing it.
• B. Missed fires — for every enabled task, computes the previous expected fire time via croniter. If task_runs has no entry within a 15-min grace window past that time, flags it. Applies to ALL cadences (daily, weekly, monthly).
• C. Execution failures — task_runs rows with status failed or timeout in the last hour.
• D. Alert channel — counts Telegram send failed entries in recent scheduler.log tail.

Implementation: ~/workspace/scripts/scheduler_watchdog.py. Deployed 2026-04-19 after uncovering a mass scheduler drop (15 tasks stopped firing together Apr 16-17) that had been invisible because the scheduler's own alerting was also broken. See MC-957, MC-958, MC-959, MC-960.