Persistent Luci watchdog — liveness keep-alive + context escalation ladder

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).

Runs every 5 minutes. Two independent jobs each tick:

1. Liveness keep-alive (MC-3763). Checks persistent_luci.status(). If the luci-persistent orchestrator session is dead, restarts it via ensure_started() and Telegram-alerts. A 2-strike grace (restart only on the 2nd consecutive offline observation) avoids fighting the daily 03:00 rotation while the fresh session boots. The strike counter is persisted to /tmp/luci-persistent-keepalive.json so systemd_watchdog.py corroborates orchestrator health hourly from the same source. Closes the post-OOM / post-reboot dead window from ~30 min (lazy heal) to ~5–10 min.

2. Context escalation ladder. Captures the luci-persistent tmux pane, parses the token counter from the Claude Code TUI footer, and escalates:

• >=700k tokens (soft) — sends /compact so old turns are summarised (defers if the pane is busy).
• >=820k tokens (critical) — sends /compact with a forced interrupt (does not defer — waiting risks overflowing the hard threshold).
• >=900k tokens (hard) — clear_persistent_luci(): sends /clear and then re-injects the continuity snapshot as the fresh session's first message, so orchestrator working state survives the context wipe (MC-3766).

Cooldown is 5 minutes by default but shortens to 2 minutes above the soft threshold so a fast climb gets enough attempts to escalate. Logs every decision to ~/workspace/logs/persistent-luci-watchdog.log.

Implemented for MC-2364 after Persistent Luci hit 201.2k tokens on 2026-04-28 and stopped responding cleanly. Liveness keep-alive added for MC-3763 after the orchestrator-robustness eval found the session had no real liveness supervisor.