bug: Watchdog not running after session resume — 11-minute stall with no detection

[PureWeen]

Problem

After app restart, a resumed session can get stuck with IsProcessing=true and no watchdog running. The session stalls indefinitely until the user manually aborts.

Evidence (2026-04-04)

20:31:51 [RESUME-QUIESCE] 'PolyPilot' — abort + 30s quiescence
20:32:35 [COMPLETE] session completed normally (gen=1)
20:35:11 [SEND] new prompt (gen=2)
20:35:14–20:36:26 — active tool rounds (turn_start/turn_end every ~10s)
20:36:26 [turn_start] — LAST EVENT
  ... 11 minutes of silence, ZERO watchdog entries ...
20:47:01 [ABORT] user manually aborted

No [WATCHDOG] entries for the PolyPilot session between 20:36 and 20:47. The watchdog was not running.

Expected

The watchdog should fire at 120s (no tools) or 180s (HasUsedToolsThisTurn) after the last event, detecting the stall and clearing IsProcessing.

Root Cause (suspected)

The watchdog timer may not be started (or may be killed) during the RESUME-QUIESCE → COMPLETE → new SEND cycle after restart. The StartProcessingWatchdog call in SendPromptAsync may be racing with the quiescence cleanup.

Related but distinct from

• bug: Zero-idle session fallback still skips for tool-using sessions (PR #470 not merged) #494 — TurnEnd fallback skips for tool sessions (watchdog IS running, just slow)
• fix: queue bridge prompts during session restore (#450) #488 — IDLE-DEFER timeout fix (watchdog IS running, wrong timeout tier)
• Workers stop when relaunch.sh rebuilds the app mid-orchestration #400 — Workers stop on relaunch (different failure mode)

Repro

1. Have PolyPilot running with an active session doing tool calls
2. Restart the app (relaunch.sh)
3. Session resumes, starts a new turn
4. If the CLI stops sending events mid-turn, the watchdog never fires