Design Rationale

1. Per-task instead of global watchdog

Hardware watchdogs detect total system freeze. microwdt detects per-task stalls — you know WHICH task hung, not just THAT something hung.

2. Three-state model (OK → LATE → STARVED)

LATE = missed one deadline (warning). STARVED = missed max_misses (critical). This gives gradual escalation — log a warning first, reset only if it doesn't recover.

3. Kick-based (not timer-callback based)

Tasks report liveness by calling mwdt_kick(). This is simpler and more natural than registering timer callbacks. The task knows when it completed meaningful work.

4. Edge-triggered callbacks

Like microhealth: fires only on state transitions. A task stuck at LATE for 100 checks = 1 event, not 100.

5. max_misses = 0 for warn-only

Non-critical tasks (display, LED) can miss deadlines without triggering a reset. Set max_misses=0 and they stay LATE forever without escalating to STARVED.

6. auto_reset is per-task

A sensor task stalling might just need a warning. An MQTT task stalling might need a full system reset. The decision is per-task, not global.

Decision	Gains	Costs
Per-task	Know which task stalled	One kick call per task per cycle
Three states	Gradual escalation	Slightly more complex than bool
Kick-based	Natural for task loops	Task must remember to kick
Edge-triggered	No spam	Misses sub-check oscillations
Per-task auto_reset	Granular control	One more config per task