Signal fatigue tracking: suppress repeatedly ignored signals

[tyejcoleman]

Problem

When the same signal fires every tick for days (e.g., CheckPendingWork for a task that stays open), the user gets desensitized. The current dedup system prevents spam within a short window:

• Fingerprint dedup: 2 × SignalCooldownNormal window (~30 min for act mode)
• Entity overlap dedup: 1h window, 85% overlap threshold
• Content hash dedup: 1h window

But none of these track fatigue across days. A stale task that fires CheckPendingWork every tick for a week will keep appearing once each cooldown expires.

Proposal

Add a signal_fatigue table:

CREATE TABLE signal_fatigue (
    entity_id TEXT NOT NULL,
    agent_id TEXT NOT NULL DEFAULT 'default',
    signal_type TEXT NOT NULL,
    memory_id TEXT NOT NULL,
    fire_count INTEGER DEFAULT 1,
    act_count INTEGER DEFAULT 0,      -- times user acted on it
    first_fired_at TEXT NOT NULL,
    last_fired_at TEXT NOT NULL,
    suppressed_until TEXT,
    PRIMARY KEY (entity_id, signal_type, memory_id)
);

Behavior changes

1. Recording: In evaluateShouldAct(), after signals pass confluence, upsert fatigue rows for each signal's memory IDs
2. Fatigue escalation:
   • fire_count 1-3, no user action: normal behavior
   • fire_count 4-6, no user action: double the cooldown for this signal+memory pair
   • fire_count 7+, no user action: suppress entirely (set suppressed_until to +48h)
3. Reset on engagement: When checkResponseTracking() detects a user response, reset fire_count and clear suppressed_until for all signal+memory pairs in that tick's fingerprint
4. Check in decision pipeline: In evaluateShouldAct(), filter out signals whose memory IDs are in the fatigue table with active suppressed_until
5. Cleanup: Remove rows where last_fired_at is older than 30 days (in decay job)

Where it fits in the pipeline

Insert between step 6 (signal classification) and step 7 (confluence scoring) in evaluateShouldAct(). Fatigued signals get removed before confluence is calculated, so they don't contribute to the score.

Files to modify

• storage/sqlite_heartbeat.go — new table, upsert/query/reset/cleanup
• storage/sqlite_migrate.go — migration
• heartbeat_decide.go — filter fatigued signals before confluence, reset on response
• watcher.go — record fire events after tick

Constraints

• Only applies to signals tied to specific memory IDs (CheckPendingWork, CheckDeadlines, CheckStaleMonitors, CheckDecaying, CheckConflicts)
• Does NOT apply to aggregate signals (CheckSentiment, CheckPatterns, CheckMemoryVelocity) since they don't have stable memory IDs
• Suppression cap: 48h max, then re-evaluate
• User action on ANY signal in the same tick resets fatigue for all signals in that tick (assumes engagement = attention)