Summary

gbrain sync --break-lock --max-age N exists and self-heals wedged sync locks.
The cycle lock (gbrain-cycle in gbrain_cycle_locks) has no equivalent CLI
escape hatch. When a dream/autopilot cycle gets stuck (cancelled job, embed phase
ignoring abort signal, timeout), the cycle lock stays held indefinitely and every
subsequent gbrain dream skips with cycle_already_running.

This is documented in TODOS.md as a known issue. Garry hits it daily on his
production brain.

Reproduction

1. Run gbrain dream — starts an autopilot-cycle job
2. Cancel the job mid-run (or let it timeout during embed phase)
3. The gbrain_cycle_locks row stays held (ttl_expires_at in the future,
   or the autopilot daemon refreshes it before TTL expires)
4. Every subsequent gbrain dream → cycle_already_running → skipped
5. gbrain sync --break-lock only targets gbrain-sync, not gbrain-cycle
6. No CLI command exists to break the cycle lock

Current workaround

Manually DELETE FROM public.gbrain_cycle_locks in the database, then
immediately run gbrain dream --phase X for individual phases. Race the
autopilot daemon which re-acquires the lock within seconds.

Proposed fix

One or both of:

Option A: gbrain dream --break-lock [--max-age N]

Mirror the gbrain sync --break-lock --max-age N pattern for the cycle lock.
Clears any gbrain-cycle lock in gbrain_cycle_locks where last_refreshed_at
is older than N seconds, then proceeds with the dream cycle.

Option B: Autopilot self-heals stale locks on startup

When the autopilot daemon starts (or on each tick), check if the held lock's
last_refreshed_at is older than 2× the TTL. If so, break it and proceed.
This makes the autopilot self-healing without operator intervention.

Option C: Embed phase respects abort signals (TODOS.md fix)

The root cause from TODOS.md: "Embed phase ignores signal.aborted between
batches today. Job wall-clock timeout fires → handler keeps running → cycle's
finally block unreachable → gbrain_cycle_locks row stays held indefinitely."
Adding signal.aborted checks between embed batches would prevent the lock
from getting stuck in the first place.

Environment

• gbrain: v0.41.20.0
• Engine: postgres (Supabase EU/Ireland, Session pooler)
• Schema: v103
• Deployment: Railway (gbrain serve --http)
• Autopilot: via Minions queue (autopilot-cycle handler)

Impact

The cycle lock bug means the brain never self-maintains. No backlinks, no
orphan resolution, no enrichment, no pattern extraction. The brain score
degrades over time. The whole value proposition of "the brain gets smarter
while you sleep" breaks.