fix(memcard): preserve watchdog timeout reasons across completion

This commit is contained in:
2026-10-14 10:26:30 +01:00
parent e0e01ed1de
commit bad942fb25
4 changed files with 314 additions and 25 deletions
@@ -236,6 +236,32 @@ not proof of rollback. If cessation is unconfirmed, isolate mount/resources,
block new work and retain source/scratch. Late completion belongs to original
reconciliation. Never kill threads unsafely, free buffers in use or create a
replacement worker. Report restart requirement when host recovery is impossible.
The first watchdog cancellation is recorded independently of the request result,
before signaling cancellation. Repeated polls preserve that cause. At 500ms the
request remains Pending with no terminal reason until the worker returns or the
1s escalation occurs. Removal or session termination that canceled first does
not become a watchdog timeout merely because a later poll crosses 500ms.
On completion before escalation, the worker's durable outcome determines
Success, Failure or Indeterminate. A recorded watchdog cancellation supplies
Timeout (guest reason 15) independently of that state; native diagnostics retain
the original worker storage error. For example:
- A save canceled while queued at 500ms, then completed at 501ms without entering
publication, reports Failure/15. The durable slot is unchanged.
- A save already inside publication at cancellation can finish successfully and
report Success/15. Its saved snapshot becomes clean; cancellation is not rollback.
- An uncertain publication reports Indeterminate/15 and requires reconciliation.
- At 1s without confirmed cessation, RestartRequired reports Failure/15 before
publication entry or Indeterminate/15 after entry. Late completion returns card
ownership for reconciliation but cannot rewrite this terminal guest result or
mark the snapshot clean. Releasing the result does not release in-use resources.
A deferred bootstrap save that has not reached the worker expires as Failure/15
at 500ms. Memory-only commits remain MemoryOnly with reason 0 and perform no IO.
The status/state/reason result slots and request identities do not change.
Fast-profile claims require measurements under asset load, including tail latency.
## 8. Player choice and lifecycle