Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/orchestrator/factory.ts`:
- Around line 3149-3152: In the babysitter critical-section exit flow around
`#finishBabysitterCriticalSection` and deferred wake processing, persist the
cleared critical fence immediately after consuming the end signal and before
submitting any deferred PTY input or wakes. Preserve the existing deferred
submission behavior, but ensure `#persistBabysitterCriticalFence` runs before that
processing so a crash cannot restore a permanent fence.
- Around line 4137-4149: The delivery-failure handling near the wake-delivery
path must check whether the wake was cancelled before restoring kinds or
persisting its session. Update the logic around `#cancelBabysitterWake` and the
failure branch to skip session recreation and related recovery for cancelled
state, preserving cleanup for active wakes.
- Around line 4495-4498: The readiness handling around the babysitter-enabled
guard must reject stale signals after PR ownership is cancelled. Before
resolving the agent, require that the babysitter still has active ownership and
that `#readBabysatPrSnapshot`() returns an authoritative snapshot; otherwise
cancel or terminate the babysitter and return. Preserve the existing non-OPEN
snapshot cancellation behavior.
- Around line 4505-4517: Update the ownership establishment flow around
`#inFlightIssueForPrSnapshot` and the babysitter reservation at the referenced
ownership block so a first match cannot claim ownership from title/body
references alone. Require either authoritative published-PR identity or a strong
head-branch association before creating the initial babysitter ownership; reject
weak same-repository references such as only “Fixes AR-420” when no existing
owner is present, while preserving established-owner mismatch handling.
- Around line 4331-4343: Update the babysitter wake transition around
`#recordPendingBabysitterWake` and `#flushBabysitterWake` so failures are handled by
recovery logic covering the entire flush sequence. On rejection, restore
deliveringKinds into kinds, schedule a retry through `#scheduleBabysitterWake`,
and ensure the promise chain ends with a catch so the ignored finally promise
cannot reject unhandled. Apply the same recovery to the corresponding wake path
noted at the additional location.
- Around line 3118-3120: Update the issue-key parser/normalizer feeding
durableIssue and babysitterCritical to accept GitHub keys in
owner/repository#number format, such as AgentWorkforce/hoopsheet#10, alongside
existing AR and numeric keys. Preserve normalization and ensure the resulting
key is recognized by both destructive-operation fence checks, including the
corresponding logic near the alternate location.

In `@src/state/file-state-store.ts`:
- Line 566: Update the persisted-record validation around candidate.critical so
missing or undefined values are rejected, requiring critical to be a boolean
rather than accepting undefined and restoring false. Apply the same strict
validation to the corresponding check at the other referenced location, while
preserving existing handling for valid boolean values.

In `@src/triage/schema.ts`:
- Around line 18-27: Remove ownedPullRequest and pendingPullRequestWake from
TriageDecisionSchema, keeping them available only on the internal AgentSpec
shape. Update LlmTriage or the decision-to-agent-state storage path to strip or
otherwise prevent these fields from parsed model output, while preserving
babysitter lifecycle code as the only source that sets them.

---

Outside diff comments:
In `@src/orchestrator/factory.ts`:
- Around line 7676-7679: Update parsePullSnapshot so fallbackNumber from the
canonical path is authoritative, rather than allowing payload.number to override
it. Validate the payload number when present and return undefined if it is not
an integer or disagrees with fallbackNumber; otherwise continue using
fallbackNumber for the resulting PullSnapshot.