fix(codex): late-child cancellation waits on its serial reader #331

What: A late Codex child outside an explicit cut-in can block the native reader for ten seconds while waiting for its own interrupt reply.
Why: Its turn-start handler awaits a request whose acknowledgement the same serial reader must parse.
So what: Reproduce and repair the ordinary lifecycle path separately; #285 removes the request only for its governed urgent cohort.

Outcome and proof

CodexAdapterV2.emitSubagentProviderTurnStarted awaits turn/interrupt under turnTerminalizationPermit. The actual effect-codex-app-server protocol/client await each notification handler before reading the next input line. No native CLI failure is claimed. Reproduce using the actual stdio client: gate the interrupt answer, send a sentinel behind it and require that sentinel to be consumed without advancing a timeout. Register the child before bounded cancellation outside the reader; preserve native outcomes, ordinary Stop, goal continuations and exact identity.

This advances truthful Promachos/Drafter work in the Objective. This is source-level evidence on555f0a00198b7aba75d8e8ce13af9c7d983712f4, independently inspected after Opus R7 identified it, not a native runtime reproduction. Searches of open/closed Codex issues found no dedicated report; #299 concerns soft-restart subscriber handover, a related but different mechanism.

Non-goals and blockers

No expansion or waiver of #285, shared app-server kill, version, deployment or live data mutation. A faithful reproduction and current upstream search are required before selecting implementation. Use actual adapters/services and native I/O gates, Deferred/persisted milestones and no sleeps or polling.

Requirements and who asked: Global Work requires recording adjacent problems discovered by #285's design; the user requires truthful native work and useful Stop controls.
Deleted: No new cancellation architecture, public intent, liveness registry, polling barrier or unsupported OS-exit promise.
Bottleneck: The handler and its awaited reply share one serial reader.
Checked myself: Actual late-child branch, permit/timeout, protocol/client notification ordering, R7 D-1 and related issue searches; runtime proof remains pending.