code-mode: clean up terminal cell dispatch gates

[cconger]

Why

A yielded cell can complete in the background while its terminal result remains buffered. Waiting for another observation or actor shutdown before cleaning Core's dispatch state leaves one gate allocated per unobserved completion. Terminal notification can also race ahead of Core installing that gate.

What

• Reintroduce cell_closed(&CellId) -> () on the runtime and session delegate contracts as a synchronous, non-blocking terminal notification.
• Emit it once after terminal completion commits, even when the ObserveOutcome remains buffered.
• Deliver an already-attached ObserveOutcome::Completed or ObserveOutcome::Terminated before invoking terminal cleanup.
• Keep actor shutdown as an idempotent fallback for termination paths.
• Remove Core's ready dispatch gate when the notification arrives.
• Linearize close-before-ready by retaining a lightweight closed marker; a later readiness registration consumes the marker instead of reopening a gate.
• Cover background completion without another observation and the close-before-ready ordering.

Host/Core sequence

sequenceDiagram
    participant Core
    participant Host
    Host->>Host: commit terminal CellEvent
    opt observer is attached
        Host-->>Core: ObserveOutcome::Completed or ::Terminated
    end
    Host-->>Core: cell_closed(&CellId) -> ()
    alt dispatch gate already exists
        Core->>Core: remove gate
    else readiness is not registered yet
        Core->>Core: retain closed marker
        Core->>Core: later readiness consumes marker
    end
    Note over Host,Core: cell_closed has no acknowledgement or retry

Loading

Stack boundary

This notification releases Core dispatch state; it does not acknowledge delivery of the buffered wire outcome. Observation replay storage is addressed separately by #29398 and #29401.

Validation

• just test -p codex-code-mode
• just test -p codex-core code_mode
• Core regressions cover terminal notification, background completion without another observation, and close-before-ready.

Stack parent: #29292.

[jif-oai]

This still has a close-before-ready race...
A fast yield_control(); ...complete... cell can call cell_closed before mark_cell_ready_for_dispatch, so this removes nothing. Then, readiness creates a fresh gate, and the initial observe only returns thebuffered yield

[jif-oai]

Vacant is not necessarily “closed before ready.” InitialObservationGuard already closes the ready gate before asynchronously terminating; the later cell_closed (or the guard if runtime close wins) lands here and installs a marker after the only readiness call has passed.
Nothing consumes it, so cancelled initial observations still leak one entry each.

Can we linearize the ready, terminal, and consumer-finished events and cover this cancellation path?

[jif-oai]

A successful commit can still produce CompletionDelivery::Rejected: terminate() may claim CellPhase::Completed between the commit await and deliver_completion. Sound pretty racy to me