code-mode: release acknowledged terminal observations

[cconger]

Why

An ObserveOutcome::Yielded value is implicitly acknowledged by the next generation. ObserveOutcome::Completed and ObserveOutcome::Terminated have no successor, so their replay payload otherwise remains retained until the session ends.

What

• Add ReleaseObservationRequest { cell_id, generation } and release_observation(ReleaseObservationRequest) -> Result<(), String> to the wire/session contract.
• Allow release only when the retained ObserveOutcome is Completed or Terminated.
• Do not release Yielded or Missing outcomes; reject them as non-terminal.
• Reject pending, failed, or future generations.
• Make repeated or older releases idempotent.
• Replace released terminal output with a lightweight generation tombstone.
• Have Core send the acknowledgement only after reading a terminal ObserveOutcome.
• Preserve the result and log a warning if release fails.

Host/Core wire sequence

sequenceDiagram
    participant Core
    participant Host
    Core->>Host: ObserveRequest { cell_id, generation, ... }
    Host-->>Core: ObserveOutcome::Completed or ::Terminated
    Core->>Core: consume terminal outcome
    Core->>Host: ReleaseObservationRequest { cell_id, generation }
    Host->>Host: replace payload with generation tombstone
    Host-->>Core: Result<(), String>
    opt release returns Err(String)
        Core->>Core: warn; consumed outcome remains valid
    end

Loading

Stack boundary

release_observation(ReleaseObservationRequest) acknowledges transport delivery of terminal output. It is separate from resume(PendingGeneration), which advances a pausable runtime frontier.

Validation

• just test -p codex-code-mode
• just test -p codex-code-mode-protocol
• Focused Core regression verifies release after terminal result consumption.

Stack parent: #29398.

[jif-oai]

A yielded generation also has no successor after wait(... terminate: true) no? should we retire this too?

[jif-oai]

This destructive ack happens before Core commits N + 1. This is racy

[jif-oai]

None cannot be an idempotent old release here because released generations remain as tombstones. This accepts any generation for an unknown or never-observed cell and hides wire desync. Should we reject this case instead?