code-mode: expose transport-neutral session runtime

[cconger]

Why

An out-of-process code-mode host needs the same transport-neutral runtime used by the in-process adapter, without depending on actor internals. Cancellation also must not let a non-cooperative callback wedge cell termination or session shutdown indefinitely.

What

• Expose SessionRuntime, its request/event types, cell capability types, and delegate contract as public crate API.
• Keep the protocol adapter layered on that public runtime API; the wire ObserveOutcome and TerminateOutcome types from code-mode: expose create and observe operations #29291 do not change here.
• Give canceled callback tasks a bounded 100 ms grace period, then abort their local tasks.
• Remove the old actor-close cell_closed delegate callback from both runtime and session delegate surfaces at this layer.
• Add whole-value delegate-forwarding and event-mapping tests for the public adapter contract.

Host/Core callback sequence

sequenceDiagram
    participant Host as SessionRuntime host
    participant Core as Core delegate
    alt nested tool callback
        Host->>Core: invoke_tool(NestedToolCall, CancellationToken)
        Core-->>Host: Result<JsonValue, String>
    else notification callback
        Host->>Core: notify(call_id, CellId, text, CancellationToken)
        Core-->>Host: Result<(), String>
    end
    opt cell/session cancellation
        Host-->>Core: CancellationToken is canceled
        Host->>Host: wait up to 100 ms
        Host->>Host: abort unfinished callback task
    end
    Note over Host,Core: cell_closed is absent in this commit

Loading

Stack boundary

Removing cell_closed is intentional here: actor shutdown is not the terminal-state signal Core needs. #29310 reintroduces it with precise semantics—once the terminal outcome is committed, even if its result remains buffered.

Validation

• just test -p codex-code-mode
• just test -p codex-code-mode-protocol
• Contract tests cover delegate field mapping, callback cancellation, runtime events, errors, and image details.

Stack parent: #29397.

[jif-oai]

Nit: Cancellation can still wedge shutdown if it arrives after we enter DrainNotifications

[jif-oai]

Could we make those 2 wait concurrent or share one deadline? Sounds odd to run them serially... mainly in a tokio runtime