fix(acp): drain message events before returning end_turn

[truenorth-lj]

Issue for this PR

Closes #25421

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Agent.prompt() returns stopReason: "end_turn" as soon as sdk.session.prompt() resolves. But trailing message.part.delta events for the assistant's final text deltas are still in the SDK event stream at that moment — they get processed by runEventSubscription and forwarded as agent_message_chunk to the ACP connection AFTER the prompt RPC reply has been written. On the wire you see chunks landing 5–50ms after id:N result:end_turn.

Why the fix works: runEventSubscription is a sequential for await that awaits each handleEvent call (which awaits connection.sessionUpdate). The message.updated event with info.time.completed set is emitted AFTER all message.part.delta events for the same message. So if prompt() waits for that completion event before returning, every prior chunk has already been flushed to the ACP connection.

The patch adds a per-messageId waiter resolved by a new case "message.updated" handler, and awaits it after sdk.session.prompt(...) resolves in both non-compact branches of prompt(). A 5s timeout falls through if the completion event never arrives, so a dropped event can't deadlock the turn.

The compact path doesn't carry an assistant message id, left untouched.

How did you verify your code works?

• bun run typecheck clean
• Existing test/acp/event-subscription.test.ts (11 tests) all pass — no regressions
• End-to-end on a real ACP client (TrueNorth's tn-claw sidecar): rebuilt the binary, captured WS frames in DevTools, confirmed end_turn is now the last ch:acp frame for the turn (3s quiet window observed empty)

Pre-fix wire trace:

... agent_message_chunk frames ...
{"id":2,"result":{"stopReason":"end_turn",...}}
{"method":"session/update","params":{...,"update":{"sessionUpdate":"agent_message_chunk","content":{"text":"\n\nDone. Created `/path/to/file`"}}}}   <-- arrives 10ms after end_turn

Post-fix: the trailing chunk above is gone — end_turn is the final frame.

Happy to add a unit regression test using the existing createFakeAgent harness if reviewers want.

Screenshots / recordings

N/A — backend protocol fix. Wire trace included above.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[truenorth-lj]

@kitlangton — would you have a moment to look? This is a re-open of #25422 (auto-closed by the bot 2h after open because the description didn't follow the PR template; same branch / same commit, just retemplated).

The fix is small (+53/−0, no deletions) and verified end-to-end on a real ACP client: rebuilt the binary, captured WS frames, confirmed end_turn is now the last ch:acp frame for the turn (3s quiet window observed empty post-fix; pre-fix had agent_message_chunk arriving 5–50ms after the RPC reply).

Happy to add a regression test using the existing createFakeAgent harness in test/acp/event-subscription.test.ts if you'd like one before merge — left it out to keep the diff focused, but the harness already supports event injection so it's straightforward.