Conform to Responses API resumable streaming and cancellation

[jayy-77]

Motivation

[#1356](Responses API: implement resumable streaming and cancellation)

Changes

Why It Works

Test Plan

Manual Testing

Automated Testing

[AlexCheema]

Review Findings

Critical Bugs (2) — will crash at runtime

1. Missing ResponsesStreamEvent import in responses.py
src/exo/master/adapters/responses.py:194 — the new on_event parameter uses Callable[[ResponsesStreamEvent], None] but ResponsesStreamEvent is never imported. This will raise NameError at runtime when generate_responses_stream() is called.

Fix: Add ResponsesStreamEvent to the import block from exo.shared.types.openai_responses (line 10–32).

2. Wrong kwarg for TaskCancelled in cancel_openai_response
src/exo/master/api.py:1340 — TaskCancelled(command_id=command_id) passes command_id, but TaskCancelled requires cancelled_command_id (it's a separate field from the auto-generated base command_id). This will raise a Pydantic validation error at runtime when cancelling a response.

Fix: TaskCancelled(cancelled_command_id=command_id)

---

basedpyright strict-mode errors (2)

3. Unused import defaultdict in responses_store.py
src/exo/master/responses_store.py:11 — from collections import defaultdict is imported but never used. basedpyright reportUnusedImport.

4. Protected member access _new_event in responses_store.py
src/exo/master/responses_store.py:58 — ResponsesStore.mark_cancelled() accesses stored._new_event, which is a protected member of _StoredResponse. basedpyright reportPrivateUsage.

Fix: Add a public method like notify_waiters() to _StoredResponse and call that instead.

---

Summary

Bugs 1 and 2 are blockers — both will crash at runtime. The basedpyright issues will also fail CI (uv run basedpyright must pass with 0 errors per project rules).