test: add comprehensive e2e tests for interrupt/cancellation

[emal-avala]

Summary

Follow-up to #106. The initial e2e test only verified cancellation during a single turn with a hanging provider. This PR adds coverage for the specific regression scenarios the original fix was designed to prevent — most importantly the back-to-back-turns test that directly exercises the stale-token-clone bug.

New tests

• cancel_works_across_multiple_turns — the actual cannot interrupt it #103 regression. Cancels turn 1, then runs turn 2 and verifies it's still cancellable via the same shared handle. Pre-fix this would hang forever on turn 2 because the signal handler held a stale clone.
• cancel_does_not_poison_next_turn — pre-cancels the shared token, then runs a completing turn to verify run_turn_with_sink resets to a fresh un-cancelled token at turn start.
• cancel_before_first_event_interrupts_cleanly — race-condition edge case where cancel fires during or before the first rx.recv().
• cancelled_turn_emits_warning_to_sink — verifies the sink is notified via on_warning(\"Cancelled\"), confirming the UI feedback path works.

Refactoring

• Hoisted the HangingProvider mock out of the test function so it can be shared
• Added a CompletingProvider that emits text + Done normally
• Added build_engine and schedule_cancel helpers to keep new scenarios concise

What's still not covered (out of scope here)

• The actual spawn_escape_watcher thread reading real crossterm events (requires a pty-based test)
• The install_signal_handler SIGINT path (would require sending real OS signals)
• Terminal raw-mode cleanup on EscapeWatcherGuard::drop

These remain manual-test items in the README checklist for #106.

Test plan

• All new tests pass locally / in CI
• Refactored existing test still passes

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.