Stabilize abort task follow-up handling

[aibrahim-oai]

What changed

• The interrupt path now cancels the running task before clearing pending turn state.
• Follow-up model requests are suppressed once cancellation has won the race.
• The tests were rewritten to assert on stabilized outbound request counts instead of sleeping and then checking whatever state happened to exist.

Why this fixes the flake

• The flake was partly a real bug and partly a test-design issue.
• In production, clearing state before task cancellation finished still left a window where a background path could emit an unwanted follow-up /responses request or record the wrong history entry.
• In tests, fixed sleeps made that race show up nondeterministically because the assertion sometimes ran before cancellation had fully propagated.
• Reordering cancellation before cleanup removes the underlying bug, and polling request counts makes the tests wait for the actual invariant: no extra requests after interrupt.

Scope

• Production logic change with matching test coverage.

[charley-oai]

@codex review this