fix: remove zombie registry entries when ActiveTask start fails

[ez-lbz]

What changed

1. Remove registry entries synchronously when ActiveTask.start() fails

Problem: ActiveTaskRegistry.get_or_create() (src/a2a/server/agent_execution/active_task_registry.py) inserts the ActiveTask into _active_tasks before calling start(). If start() raises (e.g. the task is already in a terminal state), the only cleanup was the fire-and-forget asyncio.create_task(self._remove_task(...)) scheduled via _on_active_task_cleanup inside ActiveTask.start(). That cleanup runs asynchronously and may not execute before another request checks the registry, leaving a zombie entry for a task that never started — and a concurrent get_or_create could return the half-started task.

Fix (src/a2a/server/agent_execution/active_task_registry.py):

• Wrapped the start() call in get_or_create() with try/except; on any failure the entry is removed synchronously under _lock before the exception propagates.
• The removal is idempotent (pop), so the fire-and-forget cleanup task scheduled by ActiveTask.start() remains harmless if it runs later.

Testing

• ./.venv/Scripts/python -m pytest tests/server/agent_execution/test_active_task_registry.py -q → 6 passed (includes new test_get_or_create_failed_start_removes_zombie_entry, which seeds a completed task so start() raises InvalidParamsError, then asserts the registry no longer contains the task).
• ./.venv/Scripts/python -m pytest tests/server/agent_execution/ -q → 73 passed.
• ./.venv/Scripts/python -m pytest tests/server/request_handlers/test_default_request_handler_v2.py -q → 59 passed (registry consumers unaffected).
• ./.venv/Scripts/python -m ruff check on modified files: clean.
• Behavior change: none for the success path; on a failed start() the registry is now guaranteed not to retain the task.

[github-actions]

🧪 Code Coverage (vs main)

⬇️ Download Full Report

	Base	PR	Delta
src/a2a/server/agent_execution/active_task_registry.py	97.01%	97.26%	🟢 +0.25%
Total	93.06%	93.06%	⚪️ 0.00%

Generated by coverage-comment.yml

[mykytanetipa]

Thank you for this contribution. Could you rebase this onto latest main? It's built on an older base (_lock is now threading.RLock, and get_or_create was restructured)

[mykytanetipa]

async lock doesn't make a lot of sense here for single sync pop with no await, you've likely kept it for consistency with codebase which already moved to threading.RLock.

[ez-lbz]

Rebased onto latest main — the zombie-entry cleanup now lives in the restructured get_or_create (sync _lock + start() outside the lock): on any start() failure the just-inserted entry is removed synchronously under the RLock before the error propagates, so a subsequent lookup can never observe a never-started task. Registry tests pass locally (tests/server/agent_execution/: 77 passed).