Raise TaskFailedError from gather when a task exhausts retries

[tcdent]

Summary

• When a task exhausts its retries, the pool now writes a terminal TaskFailure record (task name, agent id, error, attempt count) to the task's result key, following the same key/serialize/set convention as the success path.
• ax.get_result recognizes the record on its next poll and raises ax.TaskFailedError within ~0.5s, instead of polling into the void and raising a generic TimeoutError after the full timeout (default 300s).
• ax.gather propagates the first failure via asyncio.gather — a single fail-fast contract, no return_exceptions path. Remaining tasks keep running in their workers and their results stay retrievable, so re-gathering survivors resolves finished tasks instantly.
• TaskFailedError is exported on the ax namespace.

Design notes

• Retries reuse the task's agent_id, so the failure record can only land after the retry budget is truly spent — a retry that succeeds still writes a normal result to the same key.
• Raising does not cancel in-flight work; tasks in a gathered group are decoupled processes, and orphaned survivors are accepted by design.

Testing

• get_result raises TaskFailedError with the full field set on the first poll (no timeout wait).
• gather propagates a failure among successful tasks.
• The pool's give-up branch stores a record that round-trips through backend.deserialize.
• Full suite: 300 passed, 1 skipped.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LC1fRUucUMCBWKwJFtq5YE