[klaidliadon]

Goal

An operator can see a task failing and retrying while it happens, instead of only after it runs out of attempts. Each failed attempt that will be retried now logs a warning with the attempt number, the budget, the backoff before the next try, and the error.

Problem it solves

• Retries are invisible. Today the executor logs a task failure only when it parks the run: a permanent error, or the last attempt in the budget. A task that fails twice and then succeeds leaves no log line, and neither does one that is still working through its budget. The only trace is attempt_count and last_error on the task row.
• Fault-injection tests have nothing to search for. A test that injects a transient fault and expects the task to retry can only check the database afterwards. With a warning per attempt, it can check the logs.

The warning fires once per failed attempt, for both the in-process backoff and the longer backoff that releases the claim. Parking keeps its existing alert-level line, so the last attempt logs once, not twice.