[tokibito]

Second half of #65. Each restart under --workers was one record with its delay, and nothing said the whole fleet was down. This adds a summary of the supervisor's state to its log records.

What it does

• When every worker has exited with an error within ten seconds of starting (the existing threshold for the restart back-off) and none has since run longer than that, the supervisor logs one Every worker is failing to start WARNING with workers and each slot's last exit_codes. One worker failing on its own does not trigger it; its exit records already say so.
• Once a worker runs for ten seconds again, or exits with code 0, the supervisor logs one Workers recovered INFO with worker_index, pid, failing_for in seconds, and recovered_by (uptime or clean_exit). The WARNING can be logged again if every worker starts failing again later.
• Supervisor finished gains restarts, the number of workers started in place of one that exited, and abnormal_restarts, the number of those that followed an exit with a code other than 0 or --empty-exit-code. This is the same rule that makes the Worker process exited record a WARNING.

Things to look at

• A clean exit counts as a recovery. A worker that reaches --max-tasks in less than ten seconds exits with 0 and is restarted at once, so it never runs for ten seconds; without this the state would read "every worker failing" while that worker was processing tasks. The check runs before the restarts in each loop iteration, so that pid names the process that exited rather than its replacement.
• A slot added with SIGTTIN is not a recovery until it has run for ten seconds or exited cleanly itself; until then it has only not failed yet.
• _report_exit() now takes its abnormal-exit rule from a shared _abnormal() helper; the existing tests pass unmodified.

Compatibility

Added records and fields only. No logger name, message, existing field or stdout output changes. Existing projects need no changes. The CHANGELOG entry is under Unreleased.

Rebased onto main after #71: the supervisor sets the worker index and counts the restarts in the same _spawn(), and the CHANGELOG has both entries.

Tests

tests/test_supervisor.py gains TestSupervisorState: the WARNING logged once when every worker fails, not logged when only one does, cleared by a worker running past the threshold and by a clean exit (with the exited process's pid), not cleared by a newly added slot, and the restart counts for failing workers, clean exits and a run-once supervisor.

Run on Python 3.13 / Django 6.0 (SQLite), Python 3.14 / Django 6.1 with -W error::DeprecationWarning, and the whole suite against PostgreSQL 16.

Refs #65