[janbjorge]

Fixes #1027.

Stacked on the supervisor refactor. That PR introduces the list. This one fills it.

Summary

• A shutdown flush that fails keeps the job ends on QueueManager.unwritten_job_logs. Copy the list off before the next run, and write it with log_jobs on a live connection. A write after another worker owns the job is ignored.
• pgq run --restart-on-failure writes those ends on the next connection before it dequeues. A new process does not have the list, so those jobs still run again, as ADR-0004 allows.

Testing

• uv sync --frozen && uv run ruff check . && uv run ruff format . --check && uv run lint-imports && uv run mypy . && uv run pytest passed
• test_run_keeps_job_end_when_shutdown_flush_fails keeps a successful end when the exit flush raises, and a later log_jobs removes the row
• test_restart_writes_job_end_the_first_run_could_not writes that end on the restarted connection, and the handler does not run a second time
• uv run ruff check, uv run mypy on the touched files passed

Checklist

• I have read the Contributing Guide
• I have added or updated tests
• I have updated documentation if necessary

[janbjorge]

Superseded: #1062 landed the same buffer hand-off, and the rest is ported in #1069.