Internal: bin/run-schedule.php can stay running as the job runner. It is turned on with JOB_RUNNER=on in the server's .env; until then jobs start once a minute, as before. The manual step "Turn on the job runner" has the checks.
While it stays running, a page that saves work for a job tells the job runner at once, and the job starts in well under a second. Automations is the first job started this way. The row in the database stays the record of the work, so a message that is lost only means the job finds the work at its next look, 5 seconds later at most. The crontab line is unchanged: it starts the job runner again if the job runner ever stops.
A fault in the old arrangement is fixed for both ways of working: a job that ran past the end of its minute stopped every other job from starting until it ended.
Each run records what started it, and a failed run is kept for 90 days with the last lines it printed. The operator gets an alert email when a job fails 3 runs in a row, runs far too long, or has work waiting over two minutes.
The admin Scheduled jobs page says first whether any job needs attention and why, then lists every job with its state, last run and next run. Each job has a page with its failed runs and Run now.
Built from docs/specs/JOB_RUNNER.md; the contract is docs/SCHEDULED_JOBS.md. Two independent reviews were made and what they found was fixed. It has run on a Mac with stand-in jobs, not yet on the server.