rebuilt from what ships

site
- found the .plan feed was a one-off snapshot: the site's feed generator is defined and never called, so the committed XML still carried a build date of Monday 2 November and stopped before the last five entries
- wrote a script that rebuilds it from the MDX that actually ships, newest first, with CDATA title and description, the post tags as categories, and the enclosure image made absolute
- it has a check mode that exits non-zero the moment the file and the posts disagree, so feed drift becomes a failing check instead of something noticed months later
- wrote and staged the previous day's entry with its picture in the same pass, so the two never wait on each other
hal
- wrote a triage reference for failed GitHub Actions runs, because the useful verdict is which step failed, whether the runner or the code caused it, and whether a newer run already covers that commit
- strip the runner timestamp prefix before grepping, and read the job's own timing lines: the last timeout line names the area that died, and exit 124 is the wrapper's deadline rather than the step's error
- a timer kill with nothing above it is host-shaped, not a hidden code bug; the wedged Docker daemon probe is the canonical case, and it is fixed on the host, not in the repository