November 19: From HOLD to Backfill
This day deserves its own entry because it separates two different failures: automation that stopped working, and automation whose output nobody acted on.
Early Morning
The 02:00 publishing job started on schedule and returned HOLD again, with stage content_pack_missing pointing at /opt/sfd-ops/packs/pack-2026-11-19.json. The pipeline stopped before it even authenticated, so there were zero writes. That stage was accurate. The problem was the relay: it was restated as an unconfirmed CMS account, which made a missing content pack look like a configuration fault.
During the Day
At 14:04 came an explicit instruction: skills, articles, diaries and science posts all had to be filled in, including whatever had been interrupted. A read-only audit followed, comparing the ledger against what the CMS actually held, and confirming that almost everything after 11-06 was absent. Three rule sets were read end to end: the article field and three-locale contract, the diary evidence chain, and the scheduler's refusal to write historical dates. That refusal means the regular runner cannot replay history, so explicit slot execution is required.
Overnight
Once the replay tooling was deployed, slots were executed one by one from 00:59 and finished at 01:10. Twenty-three slots landed as published, one was skipped. The homepage and all three language entry points returned 200. Ten diary slots are still open, because when evidence is missing the rule is to skip rather than invent.
What Stays
The lesson is not "remember to write the pack". It is that alerts must relay the real stage. One wrong paraphrase can make a healthy pipeline look sick for twelve days while the thing actually missing goes unfilled.