fix(spine): the daily journey was watched by a clock that could not see it #5593

Two defects, one subject: the check covering register → repo → PAT → push → PR → merge, which is the only end-to-end proof this platform has.

1. doctor:journeys could never notice it stopped

That check takes MAX(checkedAt) across every journey:* row. The five-minute journeys keep it fresh, so a daily journey folded into the same bucket can never go stale.

Measured 2026-08-31: journey:first-run had not run for 9 hours while doctor:journeys read "newest row 1s old (max 15m)". Had it died permanently, that watcher would have stayed green forever.

A daily check measured against five-minute siblings isn't being watched — it's being averaged out. It now has its own row and its own budget: 30h, not the 3×-cadence used for the fast checks, because three days of blindness on this one is too long and a full day with no run is exactly the thing worth knowing.

2. "Daily" never described reality

The interval gate read _lastFirstRunJourneyAt — a module variable, so it reset to 0 on every restart, and autodeploy restarts this container whenever main moves.

On 2026-08-30 the journey ran six times in four hours, once per deploy, and the 24h schedule had almost certainly never fired on its own. AlecRae documented the same shape from the other end: a daily job that never ran at all because the process restarted every 15 minutes.

An in-process timer in a frequently-restarted process is not a schedule. synthetic_checks already records every run, so the real answer was one query away. The in-process value stays as a fallback: if the read fails we keep today's behaviour rather than either spamming journeys or silently never running one.

Tests

Pin the bug itself — a stale daily journey beside fresh siblings must go red while doctor:journeys stays green — plus the two ways this becomes noise: 26h is normal rather than dead, and never-run is its own state, not fresh.

Verified by mutation: folding the daily check back in with its siblings fails the suite.

Note

Neither defect was reported by anything. Both were found by asking why a number looked wrong — the same way the rest of today's were.