fix: size status-light stale threshold to observed prober cadence

[sr320]

Follow-up to #2484.

The problem

The status lights were spending most of each cycle showing unknown even though gannet and klone were reachable the whole time.

I set the 30-minute stale threshold against the interval the GitHub Action requests (*/15) rather than what GitHub actually delivers. Measured across the last 20 runs on this repo:

gaps (min): 57 71 58 57 47 35 46 43 68 57 45 37 37 48 38 44 55 41 30
min=30  median=46  max=71   (the cron expression asks for 15)
gaps exceeding the 30-min threshold: 19/19

Every single interval blew through the threshold. The lights were green for 30 minutes after each run landed, then grey for the remaining ~16 minutes on average, and up to 41 minutes in the worst gap.

The change

• Stale threshold 30 → 90 minutes, clearing the worst observed gap with headroom.
• Corrects the page text, which claimed "Checks run every 10–15 minutes" — not true of the current external-only setup.
• Records the measured cadence in scripts/README.md so the next person to touch the threshold knows what it is sized against.

Verification

Built locally and pointed at the live published status data:

• At 45 minutes old — the window where the old code showed grey — gannet and klone now render up, with raven correctly unknown since no in-network prober exists yet.
• A 95-minute-old reading carrying up: true still renders unknown, so a genuinely dead prober remains visible rather than silently reporting everything green.

Caveat

This is a mitigation, not the real fix. It buys tolerance for a slow backstop prober; it does not make the data fresher, and a dead prober now takes 90 minutes to surface instead of 30. Once the in-network cron is running every 10 minutes it becomes the freshest source, the merge logic prefers it, and this ceiling stops mattering in practice. Raven will keep reading unknown until then — setup steps are in scripts/README.md.

🤖 Generated with Claude Code