Gate production on CI, and record the source link breaking twice

[Parshkov]

What this changes

Documentation only — but it records two production changes made today.

Production is now gated on CI. Until this morning every push to main deployed within about 25 seconds whether the tests passed or not. CI itself only came into existence today, so production had never actually been gated by it. Wait for CI is on (checkSuites: true).

The Railway↔GitHub source link broke for the second time. The runbook already describes this class of breakage from 2026-09-04 and prescribes a fix that does not apply to today's occurrence, so it is now recorded as recurring, with the symptom that distinguishes the two.

Evidence

Verified end to end on the 05d7d3c merge (this repo, today):

time (UTC)	state
06:40:40	merged → deployment created, WAITING
06:52:50	CI run green → DEPLOYING
06:53:06	SUCCESS, new release serving

The site served the previous release for all 12 minutes, so the gate costs latency and not availability. That transition is the part worth having evidence for: a gate that holds but never releases is indistinguishable from a broken one until it bites.

On the source link

Symptom: a red GitHub Repo not found where the branch picker belongs, directly under the static heading Branch connected to production. That reads as a contradiction and is not one — the heading describes the section, the red box is the failed control.

What was still working: the repository name rendered above it (Railway serves that from its own database) and pushes still deployed (an already-registered webhook). Only the live GitHub API lookups failed.

The 2026-09-04 diagnosis (NO_INSTALLATION — the GitHub App was not installed) does not apply: opening the repo picker listed Parshkov/Resonance among the others, so the App was installed and healthy. I had guessed a stale installation token and that guess was wrong; the picker loading is what disproved it.

Fix: re-select the same repository via the pencil icon — not Disconnect — then reload the page. The reload matters, and is why this is written down: without it the widget keeps showing the stale error and the fix looks like it failed.

Checklist

• CI is green — 3/3: test suite 11m56s, benchmark gates 38s, lexicon 8s.
• benchmark/ gold is unedited.
• No code changed; engine untouched, no policy version move, no ADR.
• Claims recorded are measured, with timestamps.
• No credentials, tokens or private human context committed.

🤖 Generated with Claude Code