collect-models: the credential write is still missed after #1357, and it costs the PR lane its whole E2E run (recurrence of #1355)

[rafaelgiln]

Spun out of #1334's PR validation (#1369). Not a new mechanism — a recurrence of #1355 after #1357 shipped.

Symptom

pr-validation.yml's E2E job exits at its shared Collect models pre-flight, so the lane never runs the PR's specs at all — its step list goes Collect models → Run destructive lane, with no impacted-specs step in between.

build axis: ✅ openai
build axis: ✅ anthropic
build axis: ✅ google
⚠️  collect-models: no credential write observed for provider "anthropic" within 180s of clicking Save.
    The panel stays busy while a write is in flight, so the NEXT provider is what pays for this (#1355).
⚠️  collect-models: no credential write observed for provider "google" within 180s of clicking Save.
1 failed  →  Process completed with exit code 1

All three keys validate (build axis: ✅ on each) — this is not a drained or revoked credential. The first attempt surfaced the downstream form of the same thing: anthropic — no models collected from the providers panel, which collect-models.spec.ts:166 classifies as a non-billing hard failure and fails on, rather than the billing path that only warns (#952/#980).

Occurrences (2026-08-07, all on 1.12.0.dev19)

Run	Branch	Outcome
31175663512 11:51	chore/quarantine-daily-1361	no credential write observed for provider "anthropic" within 180s
31188034419 14:31	test/issue-1334-oc-credential-axis	anthropic — no models collected from the providers panel, job exit 1
same run, re-run 14:44	same	anthropic and google both stalled, job exit 1
same run, 3rd attempt 14:55	same	passed (5.0m), all three providers collected

Two independent branches, and 2 of 3 attempts on one of them. The provider that stalls moves between attempts (anthropic; then anthropic + google) and a plain re-run clears it — the load-dependent shape, not a consequence of any diff.

Preliminary read (descriptive — NOT a verdict)

#1355 was "clicks a Save button still aria-busy from the previous provider", and #1357 replaced a fixed 60 s clock with waiting for the credential write, measured at 103 s. The warning text above is #1357's own — it fires when that wait expires, now at 180 s. So the mechanism #1355 described is still reachable; what is unmeasured is whether 180 s is simply short on a loaded runner, or whether the write can fail to be observed at all (in which case a longer clock never fixes it, which is the lesson #1357's own header records).

Not claimed here: that the threshold is the answer. Raising a timeout because a timeout expired is the move #1357 argued against.

Why this matters beyond one red

Collect models is a hard gate on this lane. When it exits 1 the PR gets no E2E coverage of its own change, and the only signal is a red check whose cause is three steps away from the diff. #1369 merged on an ambient-red justification for exactly this (comment) — that reasoning was sound for one PR, but it is not a mechanism, and the next PR pays the same cost.

Investigation directive

Instrument before tuning. (a) Establish what the 180 s wait actually observes on a stalling run — is the POST|PATCH /api/v1/variables/ never issued, issued and never answered, or answered after the deadline? The distinction decides whether any clock helps. (b) Only then decide between a longer deadline, a per-provider retry, and making the sweep resilient to one provider stalling (the other two collected fine in every occurrence above). (c) Consider whether a Collect models failure should abort the PR lane at all: daily-stable.yml deliberately does not let it kill the shard (#980), and the asymmetry is undocumented.

Deliverables (Done when)

• Root cause of the missed write confirmed with evidence from a stalling run, not inferred from the deadline expiring.
• The PR lane no longer loses its entire E2E coverage to one provider's Save stalling — by fix or by an explicit, documented decision that it should.
• The daily-vs-PR asymmetry on a red Collect models is either aligned or written down as intentional.
• Covered by npm run test:scripts / npm run test:units where the logic is ours.