ops(llm): OpenRouter has been quota-exhausted on every call since 2026-09-12

[FlyM1ss]

Every platform-credits LLM call on prod attempts openrouter first, gets provider_quota_exhausted, and fails over to commonstack. From credit_llm_reservations, attempt 0 only:

date	openrouter attempt 0	outcome
09-22	47	all quota-exhausted
09-20	191	all quota-exhausted
09-19	42	all quota-exhausted
09-18	243	all quota-exhausted
09-17	94	all quota-exhausted
09-16	69	all quota-exhausted
09-15	196	all quota-exhausted
09-14	43	all quota-exhausted
09-13	273	270 quota-exhausted, 2 settled, 1 invalid
09-12	142	118 quota-exhausted, 24 unavailable

1,313 quota-exhausted first attempts in 11 days, against 2 OpenRouter successes total. It has been failing on essentially every call since at least 2026-09-12.

Measured cost: a median 12.0s per call (attempt-0 → attempt-1 gap on run agent_20260922_202119_fc7bd724; n=47, min 11.0, max 14.0) — 16% of every decision bar, on every LLM backtest the platform runs.

It is invisible because PROVIDER_QUOTA_EXHAUSTED is in _PLATFORM_FAILOVER_CATEGORIES (infrastructure/llm/execution/service.py:61), so runs still succeed. This is the fail-closed-is-not-fail-visible shape from CLAUDE.md: a dead primary credential and a healthy one produce identical output.

• Top up / rotate the OpenRouter credential — operational, no code change.
• Warn when a provider's failover rate stays near 100% over a window. A per-call category cannot report a wholesale credential outage; that is what let this run for 11 days.

Adjacent, same data: commonstack answers response_invalid on 9/47 calls of that run (19%) and 37/191 on 09-20. Those bars fall back to rule-based and do not count toward H6 LLM coverage. Worth a look if it persists.

[FlyM1ss]

Filed from one diagnosis of prod run agent_20260922_202119_fc7bd724: #522 (why it timed out) · #523 (the dead primary provider) · #524 (why nothing is visible afterwards) · #525 (the refund log flood).