Stop observer reasoning from consuming the output budget

Observer requests to OpenRouter, the cmem.ai gateway, and OpenAI-compatible endpoints now disable reasoning by default, so reasoning tokens no longer fill the output budget and leave a session with zero observations.

• An empty reply with finish_reason=length is logged and retried once with reasoning forced off. That retry counts against the existing paid-send budget and is bounded by the per-attempt timeout.
• Billed usage keeps the retry's prompt size and adds completion tokens and cost from both attempts.
• Health status names the active OpenRouter credential, including a custom OPENROUTER_BASE_URL.
• OpenAI "Unsupported parameter" responses drop the reasoning field and remember the endpoint. Mandatory-reasoning and invalid-effort refusals are left in place.
• Empty replies that finish with stop stay accepted.

Ships the observer repair from #5029.