fix: scope router metadata to the request that produced it - #2582

Problem

endpointOai creates one OpenAI client per model and stores the router metadata read from response headers (X-Router-Route, X-Router-Model, x-inference-provider) in a single closure variable. Every request through that model shares it, and each request reads it back when its own stream ends (openAIChatToTextGenerationStream / openAIChatToTextGenerationSingle). Two consequences:

• Overlapping requests: a generation that finishes later reports the provider/model from whichever response arrived most recently — possibly another user's — so the provider badge on the message is wrong.
• Stale metadata: a response with no router headers replays the previous response's metadata, because the variable is never reset.

Fix

Build the SDK client per request, so the headers captured by the custom fetch are scoped to the request that produced them. Constructing the client does no I/O and nothing is pooled on it, so there is no cost to creating it per call. The three call sites (completions, streaming and non-streaming chat_completions) now take openai / getRouterMetadata from the per-request client.

Tests

New endpointOai.spec.ts with two cases that drive the real SDK against a stubbed fetch:

• two interleaved requests whose headers arrive in one order and whose bodies finish in the other — each must report its own provider;
• a header-less response following one with headers must report no provider.

Both fail on main (expected 'beta' to be 'alpha', expected 'alpha' to be undefined) and pass with this change.

npm test (1418 tests), npm run check and npm run lint are clean.

🤖 Generated with Claude Code