Chat completions: Novita returns sla_metrics per chunk and provider_specific_fields (upstream discriminator) #128

Found while working on nan-devops#665. On /v1/chat/completions, Novita-served models (deepseek-v4-flash, glm5.3*) return:

• sla_metrics in every streamed chunk;
• provider_specific_fields in the non-streamed response.

Neither field comes from any other upstream, so a caller can tell which models are served by Novita. The #637 allowlist only covers the Responses API. Extend the chat success and streaming scrub, in both hooks, to drop these fields, preferably with an allowlist based on the OpenAI chat schema. Also check the edge Worker, which does not touch 2xx bodies by design.

Also noted from #660: in litellm 1.83.14 the router never reads a per-deployment allowed_fails, so the allowed_fails: 1 on the deepseek rows was never active; the router-wide value applies (10 community / 20 enterprise). docs/runbooks/deepseek-upstream-outage.md says otherwise and needs correcting.

🤖 Generated with Claude Code