Responses API: remaining per-upstream discriminators after nan-devops#637 #122

After nan-devops#637, the Responses API body and event fields are scrubbed (allowlist, nested model, provider-only fields, resp_ ids encrypted). These discriminators remain, so a caller can still tell which upstream serves a model (owner requirement: they must not):

1. Output item and delta ids keep each upstream's format. Novita/DeepSeek rs_/msg_0217…ffff<ipv4> embed an upstream-internal IPv4 (172.21.x.x). The vLLM delta item_id is the raw upstream resp_<16hex>, unencrypted. On glm the message item id equals the upstream response id. Needs a reversible mapping on the request path, so that previous_response_id and item_reference still work.
2. Reasoning item shape differs per upstream. deepseek {id, summary, type, status}; glm content[output_text, annotations] + status; vLLM summary: [] + content[reasoning_text] with no status; gemma4 message content carries logprobs: null. The "identical output item key sets" test uses stubs, so add real-shape fixtures.
3. Encrypted response id length varies per upstream and model_id (over 1000 chars on deepseek). On glm, response.created and response.completed carry different ids.
4. Error bodies differ per upstream. A made-up previous_response_id returns invalid response ID … trace_id on Novita and Response with id '…' not found. on vLLM.
5. Stream shape differs. vLLM groups emit only output_text.delta and completed, and sequence_number is present only on some upstreams. This is behavioural, so it can only be closed with a normalising stream re-emitter.

🤖 Generated with Claude Code