feature: improve request/response transparency in /v1/chat/completions

[m-misiura]

Did you check the docs?

• I have read all the NeMo-Guardrails docs

Is your feature request related to a problem? Please describe.

Teams using NeMo Guardrails as an inference proxy via /v1/chat/completions might encounter issues where the server silently drops or rejects valid OpenAI-compatible parameters and response fields:

• request filtering: strict Pydantic schemas reject unknown fields, so newer OpenAI parameters (e.g. tool_choice, response_format, reasoning) are rejected rather than forwarded to the upstream LLM.
• response filtering: extension fields from upstream LLM responses (e.g. provider-specific metadata) may be dropped.
• header forwarding: auth headers and request-correlation headers (e.g. X-Request-Id) are not propagated to/from the upstream LLM.
• streaming fidelity: SSE chunk structure may not be preserved end-to-end.

Describe the solution you'd like

To be conformant, we need to:

• forward unrecognized request parameters to the upstream LLM rather than rejecting them (pass-through mode)
• preserve upstream response fields that the guardrails layer does not need to modify
• propagate relevant headers (auth, correlation IDs) between client and upstream LLM
• ensure streaming SSE chunks maintain their structure through the guardrails pipeline

Describe alternatives you've considered

• dedicated proxy mode that bypasses Pydantic validation for unknown fields — more transparent but harder to validate.
• document that /v1/chat/completions is not a transparent proxy and should not be used as one — accurate today, but limits adoption.

Additional context

No response

[m-misiura]

cc @Pouyanpi @tgasser-nv

I appreciate that this issue actually covers more than a single feature, so please let me know if it needs to be broken down further. As always, I would be like to contribute where possible so feel free to assign me to the subsequent work