refactor: hoist provider schema handling for Responses, Messages, and Chat Completions

[zhongxuanwang-nv]

Overview

Hoist provider schema handling for OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages into a single codec-layer path, so exporters and plugins consume normalized provider data instead of re-implementing provider parsing from raw JSON. (RELAY-325)

Opened as a draft for review. Local validation: cargo fmt, cargo clippy --all-targets -D warnings, and the pre-commit hooks pass; targeted unit + integration tests are green (resolve, codec, observability::manual, otel, openinference, atif, types, adaptive request-surface, plus codec/pipeline integration). The full cross-language matrix is left to CI — no binding or public-API surfaces changed.

• I confirm this contribution is my own work, or I have the right to submit it under this project's license.
• I searched existing issues and open pull requests, and this does not duplicate existing work.

Details

• codec::resolve (new) — the canonical detect-then-decode entry point. ProviderSurface + detect_request_surface (priority order hoisted verbatim from adaptive's resolver) / detect_response_surface (strict; ambiguous or empty payloads → None, avoiding the codecs' permissive {} decode) + fail-open normalize_request/normalize_response.
• Event::normalized_llm_request/response — annotation-first Cow helpers over the detector; annotated_*() left unchanged.
• OpenAIResponsesCodec — response-only extension to decode top-level output_text and item-level output[].type == "output_text" (structured output keeps precedence). Request input_text deliberately left out (strict request parse).
• observability::manual (new) — the otel/openinference non-provider usage/cost fallback de-duplicated into one module, with the USD-only vs emit-any-currency policy parameterized. otel.rs and openinference.rs each shrink ~185 lines.
• ATIF — reads normalized message text from annotations when present (multimodal falls back to the raw extractor); tool calls and metrics stay on the raw path so provider-specific extras (e.g. provider_data) are preserved.
• adaptive — resolve_request_surface_from_request delegates to codec::resolve::detect_request_surface (function name, AcgError::Internal message, and priority order preserved).
• Fixed the misleading "global codec registry" docstring in codec/traits.rs; added an annotation-vs-raw-fallback note to provider-response-codecs.mdx.

Where should the reviewer start?

crates/core/src/codec/resolve.rs (the new single path), then crates/core/src/observability/atif.rs — the annotation-first / raw-fallback consumer is the trickiest equivalence surface (the raw parser is intentionally retained where the normalized form would be lossy: tool-call extras and multimodal content).

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• Relates to RELAY-325

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Enterprise

Run ID: 43784b3a-5c41-4fbb-b2b8-c098a62a7038

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

Fern docs preview: https://nvidia-preview-pull-request-288.docs.buildwithfern.com/nemo/relay (​https://nvidia-preview-pull-request-288.docs.buildwithfern.com/nemo/relay​)