fix(observability)!: attribute ATIF steps to effective response model

[bbednarski9]

Warning

BREAKING CHANGE: [Behavioral] ATIF LLM steps now prefer the model reported by the normalized provider response over the requested or profile model. Consumers that group, compare, or price trajectories by step.model_name may observe a different model for routed or cross-protocol calls. When no effective response model is available, the existing attribution fallback remains unchanged.

Overview

Correct ATIF model attribution for LLM calls whose effective provider model differs from the model requested at the beginning of the Relay lifecycle.

This generic exporter fix is intentionally independent of the Switchyard plugin. Any router, proxy, or provider translation layer that records the effective model in AnnotatedLlmResponse receives the same ATIF behavior.

• 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

ATIF previously preferred the model recorded on the LLM start event. For routed requests, that value can be a client alias, profile model, or originally requested model rather than the backend that actually produced the response. The normalized end response already contains the effective model, but ATIF did not use it for step.model_name.

This PR:

• Adds one shared effective-response model selector.
• Uses the normalized response model for paired start/end LLM spans.
• Uses the same selection for standalone LLM event conversion.
• Falls back to the existing event-model extraction when the response does not identify a model.
• Adds a regression where the request is attributed to qwen3:4b but the normalized response is produced by qwen3.6:35b.

No ATOF event, provider response, routing decision, cost calculation, or plugin behavior changes. Only exported ATIF step.model_name attribution changes.

Validation:

• cargo fmt --all -- --check
• cargo test -p nemo-relay test_exporter_prefers_effective_response_model_over_requested_profile_model
• cargo test -p nemo-relay observability::atif::tests — 88 passed.
• cargo clippy -p nemo-relay --all-targets -- -D warnings
• git diff --check

Where should the reviewer start?

1. crates/core/src/observability/atif.rs — model selection and fallback behavior.
2. crates/core/tests/unit/atif_tests.rs — routed-model regression fixture.

The central design decision is that ATIF describes the model that produced the response, while the requested model remains available in the original ATOF lifecycle input and optimization metadata.

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

• Relates to NVIDIA/NeMo-Relay #369, where this correction was originally discovered and has now been removed from that diff.
• Relates to NVIDIA/NeMo-Relay #389, which records effective model and optimization evidence on the normalized LLM response.

[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: 729cde13-0c3b-429d-ab51-f4614904022e

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.}