Replace retired answer model defaults with Nemotron 3.5 Lightning

[jioffe502]

Description

Replace the retired NVIDIA-hosted Super-49B generator and judge defaults with Nemotron 3.5 Lightning:

• Hosted LiteLLM and evaluation defaults use nvidia_nim/nvidia/nemotron-3.5-lightning-30b-a3b.
• The optional Helm and Compose answer NIM uses nvcr.io/nim/nvidia/nemotron-3.5-lightning-30b-a3b:2.0.9-variant, advertises nvidia/nemotron-3.5-lightning-30b-a3b, enables the nemotron_v3 reasoning parser, and uses the one-GPU chart topology.
• NIM Operator profile selection remains automatic so it can select a compatible BF16, W4A16, or NVFP4 profile for the target GPU.
• Super-49B remains supported as an explicit self-hosted or local override; this PR removes it only from defaults and hosted examples.
• Coupled tests, runnable examples, and operator documentation reflect the new defaults.

Answer generation remains opt-in, and this does not change retrieval behavior.

Qualification status

This PR is a draft because Nemotron 3.5 Lightning has not yet been exercised through these NRL runtime paths. The completed validation confirms code, configuration, and deployment-template wiring; it does not qualify Lightning as a production default.

Before marking this PR ready for review:

• Run a hosted API smoke using the new default model identifier for both answer generation and an LLM judge call.
• Run a small end-to-end answer-generation evaluation to confirm output quality and judge-response parsing.
• Start the self-hosted NIM on a representative NVIDIA-verified GPU and smoke-test /v1/chat/completions with the default one-GPU topology, automatic profile selection, and nemotron_v3 reasoning parser.
• Confirm required CI checks complete successfully.

If self-hosted qualification is unavailable, narrow this PR to the hosted defaults and leave the Helm and Compose self-hosted defaults unchanged.

Validation completed

• 152 passed, 26 subtests passed across focused LLM, service-answer, Helm, and CI-workflow tests.
• Repository commit hooks: trailing whitespace, EOF, large-file, Python AST, debug statements, Black, and Flake8.
• Docker Compose config validation for local judge and service mode.
• Strict MkDocs build.
• git diff --check.

Not yet validated

• Hosted Lightning generation or judge request.
• Self-hosted Lightning NIM startup or inference.
• End-to-end answer quality, judge quality, latency, and GPU-memory behavior.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• Nemotron 3.5 Lightning runtime qualification is complete.

[greptile-apps]

Greptile Summary

The PR replaces retired Super-49B answer-generation and evaluation defaults with Nemotron 3.5 Lightning across hosted and self-hosted configurations.

• Updates LiteLLM, judge, service, evaluation, and skill-evaluation defaults.
• Changes the optional Helm and Compose answer NIM to a one-GPU Nemotron 3.5 Lightning deployment.
• Retains documented Super-49B overrides and updates tests and operator documentation.

Confidence Score: 5/5

The PR appears safe to merge because no blocking failure remains in the follow-up review scope.

No blocking failure remains.

Important Files Changed

Filename	Overview
nemo_retriever/helm/values.yaml	Replaces the optional answer NIM defaults with Nemotron 3.5 Lightning, one-GPU resources, automatic profile selection, and its reasoning parser.
nemo_retriever/src/nemo_retriever/models/llm/clients/litellm.py	Changes the provider-neutral generation client's hosted default model to Nemotron 3.5 Lightning.
nemo_retriever/src/nemo_retriever/models/llm/clients/judge.py	Changes the judge's hosted default model while retaining its reasoning-sized output budget.
nemo_retriever/src/nemo_retriever/service/config.py	Updates the opt-in service answer-generation model default.
nemo_retriever/dev/compose/service-mode.compose.yaml	Moves the optional Compose answer service to the one-GPU Nemotron 3.5 Lightning NIM.
nemo_retriever/tests/test_helm_answer_llm_generation.py	Updates rendered-chart assertions for the new answer NIM identity, resources, and reasoning configuration.

_{Reviews (2): Last reviewed commit: "fix: replace retired answer model defaul..." | Re-trigger Greptile}