feat(sdk): onboard nemo evaluator sdk

[wprazuch]

Copies the public nemo_evaluator_sdk tree from NVIDIA-NeMo/Platform into NEL as src/nemo_evaluator/sdk/, minus the execution/ orchestrator submodule (NEL owns the evaluation loop — see AGENTS.md design principle).

What's included:

• metrics/ — BLEU, ROUGE, F1, ExactMatch, ToolCalling, LLMJudge, Remote, RAGAS (11 variants) + aggregation, base types, template rendering
• values/ — metric config types, result types, score types, secrets
• resilience/ — scheduler with retry/rate-limiting, policies, classifier
• datasets/ — multi-format loader (JSON/JSONL/CSV/Parquet/Feather/Arrow)
• inference.py, agent_inference.py — OpenAI-compatible async client with preprocessing/postprocessing hooks
• templates.py — Jinja2 prompt rendering
• structured_output.py — structured output mode detection/validation
• enums.py, constants.py — type definitions

Changes from upstream:

• Import rewrites: nemo_evaluator_sdk.* → nemo_evaluator.sdk.*
• init.py: drops the execution.evaluator.Evaluator export
• metrics/remote.py: inlines the _run_sync helper (previously lived in execution.metric_execution) to avoid pulling in the full orchestrator

Dependencies: added via new [sdk] optional extra (jsonschema, jsonpath-ng, pyarrow, pandas, openai, sacrebleu, rouge_score). Python floor bumped from 3.10 to 3.11 to match SDK constraint.

Tests: ported 10 test files under tests/test_sdk/ covering inference, agent_inference, llm_judge, exact_match, remote, values. Skipped tests that depend on the execution orchestrator (bleu/rouge/f1/number_check/ string_check/tool_calling/api) and test_params.py (depends on nmp internal module). All 148 ported tests pass.

Verified: pip install -e .[sdk] succeeds, imports resolve, tests pass. This is Approach 1 (verbatim copy). A follow-up PR will distribute the SDK into NEL's existing scoring/, metrics/, adapters/, engine/, environments/, and config/ modules per the RFC Phase 1.5 design.

[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: Pro Plus

Run ID: f884d49a-9768-460c-b555-227df8d7988c

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
• Commit unit tests in branch wprazuch/sdk-onboarding-approach1

---

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