fix(evaluations): normalize dictionary results in batch runs

[ningyixi]

What does this PR do?

Batch evaluation accepts dictionary evaluator results but passes them straight to _create_score(), which expects Evaluation attributes. Returning {"name": "quality", "value": 0.75} therefore fails the item with 'dict' object has no attribute 'metadata'. A composite evaluator returning a dictionary logs the same error and produces no composite score.

Normalize item and composite results with the existing _normalize_evaluator_result() used by experiments. This converts dictionaries, lists of dictionaries, and mixed dictionary/Evaluation lists into Evaluation objects before score creation, while preserving score metadata, comments, type, and config ID.

Ten regression cases cover sync/async item evaluators and composites through the batch runner, with local observation fixtures and assertions on the score calls. All ten fail before the runtime change.

Type of change

• Bug fix
• New feature
• Breaking change
• Refactor
• Documentation update
• Tooling, CI, or repo maintenance

Verification

uv sync --locked --python 3.13
uv run --frozen pytest tests/unit/test_batch_evaluation.py tests/unit/test_experiment.py -q
uv run --frozen ruff check .
uv run --frozen ruff format --check .
uv run --frozen mypy langfuse --no-error-summary
uv run --frozen pre-commit run --all-files

The focused suite has 59 passing tests. Ruff, formatting (575 files), mypy, and all pre-commit hooks pass.

The full unit suite also passes: 932 passed, 2 skipped on Python 3.13.5, using four workers. The existing prompt fixture needs LANGFUSE_PUBLIC_KEY and LANGFUSE_SECRET_KEY; a first run without them had 18 fixture setup errors (914 passed). Rerunning with dummy local test credentials and LANGFUSE_BASE_URL=http://localhost:3000 passed the whole suite. No real credentials were used.

No Langfuse server or live provider tests were run. Provider instrumentation and generated API files are unchanged.

Checklist

• I self-reviewed the diff using code_review.md.
• I added tests for behavior changes.
• Existing evaluator/composite documentation already describes dictionary results; no examples or environment contracts change.
• I did not hand-edit generated files.
• I did not commit secrets or credentials.

Developed and checked with Codex assistance.

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues were found.

Summary

Batch evaluations now use the existing _normalize_evaluator_result() helper before creating item and composite scores.

• Batch runs turn dictionary evaluator results into scores.

_{Reviews (1) · Last reviewed commit: "fix(evaluations): normalize dictionary r..." · Reviewed by Greptile}

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

All committers have signed the CLA.