fix: make EvaluationReport.to_file() reject non-strict JSON

[serenearyal]

Description

EvaluationReport.to_file() had the same three problems that #383 fixed for Experiment.to_file():

• json.dump ran with defaults, so a NaN or Infinity score was written as the bare literal NaN. Python's lenient json.loads reads it back, but strict parsers (jq, JavaScript JSON.parse, serde) reject the file. overall_score: float and scores: list[float] make NaN reachable from ordinary aggregation.
• The destination was opened with mode "w" before serialization, so a failure mid-write truncated the file. On main, a valid 224-byte report was cut to 106 bytes of invalid JSON after a save with a non-serializable case value. If to_dict() itself raised, the file was left at 0 bytes.
• The file was opened without encoding=. This was harmless while ensure_ascii defaulted to True, but it becomes a real bug on non-UTF-8 locales once the writer emits UTF-8 text like fix: make Experiment.to_file() reject non-strict JSON instead of writing invalid files #383 does.

Changes

• EvaluationReport.to_file() now mirrors Experiment.to_file(): serialize first with ensure_ascii=False, allow_nan=False, encode to UTF-8, and write the bytes only on success. NaN, Infinity, and unpaired surrogates raise a ValueError with the original error as __cause__. Non-serializable objects raise TypeError before the file is touched. The Raises: docstring is extended to match.
• EvaluationReport.from_file() reads with encoding="utf-8" so the read side matches the write side, the same as Experiment.from_file().
• Four regression tests in tests/strands_evals/types/test_evaluation_report.py: NaN rejected with the existing file intact, unpaired surrogates rejected without writing, non-serializable value rejected with the existing file intact, and a UTF-8 round trip.

This is a plain mirror of #383 rather than a shared helper. Happy to extract one if you prefer; the issue notes this is the third copy of the pattern.

Related Issues

Fixes #384. Same pattern as #383 (Experiment.to_file), which also covers #382.

Documentation PR

None needed. Behaviour change is limited to raising on data that could not produce valid JSON.

Type of Change

Bug fix

Testing

• Added four regression tests; all fail on main and pass with this change.

• hatch test on Python 3.10, 3.11, 3.12, 3.13, 3.14: 2144 passed on each.

• hatch fmt --formatter, hatch fmt --linter, mypy, and pre-commit run on the changed files: clean.

• hatch run test-integ not run: no integration test exercises EvaluationReport.to_file(), and the suite needs AWS credentials I do not have locally.

• Reproduced the truncation and NaN cases with a scratch script on main, and confirmed the fixed writer raises before touching the file.

• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have reviewed and understand every line of code in this PR, including any generated by AI tools, and I can explain why it works
• My change is focused and reasonably small; I have split unrelated work into separate PRs
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.