[BUG] EvaluationReport.to_file() writes RFC-8259-invalid JSON (literal NaN), has no explicit encoding, and can leave a truncated file #384

Checks

• I have updated to the latest minor and patch version of Strands and evals
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.53.0

Strands Evals Version

main @ 166aa31

Python Version

3.13.14

Operating System

Amazon Linux 2023 (aarch64)

Installation Method

git clone

Steps to Reproduce

EvaluationReport.to_file() (src/strands_evals/types/evaluation_report.py:324-325) is the sibling of Experiment.to_file() and has the same bug family that #380 reported and #383 fixes for experiments — plus one more:

• json.dump is called with defaults, so allow_nan=True: NaN/Infinity scores are written as bare NaN — not valid JSON per RFC 8259 §6.
• open(file_path, "w") passes no encoding=, so the file's bytes are locale-dependent on non-UTF-8 systems.
• It serializes into the already-open handle, so a mid-serialization failure truncates or destroys an existing file — same shape as [BUG] Experiment.to_file() truncates the destination before serializing — a serialization error destroys the existing file #382.

import json
from strands_evals.types.evaluation_report import EvaluationReport

r = EvaluationReport(overall_score=float("nan"), scores=[float("nan"), 1.0],
                     cases=[{"name": "a"}], test_passes=[True, False])
r.to_file("report.json")   # "succeeds"

txt = open("report.json").read()          # contains:  "overall_score": NaN
json.loads(txt, parse_constant=lambda c: (_ for _ in ()).throw(ValueError(c)))  # ValueError: NaN

Reproduced on main @ 166aa31 (and unchanged on #383's head 8268555 — correctly out of that PR's scope).

Expected Behavior

The report file is valid, strict, UTF-8 JSON — or the write fails loudly before touching the file.

Actual Behavior

The file contains the literal NaN, which Python's lenient json.loads reads back but any strict parser (JS JSON.parse, jq, serde, ...) rejects. overall_score: float and scores: list[float] make NaN reachable from ordinary aggregation (custom evaluators returning NaN), and the report is exactly the artifact meant to be machine-read downstream.

Additional Context

Found while reviewing #383. Same family, lower priority: LocalFileTaskResultStore.save (src/strands_evals/local_file_task_result_store.py:38-39) uses write_text(result.model_dump_json(indent=2)) with no encoding=, and pydantic also emits bare NaN.

Possible Solution

Mirror #383: serialize first, then write bytes —

data = json.dumps(self.to_dict(), indent=2, ensure_ascii=False, allow_nan=False).encode("utf-8")
with open(file_path, "wb") as f:
    f.write(data)

and extend the Raises: block accordingly. This would be the third call site of that exact pattern, so it may be the right moment to extract a shared strict-dumps helper.

Related Issues

#380, #382, #383