fix(batch): one truncated run-meta.json no longer discards the batch summary

[vaibhavdabas16]

What does this PR do?

Fixes #303: a single truncated run-meta.json currently destroys the machine-readable results for an entire batch, and the truncation itself is made impossible in the first place.

The crash. print_run_stats() parsed every run-meta.json with a bare json.loads() (batch.py:450). async_main() calls it at batch.py:753, one line before write_summary_json() at :754 — so one malformed file raised JSONDecodeError out of the function and took batch-summary.json down with it. Every other job in the batch lost its recorded results, even though those runs succeeded.

The truncation. write_run_meta() (metadata.py:288) wrote with a plain write_text(). A run killed mid-write — Ctrl-C, OOM, host shutdown — leaves a half-written file on disk, which is precisely what the reader above then chokes on.

Both halves are addressed:

• New clawbench.utils.jsonio with write_json_atomic() (same-directory temp file, flush + fsync, then os.replace) and read_json_or_none(). os.replace replaces atomically on POSIX and Windows, so an interrupted write leaves either the previous file or no file — never a truncated one.
• write_run_meta() and write_summary_json() now write atomically.
• print_run_stats() reads via read_json_or_none(). It also rejects valid JSON that isn't an object — a bare list previously reached .get() and raised AttributeError — prints a warning naming the offending file, and falls back to the directory name for that one run.

Reporting now degrades to a single warning line for the affected run instead of losing the whole batch.

Corpus

• v2
• v1
• both
• not applicable

_{Host-side runner/reporting change; no task data involved.}

Test plan

• New tests/test_batch_stats_resilience.py (11 tests): atomic round-trip and parent-dir creation, no leftover temp files, previous file survives a failed serialization, read_json_or_none() against truncated / empty / garbage / undecodable input, and two print_run_stats() regression tests covering a truncated run-meta.json and a non-object one.
• Confirmed the regression tests genuinely catch the bug: reverted batch.py and metadata.py to main and re-ran, which fails with JSONDecodeError and AttributeError respectively; both pass with the fix applied.
• Full suite: 203 passed, 3 skipped. The one unrelated failure, test_host_tasks.py::test_checked_task_json_files_parse_and_validate[v1-lite], reproduces identically on a clean main checkout on this machine — the v1-lite task files are git symlinks (mode 120000) that Windows checks out as plain text.
• ruff check and ruff format --check clean on all changed files. Removing the last json. uses from batch.py left import json unused, so it is dropped (F401).

Related issues

Fixes #303.

_{One note for maintainers, deliberately left out of scope: eval/rescore.py reads run-meta.json with the same unguarded json.loads() at lines 58, 102 and 279. Happy to follow up in a separate PR if that is wanted.}

[Perry2004]

Thanks for the contribution. I'll take a look.