grade: add --write to persist the score back into the trace

[saikrishnarallabandi]

Problem

cmd_grade recomputes a full grading_result — dimensions, task_score, passed — and then only prints it. The trace file is never updated. _append_grading_to_trace exists and is used by the run/batch path, but cmd_grade does not call it.

So claw-anything grade cannot be used to re-score an existing trace. It looks like it worked (it prints the new scores, exits 0), but nothing persists.

Why this matters

Adding an LLM judge to an already-finished --no-judge run should be cheap: the judge reads audit_data from the trace, so no mock services and no agent re-run are needed — only judge tokens.

We tried exactly that across 226 finished traces. Every invocation printed a fresh judged score and exited 0. Nothing changed: 0 of 148 completion values differed from the un-judged originals. The re-grade was a no-op we only caught by diffing against a backup and comparing file mtimes.

With --write, the same operation works and is dramatically cheaper than re-running the agent — for us it avoided re-spending ~27M tokens on a metered model purely to attach a judge.

Reproduce

# a trace produced with --no-judge
claw-anything grade --trace run/T04_x/T04_x.jsonl --task benchmark/tool/T04_x
grep -c grading_result run/T04_x/T04_x.jsonl   # unchanged

The printed completion: differs from the one stored in the file.

Fix

Add an opt-in --write flag. When set, drop any existing grading_result lines and append the fresh one, so a re-grade replaces rather than stacks (two grading_result events in one trace make downstream "last wins" readers ambiguous).

Default behaviour is unchanged — without --write, grade still only prints.

Verified: a trace re-graded with a judge goes completion 0.163 → 0.628, and the file reflects it.

Environment

OS	Ubuntu 22.04.5 LTS, kernel 6.8.0-124
Python	3.12.10
Docker	29.1.3 (grade run inside the trial image)
Judge	gpt-4o-mini via https://api.openai.com/v1
Base commit	verified on f9a0f63 and current main (e19dce2)

Note

This is the only one of my three PRs that adds a flag rather than fixing a bug — happy to rename it or gate it differently if you would prefer --in-place, or for it to be the default.