feat(eval): first-class artifact grading — measure an ACTING persona by her HANDS (workspace file), not her mouth

[joelteply]

The eval grades settled.spoken — her spoken answer. But when a persona ACTS (writes code to a file, compiles,
fixes), her code is in the FILE and her spoken text becomes a summary — so the act→verify loop scored near-zero
not because she failed but because her work was invisible to a speech-grader. (Confirmed: solo WITH a verify
directive scored 25% vs 75% spoken, purely because she filed the code.)

Fix (first-class, replaces the /tmp/grade.py hack): a solution_file field on EvalTask. When set alongside
test, the grade reads HER FILE via new gym_grader::test_grade_file and runs the SAME harness as a spoken
answer (grade_rust: strip her main, append the task's test, compile, run). The file resolves relative to the
core cwd — the workspace root code/write sandboxes to — so a relative sol_x.rs lands exactly where the grader
reads. Fails LOUD (never a silent pass) on a missing file ("never wrote it" = acts=0), an empty file, or a path
escaping the workspace; the file is removed after grading so a stale artifact can't false-pass a later task.

Committed reproducible gym docs/genome/hard-rs-acted.jsonl (8 tasks: verify-prompt + solution_file + hidden
test) — no /tmp fixtures, no external grader script; the eval grades her file natively. Unit test covers
pass-correct / clean-up / fail-loud-missing / refuse-traversal.

This is the seam that finally makes the act→verify loop — and the team's reviewer catching a real compile error —
show up in the score. Runs detached via #86; survives the client via #85.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo