evals: make field leaves stale scored output and field_report.md on rerun #170

Follow-up from PR #12, review thread #12 (comment). This is the same class as #111, which covers the stale report.md on #20.

Problem. When make field reruns into an existing FIELD_OUT, evals.field score removes only the four row files (ROW_FILES, experiments/evals/evals/field.py). Two derived outputs can survive:

• If a rerun without TRUTH follows a run with TRUTH, the old FIELD_OUT/scored output from the scoring harness stays.
• If a rerun fails after the cleanup, the old field_report.md stays.

Either can be mistaken for the current capture's results.

What a fix needs. At the start of score, remove or stage every derived output (field_report.md, scored/ and measure-lab.json) together with the row files. Better, write the whole results folder to a staging folder and swap it in only on success, as models.run.publish does. Then test a rerun without TRUTH and a rerun that fails after cleanup.