v5.1.2 — Fix Critical Held-Out Contamination + 2 More

What's Fixed

3 findings from 5th Codex adversarial review.

• CRITICAL: Auto-guard no longer contaminates held_out — --auto-guard-failures now only guards failures from the train split. Held_out failures stay unseen. Previously, held_out misses were cloned into train, inflating next iteration's held_out scores.
• HIGH: Guards actually work with dict inputs — run.inputs stored as original dict (not str() repr). json.loads() no longer fails silently on Python dict reprs. Errors surfaced to stderr instead of swallowed.
• HIGH: Archive copied before proposer spawn — Moved worktree file propagation (.evolver.json, .env, evolution_archive/) to step 3 (proposer bootstrap). archive_branch lenses now have data to work with.

Codex Review Running Total

5 adversarial reviews, 17 total findings, 17 fixed.

🤖 Claude (author) + Codex/GPT-5.4 (reviewer)