fix: use WAA server for /evaluate instead of fragile socat proxy

[abrichr]

Summary

• The /evaluate endpoint is already available on the WAA Flask server (port 5000), accessed via a single reliable SSH tunnel (local:5001 → VM:5000)
• The separate evaluate chain (local:5050 → VM:5051 → socat → docker exec → container:5050) was fragile — socat died mid-trial causing 6/7 infrastructure failures in Core4 Trial 1
• Default --evaluate-url to None so it falls back to --server URL (one port, one tunnel, no socat)
• Remove _setup_eval_proxy() and port 5050 SSH tunnel from run_dc_eval.py
• Make done-gate non-fatal when evaluate returns infrastructure error (accept "done" instead of forcing agent to continue pointlessly)

Root cause analysis

The Codex session (commit 9071fca) completed all 40 runs because the socat proxy happened to stay alive. Our trial (HEAD) had 6/7 infra failures because socat died mid-run. The evaluate code itself is identical between versions — the fragile proxy chain was the sole cause.

Port 5050 inside the container is the same Flask app as port 5000 (confirmed: curl http://localhost:5050/probe returns 200). There is no reason to maintain a separate proxy chain.

Test plan

• All existing tests pass (926 passed, pre-existing failures unchanged)
• Run Core4 ZS-only trial to validate 0 infrastructure failures
• Verify 0e763496 scores 1.0 (confirms evaluate works via server_url)

🤖 Generated with Claude Code