[eval] Repair TaskTrove v4.8 campaign contracts

[penfever]

Repair the adversarial verifier so malformed reward output fails closed, and
raise the C++, PHP-large, and PHP-v2 task memory limits to 4096 MB. Add the
reproducible TaskTrove v4.8 assembly and publication paths used for the repaired
sources.

Track the GLM 5.2, Qwen3.5-122B, and Qwen3-Coder campaign configurations needed
to reproduce the v4.8 evaluations. Remove the unsupported Qwen3-Coder vLLM swap
argument. Warm resumes now delete and refire trials terminated by Together
service unavailability or agent SIGKILL instead of preserving them as completed
infrastructure failures.