tasks_policy: long-context policy-compression tasks

v3: multi-turn env, thinking tokens, cross-family Qwen->Llama, multi-step GRPO

tasks_tough: add 42 more tough scenarios + baseline profiler

tasks_tough: add 10 domain-classifier tough scenarios (seed batch)

eval: use merged v1+v2 task bank (same fix train_grpo.py already had)

Pre-launch fixes: disable Qwen3 thinking, strip think blocks, degenerate-short guard

Fall back from Qwen3.5 -> Qwen3 family (transformers==4.56.2 compat)

v2 stack: Qwen3.5-2B agent/target, Qwen3.5-9B judge, hard tasks, additive reward

eval: --push-to-hub uploads eval JSONL to adapter repo under evals/

Switch agent loading to Unsloth FastLanguageModel + fix padding side

Initial commit: Prompt Golf environment for OpenEnv