feat(train): show current task_id in heartbeat lines so round-robin coverage is visible at a glance

fix(collector): persist task cursor across PPO updates - was resetting to 0 each update so only first IC_ROLLOUTS tasks were ever trained on; now round-robins full shard (3 rollouts x 60 updates = 180 visits over 127-task shard = full coverage); also log first/last 5 task ids at startup so coverage is visible

fix(train_lib): wrap unsloth for_training/for_inference in try/except — peft+unsloth compat bug deletes _flag_for_generation; falls back to model.train()/eval() (correct, slightly slower); add IC_DISABLE_UNSLOTH escape hatch

feat: hand-written 3 Kaggle ipynbs with unsloth install attempt; train_lib disables actor grad-checkpoint + filters Phi3 warning spam (2x faster + clean logs)

fix: drop trust_remote_code so transformers uses native Phi-3/Qwen3 (custom modeling code is incompatible with t>=4.48 DynamicCache API)

feat(train): make HF_TOKEN optional + continuous live ETA heartbeat during rollouts

feat(kaggle): use user-pinned mount paths Phi-3.5-mini (actor) + DeepSeek-R1-0528-Qwen3-8B (critic); train_lib: trust_remote_code + adaptive LoRA target_modules for fused-qkv archs

feat(kaggle): 3 sharded training notebooks + local-critic mode + LoRA merge

fix(actor): fallback to plain HF transformers when unsloth fails to import

feat(train): warm-start adapter loading + IC_TASK_MODE (curated/all/hard) + phase2/phase3 launchers

feat(train): live tqdm+ETA, per-update HF Hub checkpoint upload, HF Jobs entrypoint + launcher

fix(colab): silence Qwen max_length FutureWarning, real 120-update training run, validated task IDs

feat(phase8-10): saboteur+slack+replay, 381 scenarios, Colab Unsloth+Qwen72B critic, push helpers