fix: dtype kwarg (torch_dtype deprecated), vLLM max_model_len=4096

fix: max_completion_length 512→128, firewall circuit-breaker

fix: drop BnB 4-bit, load attacker in plain bf16

fix: dtype kwarg + cast lm_head/embeds to bfloat16 to fix GRPO generate

fix: add torch_dtype=bfloat16 to prevent Float/BFloat16 mismatch in GRPO

fix: GRPO batch_size must be divisible by num_generations (1→4)

fix: skip Unsloth in GRPO trainer (grpo_accumulated_loss signature mismatch)

fix: stub GuidedDecodingParams for vLLM 0.19+ / TRL compatibility

Phase 5: training pipeline — client, GRPO trainer, eval, baselines (23 handcrafted attacks)

Phase 0: bootstrap