fix(notebook): rebuild for free CPU Colab — judge-friendly reproducer

Update demo space

docs(readme): add demo links, honest fallback, create demo space

fix(reward): align prompt with SFT, soft format, real recursion signal (A.3 + A.4-bis)

feat(train): flip Path 3 -> Path 1 — production 1.5B GRPO on a100-large

docs(D): wire training evidence - link plots from HF model repo, add W&B run

deps: bnb>=0.48.1 for vllm 0.18 4-bit gate

fix(grpo): TRL 1.2 stock path + LoRA double-wrap fix + smoke memory caps

fix(job_train): force CUDA_VISIBLE_DEVICES=0 to dodge TRL multi-GPU entropy bug

fix: heredocs use initialize_config_dir absolute path (no caller file = relative breaks)

fix(jobs): bump bitsandbytes to 0.47 (cu128 binary) so GPU dequant works in HF Job runtime image

fix(job_train): drop flash-attn install to avoid build-time OOM (R3 mitigation)

fix(job_train): pip install huggingface_hub before downloading data files (R1 parity with job_smoke.sh)

config: scripts/job_train.sh \u2014 Path 3 (qwen_0_5b_smoke + GRPO max_steps=50) for sanity-check run before burning 1.5B credits

feat: phase 1 complete — smoke green on HF Jobs, training scripts, plot generator, Colab notebook, submission preflight

clean repo without secrets or data