Hackathon submission: new README (3-5 min read), BLOG.md narrative, frontier baselines, design-principles framing

Submission deliverables: Colab notebook + per-step CSV + trainer log upload

Phase M Stage B: align fully with kube-sre-gym (sid-rp) patterns + A100 prep

Phase M Option 3 v3: 3 fixes lifted from kube-sre-gym (sid-rp) GRPO patterns

Phase M Option 3 retry: vLLM debug logging + reduce gpu_memory + max-steps 30

Phase M Option 3 hotfix: re-add vllm to PEP 723 deps

Phase M Option 3: vLLM single LoRA, no SFT adapter

Phase M Stage B: Option A coerce dict->list + bump to 200 steps

Phase M smoke test wrapper: 20 steps for fast validation

Phase M Stage A v2: fix 3 compound bugs preventing GRPO learning

Phase M Stage A: revert wrapper + trainer to 4940a15 (last known-working non-vLLM state)

Phase M Stage A: revert --use-vllm, ship non-vLLM path

Phase M debug Approach 2.5: pin peft>=0.18.0,<0.19 for SFT adapter loading

Phase M debug Approach 2: revert vllm pin + enforce_eager=True

Phase M Stage A vLLM Approach 1: pin vllm==0.6.3 (avoid v1-engine graph bug)

Phase M Stage A vLLM: enable use_vllm + fast_inference for ~3x speedup

Phase M Stage A: --max-steps 100, fix Track-IO wiring

Phase M-prep fix: append /final to sft_path in GRPO wrapper

Phase M-prep fix: drop --system from uv pip install (UV-venv vs system-Python mismatch)

Phase M-prep fix: capture env-server logs + extend boot timeout in GRPO wrapper

Phase M-prep fix: align wrapper script args with actual training script CLIs

Phase M-prep fix: use 'uv pip install --system' in wrapper scripts

Phase M-prep: HF Jobs UV wrapper scripts for L4 training (Option A)

Phase X1-X3: strip handoff_notes from SFT data + add schema guardrail