Add experiment results: Qwen3-4B outcome-only baseline on 4xB200

Add .claude/ to .gitignore

Clean up repo: remove stale docs, add skill config

Skip non-scalar values in TensorBoard logging

Set VLLM_ALLOW_INSECURE_SERIALIZATION in engine init, revert dtype workaround

Fix torch.dtype serialization for vLLM 0.17 msgspec RPC

Add VLLM_ALLOW_INSECURE_SERIALIZATION=1 for vLLM 0.17 compat

Fix SamplingParams compatibility with vLLM 0.17.1

Fix vLLM version check and add missing code.json schema

Fix UnboundLocalError: remove redundant import os in _init_logging

Update coach ablation examples: use gemini-2.5-flash-lite as degraded coach

Add B200 coach script with coach model ablation support

Fix model name: Qwen/Qwen3-4B (plain, not Thinking variant)

Add 4×B200 adapted outcome-only baseline

Add outcome-only baseline for MAPPA ablation