Add simulation disclaimer banner

Full-screen UI + OpenEnv API tab (reset/step/state/stop)

Redesigned UI: dark theme, Plotly charts, 4 tabs, professional layout

Fix: simulate action effects on next state so AI wins on latency reduction

Deploy interactive simulation demo (Gradio, free CPU)

Fix merge: fall back to warm-start adapter from HF when GRPO skipped

Skip GRPO: merge warm-start and push to HF

Fix: batch_size=4 so num_generations=4 divides evenly

Fix: max_length -> max_seq_length for trl 0.15.2 (verified all configs locally)

Fix: trl==0.15.2 (has GRPO, no vllm/FSDP dep)

Fix: pin trl==0.12.2, verify imports during build

Fix: pin trl<0.17 for FSDP compat, skip world model (already done)

Install CUDA PyTorch in slim image for A100 GPU

Revert to python:3.10-slim (was working) + health server prevents timeout

Fix: install PyTorch with CUDA 12.1 support

Fix: use CUDA base image for GPU support

Fix: add health server on port 7860 to prevent timeout

Fix: batch_size=16, 10K samples, unbuffered output, 2 epochs

Fix: set HOME/USER/TORCH env vars for uid 1000

Fix all: writable /tmp cache, no login(), proper permissions

Fix: set HF_HOME to writable directory

Fix Dockerfile: read HF_TOKEN from env correctly

Use HF_TOKEN secret in Dockerfile

Add Dockerfile and training script

initial commit