fix: install nvidia-cudnn-cu12 alongside CUDA torch (removes --no-deps)

fix: force-reinstall CUDA torch before gradio import; populate logs on page load

fix: eliminate API leaks in training loop, set 30k steps for A100

Fix CPU training + 1902ms step speed bottleneck

fix: push reward curve + logs every 10k steps, not just at training end

perf: add encode+step benchmark to logs, validate CUDA speed

perf: force ST to CUDA, disable LLM tasks, reduce to 100k steps

fix: force-install CUDA torch in app.py before training starts

pre-deployment fixes for HF Spaces

Remove gr.Timer: caused C-level crash with no Python output

Fix logs not showing: use file-based state instead of in-memory globals

Add auto-refresh timer and dark theme CSS

Fix theme/css: keep in Blocks(), not launch()

Improve Gradio UI: fix theme, add CSS, clean up layout

Fix create_repo naming error with huggingface_hub 1.x

Add startup diagnostics

SpindleFlow RL — periodic push + log persistence