Add few-shot prompt and shaped rewards

Revert to Gradio 5.49.1 and ASCII logs

Disable Gradio queue for buttons

Replace timer polling with manual Refresh Log

Disable queue for UI polling and buttons

Initialize log immediately on Start

Fix resume adapter training (no peft_config with PeftModel)

Fix code block extraction and increase completion length

Persist LoRA adapter via HF dataset repo

Switch to correctness-gated GRPO LoRA with persistence

Add VLIW simulator for cycle-count based rewards

Add GRPO training with proper state management

Simpler test version for model loading

Add training functionality with GRPO

Minimal test version

Simplified version without simulator

Full training app with verified imports

Debug version to check imports

Initial commit: VLIW kernel optimizer via RL