Maximize environment: curriculum task, metrics endpoint, 5 bug fixes, notebook fix

Add long_horizon/personalized tasks + GitHub-hosted training curves

Implement all audit fixes: base class, curves, README, code quality

Remove show_download_button — unsupported in this Gradio version

Fix training curves: switch gr.Plot → gr.Image with PNG bytes

Fix /web 404: guard matplotlib import and harden Gradio mount

Update training curves with real Colab data

Add Training Results tab with GRPO reward curves for all 3 agents

Overhaul UI + README for submission

Fix pipeline UI: total regex case-insensitive, deduplicate invoice IDs

Wire trained LoRA agents into pipeline demo UI

Auto-seed Regulator tracker on startup — pipeline demo works immediately on cold start

Add Multi-Agent Pipeline tab — live 5-agent episode trace

Add Generator adversarial GRPO training + /generator/score endpoint

Add 3 novelty upgrades: predictive Regulator, compound fraud, confidence calibration

Add multi-agent architecture: Regulator, biased Generator, Auditor rewards

Add Gradio web UI mounted at /web for interactive agent testing

Add WebSocket /ws endpoint required by openenv-core GenericEnvClient

Fix: score formatting was rounding 0.9999 to 1.000 in stdout logs

Fix: clamp all remaining hardcoded 0.0/1.0 score returns

Fix: clamp all task scores to strictly open interval (0, 1)

Add adversarial, negotiate, supply_chain tasks + dynamic difficulty + richer rewards

Fix concurrent state conflicts with session-based environment registry

Add expert fraud audit task and improve inference feedback loop

Fix openenv validate issues

Add full invoice processing pipeline environment