perf: preload both models in parallel background threads for instant UI + fast first inference

perf: implement lazy loading for LLM models to enable instant Gradio UI startup

fix: improve semantic overlap validation, strip task prefixes, and refine fallback prompts

fix: resolve port 7860 conflicts from zombie processes during restart

fix: download models at build time to prevent HF Space startup timeout

Use prebuilt CPU wheel index for llama-cpp-python

Bump llama-cpp-python to 0.3.29 to support minicpm3

Revert llama-cpp-python to 0.2.90 to support Nemotron model architecture

Install git in Dockerfile to satisfy HF Dev Mode build requirements

Fix requirements.txt structure and pin llama-cpp-python to CPU pre-built version 0.2.62

Update requirements.txt

Create packages.txt

Update requirements.txt

Install build-essential for llama-cpp-python

Refactor Hugging Face sync workflow

Change colorFrom from black to gray

Update Hugging Face repository ID in workflow

docs: update setup instructions and clarify repetition mitigation strategy in README

refactor: update grammar, remove unused project files, and clean up workspace configuration

refactor: remove experimental test files and enable production mode in Dockerfile

feat: increase max_tokens to 64 and add model download instructions to README

feat: add extensive test suite and debugging utilities for Gradio app API validation

feat: implement event handler wrappers and add test suite for Gradio components and app integration

refactor: update startup command to use app.py instead of uvicorn and add HF_TOKEN note to documentation

refactor: replace FastAPI/Static setup with custom Gradio UI and update task generation with grammar-enforced inference and semantic repetition checks.

perf: bound session trace memory

docs: update project documentation to reflect fine-tuned model integration and remove standalone model dependencies

refactor: limit task history to three items and add explicit constraints to prevent fallback repetition, with added test script

fix: refine system prompts to enforce physical actions and single-sentence output while adding repetition detection for model responses

feat: improve task generation logic with line-by-line parsing, refined system prompts, and enhanced state tracking for raw task history

feat: implement MiniCPM fallback to break down rejected tasks into atomic steps and add verification script

refactor: expand and diversify dataset entries while removing deprecated testing scripts

feat: update training pipeline to use full-precision merge and add volume verification test

refactor: remove post-processing constraints in app.py and update model training to export HF merged format via llama.cpp conversion

feat: create export directory before saving GGUF model

feat: update GGUF export path and implement volume commit logic for training artifacts

feat: disable model checkpoints during training in modal_train.py

feat: implement Unsloth training pipeline with Modal volumes and update documentation on model persona mitigation

refactor: optimize inference with async locking, prompt engineering, and input post-processing while adding comprehensive testing suites

feat: add environment variable support and expand grammar allowed character set

Merge branch 'main' and remove venv from tracking

chore: upgrade python version, add git dependency to modal image, and update gitignore entries

Delete __pycache__ directory

Delete .venv directory

feat: harden Step-Zero demo runtime and trace export

feat: Generated proprietary Fog of War SPA interface using OpenAI Codex

feat: initialize project configuration and implementation plan for Step-Zero