Restore native inference, card-bound readings and paged dialogue.
Provide manual, checksum-verified model setup instead of Git LFS so
source checkouts work without uploading multi-gigabyte weights. The
game stays offline; complete release archives still bundle the model.
- Offline neural worker based on llama.cpp (fortune-engine.exe) with local Qwen3-4B model support
- Native assembly Tarot 3-card spread divination and teacher grade divination mode
- Text input dialog in MODE_FORTUNE communicating over anonymous pipes
- Clean process lifecycle management, WM_CHAR dispatch and fallback handling
- Test suites and packaging tools for offline release
