Read GEMMA_SERVER env var; log it in __main__ block

Route /api/pick_move to Modal T4 GPU when GEMMA_SERVER is set; local CPU fallback retained

Drop round-end flavor subtitle from RoundOverlay

Speed up Gemma: shorter prompt, max_new_tokens=18, end_of_turn stop, LRU cache(64)

Fix NameError: WEIGHTS_PATH -> ADAPTER_DIR in __main__ block

Fix BatchEncoding RuntimeError in _generate (transformers 5.x compat)

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload Dockerfile with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload gemma_npc.py with huggingface_hub

Upload app.py with huggingface_hub

Upload folder using huggingface_hub

fix: TDZ crash - move isLocalDev above callbacks that reference it in deps

fix: rebuild frontend - throttle, immediate fallback, rAF polling, match chronicle, duel lifecycle

fix: LayerNorm, inference_mode predict(), BN->LN remap

fix: decouple Modal, strategy cache, duel lifecycle, health ready

fix: upload binary PyTorch weights (was corrupted by CRLF in git)

remove: root tiny_fighter.pt from tracking

remove: binary file from git tracking (was corrupting Space build)

fix: add .gitattributes for binary files, Dockerfile --chown for weights

fix: add .gitattributes to prevent binary file corruption on Windows

chore: remove binary model weights from git (upload via HF API)

feat: hybrid two-tier architecture - Tiny Fighter (142k CPU) + Nemotron (Modal)

fix: support frontend sequence format in Modal /strategize endpoint

feat: Nemotron Duel - fighting game AI with SFT training on Modal A100

Initial commit