Add idle self-play training

Try /data storage bucket with runtime fallback to ./ai/ if write fails

Drop bucket mount approach, persist via Hub API pushes on each game save instead

Seed model weights to DATA_DIR from git before training (fixes ENOENT /data/model_weights.json)

Fall back to ./ai/ if /data is not writable (EACCES on mounted buckets)

Add logging for DATA_DIR and game saves to debug recording

Persist game data and trained weights via HF Hub API (no bucket needed, uses built-in HF_TOKEN) + defer server start until init complete

Fix loadGames is not defined (missing import after fork refactor)

Auto-detect HF /data persistent storage; no env var needed

Store training data and trained weights in DATA_DIR (configurable, survives HF Space restarts); fall back to git-tracked model_weights.json

Run AI training in child process (fork) so server stays responsive during training

Auto-train AI every 5 games: background training, games-since-last-train counter, UI shows pending status

Train AI from human games: game recording, JS backprop training, value head init, Train AI button, record toggle

Add AI mode: neural network opponent with MCTS-guided self-play training

Update server.js

End game when all macro squares are owned (majority wins, tie=draw)

Fix: lines through owned boards still count; only award unowned boards in the line

Revert: owned board cells are NOT nulled in global line detection grid

Remove board-sending mechanic, restore free-play; keep owned-board cell nulling in global line detection

Fix: null out owned board cells in global line detection; add board-sending mechanic

Update server.js

Upload 9 files