feat(training): add early stopping, cosine LR schedule, and push-to-hub

fix: remove duplicate-tail dead code; harmonise app label vocabulary

feat: expand to 81 ciphers — improved heuristic, 12 new cipher generators, focal-loss training

Improve training: weighted loss, blind eval split, roberta-base defaults, 512 tokens

Fix singleton class crash: normalise wallis_ciphers typo, guard train_test_split

Fix Trainer API: tokenizer -> processing_class (transformers 5.x)

ADA/mobile UI, pytest pythonpath, train_transformer PyArrow fix, enriched Explain Mode

Initial Cipher Detective AI public foundation