Typo generation

Each source item is treated as exactly one word. For every configured typo distribution and every source word, MULTYPO is sampled generation_attempts_per_word times.

MULTYPO itself receives typo_rate=1.0 for each attempt because every sampling attempt is intended to corrupt the supplied word. The project therefore does not expose the old custom typo_rate argument that previously scaled the number of iterations.

The low-level generator uses insert_typos() directly rather than the sentence-oriented insert_typos_in_text() wrapper. Sentence tokenization and NLTK resources are therefore unnecessary for this workflow.

Raw samples are aggregated by noisy form. A noisy form that maps to exactly one target word becomes a candidate; one that maps to multiple target words is recorded as an internal clash and removed before AutoCorrect2 processing.