and example prompts.
The LLM sometimes embeds late-arriving characters
(Mom:, Wizard:) inside another speaker's turn
instead of starting a new Speaker N line. Parser
now detects both forms of tags, splits them
correctly, and assigns each unique character its
own speaker number while preserving explicit
Speaker N slots. Adds 15 unit tests covering the
regression and 5 example prompts for manual
verification.
