Action prompts let LaTeX leak into speech text, so TTS reads "frac" aloud

`slide-content/system.md` (rightly) asks for math to be formatted as `LatexElement`, but the action prompts (`slide-actions/system.md`, `interactive-actions/system.md`) say nothing about speech `text`, so the model carries the same notation into narration: `\frac{7}{10}`, `x^2`, `3/8`. TTS providers read the symbols literally ("frac", "caret", "three slash eight"). Observed on three generated lessons (Turkish, grade 5 fractions): every narration line with a fraction was spoken as LaTeX.