Read markdown aloud as plain speech #243

Story

As a family member, I want spoken answers to sound like speech, so that the agent never reads out "asterisk asterisk" or a URL.

Context

The speech server (Piper) reads markdown marks aloud. For one sentence the spoken length was: plain 2.0 s, with **bold** 3.9 s (read as "asterisk asterisk"), with a link 4.1 s, with a heading 2.7 s, with emoji 2.7 s. The agent's speaking instructions and the say help line ask for plain sentences, but a model does not always follow them, and a later routine (a morning briefing) will pass longer text.

Proposal

ai say turns its input into speakable text before it is spoken, as a pure function: emphasis marks go, a heading becomes a sentence, list items become sentences, a link becomes its text, a bare URL becomes its domain, inline code keeps its content, emoji go, a table becomes one sentence per row.

Definition of done

• A module test feeds markdown with each of those elements and expects the plain sentences.
• Plain text passes through unchanged.
• On a running instance, ./stack ai say "Good morning. **Three emails** came in." takes as long as the same sentence without the marks.