Format speech markup only where transcript text is rendered

When assistant speech text doubles as transcript text, TTS control markup can leak into the visible conversation. Convert only the known markup pattern at the display boundary, leaving the speech input and stored transcript intact; this preserves audio behavior and avoids rewriting historical records. Limit the conversion to assistant turns and a specific grammar, such as grouped digit spans separated by pause tags, so unrelated user text and other markup are not silently changed. If several markup forms appear, inspect their source contract before broadening the formatter.

Applies when: one assistant text value serves both speech synthesis and a human-readable transcript. Keywords: transcript display TTS markup phone number speech tags rendering