Voice mode: markdown renders briefly then reverts to raw text when response is interrupted

During a voice-mode conversation, the assistant produced a long response containing markdown (headings, tables, inline code). While voice mode is speaking, the transcript appears to be held in an unrendered/raw markdown state, and rendering is only applied once speech playback completes. The reporter interrupted/cancelled the response mid-speech. At the moment of cancellation the markdown was briefly seen applied (rendered), and then it reverted back to raw markdown source text, which is what remained on screen. This suggests the render pass is gated on voice playback finishing: the cancel path triggers the render, but some subsequent voice-mode logic re-formats or re-flushes the message back to its raw form, undoing the render. The impact is that any markdown-heavy answer delivered in voice mode becomes hard to read on screen if the user stops playback early — tables and headings appear as literal pipe characters, hashes, and backticks. This is significant for workflows where the user is deliberately asking for on-screen structured output (tables, reference material) while in voice mode, since stopping the speech is the natural way to switch from listening to reading.