Mark dictated messages so Codex can account for transcription errors

When I use Codex Desktop’s dictation, the audio is transcribed into text and then submitted as an ordinary user message. Codex therefore cannot distinguish typed text from speech-recognized text.

This matters because dictation can misrecognize filenames, commands, model names, numbers, and technical terms. The model should be able to treat voice-derived text more cautiously.

Requested behavior:

• mark a message as voice-dictated in the conversation metadata;
• let Codex use that signal to flag or clarify suspicious technical terms;
• optionally offer conservative, context-aware cleanup before sending;
• preserve the raw transcript and let the user review changes;
• never silently alter commands, paths, identifiers, or user intent.

This is narrower than #13714: the core request is provenance metadata and cautious interpretation, not necessarily automatic rewriting. Related: #42078.

Start by tracing Codex Desktop dictation into conversation metadata and message submission; the issue names no file, test, or entry point. Done means voice-derived messages retain provenance and raw transcripts, with any clarification or cleanup remaining conservative, reviewable, and non-destructive.