voice: tag dictated messages so the model knows input was speech-to-text

Description

When using /voice dictation mode, the transcribed text is sent as a plain user message with no metadata indicating it came from speech. The model has to guess from speech patterns (repeated words, filler, self-corrections) whether input was typed or dictated.

Use Case

If the model knows a message was dictated, it can:

• Parse through disfluencies instead of treating them as typos
• Extract user intent more accurately from spoken structure
• Avoid pedantic corrections on speech artifacts
• Adapt response style accordingly

Proposed Behaviour

Wrap dictated messages with a tag or system hint, e.g.:

<voice-dictation>transcribed text here</voice-dictation>

Or include it as metadata in the message payload that the model can read.

Current Workaround

The model infers dictation from speech-to-text artifacts (repeated words, self-corrections, filler phrases like "so basically"), but this is unreliable and requires the model to spend tokens on guesswork.