EPIC: Speech and contextual interaction

Programme: #1

Own local speech-to-text, command/dictation separation, foreground semantic context, correction vocabulary and optional read-aloud/voice output.

## Child issues
- [ ] #65 VOICE-01 — Local speech-to-text
- [ ] #66 VOICE-02 — Semantic commands and dictation
- [ ] #67 VOICE-03 — Read-aloud, multilingual and optional wake word

## Cross-cutting gates
Microphone state is visible, audio retention is opt-in, risky actions always use trusted confirmation, contextual references resolve to selected/shared objects rather than unrestricted desktop surveillance, and keyboard stop remains available.

Start with child issues #65, #66, and #67, which split the epic into local speech-to-text, semantic commands and dictation, and read-aloud or multilingual voice output. Review the cross-cutting gates in this issue, including visible microphone state, opt-in audio retention, trusted confirmation, bounded contextual references, and keyboard stop; the work is done when the child scope and these safeguards are addressed.

Written by the indexing model from the issue text.