[codex] Make voice input opt-in for MVP image

[realadeel]

Summary

Makes the default agentos-image-minimal image text-first by removing the microphone/ASR pipeline from the default install set and making voice input opt-in through AGENTOS_VOICE_INPUT_INSTALL.

The image still keeps the core agent control plane, local LLM packages, and optional TTS output. README guidance now describes microphone input as an adapter rather than an MVP dependency.

Impact

• Default builds no longer install voice-pipelined, whisper-cpp, whisper-model-tiny, python3-webrtcvad, or speexdsp unless explicitly enabled.
• Text input, brokered actions, inference routing, audit logging, and TTS remain available for MVP trials.
• README quick-start health checks no longer expect voice-pipelined in the default image.

Validation

• python -m pytest tests/ -q -> 114 passed