INGRID can hear you now

INGRID has been able to speak for a while. This milestone is about the other half: actually listening.

The full loop now works end to end, live-verified on real hardware — not just passing tests against a faked HTTP response:

• Real microphone capture. Push-to-talk recording runs through Windows’ own audio stack, with a temporary diagnostics view built specifically to answer one question honestly: is the audio actually reaching the recognizer, or failing silently before it gets there? That diagnostics work paid off immediately — it caught a genuinely disconnected headset defaulting silently to a near-mute onboard microphone, which would have looked like a “recognition” bug if we’d only looked downstream.
• OpenAI speech recognition. What you say gets transcribed by OpenAI’s Whisper, with Windows’ own local speech engine as an automatic fallback if no key is configured or the request fails — voice conversation still works with no internet connection, just less reliably.
• Marin, an OpenAI voice. INGRID’s spoken replies now use Marin, a natural OpenAI voice, through the same reasoning pipeline typed chat already used. Getting real audio playing correctly took fixing a genuine, live-discovered bug: OpenAI’s streaming response format uses a placeholder header value our audio library didn’t know how to read, silently failing playback until we found and patched it.
• One shared state, not a guess. The reactor emblem and avatar reflect whichever phase of the conversation is actually happening — listening, thinking, speaking — driven by the same live signal bus the rest of BlueCrow’s presence work already established, so nothing here is a separate visual system that could drift out of sync.

None of this replaces typed chat — it’s an addition, entirely optional, and it degrades honestly (to the local fallback, or to a clear error) rather than pretending to work when a step upstream has failed.