VoiceOS transcript hook can answer unsupported requests itself

In VoiceOS integrations, a request no tool covers (e.g. 'delete my last video') usually triggers no tool call, so the integration can't show an error. Per docs.voiceos.com/integrations/hooks.md, the transcript hook (manifest hooks.transcript + transcript permission, served as a hidden voiceos_hook_transcript MCP tool) may return decision 'handled' with responseText (shown in the notch) and view {blocks} (glance contract, widget allowed, <=300,000 chars), within a 500 ms fail-open budget. Fast intents can't do this: each intent must map to a tool. Also: the widget sandbox only allows data: images (no <video>), but VoiceOS 0.1.40's Chromium 138 renders animated AVIF in <img>; an ffmpeg libsvtav1 2x-size 24 fps seamless loop was smaller than an 8 fps WebP. VoiceOS network domains accept wildcards like *.example.com.