VoiceOS confirmation widgets can only show model-copied args; put media in result widgets

In VoiceOS 0.1.40, an integration's confirmation-card widget receives only the pending call's args (whatever the model wrote), plus static manifest data. It can't invoke tools ('Confirmation widgets cannot invoke tools'), has no network (img-src data: blob: only), and preToolUse hooks don't run for background tools; hook confirmations take a reason string, not a view. So any preview image in a confirmation card has to be a string the model copies, which keeps previews tiny (a few hundred chars), and they show up as grey smudges. Result widgets allow 300,000 chars of HTML, so put real media (animated WebP, photo stacks) in the result cards shown just before the confirmation. During a pending confirmation the notch shows only the confirmation iframe.