Problem

Model discovery now carries modality metadata such as Text and Image, and OpenAI Codex OAuth discovery can resolve input_modalities from the live Codex catalog. Operators still cannot see those modalities in the main model selection surfaces:

• netclaw model discover <provider> shows model ID, context window, and cost, but not input/output modalities.
• netclaw model list shows assigned model ID and context window, but not persisted input/output modalities.
• The TUI model manager and picker should make modality support visible before assignment and in role overview where practical.

This matters for image-capable models: users should be able to confirm whether a discovered or selected model accepts image input without checking daemon status after assignment.

Desired behavior

• netclaw model discover <provider> displays input and output modalities when the provider returns them.
• netclaw model list displays persisted input and output modalities for assigned roles when present.
• The model manager TUI exposes discovered model modalities in the picker/details flow.
• The TUI role overview exposes configured role modalities where layout allows.
• Unknown modality metadata should render as - or unknown, not as a guessed capability.

Notes

OpenAI Codex OAuth discovery currently resolves modalities from the live input_modalities / output_modalities fields and persists them when a model is selected. This issue is about surfacing that metadata in the operator-facing model selection UI.