Issue draft (short) — media-understanding chosen model should be exposed to the assistant as runtime telemetry

Summary

When a user asks which model analyzed an image/audio/video input, the assistant may answer with a guessed model name if the current turn only contains preprocessed media description text ([Image], Description:) and not the actual chosen media model/provider.

In our Discord test, the assistant replied with openai-codex/gpt-5.4-mini even though the current turn did not expose actual media-understanding telemetry for that turn.

Expected behavior

Either:

1. the actual chosen media model/provider is exposed to the assistant for the current turn, or
2. the assistant should say it cannot confirm the analysis model for the current turn

Proposed fix

Include the chosen media model/provider in the generated inbound media-understanding body, e.g.

[Image]
User text:
...
Description:
...
Analysis model:
ollama/qwen3.5:397b-cloud

This lets the assistant answer based on runtime telemetry instead of guessing.

Root cause

The media-understanding path already knows the chosen provider/model internally, but the inbound body only exposed the description text, not the actual chosen model. That made it easy for the assistant to hallucinate/guess a media model.

Local validation

• patched local apply-41Abj0q9.js to append Analysis model: per media output
• re-tested in Discord #test
• inbound body included:
  • Analysis model: ollama/qwen3.5:397b-cloud
• assistant then reported the same value correctly

Notes

• This is not mainly a conversation-model fallback issue
• The main issue is guessed media model reporting when current-turn telemetry is absent
• Reviewer result on local patch: GO with caveats