Resuming a session has strange ux with regards to model and reasoning

### Describe the bug

If you have a session that's was done on `gpt-5.6-terra` with `medium`, and you `--resume` it, with either `--model=gpt-5.6-luna --reasoning-effort=xhigh` or `gpt-5.6-luna + xhigh` in your `~/.copilot/settings.json`, the `reasoningEffort` is honoured, but the model is not, and your resumed session will be on `gpt-5.6-terra xhigh`

Feels like this should be consistent? It should either:

- Explicitly honour the original session, keeping the settings the same
- Explicitly honour your copilot CLI args and/or settings.json

I would rather the latter personally, but raising this as a UX bug.

This also affects sessions resumed via copilot-sdk.

### Affected version

1.0.77

### Steps to reproduce the behavior

_No response_

### Expected behavior

_No response_

### Additional context

_No response_

Start at the --resume flow and trace how --model, --reasoning-effort, and ~/.copilot/settings.json are applied when restoring a session. Check the corresponding copilot-sdk resume path as well. Done means the chosen precedence is consistent for both CLI and SDK sessions, with regression coverage for model and reasoning effort.

Written by the indexing model from the issue text.