### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

When trying the upgrade from `v1.81.0-stable` to `v1.81.3-stable` (released few hours ago) and running our integration tests, we have detected that all the pass-through tests related with gemini-3.0 (flash and pro) and imagen-4.0 have stopped working.

They have been working perfectly since long ago and, in fact, downgrading the testing instance back to `v1.81.0-stable` makes the tests to be back to pass.

More yet, we have detected that the problem happens **when you've some default fallback model defined**, because with the `v1.81.3-stable` version it seems that the fallback is always used when trying the vertex_ai pass-through. If I disable the default fallback, then the tests are back to pass.

We also have other fallbacks (anthropic) in the tests are those are not affected.

### Steps to Reproduce

1. Setup a `v1.81.3-stable` instance, and ensure that some gemini 3 (preview), or imagen 4 model is configured. This is our config: ``` - model_name: "gemini-3.0-flash" litellm_params: model: "vertex_ai/gemini-3-flas…