Can you attach a screen shot of how the render test would look without the fix to make sure the render test covers this scenario?

Is the code coverage report correct by highlighting that the changed code is not covered in tests?

Render test screenshots
On main branch:

On current branch (correct)

For the coverage report, this is a bit more complex.

I assume the coverage report is based on unit tests. Unit tests check only logic responsible for adding and omitting tiles. They do not test matrix computation. This is because I believe tests should be easy to verify, it should be clear if they are correct, while in such tests we would compare some magic numbers in 3D matrices.

This logic is however tested in render test - but only a part of it. For the matrix there are 3 different scenarios, depending on a image and terrain tile zoom level (might be higher, lower or the same). I tested only the case which was relevant for the reported issue. My assumption was that I shouldn't add too many render tests, I didn't want to make them too granular covering very narrow edge cases.

Let me know if and how I can improve that. Should I add unit or render tests? Or current state is ok?

Both render tests and unit tests are added to the coverage report, I think you might be able to find it in codecov site, unfortunately it doesn't get printed here due to a big on their side.
The requirement in general is that new code that was added will be covered (hopefully fully) with tests.

As far as I can tell from the code coverage report, the last else was not covered before this PR and seems like it is still not covered after this PR.
I'm not sure I know how to get to this code path, but since you have changed it, I think it would be wise to try and get to it using either a render test or unit test.
https://github.com/pstaszek/maplibre-gl-js/blob/4884e54f27a509b3b6535f0b7623fee6023d586e/src/source/terrain_source_cache.ts#L239

Let me know if there's anything else I can help with.