will runtime type checking go beyond function parameters and return type?

Great project, helps me understand DL code a lot.

I used it like this:

Patch_embed = Float32[torch.Tensor, f"B {PATCH_H} {PATCH_W} {PATCH_EMBED_DIM}"] 
Mlp_mid = Float32[torch.Tensor, f"B {PATCH_H} {PATCH_W} {MLP_HIDDEN}"]

...

    @jaxtyped(typechecker=beartype)
    def forward(self, x: Patch_embed) -> Patch_embed:
        x: Mlp_mid = self.act(self.lin1(x))
        return self.lin2(x)

but turns out, it doesn't do runtime type check on this x: Mlp_mid = self.act(self.lin1(x)) line. And this makes me feel insecure.
So, my question is, will this feature be added in the future? Or is it in confilct with some design intention?

BTW, I mainly use it when I am trying to understand others code. But can I include it in production? How much does it slow down the training and inference?

The report names @jaxtyped(typechecker=beartype), local variable annotations, and PyTorch examples, but no source file or test. Start by reviewing the runtime-checking behavior around that decorator and local assignments. Done would require a clear decision on whether local annotations should be checked, plus documented production overhead and design intent.