From #1076

We now have 5 texture bind point type and 3 sampler bind point type. Now that the matrix has grown so much, the claim "people just know what bind point to pick" is more difficult too justify. In particular, our matrix is now significantly larger than it is for any native API.

We should investigate just how costly deferring certain kinds of validation is.

I think there could be three different results from this analysis:

1. Deferred validation is too expensive, we shouldn't allow authors to trigger it, no way no how.
2. Deferred validation can get expensive, but for simple cases it's not, and the benefit of allowing authors to specify less information up front is sometimes larger than the runtime cost of deferred validation, for some apps
3. Deferred validation is somehow super cheap, and we can just always defer it and not require any of this complicated stuff up front