Currently, the specs say:
"User agents must filter the suggestions to hide suggestions that the user would not be allowed to enter as the input element's value, and should filter the suggestions to hide suggestions that would cause the element to not satisfy its constraints."

Which can be summarized to:
1. filter not allowed value
2. filter invalid values

I do believe that we should not make the filtering mandatory. For example, in the case of a textfield, filtering the values sound useless given that not allowed value will be sanitized and invalid value will just have to be corrected.
However, I agree that, for some widgets, invalid and not allowed values should be filtered. For example, for <input type='number'> and <input type='range'>, we agree that "foo" wouldn't be a good suggestion here.

So, I think we should make the quoted sentence weaker and remove the "must" to at least a "should" and even add that, "Depending on the widget, user agents should filter [...]".