This would be great. I'd have a use for this in unicode-collation; if we could quickly determine that the text is already normalized and avoid the expensive normalization step (which accounts for about 1/3 of my run time according to profiling and benchmarks), that would help us achieve parity with text-icu.

We can do that, its just a matter of putting some time on it. In the meanwhile we can try to figure out how fast unicode-collation can go with the current impl itself. I guess there will be some avenues to optimize irrespective of this.

If you have any suggestions for optimizing, I'd be glad to hear them.