Reject any mixed-script confusable (and for certain types of inputs, like email addresses, validate sub-components separately -- someone might get to choose their local-part but not their domain).

For strings which aren't mixed-script confusable, if you need to use them as identifiers and perform comparisons with them, normalize to NFKC and case fold. Doing so will eliminate any ligatures, stylistic variants, composed versus decomposed forms, or other things that people use to call Unicode "weird".