Find invisible characters, bidi overrides and script confusables hiding in text.

Scans for characters that are present to a tokenizer but absent to a reader: the Unicode TAG block, zero-width marks, bidirectional overrides, variation selectors, and letters from other scripts that look like Latin ones. This is the mechanism behind instructions concealed in tool descriptions and in text a human has already approved — an evaluation of MCP metadata surfaces found all eight tested techniques delivered a payload into model context, while a representative string-matching sanitizer flagged none of twenty-five benign descriptions. Reports each suspect with its code point and offset, and returns a cleaned copy.