So, I went through the exercise of writing my own implementation based on a reading of the document, but running the tests from the sample-implementation code.   I hope these are helpful.

1)
The spec for text-unfold requires the mapper and make-final to return a character, string, or text.
However, several test cases rely on support for a return value of #f also which is treated as an empty string or text.   ((and (null? x) "G")).

2)
In the spec for textual->utf{8,16,16be,16le}, third paragraph, "and the rest of those BYTEVECTORS are encoded using that endianness.".    I think this word should be codepoints or characters, not bytevectors.

3)
In the spec for textual-fold and textual-fold right, "The LEFT-FOLD operator maps..."    This should be TEXTUAL-FOLD, and for consistency with the later description of textual-fold-right, the sentence should read "The TEXTUAL-FOLD procedure maps the kons procedure across the given text from left to right:".

4)
The single test case for textual-titlecase fails if string-titlecase from the current sample implementation of (SRFI 129) is used in the implementation.    In that case, "you're" becomes "You'Re".    When I dug a little further, 1) this is 100% consistent with SRFI 129's documentation, and 2) that the Unicode Text Segmentation annex states that "the use of the apostrophe is ambiguous".   Upshot -- maybe it would make everyone's lives easier to use a test phrase without an apostrophe.