Now that the solution is out I'll clarify the thing I'd been cryptically alluding to. Somehow the two possible grid fills had to be disambiguated, and the solution notes have explained that the correct route was to realise only one fill could lead to an instruction:
https://www.listenercrossword.com/Solutions/S2023/Notes_4751.html
That is what I did, but in non-numeric puzzles I feel like I'm doing something unintended when I deduce the rest of a half-uncovered message and use the information to work backwards (eg to work out which extra words to remove from remaining clues). I always do it, of course (I need all the help I can get!) but I hadn't realised it was intended.
What I'd expected/assumed, and was failing to ask clearly about, was that a subtlety lay in the last sentence of the preamble: the different solution "satisfies all the clues", but that doesn't mean it has to satisfy the rest of the preamble -- specifically, the restriction on there being no duplicate entries. So I'd assumed that at some point in the correct solve path there must have been a deduction based on avoiding a duplicate entry, thus taking the solver down the path to the instruction-giving solution. And that following the instruction would take us to the end of the other solve path, where the duplicate was allowed.
So I was looking for a pair of duplicate entries to confirm I'd done things correctly, but I couldn't remember making the relevant deduction and had assumed I'd cheated with the message deduction. (Actually I had a partial message and deduced the likely missing letters, which gave me something to check the remaining digits against later. So a bit of "cheating" after all...)
Thanks again Oyler!