Thanks for your input, I'm feeling more comfortable that I correctly understand the problem now and it's "just" a collation related issue.
I recommend running amcheck on all indexes, or at least all possibly-affected text indexes.
Will the amcheck reliably identify all issues that may arise from a collation change?
I don't enough about the details of B-tree index to know whether the problems are only "obvious" if they happen to interact with some boundary, or whether it will always be evident with the basic amcheck, ie
bt_index_check(index regclass, false)
and therefore I can just use that to identify bad indexes and recreate them, or should I recreate all btree indexes involving text fields?
We planned to do a dump/restore upgrade to PG14 in a month or so (already done in dev). I am wondering whether it will be less work overall to bring that forward than rebuild these indexes...