If you alter one of the built-in text search configurations, the modifications to it will not get dumped by pg_dump, and thus won't get propagated by pg_upgrade leading to silent behavior changes in the new cluster (as well in any other type of restoration from pg_dump output)
I would say it is bad practise to alter one of those anyway, rather than copy and alter the copy. But I wasn't aware until now of just how bad of an idea it was. I think that this needs to be warned about in the docs for ALTER TEXT SEARCH CONFIGURATION. If I were monkeying around in $SHAREDLIB/tsearch_data, I'd expect traps like this, but not when executing things from SQL after reading the docs on the command.
Cheers,
Jeff
