SRFI 125 states that
(hash-table-update! hash-table key updater)
"is semantically equivalent to, but may be more efficient than, the following code":
(hash-table-set! hash-table key (updater (hash-table-ref hash-table key)))
It seems unnecessary difficult for an implementation to provide a more efficient implementation of the latter if the updater procedure is allowed to mutate hash-table arbitrarily.
In its abstract, SRFI 125 says: "Does not guarantee that whole-table operations work in the presence of a concurrent mutation of the whole hash table (values may be safely mutated)."
If this restriction applies to hash-table-update! as well, the problem raised above won't occur. However, with this restriction, hash-table-update! would not be strictly semantically equivalent to the combination of hash-table-set! and hash-table-ref.
What has been the intent of the SRFI authors? It is possibly a good idea to spell out such a restriction/to clarify hash-table-update! in a post-finalization note.