For what it's worth, I just tried the attached little stress test on
an updated C port of `decode-char' in order to check whether it
returns equivalent results. It does. (Well, except intentional
differences like that `ucs_to_internal' throws an error where
`decode-char' returns nil.)

Basically the test runs through all positive integers up to MAX_CHAR
and inserts an alist into a temp buffer with each car being the
integer and each cdr being a character in the \u syntax (e.g.
`?\u3b1'). It then reads that alist again and checks whether
`decode-char' on its car is `eq' to its cdr. I tried it with and
without `utf-translate-cjk-mode' and with and without
`utf-fragment-on-decoding'. Since all tests succeed, ucs_to_internal
and `decode-char' are functionally equivalent on all supported
characters.

The test: