The implementation notes (section 19.3) say that for Unix pathname namestrings "Name and type may be STRINGs of any LENGTH (consisting of printing CHARACTERs, except "/")".
The documentation might be said not to describe all Clisp builds: the character repertoire that may occur in a Unix namestring turns out to be the set of characters whose encoding under CUSTOM::DEFAULT-FILE-ENCODING contains octets satisfying the expression VALID_FILENAME_CHAR in config.h.
(What might be a separate issue is that for reasons I haven't figured out, on one OSX host where I've built Clisp without any specific configure options or config.h editing, VALID_FILENAME_CHAR excludes octets above #d127, even though the resulting Clisp supports the Unicode character set. Is there any reason for VALID_FILENAME_CHAR to do this when building a Clisp with Unicode characters?)