I built a workbench estudio and tried to find out what was wrong using the
debugger.  I stopped in {ENCODING}.convert_to and found that the
source code page is 646 and destination code page is UTF-32.
Then I stepped into {ENCODING_IMP}.convert_to.  The call to set
`l_converted' by calling `pointer_to_string_32' appeared to work correctly.
I could look at `l_converted' in the debugger and it looked reasonable.

Then the code looked at the byte order mark (BOM) and decided the
string was little endian.  Since the destination code page was no-endian
and since the platform was not little endian, the code removed the BOM and called
`string_32_switch_endian'.  After this call, none of the characters were
valid unicode characters.

So the problem seems to be that the routine thinks the value returned from
`iconv' and placed into a STRING_32 starts with a BOM that indicates little endian.

I don't yet know why it (mostly) works on Solaris 9.