For as long as I have been wrestling with this problem, I believe that I have in fact accumulated enough evidence to conclude that it is memory-related and that all of the intended files are being accessed.

The program, when it runs in CGI (and not when it runs in the hosting-company's test-rig, or from the command line), can be shown to be running under a ulimit of approximately 38 megabytes. It can furthermore be shown that the aforesaid limits exist only in that case. When run, in any environment, under the same conditions, the failure scenario can be reproduced.

I grant that you are entirely correct in your theory, as it applies to a great many cases, and I would stress “to those who come after us” that the recommendation you make should be carefully considered. But I feel comfortable now in saying that the root cause in this case really is “memory.” I say that in part because I have indeed demonstrated that the correct .so’s are being found, in the correct locations, and loaded (but they do not initialize). (And furthermore, I have been able to capture the $@ output from their attempt to do so.)

Thank you, nevertheless, for the important suggestion:   it is what one should consider first; and in fact, I did so.