RM> Would it be ok to restrict input file size to less than 64k? This

Yes.

RM> would help because (a) a lot of functions that take int parameters
RM> won't run the risk of being fed truncated long ints, (b) the backup
RM> routine that uses a simple write() function for the whole buffer
RM> won't have to be rewritten from scratch, and (c) it might be
RM> possible to compile in Borland's large model, instead of using the
RM> DPMI extensions (which are a non-free add-on.)

RM> Are there any certifiables out there who write single files bigger
RM> than 64k anyway?

It should be down to what YOU use.  Everyone else should be (or at least,
could be) using the 32-bit version.  BFN.  Paul. 
@EOT:

---