In addition - the HL test_table also fails to do the right thing, but the failure is not checked.

I'm assuming what's happening is the compound type in memory has some field alignment that doesn't line up with the compound stored in the file and when they use H5Tget_member_offset to get the offset of a particular field, they're assuming it's aligned at the same spot as the in-memory compound, which may not be the case on some platform vs. another. I could be completely wrong, but that's my first intuition.
/* get the member offset */
member_offset = H5Tget_member_offset(tid_3, (unsigned)i);

if (H5TBget_field_info(loc_id, dset_name, NULL, NULL, src_offset, NULL) < 0) is used before this in a call;
/* get the fill value attributes */
if ((H5TBAget_fill(loc_id, dset_name, did_1, tmp_fill_buf)) < 0)

Replacing the src_offset from the H5TBget_field_info fixes the problem