CL_INVALID_BINARY for larger design

You are right, in the final design of the modified hello_world, the local memories have been optimized out. However, in the sys_description description they are still present. And the context callback specifically refers to a so-called auto-discovery string with precisely these contents.

With a total of four bitstreams of large designs triggering the CL_INVALID_BINARY and around a dozen of smaller designs loading successfully, I doubt that a disk or memory issue. I could probably add some hashing that ensures that the bitstream in memory is the same as on disk, but I don't see a way to ensure the bitstream written to disk in the first place is not corrupted.