You may have used file, a command line tool that identifies files by looking at magic numbers, special values in special locations of the file. Or maybe other tools that use the derived libmagic, like file managers or webservers. The results are like “Blender3D, saved as 32-bits little endian with version 2.49.0002” for .blend files (if it says something about gzip instead, run with -z).
The file format evolves (around version 2.62 something changed, previous versions than 2.42 seem to be detected fine), so the recipe in charge of detecting the .blend details needs updates, and files to be tested against. Best samples are .B.blend / startup.blend / userpref.blend as figuring what version saved them is easy. Other files are OK if you don’t mind making them avaliable (at least to me), and in reality I don’t need much more than the first bytes, to end of GLOB chunk, which seems to be less than a KB for some versions and first 128KB for others (extra data added before GLOB) from the examples I already have. If you don’t want to share the full file and can cut just the start (dd if=somefile.blend of=cutfile.blend bs=128K count=1), it should work too for this purpose.
Most interested in files saved as big endian (as in PowerPC processors, probably running OSX or Linux) 32 and 64 bits, and little endian 32 bits (as in Intel/AMD). That is the catch, uncommon/disappearing architectures. Older than 2.42 are not needed, and the more problematic and interesting seem to start around 2.62.
Thanks in advance.