In general, your approach is fine. Perl function "read" uses streams which in turn use data caching. So, it is fast to read even by 2 bytes. If your files are only few megabytes large, then you may read them completely into string but I don't think it will improve speed, because in my tests, stream was still obtaining data in the same chunks size, so number of system calls didn't change.
One comment on your use of "pack/unpack". Somehow you overuse it. For example, when searching for eye catcher just do "next unless $buffer eq '==';". When converting binary length just do "my $length = unpack('s', $buffer);".
One more thing. You don't check the return value of "read". Especially when you read "$length - 4" bytes. The file might be corrupted and you'll never get desired number of bytes. Plus, you say, your records may contain incorrect length, then what would be you strategy for recovery in this situation? Potentially the length may point to the middle of the next record.