I was slightly vague in my oroginal post. I didn't mean "match completely the soruce data", I meant "match a part of the source data", as a standard rsync/Exchange-style block-deduplicator would do.
Per the challenge, it need not work on every file, only the one the challenger chose to provide. If the challenger is not suspecting this attack, he won't know to make sure to pick one of the many source files that are resistant to the attack.
Of course the compression is impossible in general, if the task is sufficiently rigorously described, which it wasn't in this case.