It cannot. The Google implementation must effectively be done on a blob as the result would not be usable in a structure specific document. What is more, things that require a block chain (like git) are NOT covered with this current attack as both the source and resulting have to be worked on.
Currently the attack vector only works when you can get both documents to "work towards each other" to produce a valid identical SHA1 value.
This was a fixed prefix collision attack. That means they can make two documents (P | A | anything) and (P | B | anything) for a fixed P, and they can find A, B, such that A and B are different but
SHA1(P | A | anything) = SHA1(P | B anything)
The Merkle-Damgård construction (used in MD4, MD5, SHA1 and SHA2 but not in SHA3 and some other modern hashes) invariably means length extension is possible, if you can collide two documents then you can add a suffix to both and also get a collision.
This is how there's already a web site where you feed it images and it makes a "different" colliding PDF, it's just using Google's result with a different suffix after the 128-byte collision near the start.
Currently the attack vector only works when you can get both documents to "work towards each other" to produce a valid identical SHA1 value.