The point is that evidence is an agreement between the two sides in a case, and it's not an absolute thing.
If you have the original document that was signed by both parties, great. If you have a scan (using a lossless compression format) of the document and proof that the original was destroyed, great. If you have a scan but no proof of destruction, still great. If you have a photograph of the document and no proof, still great. If you have a vague recollection of what was in the document, still great. All of these are "great" if the other side accepts that they are accurate depictions of the original. If they don't accept that, then there's an argument about what the original document contained and the provenance of the evidence, and only then does the actual quality matter. Original document with wet signature is hard to argue with (but not impossible - wet signatures can be forged). The further away from that, the easier it is to argue that the document presented is not accurate and should not be accepted as evidence.
Knowing that it's possible to use collisions to create false evidence doesn't matter if no-one contests that the evidence is false. It only becomes significant if one side says that the document has been tampered with, and that's not that common. The side claiming it was tampered with would have to present their version of the document, and their version of events that allowed the document to be tampered with, and so on. The judge would make a ruling about which version of the document was considered the "real" one and the case would continue. Obviously there are edge cases where the whole trial verdict hinges on which version of the document is the correct one, but they're edge cases. And in those cases you could-re-hash the documents involved and double-check with one was right, etc.
In the OP's example, where a letter of recommendation has the same hash as a authorisation letter, this is only going to matter if one side says the accused was authorised and the other says they weren't. The authorisation letter will be produced by one side, and the recommendation letter produced by the other, and there'll be an argument about which was the original document. The fact that they have the same hash isn't really relevant. It's a minor point of interest given that these are two clearly different documents saying different things.
In the specific cases for the SFO that I worked on, the SFO descended on the accused's offices like locusts, sweeping every single document into carefully numbered bags. We scanned the documents in secure facilities, stored the originals in secure facilities, stored the resulting images in secure storage, and deleted any cache or copies. My professional opinion is that it would be impossible for anyone to create two documents prior to the SFO's investigation that would create an intentional MD5 collision in the evidence used in court. And, even if they somehow did, it wouldn't matter because both documents would be in evidence bags in storage and could be recovered to be examined by the court.
Obviously, from a black/white technical point of view, using a better hash algorithm would be better. But I can see why the legal profession is reluctant to adopt the new thing; it's a hassle and it will only affect a tiny amount of cases, if any.