Multihash seems to argue that by using an TLV encoding on the hash, that will make it clear that the hash length is dynamic, so tooling will design with that in mind.
Simply assuming that you can rev the version of your protocol in the future doesn't communicate to outside developers that, for example, hash lengths in your protocol could change in the future.
I'm not sure I agree with Multihash's argument, though, and on average I agree with your methodology in systems I design. Keeping the crypto fixed per version makes implementations and tooling simpler and thus easier to debug/less chance of [INSERT WORD]Bleed. That outweighs the supposed future proofing of things like multihash, in my mind, and I also feel a lot of designers won't take the hint even when presented with a TLV encoded hash.
Is the suggestion that applications, which have now switched to SHA-256, should serialize this using Multihash, in order to be able to easily switch to a new hash function in 15 years, when SHA-256 is broken?
I think, for most cases, it makes more sense for a protocol/application to just define that v1 uses SHA-256, and, if this ever breaks, we'll update the protocol/application version (rather than change the protocol's hash serialization format). Unless you explicitly need to support many different hash functions.
For example, let's say your app needs to securely hash many files, and store all these hashes somewhere. Would you rather serialize thousands of hashes, all of the same type, each with a pre-pended identifier and length? Or would it make more sense to put the hash function identifier in a single place, and use this for all the hashes? So perhaps Multihash would benefit from separating the versioning scheme from the serialization format of the hash itself?