> I'd rather see the xpubs shared in the global section of the file,
> with the restriction that they must/should only include the hardened
> prefix of the path. The existing bip32 derivation path included in
> individual inputs and outputs be merged in as needed.
> After all in a typical PSBT, we would expect the same master keys
> to be used on all inputs, and at least one output, and there might
> be as many as 20 co-signers. No need to repeat all that information.

I agree that it makes sense to put xpubs to the global scope.
But I am not sure that restricting xpubs to have only hardened derivation is necessary.
People may want to share non-hardened xpubs with co-signers and keep parent xpub on there watch-only wallet.
For example, in bip45 cosigner_index is not hardened, and sharing top level xpub is not necessary.

> Even with this additions to the PSBT format, I think PSBT-signing
> devices still need to store the xpubs of their co-signers. It's not
> possible to safely show an incoming address to the user without a
> full understanding of the other keys in a "multisig wallet". Also,
> it represents data that should not change between PSBT instances
> (ie. tomorrow's co-signers should match today's).

I would like to keep hardware wallets state-less, otherwise wiping and recovering them would be problematic.
At the setup phase the user can verify a multisignature address (or xpub) on the screens of all devices, 
after that we just need to verify that xpubs in the inputs and in the change output are the same.

Andrew, regarding your question: 
> Just for clairty, by xpub, do you mean the extended serialization format 
> defined in BIP 32 or the Base58 check encoded string of that serialization?

As PSBT is a binary format I think it makes sense to use extension serialization format without any encodings. 
I am not sure if we need the whole xpub or keeping chain_code and public_key is enough, but I would suggest to keep other data
just in case. For example, keeping prefix that defines if the key is used for testnet or mainnet may be useful.

Best regards,
Stepan