Hey Adam,
 
Mike had posted a detailed response on the topic on why its complex
and becomes bandwidth inefficient to improve it usefully.

To clarify, we could improve privacy and still preserve usefully high performance, it's just a lot of complicated programming work. You need to find out from the OS how much bandwidth you have to play with, for example, and do all the very complex tracking to surf the wave and keep yourself in roughly the right place.

The basic summary of which I think is that its not even intended to
provide any practical privacy protection, its just about compacting
the query for a set of addresses.

The original intent of Bloom filtering was to allow both. We want our cake and we want to eat it.

The protocol can still do that, with sufficiently smart clients. The problem is that being sufficiently smart in this regard has never come to the top of the TODO list - users are always complaining about other things, so those things are what gets priority.

It's not IMO a protocol issue per se. It's a code complexity and manpower issue.
 
Its seems surprising no one thought of it
that way before (as it seems obvious when you hear it) but that seems
to address the privacy issues as the user can fetch the block bloom
filters and then scan it in complete privacy.

And then what? So you know the block matches. But with reasonable FP rates every block will match at least a few transactions (this is already the case - the FP rate is low but high enough that we get back FPs on nearly every block). So you end up downloading every block? That won't work.

Eventually, wallets need to stop doing linear scans of the entire block chain to find tx data. That worked fine when blocks were 10kb, it's still working OK even though we scaled through two orders of magnitude, but we can imagine that if we reach 10mb blocks then this whole approach will just be too slow.

The main reason wallets are scanning the chain today (beyond lack of protocol support for querying the UTXO set by script), is that they want to show users time-ordered lists of transactions. Financial apps should show you payment histories, everyone knows this, and without knowing roughly when a tx happened and which inputs/outputs were mine, providing a useful rendering is hard. Even with this data the UI is pretty useless, but at least it's not actually missing.

By combining Subspace and BIP70 we can finally replace the payments list UI with actual proper metadata that isn't extracted from the block chain, and at that point non-scanning architectures become a lot more deployable.