BlindBit does exactly this. I will make sure that this is emphasised in the Specification.
Not yet. I will try to get some numbers.
The main idea was to reduce bandwidth. I wanted to condense the block in such a way that only the relevant information is included. Downloading the entire block would probably include quite a bit of information that the light client will never need. I believe, having raw block data would also require the light client to do some extra work with regards to parsing and finding eligible transactions, right? With the current specification the client can directly use all the information and has to do minimal work on its own side. With this method the client still finds all labels and can immediately spend as well.
Also in general simplified UTXOs might not be the correct wording for this. The basic idea is that it’s a data structure which contains all necessary information to find and properly spend a UTXO.
This is probably an artefact from testnet, as there are a lot of blocks were no tweaks exist and we can save bandwidth by not requesting filters. On mainnet this is not the case. Apart from that I can’t think of a good reason. I will merge those two steps into one.