I do see the consensus capture argument by miners but in reality isn't this attack scenario have a lot of assumptions on topology an deployment ?
For such attack to succeed you need miners nodes to be connected to clients to feed directly the invalid headers and if these ones are connected to headers/filters gateways, themselves doing full-nodes validation invalid chain is going to be sanitized out ?
Sure now you trust these gateways, but if you have multiple connections to them and can guarantee they aren't run by the same entity, that maybe an acceptable security model, depending of staked amount and your expectations. I more concerned of having a lot of them and being diversified enough to avoid collusion between gateways/chain access providers/miners.
But even if you light clients is directly connected to the backbone network and may be reached by miners you can implement fork anomalies detection and from then you may have multiples options:
* halt the wallet, wait for human intervention
* fallback connection to a trusted server, authoritative on your chain view
* invalidity proofs?
Now I agree you need a wide-enough, sane backbone network to build on top, and we should foster node adoption as much as we can.