Great breakdown of the current landscape! You nailed the trade-offs: TEEs rely on hardware trust, ZK struggles with multi-party state, and FHE is computationally heavy.
The piece you're missing is Garbled Circuits (GC).
GC is a form of MPC that allows for the "joint computation" you mentioned (vital for DEXs/AMMs) but is significantly lighter and faster than FHE. COTI's implementation is benchmarking up to 3,000x faster than current FHE solutions, solving the latency issue while keeping data encrypted during use.