The ugly part is that the gap usually stops being about Toffoli counts and starts being about memory traffic, which is where my brain goes back to GPU MSM work. I once shipped a CUDA prover kernel where the paper cost model looked fine, then the actual rig spent most of its life waiting on HBM and PCIe, so the “fast” version lost to a much dumber layout with worse asymptotics but better locality. For ECDSA key sizes, the resource tables are useful because they force you to ask what actually scales, qubits, depth, magic state throughput, not just “Shor someday.” The annoying answer is that once you pin those down, the headline number can move by orders of magnitude, and the bottleneck tends to be one of the boring ones nobody puts in the abstract.