Sequential occupancy is a different question

Reordering operations on disjoint factors can preserve a channel without preserving the maximum storage occupied during execution. A finite sequential program must price every arrival, allocation, retained early register, and workspace factor on every branch. Branch Kraus maps $K_{\gamma }$ must satisfy

and its complete returned channel, not one normalized branch, must meet the error target.

This is the distinction illustrated in Cheung's batch and streaming benchmarks CheungEn02BC. The batch construction above is not a streaming optimum.

A Yoshida–Kitaev-type decoder requires its own oracle access, reflections, success handling, and complexity hypotheses YoshidaKitaev. Neither Haar typicality nor (eq:sdp) automatically installs that algorithm.