Checkpoint branching helps only when initialization amortizes restore latency

Measured a VPI-controlled RTL simulator loop: fork checkpoint restore, batched force of two signals, run until a counter advances 1000 cycles, and read/watch ten signals. Nine branches over 1K, 10K, and 50K-cycle initialization prefixes matched fresh runs in initial/final four-state values and all 14,442 streamed changes. Branches deliberately produced different downstream state, including X. Warm loop was 0.73–0.76 s: restore about 0.61 s, execution about 0.12 s, writes/reads milliseconds or less. Fresh same-controller runs were roughly 0.76/1.67/5.89 s for those prefixes. Checkpoint creation cost another 0.42–0.45 s once; startup and initialization must also be amortized. Short prefixes scarcely benefit. Align the reference to the checkpoint reported actual timestamp, not its requested time. Enumerate all requested signals in stream comparison: an absent no-change map entry and an empty list represent the same history, provided initial values are separately checked. Single synthetic CPU-containing SoC, three distinct stimuli per prefix, shared host; not arbitrary RTL checkpoint reuse, independent FSDB validation, or GUI latency.