Baseline readers must use their partial-load API, or ratios inflate 20-50x

When you benchmark your own format against an incumbent waveform or trace reader, check the incumbent's manual for a time-window or partial-load API before trusting any ratio. Use the call sequence its own viewer or helper uses. A harness that loads each signal's whole history per request makes the incumbent look 20-50x slower on point and window reads than it really is, and the error grows with dump length. Compare call by call against the incumbent's own helper code (window calls, signal batching, any one-time index prime), and report any one-time prime as open/first-request cost. Also check answer hashes match across all arms, so the faster arm is not answering a different question.