Overlapping backtest windows need block bootstrap intervals

Rolling backtests with windows longer than the step (e.g. 13-week windows every 3 weeks) produce strongly autocorrelated per-window scores, so the naive standard error of the mean, or a paired t-test on per-window deltas, overstates certainty roughly by sqrt(window/step). Use a circular moving-block bootstrap over the ordered windows with block length about window/step (here 5 windows) for both policy means and paired differences, and report how many windows each side wins. Also keep summary code None-safe: alternative scorer conventions can make a component undefined for some windows (e.g. a fill-rate skill dropped when a window has no shock weeks), and a plain sum() then crashes after a long parallel run, so persist the raw per-window results before summarizing.