Restart dot product after a constant subsequence?

[seanlaw]

Currently, when we traverse a diagonal, it may be best to restart the distance (or covariance) calculation after we pass a region in the time series where the time series signal dropped. I forget how this is currently being handled so we should investigate whether this is a problem.

[viknesh-ai]

I actually just traced through the code to check on this!Right now, there isn't any restart logic for constant subsequences. In stump._compute_diagonal, cov is only seeded with a direct np.dot right at the start of a diagonal ($i=0$ or $j=0$). After that, it strictly uses the $O(1)$ recurrence step. When it hits a flat region, it short-circuits the output Pearson value for that specific alignment, but cov itself isn't reset—it just keeps accumulating straight through the flat patch into the next non-constant alignment. (scrump._compute_PI does the same thing with QT_j).For a completely flat window, that's probably harmless since the added term is roughly 0. The case that worries me more is windows straddling the edge of a flat region rolling_isconstant needs exact ptp == 0, so a 'mostly flat' subsequence won't get flagged. It gets a huge inverse standard deviation multiplier applied to a cov that just came out of a long flat run.I don't have a reproduction script yet, just the code trace. Should I try building a test case (a long series with a flat patch away from a diagonal boundary) and diff stump() against naive.stump() to see if numerical drift actually shows up?If it is an issue, the fix seems pretty straightforward: re-seed cov/QT_j with a direct dot product on any constant $\rightarrow$ non-constant transition. That’s $O(m)$ per transition, so performance impact should be negligible unless the series is pathologically noisy/flat.