A rank-1 guess from the mean profiles

Before any fitting, the notebook builds a free rank-1 guess from the mean profile along each axis: a over neurons (43,), b over time (200,), c over trials (88,). A real rank-1 component, just not the best one.

Direction and size are two separate questions. As built, a ∘ b ∘ c has a relative error of about 1.00 — the mean of a product is not the product of the means, so the tensor comes out far too small. Rescale the same three directions to the best possible size and the error drops to about 0.49. The profiles did not change shape; only one number did.

That 0.49 is the honest baseline that alternating least squares has to beat.