Why it is a question
THEORY-278 now states two things. First, in LIT-974’s fifteen dataset–network cells, the MSE-trained classifier stays near the least-squares classifier of its own features from an early epoch. Second, in an unconstrained-features model whose features are re-whitened after every step, gradient flow on that central path drives the class-separation singular values ω_j of Σ_W^(−1/2)M̃ to infinity at the common rate (4t/N)^(1/4), with their ratios tending to 1. That limit is neural collapse. The continual renormalization is imposed, not measured, the theory sets weight decay to zero while every run used it, and the measured cells are balanced, ten-class and unaugmented.
What would answer it
Track the singular values ω_j of Σ_W^(−1/2)M̃ through the terminal phase of MSE training, without imposed re-whitening, on more than one architecture and with several seeds. Check that they grow at a common rate, that their ratio tends to 1, and that the limiting direction is set by the singular vectors early in training. A derivation that drops the continual renormalization or allows weight decay would answer the first part in theory. The answer is no if, near the central path, the ω_j spread apart or stall while L⊥_LS stays negligible. It is also no if the deviation from the least-squares classifier stops being negligible with augmentation, many classes or imbalance. Further closed-form results in the same renormalized model cannot settle it, nor can NC1–NC4 curves alone, which any account of collapse predicts.