What an Active-Dendrite Continual Learner Actually Learns: A Dead-Neuron Artifact, a Granularity Result That Did Not Survive Its Own Controls, and a Repair That Is a Damping Factor

This record (version 2) deposits the combined manuscript of three previously separate studies of dendritic context-gated protection in continual learning, together with the additional controls that were run after the first review of this line of work. The three original studies were written independently and published separately (Zenodo records 23685865 for S1, 23685385 for S3, and 23686689 for S4); this record supersedes them as a single, self-consistent narrative, and the earlier records are retained unchanged so that their DOIs continue to resolve to the documents they were issued for. The merge was necessary because the three sets of results could not all be true at once: S1 reported the own-segment bias as a causal repair of the missing segment-task specialization, S3 showed that the repair does not transfer to a convolutional backbone, and S4 showed that the mechanism is a saturation of a gate that is never context-dependent. Version 2 adds five controls that a review of version 1 identified as missing rather than as errors, and reports them in a new section of the manuscript, with the earlier sections, the abstract and the conclusion edited to reflect the outcomes. (i) A specificity control with four arms shows that the bias acts on whichever segment the network is using rather than specifically on the task's own segment, and that a pure constant rescaling of the trunk outperforms every bias arm. (ii) A positive control in which task-dominant gating is forced by construction shows the dominant-segment instrument reads 1.000 where the base network reads 0.055, so the paper's negative readings are informative rather than artefacts. (iii) A construction-diversity control shows the granularity null does not survive a non-self-limiting anchor: refreshed per-segment references gain about two points of learning term on MNIST while losing about twenty points of final accuracy, so the general claim that sub-neuron protection granularity is worthless is withdrawn. (iv) A comparison with contemporary replay methods (ER and DER, 40 runs) shows the audited method is not competitive with them: DER reaches 86.4 final accuracy on Permuted-MNIST against DeLP's 64.0, and 75.3 against 42.5 on Permuted-Fashion, with markedly less forgetting in both, in all twelve paired comparisons significant at p<1e-6. (v) Equivalence testing (TOST at a 0.5-point margin) and a metric-sensitivity analysis replace the manuscript's earlier reliance on failure to reject. All statistics are recomputed from the raw per-seed records with Holm correction within each metric family, 95% confidence intervals, and paired standardized effect sizes. The package contains the LaTeX source, the compiled PDF in both the Elsevier (submission) and TMLR (double-blind, anonymized) formats, the supplementary material, all figures, the bibliography, the machine-readable recomputed statistics, and the scripts that produced them. All code, per-run records, pre-registrations and dated amendments are archived in the project repository linked in the manuscript.