Repair Headroom, Not Dying Gates: When an Own-Segment Bias Helps Continual Learning — and When It Silently Reverses

A recent audit of dendritic context-gated protection (companion paper, S1) found that a one-hidden-layer active-dendrite network with winner-take-all routing fails to develop segment-task specialization, that the standard dominant-segment instrument is contaminated by dead neurons, and that a constant bias b on the own-segment dendritic input repairs the learning term (Δℓ up to +4.6 at T=20, t ≥ 8). We test whether these findings transfer beyond the bare-MLP family. They do not — and the failure mode is informative. On a convolutional active-dendrite network (4-layer conv trunk + GroupNorm + global average pooling feeding the identical dendritic gate; 160 runs, 10 seeds per cell), the repair collapses to noise (Δℓ significant in 4/12 dose cells; largest effect +1.46 vs. the MLP's +0.8 to +4.6) and final accuracy reverses (Δavg down to -2.29, t=-3.1). The dead-neuron artifact disappears entirely — not because the instrument improved, but because conv+GroupNorm mortality is exactly zero where the bare MLP's was 57%. A decisive same-architecture control (GroupNorm inserted into the MLP's feature layer, 80 runs, the only delta) separates the two candidate explanations: mortality and repair-gain decouple (GroupNorm raises mortality to 76.5%/85.5% while collapsing the gain to 17%/26% of its bare-MLP value), so dying gates are not the mediating variable; instead, gain decays as the unbiased baseline strengthens, and on both strong-feature backbones the repair buys learning only while significantly worsening retention. We conclude with a practitioner-facing boundary: own-segment bias repairs pay when the baseline is weak and its gates die; on strong-feature backbones they silently trade retention for learning, and the dominant-segment diagnostic must be paired with a mortality probe before it is trusted anywhere. Files: main.pdf (full paper), LaTeX source, highlights, cover letter, four figures, the preregistration and data-closure documents (prereg/, locked before data collection), the experiment scripts (code/, including the decisive same-architecture GroupNorm control), and 60 raw per-seed records (results/). Companion code and all records: https://github.com/zhouranxi/delp-anchor-budget . Related records: precursor https://doi.org/10.5281/zenodo.23602281 ; S1 (the repair audited here) https://doi.org/10.5281/zenodo.23637973 . This record is a timestamped preprint deposited to establish a citable public record independently of journal acceptance.