DeLP continual-learning data & code (per-seed records, pre-registrations, figures)

Version 3 replaces the earlier S3/S4 records with the corrected ones (the audit found a dropped anchor-penalty coefficient and a width-mismatched no-gate ablation arm; both experiments were re-run). Superseded files are named explicitly in bench/results/DEPRECATED_README.md inside the package. The package now holds 170 raw per-seed JSON records, 35 scripts, 9 preregistration and amendment documents, and 7 S3/S4 experiment scripts.

This record deposits the raw experimental records and the code that produced them for the DeLP (dendritic context-gated, usage-localized elastic weight consolidation) line of continual-learning studies. It exists so that the underlying measurements can be inspected and re-derived independently of whether any individual paper is accepted - the data are the primary artifact; the papers are one way of reporting them. Version 3 replaces the earlier S3/S4 records with the corrected ones: the audit found that the elastic anchor penalty had been added to the loss without being multiplied by its weight lambda, and that the no-gate ablation arm had been built with 2000 hidden units against a 256-unit dendritic arm (a 7.8x parameter mismatch). Both experiments were re-run; the superseded files are named explicitly in bench/results/DEPRECATED_README.md. After correction the S4 conclusion reverses - the earlier claim that the repair bypasses the gate does not survive, and the corrected four-arm comparison (base 63.86, constant-gate 63.62, repaired 68.67, no-gate 67.50) supports a saturation mechanism instead. The package contains 170 raw per-seed JSON records, 35 training and analysis scripts, 9 preregistration and amendment documents, 7 S3/S4 experiment scripts, and the figure sources for the S1 study. Source repository (live, canonical): https://github.com/zhouranxi/delp-anchor-budget