Update to #3538: #3466 is not evidence for pumping; clamp vs freewheel tail τ test; the blocking diode's own Cj couples charge back on every reverse edge and helps set the ratchet ceiling

Discord #neuralstan20-testing 2026-11-08, Scotchn. (1) CORRECTION: in chat I cited #3466 (Scotchn's open-loop ms tail) as support for stray-capacitance pumping. That was wrong. #3469 traced it to a likely misidentified bifilar tail pair, i.e. real freewheeling magnetizing current. The better support is #3477: one turn-off gave about 6 forward packets via the A–B interwinding capacitance. (2) TAIL τ DISCRIMINATOR: pumped clamp gives τ = R_bleed·C_node (probe-dominated; 10 MΩ×50 pF ≈ 0.5 ms, 100 MΩ ≈ 5 ms). Freewheeling magnetizing current gives τ = L/R of the loop (e.g. 30 mH / 10 Ω ≈ 3 ms), largely probe-independent. Tests: switching the probe between ×10 and ×100 stretches the clamp tail about 10× but leaves the freewheel tail alone. Adding loop R shortens the freewheel tail. A few MΩ across the node drops both the clamp plateau and its tail. (3) DIODE Cj (raised by Scotchn): the blocking diode is not one-way for edges. Its junction capacitance couples charge back on every reverse-going edge, Q_back ≈ Cj·ΔV. That is not a DC leak, but it is a per-cycle charge return, so the ratchet levels off where forward charge per pulse equals Cj charge-back plus resistive bleed. Scotchn's IDWD10G200C5 (#3483): Cj 1140 pF at 1 V, 45 pF at 600 V, 31 pF at 1500 V. Against a ~3.7–3.97 nF node, the back-coupling fraction Cj/(Cj + C_node) is about 24% at low bias but only about 1% at 600 V. So the charge-back matters most at the start of a burst, when Vbase is low, and fades as the node climbs. The climb is therefore nonlinear: slow early, then faster per pulse, unlike the constant-k model in #3538. My calculation from datasheet values, not measured. Cj also adds to C_node in the clamp τ.