ECI v7-note — Empirical Berry-Keating arithmetic correction on 10^5 Riemann zeros (v0.3)

v0.3 (2026-05-03): consolidation update incorporating (a) the v0.2 prefactor erratum -2/(2pi)^2 -> -2/L^2 (lit-triangulated Berry-Keating SIAM 41 (1999) eq. 4.20 + Bogomolny arXiv:nlin/0312061 + Conrey-Snaith arXiv:math/0509480 Theorem 4.1), and (b) reframe of the high-T fit anomaly previously reported as an A_fit ~ 1/3 plateau / 12.7sigma T-trend signal. Bootstrap-vs-jackknife sigma_A split is ~6.6x (jackknife biased low along the (A,L) trade-off valley); after bootstrap correction, A=1 is still rejected at 6-10sigma per band, but the genuine invariant is L_eff = 7.40 +/- 0.22 universal across L_mean in [9.1, 12.1], with A_fit ~ (L_eff/L_mean)^2 * 0.80 = parametric trade-off signature. No new arithmetic constant claimed. The L_eff = 7.40 = matched-filter window saddle (geometric anchor L_geom = 4*x_max*log(2) = 8.32, sympy-verified), specifically a fit-window saturation signature, NOT a physical scale. The nu = 1.094 cosine mode is consistent with the GUE-first-zero calibration artefact already disclosed in v0.1 plus prime-square cluster envelope of Berry-Keating arithmetic corrections. Headline M2 chi^2/dof = 4.17 unchanged across all reframings. v0.1 (DOI 10.5281/zenodo.19983241) remains as published; v0.3 supersedes it on the Berry-Keating prefactor and the high-T interpretation only.