Census structure of floor-attaining local models for the singlet: supporting computations for arXiv:2608.18886

Supporting computations for arXiv:2608.18886 (A. Alai, "…an unmeasured fringe…", versions 4–5), Section 7. The deposit contains the scripts, offset banks, solutions and run transcripts behind three results about the measurement-dependence-minimizing census ρ of the moiré-phase-locking mechanism at Hall's floor M = (√2−1)/3: (i) Four independent 1024-offset solutions (two samplers, two seeds, no offsets in common) share the same low-order Fourier skeleton in the station offsets, dominated by cos(4(ξ_A+ξ_B)−φ) with φ fixed to within 5°, together with a weaker relative-phase term cos(4(ξ_A−ξ_B)−ψ) at about a third of its amplitude. (ii) Under perturbation of the terrain, the skeleton follows the perturbed term while the price stays at the floor (M − M_min ≤ 5×10⁻⁸). (iii) No station-local census f(ξ_A)g(ξ_B), and no census separable in the sum and difference of the offsets f(ξ_A+ξ_B)g(ξ_A−ξ_B), reproduces the singlet correlators at tolerance ε = 0.01 from any of 14 starts on the paper's terrain; the station-local class also fails on a terrain bred to favour it, and where feasible (ε = 0.05) these classes pay 2–3× the floor. Files: census_structure_v1/v2.py, terrain_factor_v1.py, flight_time_perturb_v1.py, dither_sweep_perturb_v1.py, product_census_v3/v4.py (v1–v2 withdrawn: normalization error), breed_product_v1.py, pipeline_v1.py, RUN_PIPELINE.bat, check_folders_v1.py; the terrain greedy_results.json; the four floor-attaining solutions and two perturbation solutions (rho_*.json); the matched offset banks (banks_*.npz); the station-local test outputs, the breed and validation folders, pipeline_FINAL.txt, and all run transcripts (in census_structure_deposit_v1.zip); pixel_null_v1.py with its output; the sum/difference test outputs (product_census_v4); and a short note (census_structure_note_zenodo_v2.pdf) with the method, the result tables and the caveats. Every quoted number is produced by the named script from the named input, and each script asserts the stored floor price before reporting. Version 2 adds the sum/difference-coordinate census test (Result F), a per-pixel permutation null for the (2,−2) term, and corrections to the note (phase convention, start counts, active-atom counts, slack vs. error). Version 3 adds the bandwidth analysis of the floor-attaining census (Result I–L): the minimum correlator miss attainable by censuses band-limited to K ≤ 5 fence harmonics (certificates by value on four independent censuses: no census with K ≤ 5 reproduces the singlet to 1%), the price of band-limited censuses from K = 6 to 14, the complete (ε, K) surface for K = 6, 8, 10 and ε = 0.01–0.10, the rank-k mixture test in sum/difference coordinates (Result G), and the roles-swapped solve showing the census and its station asymmetry are independent of the solve frame (Result M); scripts, launchers, transcripts and coefficient files included. Version 4 corrects the smoothness prices of version 3, which were computed on 1024 random offsets and include a quadrature cost: on a 32×32 lattice of offsets (exact quadrature below the 16th harmonic) the floor is attained and the band-limited prices are 1.70, 1.43, 1.26, 1.11, 1.00 at K = 6, 8, 10, 12, 14; on 4096 random offsets 1.83, 1.43, 1.30, 1.22, 1.14, 1.07, 1.01 at K = 6–18. The K ≤ 5 certificates hold on all samplings (best K = 5 census misses by 1.03–1.06%). Also added: the cell-separability analysis (the census varies with the dial), the dial-blind census class and its price (1.47× / 1.20× for one / two shapes on the lattice), the smooth-and-dial-blind class, and the parametric birth-model runs. Files: census_bandwidth_deposit_v4.zip (scripts, transcripts, banks, censuses, JSON outputs) and census_bandwidth_note_zenodo_v4.pdf (the one-page note with the correction table). Version 5 adds the (ε, K) surface on the lattice; the parametric birth-model runs (a few-dozen-parameter law reaches a 2.1–2.6% miss and no further); tests of the census's fine structure — a periodic ripple (pleat) at any frequency 7–20 changes the smoothness price by < 2%, a gap in harmonics 7–9 costs 13%, sharp features do not help, a dictionary of ~6° localized features is taken up fully by the LP; the grain test on two cut terrains (the lowest harmonic the census needs tracks the terrain's heaviest station index; the highest, ≈ 13–14, does not); and a negative replication: an apparent partition of relative-phase values across (τ, scale) cells, found on the lattice with a vertex-perturbation control, does not appear on random offsets and is recorded as lattice-specific. Files: census_fine_structure_deposit_v5.zip (scripts, transcripts, cut-terrain banks and censuses in their own folders, JSON outputs) and census_fine_structure_note_zenodo_v5.pdf (the one-page note).