test: compare fixed centered reconstruction on audited square faces

docs/validation/taylor_green.md

@@ -548,3 +548,59 @@ bottom/top inner fine quartets its mean delta jx / d(uy)/dx is
 -/+ 1/(576*tau_f), with opposite adjacent-coarse redistribution.
 Keep raw values, step-zero rounding and fresh controls: sinusoidal curvature
 and finite-Mach effects prevent treating this coefficient as an exact gate.
+
+### Test-only centered straight-face comparison
+
+`centered_face_square_n128` reuses the square axis-wave POST-G initializer,
+geometry, native probes and root steps 0–8. It has four fresh cases: U128,
+U256, current projected R128, and experimental centered R128. This is a
+qualification-ineligible diagnostic; there is no legacy assessor or new physics gate.
+
+Only the candidate's private generated fine-level `blocks.cu` changes. A
+2,112-byte constant mask selects 1,792 rows (448 complete donor quartets).
+The diagonal projected correction is doubled on straight faces with exactly
+one unsupported normal axis and both legal tangent neighbors. Donor centers
+within four root cells of a corner retain the current rule, as do all other
+patterns, axial populations and rest. The mask is symmetric under transpose
+and reflection. Its hash and the actual modified source/PTX hashes are retained;
+production files, restrictions, collision, scheduling and launch arguments stay fixed.
+A complete quartet has zero summed offset, so each population's quartet mean
+is preserved. The perpendicular gradient also retains equal characteristic
+packets; the existing two-hop conservation routing therefore remains applicable.
+Host tests additionally check the mixed mask through complete two-hop/collision
+cycles, including corner grazing, and execute the actual edited kernel via C++.
+The central probes are more than 20 root cells from mask transitions; the
+reviewed conservative dependency radius is 2.5 root cells per root interval.
+
+The separate comparison subtracts each group's step-zero residual and the
+fresh uniform control at matching quadrature (U256 inner, U128 outer). It retains
+signed density and both momentum components, actual-center gradient normalization,
+parities, and sign flips. Step 1 requires bottom-inner cross-momentum magnitude
+improvement greater than its deadband, and top-inner nonworsening within it.
+Each horizontal face's step-8 RMS combines its inner and outer group means;
+both must be nonworse. All four vertical groups must remain unchanged within
+the component deadbands at every sampled root step. The norm
+`sqrt(drho²/3 + djx² + djy²)` must not worsen beyond its corresponding deadband.
+Sampled raw-pressure envelope and step-8 global gauged pressure RMS must not
+worsen beyond the pressure deadband. Nonfinite/nonpositive sampled density,
+invalid traces, source/build mismatch, or exceeded observed population envelope
+reject the comparison. A small top signal or sign reversal remains explicit.
+
+The predeclared shifted population envelope is 1/32. From per-population oracle
+allowance `6e-8 + 2e-7*1/32`, pairwise moment deadbands are
+`[1.1925e-6, 7.95e-7, 7.95e-7]`; pressure uses the density deadband divided by 3.
+Area averaging cannot enlarge this local allowance. These are practical diagnostic
+deadbands, not certified accumulated root-8 rounding bounds. The envelope is
+checked on downloaded synchronized ACTIVE populations, not hidden substeps.
+Known quadratic fine-cell O(h) terms, corner transitions, repeated coupling and
+stability remain unresolved. Passing permits review of a candidate, not promotion.
+
+Run the existing `run_case.py` with `--diagnostic centered_face_square_n128
+--flow shear_wave --re 100 --probe-trace --test-reconstruction current` for
+U128/U256 (depth 1, constant) and current R128 (depth 2, projected_linear).
+For candidate R128 use `--test-reconstruction centered_face_v1`. All require
+an exact clean `--repo`, `--expected-revision`, fresh `--output`, and either
+`--dry-run` or explicit single-device `--run-gpu`. Compare ordered receipts with
+`shear_centered.py --repo REPO --expected-revision SHA --output comparison.json
+U128.json U256.json current.json candidate.json`. Total work is 1,900,544 ACTIVE
+updates, six build sidecars, and 720 cell-time samples; scalar cap stays 1 MiB/case.