M04 independent review — scope and reproduction boundaries

Independent means a separate computational implementation within this project,
not an outside laboratory or an external replication. Physical trials: zero.

The prospective review plan fixes one independent direct solution at n20,
original index131,131758.91068838467Hz. Selection occurred before M04 results.
The main study uses18 direct solutions at six selected frequencies; these are
not a new whole-band sweep. Direct mesh differences are not known continuum
errors. Contraction over three meshes does not establish asymptotic order.

Review has four distinct layers:
1. Frozen plan/source/preflight: preserve the exact inherited engine and explain
   its input-range wrapper, compare the structured-mesh AST except the guard,
   compare every returned geometry array for inherited n1–12, and check n20
   geometry analytically without solving its FEM system.
2. Fast evidence audit: verify source/input/result hashes,18 scalar/energy/power
   records, six parent anchors, fixed historical normalizers and unchanged5%
   screens. Reconstruct stress/energy/residual from the one public full field
   with Cartesian tensor equations separate from the author's Voigt assembly.
3. One separately assembled n20/index131 solve, after the main run ends and
   after a fresh exact public-core extraction. Both implementations use SciPy
   sparse solving; algorithmic/library diversity is limited. The legacy tensor
   reference does not import the author's FEM engine. A Windows Job guard
   caps worker aggregate commit at4GiB, stops the worker after180seconds, and
   uses a210second total stopping budget. Combined resident/commit monitoring
   is sampled. No fallback point, silent retry or fabricated partial field.
4. Figure-data audit: reconstruct all six plotted values per mesh and all4800
   surface-triangle colors from the actual public displacement field. Geometry
   is undeformed. Component-complex norm is not necessarily the instantaneous
   vector peak and is not damage, strength or fracture.

The independent reference vector is retained privately, prospectively, to avoid
publishing a duplicate large field. Its exact hash, numerical comparison,
source and guarded regeneration command are public. Fast downloaded-package
replay cannot recompare that unavailable vector. It verifies the author field
and scalar records; the one bounded reference must be rerun for a new full-field
comparison. Do not add the reference solve to the main study's18 count.

All verification passes remain distinct from scientific accuracy screens.
An internally consistent calculation may fail the unchanged5% test. Nothing
here establishes measured damping, actuator coupling, a printable working
device, material fracture or patentability.

Historical reviewer output-monitoring failure
--------------------------------------------
The one reference solve completed in52.704seconds with confirmed Windows memory
and time guard. Its original runner passed a relative output path to a worker
using another working directory. The field/report landed in a nested private
folder, outside the folder watched by the8MiB output-size monitor. The complete
observed output was388,899bytes, checked retrospectively, but continuous output
monitor coverage was not achieved. No all-resource-guards-passed claim is made.

The original runner/receipt are preserved in history/reference-relative-path-v1.
The exact existing field/report were copied without transformations for review;
the first-location bytes remain retained. The comparison did not require a
second solve. Current run_reference.py resolves the path before dispatch and
has two file-only regression tests. This corrected reproduction helper is not
the source used for the historical solve. REFERENCE-PATH-INCIDENT.json binds
the original helper, guard receipt, recovered field hash and monitoring gap.
