M04 review findings

Eighteen author direct solutions completed valid on the declared discrete
models. All six n12 parent anchors reproduce;24source bindings remain unchanged.
The targeted mesh-pair5%accuracy screens still FAIL:185.56387372013752% for
n12->n16 and42.60512253254629% for n16->n20, worst at original index130.
Absolute compliance differences shrink at all six frequencies. This does not
establish asymptotic order, continuum error or whole-band accuracy.

At20kHz, pair compliance differences are0.173458% and0.097430%, with RMS stress
differences0.249974% and0.141263%; these local screens pass. The floor is the
fixed prior M03 whole-grid value,1.6121892443000918e-10m/N. At floor-active points,
unfloored differences remain published as diagnostics. No threshold changed.

The final review has121 passing numerical/provenance/factual receipt checks:
all18scalar records, six anchors and one public full field reconstructed. One
separately assembled n20/index131 reference agrees at1.1236502914e-13 in mass
norm,1.1603171306e-13 in compliance and1.0136088385e-13 in RMS stress. This is
within-project computational consistency, not an external laboratory result.

Reviewer monitoring failure remains explicitly open as historical evidence:
the52.704second reference used a relative output path and wrote to a nested
private folder outside its output-size monitor. Windows memory/time receipts
confirm those guards; complete observed output388899bytes is below8MiB only
as a retrospective check. Continuous output-monitor coverage failed. There is
no all-resource-guards-passed claim and no second reference solve. Original
helper/receipt and exact reference field are preserved. Current reproduction
helper corrects path resolution and passes two file-only regressions; it is
not the source used for that historical solve.

Exact public-core replay passes21author guarded tests,5reviewer analytical
tests and2file-only regressions,118fast consistency checks and17figure-data
checks. Fast replay does not recompare the private independent vector or rerun
the18-point matrix. That vector is deliberately not duplicated publicly; its
hash, source and bounded regeneration command are supplied. External PNGs are
explicitly unverified bytes in a core-only extraction; the preserved canonical
figure audit separately verified their actual bytes and all4800triangle colors.

All three meshes are a20mm ideal isotropic steel cube with a prescribed1Npeak
surface force and assumed damping. These records do not validate an actuator,
coupling efficiency, fracture, strength, specimen calibration or scale transfer.
Physical trials remain zero.
