M03: direct-response comparison on every original sampled frequency
=================================================================

Main execution is held for root and independent plan/code/preflight clearance.
This directory leaves all previous model and evidence files unchanged.

Two assembled-mesh workers (n10then12) produce exactly322direct point records:
the original161geometric frequencies20–200kHz on each unchanged20mm steel cube.
The bottom is fixed and the top carries1N peak uniform downward traction. Mass,
TET4 geometry and assumed Rayleigh damping are inherited exactly. Matrices and
free submatrices are reused within a mesh; no different-frequency factorization
or modal reduction is reused for direct solutions. There is no new eigenproblem.

Primary metrics compare complex scalar compliance, with the original5% threshold
and1%-of-peak denominator floor. Directn10/n12 comparison is a mesh difference,
not continuum error. Same-mesh modal144/direct compliance differences are not the
M02displacement mass-norm error. Whole-grid stress curves are diagnostic; only the
original20kHz RMS-stress gate is inherited.161points do not bound continuous peaks.

The decomposition M=D+E10-E12 preserves complex signs. All decomposition terms
use the directn12scale. The historical M02 modal-mesh percentage uses modaln12as
its denominator and is reported separately. Triangle bounds are algebraic checks,
not energy shares, causal percentages or independent physical validation.

Each point stores scalar stress/energy/residual diagnostics. Only indices0/80/160
on both meshes retain full real/imag displacement checkpoints; all6are required in
the public study package. Remaining full fields are computed but not retained.
Actual replay uses exact shared indices0/83/160 against frozen M02 records, with
no interpolation. Failed/missing anchors block every official whole-grid screen.

Requirements:64-bit Windows8+ for the tested guarded runner, Python3.11+, NumPy
and SciPy. The FEM itself is portable; this runner has no unguarded OS fallback.
  python -m pip install -r requirements.txt
  python run_m03.py --output fresh-preflight --preflight-only
After review/clearance, main command (use a NEW directory):
  python run_m03.py --output fresh-results --execute-frozen-plan 366ed2f10874ae6b56ff9787792090a812985a796463d9f63f621a0d7a0245a4

Budget:900seconds before preflight through final manifest,450seconds per mesh,
one worker,2BLAS threads,4GiB hard worker-job aggregate commit. Combined coordinator
and worker memory/resident values are sampled at0.1seconds, not a continuous hard
resident cap. Source/input mutation stops workers and invalidates acceptance.
The copied M02guard preserves the unexplained native peak-counter discrepancy
observed during allocation-denial preflight; current commit and allocation refusal
are separately checked. Do not equate that OS peak with admitted memory.

Evidence cap16MiB, with1MiB reserved for final records; atomic writes count temporary
coexistence. Every expected ID is retained, including failures/unrun points. A partial
grid cannot pass a whole-grid screen. Finalization receipt identifies timing through
serialization and any overrun; its own tiny final write is explicitly excluded.

No specimen hardness/fracture, acoustic transmission, optimized destructive peak,
actuator capability, physical trial or time-domain integration is modeled here.
This is a numerical discretization study under a prescribed load, not a build test.
