M03 latest edition: actual direct whole-grid comparison
=====================================================

The as-executed README and M03 plan are preserved prospective snapshots. Their
HOLD language describes the review gate before this run, not the current outcome.
M03-STATUS-ADDENDUM.json clarified existing stop/reporting states before execution.
M03-IO-ADDENDUM.json freezes the separately reviewed metadata-only replacement
repair. Attempt1 stopped with233 valid records/89 unrun after a Windows sharing
error; its full233 reports and four fields are in history/attempt1. Its original
n12 interrupted resource receipt was never returned and remains unavailable.
PUBLIC-TRANSFORMATIONS.json documents the sole private-path redaction and original
hashes. Original source/results remain unchanged locally. Revision2 starts the
same matrix from the beginning; the repeated233 conditions are not novel cases.

Current result: complete; 322 of322 direct
point solves recorded. These are two assembled-mesh sweeps, not322 experimental
trials. No new eigensolutions, time-domain integrations or hardware trials occurred.
Elapsed through final manifest: 288.265000seconds.
The900-second limit is a monitored stopping budget, not a hard serialization promise.
Exact raw status, memory observations and finalization receipt remain included.
Together, the first233 retained records plus this replay's322
records document555 retained direct evaluations
across attempts, not555 different frequency/mesh conditions.

All322 points numerically valid: True.
All six exact shared parent anchors pass: True.
Official whole-grid evaluation eligible: True.
Numerical consistency permits reporting the mesh/approximation screens; it does not
make a failed accuracy screen pass. Read results/M03_summary.json for every outcome.
The maximum normalized direct n10/n12 compliance difference is
165.280945% (5% criterion: False).
Same-mesh modal144/direct differences are
10.014369% for n10 and
9.841788% for n12; both exceed5%.
At20kHz only, direct compliance differs by
0.146949% and time/volume RMS stress by
0.210764%; these pass their separate5% gates.
That single frequency does not establish a qualified contiguous low band.

The six preselected full displacement fields are included at indices0,80,160 on
each mesh. Other full fields were computed but not retained. All322 scalar reports,
stress/energy diagnostics and both complete161-point curves are public in this core.
No original results or required fields have been omitted to meet the delivery cap.
The PNG is separately delivered at the path/hash listed in PUBLIC-FILES.json; its
SVG, plotted CSV, post-run plotting source and figure manifest are in this archive.
Figure/package code was added after the run and is not a runtime source dependency.

On64-bit Windows with Python3.11+ and requirements installed, from this folder:
  python run_m03.py --output fresh-preflight --preflight-only
  python run_m03.py --output fresh-results --execute-frozen-plan 366ed2f10874ae6b56ff9787792090a812985a796463d9f63f621a0d7a0245a4
Every output directory must be new. Reproduction reruns the full fixed matrix with
the same guards; timings and floating-point output bytes can differ across systems.
The three-dimensional FEM uses NumPy/SciPy; the guarded runner requires Windows.
To reproduce the figures from retained results, also install Matplotlib, then run:
  python figures-v3/make_figures.py --results results --output fresh-figures
The figure manifest records the actual plotting library versions and input hashes.

Directn10/n12 differences remain discretization differences, not continuum error.
The161 samples do not bound peaks between frequencies. Small-strain steel response
under an assumed1N load and unmeasured damping does not predict fracture, hardness,
transmission efficiency, a selected actuator, or a buildable acoustic laser.
