M04 latest edition: targeted direct spatial refinement
======================================================

The as-executed README and plan remain prospective frozen snapshots. Their
HOLD language describes the source-review gate before this completed run.
This companion records the actual outcome without editing those snapshots.

Exactly 18 of18 fixed direct-response points were
recorded: {'completed_valid': 18}. Three meshes used the same six frequencies.
No eigensolutions, new time-domain integrations, or physical trials occurred.
Elapsed through final manifest: 188.359000 seconds.
All24 source/input bindings remained unchanged. All six n12 reference replays
passed.21 author preflight tests passed with zero skips. Numerical consistency
allows the mesh-agreement screens to be evaluated; it does not make them pass.

At six selected frequencies, the maximum normalized complex-response differences
are 185.563874% for n12->n16 and
42.605123% for n16->n20. Both FAIL the unchanged5% criterion;
both maxima occur at original index130 (129876.32631524227 Hz).
Absolute differences shrink at all six points, a descriptive improvement that
does not establish asymptotic convergence, convergence order or continuum error.
The finest-pair checks still fail at indices130,131,132,160. They pass at20kHz
and63.245553kHz, two isolated controls, not a qualified contiguous low band.

The scale is max(local finer-response magnitude, fixed historical floor
1.6121892443000918e-10 m/N). This floor is inherited from the prior full161-point
n12 curve, not recomputed from the six selected frequencies. It is active for
several sensitive samples. At the finest-pair worst point, the fixed-floor
difference is42.605123% while the unfloored pointwise difference is45.960120%.
Both numbers are mesh differences, not known errors relative to continuum truth.

At20kHz, compliance differences are 0.173458% and
0.097430%; time/volume RMS stress differences are
0.249974% and0.141263%.
These separate checks pass. Other stress values are retained diagnostics, not
new stress-accuracy gates or fracture predictions.

All18 scalar/energy/residual reports and the one prospectively required n20,
index131 full unrounded complex displacement field are in this core. Other
full fields were computed transiently and require a new run to regenerate.
No mandatory field was omitted for file size. Geometry is regenerated exactly
by mesh_extension.py; only the inherited validation maximum changed12->20.
Original inputs/fem3d.py is unchanged. The extension and review boundary are
documented in M04-ENGINE-EXTENSION.json.

The two PNGs are separately delivered with hashes in publication/PUBLIC-FILES.json.
Plot source, SVG, plotted scalar/surface CSVs and figure manifest are included.
Figures use saved numerical data; the field image has undeformed true geometry
and color proportional to a component-complex displacement norm, not damage or
an instantaneous vector-peak claim. Plot/package scripts are post-run additions,
not scientific runtime dependencies. The public index hashes every payload.

To reproduce on64-bit Windows with Python3.11+:
  python -m pip install -r requirements.txt
  python run_m04.py --output fresh-preflight --preflight-only
  python run_m04.py --output fresh-results --execute-frozen-plan 0a971364838f1d1b34cd27e0c79fc1c666604385d518c7f39083e16d260d0a85
Each output directory must be new. The runner regenerates the full18-point
matrix and selected field. Floating-point bytes and timings can vary by system.
For figures, also install Matplotlib, then from the extracted root run:
  python figures-v1/make_m04_figures.py --study . --output fresh-figures
No scientific solve is invoked by the plotter. To test the exact public archive:
  python publication/smoke_package.py --archive m04-local-refinement-core.zip --output fresh-smoke

The900-second limit is a monitored stopping budget including preflight; final
metadata reports measured elapsed, not an unconditional serialization deadline.
The Windows Job Object enforces worker commit; resident/combined memory is
sampled with callback and scheduling overhead. Raw resource counters remain
included with their documented Windows peak-counter interpretation limitation.

This targeted comparison covers six frequencies, not a full band or intervening
peaks. It is an ideal homogeneous small-strain20mm steel cube under assumed
damping and prescribed1N peak traction. No actual actuator/coupling, hardness,
fracture, rock, diamond, large structure or electrical efficiency is validated.
The next useful research question is a bounded analytical benchmark or separate
discretization check of the sensitive response, rather than automatic expansion
to a larger brute-force sweep. No new study is executed by this recommendation.
