M04: targeted direct spatial-refinement study
===========================================

The main experiment is held for source/preflight review and root clearance.
All earlier source, failed attempts and published results remain unchanged.

This study uses eighteen full direct solutions: n12/n16/n20 uniform TET4
meshes at six selected original M03 frequencies. The same 20 mm ideal steel
cube has a fixed bottom, prescribed 1 N peak downward top traction, and
unmeasured illustrative Rayleigh damping. No modal reduction is used.

The six frequencies include M03's worst direct-mesh sample at 131.7589 kHz,
its immediate original-grid neighbors, and indices 0/80/160 controls. This is
an explicitly targeted follow-up, not a new whole-band accuracy claim or
true resonance-peak search. Failure to bracket a shifted resonance is not
resolved by adding favorable frequencies after seeing the results.

Both mesh-pair screens retain a 5% criterion, with denominator equal to the
larger of the local finer-mesh response magnitude and the FROZEN prior M03
full-grid n12 peak times 1%. The floor is 1.6121892443000918e-10 m/N; it is
not refitted to these six points. Unfloored values and floor-active flags
are diagnostics. Absolute-difference contraction is not a convergence
order, continuum-error estimate, or additional pass criterion.

Every point retains its complex compliance, constant-element time/volume
RMS stress, energy, power and residual checks. Only the full n20/index131
complex displacement field is retained; this one field is required in the
public package. Other full fields require rerunning the executable study.
All six n12 points must reproduce exact M03 reference values before the
official selected-set screens are eligible. Incomplete evidence cannot pass.

The inherited mesh routine explicitly limited subdivisions to twelve.
mesh_extension.py copies the same algorithm with only the upper-bound
validation extended to twenty. Original inputs/fem3d.py remains byte-exact;
M04-ENGINE-EXTENSION.json explains the source boundary and geometry checks.
Assembly, material law, damping, direct operator and energy recovery are
inherited. Geometry generation alone is not a solved frequency response.

Reproduction requires 64-bit Windows, Python 3.11+, NumPy and SciPy:
  python -m pip install -r requirements.txt
  python run_m04.py --output fresh-preflight --preflight-only
After review, obtain the SHA256 of M04-PLAN.json and use a NEW directory:
  python run_m04.py --output fresh-results --execute-frozen-plan <PLAN_SHA256>

The total monitored stopping budget is 900 seconds, including preflight;
individual mesh budgets are 90/240/540 seconds. One heavy worker runs at a
time with two BLAS threads. Windows Job Objects enforce 4 GiB aggregate
worker commit; combined/working-set checks are sampled, not hard resident
caps. The configured 0.1-second monitor sleep has callback/scheduling
overhead. Existing bounded atomic-replace recovery is preserved, including
its 0.5-second retry window; persistent faults fail closed. No scientific
solve is silently retried. Finalization reports actual elapsed and any
overrun rather than promising an unconditional hard end-to-end deadline.

The numerical output cap is 8 MiB with a 1 MiB metadata reserve. Keep all
failed, invalid, stopped and unrun rows. A missing guard receipt is explicitly
unavailable, not reconstructed from assumptions. Publication may redact only
private local paths with an original/public hash transformation record.

This is a small-strain specimen discretization test. It does not establish
material hardness/fracture, measured acoustic coupling, delivered actuator
force, electrical efficiency, buildability or physical specimen performance.
