M02: bounded numerical-resolution refinement of one steel cube
=============================================================

This independent directory preserves M01 unchanged. Exact copies of its engine,
test source and necessary input/reference files are listed in inputs/INPUT-MANIFEST.
The frozen M02 plan retains the20mm cube,1N peak distributed top traction, fixed
bottom, isotropic steel and illustrative0.01 Rayleigh damping at20/200kHz.

Three meshes n8/10/12 each receive one144-mode eigensolution. Nested24/48/96/144
cutoffs share that basis and161original frequencies. Twelve fixed direct probes
per mesh give36unreduced complex reference solves. These are not physical trials.
Direct fields are retained in small NPZ files; the n12modal basis is retained.
Earlier transient bases are retired after their reports/curves/direct jobs finish.

Numerical identities, mesh errors, cutoff errors and direct-reference errors are
separate. A fixed144-mode mesh comparison still contains modal truncation. Direct
sample comparisons isolate spatial error only at those12frequencies.20kHz agreement
does not validate a whole low-frequency band. No fracture, hardness, acoustic
coupling, device capability or specimen calibration is inferred.

Requirements:64-bit Windows8+ for the resource-limited runner; Python3.11+, NumPy,
SciPy. The inherited FEM algebra remains portable Python. This edition refuses an
unguarded non-Windows full execution. Matplotlib is needed only for plots.

Commands from this directory, always using NEW output names:
  python run_m02.py --output fresh-preflight --preflight-only
  python run_m02.py --output fresh-results
  python make_figures.py --results fresh-results --output fresh-figures

The full run clock starts before preflight, including29inherited/new tests and
serialization. Budget600s overall,120s/job; one worker,2BLAS threads. Windows Job
Object imposes a hard4GiB aggregate committed-memory cap on each worker tree,
with kill-on-close and no breakaway. A tested GO handshake occurs after job
assignment, before scientific imports/work. The real base Python executable avoids
an unassigned virtualenv-redirector child. Coordinator memory is separately
reported; combined resident/commit excess stops work on sampled0.1s observations.
Sampled resident limits are NOT continuous hard resident caps. Unknown memory
queries fail closed; exited-PID races are distinguished.

Preflight-v1 preserved a failed test of Windows PeakJobMemoryUsed<=configured cap.
A256MiB allocation was refused under a64MiB cap, yet that OS peak counter exceeded
the cap. A separate VirtualAlloc probe returned null/error1455 and unchanged
current private commit. The reason for the counter disagreement remains unresolved.
Current checks verify limit readback, allocation rejection and sampled current
memory; they retain the OS counter without claiming it measures admitted memory.
This correction changes resource-report interpretation, not scientific criteria.

Output is limited to32MiB, reserving1MiB for status/evidence. Atomic writes check
temporary-file coexistence size. Exact planned IDs are recorded once; successful
details must exist before completion is counted. Failed or interrupted jobs and
unstarted rows remain explicit. Source/input hashes are checked before and after.
The deadline is a stopping bound, not a prediction of runtime or full completion.

Resource API references, not scientific or hardware validation:
https://learn.microsoft.com/en-us/windows/win32/procthread/job-objects
https://learn.microsoft.com/en-us/windows/win32/api/winnt/ns-winnt-jobobject_extended_limit_information

Do not overwrite an existing output, change thresholds after inspecting results,
count nested cutoffs as independent eigensolutions, or interpret reviewer algebra
agreement as a waiver of failed accuracy screens.
