M03: computation checks pass; whole-grid accuracy screens fail
=============================================================

The isolated second attempt completed all322 prescribed direct-response points
on the unchanged161-frequency grid, with all six parent anchors passing and
all26 source/input bindings unchanged. Its elapsed time through the manifest was
288.265seconds. No main-run replacement retries were reported. Positive/negative
Windows reader-lock behavior was exercised separately in the21-test preflight.

The independent audit passed999 checks. Its coverage is deliberately limited:
322 scalar/report records, six independently reconstructed saved fields, and
only two separately implemented direct solutions. The two fixed midpoint solves
ran sequentially after the main sweep, taking13.704seconds total. Their full
displacement mass-norm differences were1.02992259421232e-13 (n10) and
1.397285680860875e-13 (n12). This verifies implementation consistency at those
points; it does not turn the entire grid into independently solved evidence.

The unchanged5% accuracy screens still fail:
  Direct n10/n12 complex-compliance mesh difference:165.2809447783683%.
  n10 modal144 versus same-mesh direct compliance:10.014368718164385%.
  n12 modal144 versus same-mesh direct compliance:9.841787532297232%.

At the single20kHz point, the mesh compliance difference is0.14694870125215107%
and the volume/time RMS-stress difference is0.2107643468830688%; both pass their
5% thresholds. That isolated result does not qualify a contiguous low-frequency
band or establish accuracy at the other160points.

Keep two denominators distinct. The historical M02 modal-mesh difference is
155.51541668064848% on its original Hm12-based scale. The SAME modal difference
is165.40570502037613% on M03's common Hd12-based scale. These are differently
normalized diagnostics, not two interchangeable error estimates. The signed
complex decomposition is an algebraic identity, not a positive error-budget
partition, energy distribution, or physical attribution.

The large direct mesh difference persists after removing modal truncation from
that comparison. This supports further spatial-discretization work. It does not
identify the exact continuum solution or justify selecting a destructive drive
frequency. All161samples still leave between-sample extrema unresolved.

The first attempt remains preserved:233 valid point records,89 unrun conditions,
four saved fields and no official whole-grid result. A controller progress-file
replacement raised Windows error5 at156.75seconds through finalization. A
separate deliberate sharing-lock test reproduces that class of error; the
original lock holder has not been identified. The interrupted n12 guard receipt
was not returned and has not been reconstructed. The second attempt repeats the
same322conditions; the two attempts contain555 retained main point records,
not555 distinct conditions. The two independent midpoint solves are additional
review evidence, not an expanded main condition matrix.

All277 public history members are mapped to original hashes. Only the first
summary's absolute workspace prefixes were redacted. A local comparison proves
that exact byte transformation; a public-only replay can check the resulting
public bytes and declared hash mapping, but cannot rehash the removed private
path text. No numerical value was removed from the history.

Independent means a separate computational implementation within this project,
not an outside laboratory. Both implementations use SciPy's sparse direct solver.
There are zero physical trials, zero damage/fracture predictions, and no actuator,
acoustic transmission, universal-material, or manufacturing validation here.
