E01 independent numerical review

The independently reconstructed frozen criteria agree with the saved E01
results. This is a numerical consistency review, not physical validation.
No trajectories were integrated by this reviewer and no source dynamics or
saved states were changed.

Scope and results
- All 972 declared speaker cases were checked against a direct complex 3x3
  circuit solution and the actual JavaScript browser implementation.
- All 24 acoustic output traces were reconstructed from NPZ. Fourier analysis
  of the last five periods independently checks the production least-squares
  fit. Maximum amplitude disagreement was 1.67e-9 relative; maximum phase
  disagreement was 2.80e-7 degree. Maximum reconstructed acoustic energy
  residual was 3.63e-10 of final positive input energy.
- All 16 impact traces were reconstructed using independently calculated
  60-digit Decimal cantilever roots and direct hyperbolic mode shapes. Maximum
  retained energy residual was 4.71e-6 of initial gravitational energy.
- Coarse/fine comparison fields and thresholds were recomputed. Their reported
  pass statuses agree with the saved paired results.
- The deliberately soft missing-target capture, I04, crosses the hard travel
  boundary at both resolutions. Its first crossing states remain in the raw
  files. No hard-stop force, velocity reset, successful capture or unobserved
  final rest interval is inferred after that termination.
- Only I02 and I03 satisfy the declared final-50ms numerical near-rest test
  among the eight fine-resolution impact cases. This is not proof of a real
  mechanism settling or being ready for another release.

Additional strict reviewer finding: small nonmonotonic ledger samples
The initial saved-evidence-v1 review reports FAIL because an extra reviewer
criterion required each retained cumulative-energy increment to be at least
-1e-10 J and the ledgers to be at least -1e-11 J. Fourteen of the 24 acoustic
traces violate that increment criterion. The largest decrease is 1.34244e-8 J,
or 4.42013e-7 of that trace's final positive input energy. Smaller decreases
remain in some fine-resolution traces.

The largest observed decrease occurs in the ND91-4 one-litre, 1000 Hz, coarse
case's positive supply ledger at 0.07425 s. The decrease is about 0.0000442%
of the final positive input energy. It does not exceed the original frozen
energy-balance criterion, but exact monotonicity of the saved curve cannot
be claimed.

The output samples come from an adaptive solver's dense interpolant. Small
interpolation excursions near sign changes of max(power,0) and max(-power,0)
are a plausible explanation; this is an inference, not a separately proven
diagnosis. We did not smooth, clip, monotonize or regenerate these traces.
The initial strict FAIL report is preserved. The detailed v2 review retains
all 14 failures and separately reports that the original frozen checks were
reproduced. The reviewer program exits nonzero while these strict failures
remain. No acceptance threshold was retroactively relaxed.

A future numerical revision could preserve actual solver-step ledgers and
compare them with separately integrated cumulative work on an independently
refined sampling grid. That would be a new version with fresh evidence.

Sampling limits
Acoustic output points are not every adaptive internal step. Impact NPZ data
are retained every 50 microseconds plus a failed boundary endpoint. Impact
report extrema were calculated at the finer 5 or 10 microsecond integrator
steps. The independent audit confirms retained-state identities, retained
extrema bounds and endpoint ledgers; it cannot independently recover every
unretained maximum. The website's 1 millisecond plotting trace is coarser
again and must not be used to estimate narrow peak contact forces.

Engineering limits
The driver parameters are manufacturer nominal small-signal values. Constant
inductance, an ideal sealed volume and an aggregate fitted mechanical loss
do not establish sound pressure, radiated acoustic power, wall-plug efficiency,
fracture capability or component thermal protection. Model voltage means
actual voltage at the driver terminals, not a PC volume percentage.

The impact target, buffer and guide coefficients are explicit unmeasured
assumptions. Smooth tanh guide friction omits static friction and jamming.
Real breakaway force could exceed the approximately 0.118 N weight of the
nominal 12 g moving mass. The mass must include all moving hardware and be
weighed. Model compliance is not inferred from a printable CAD shape.

Files and reproducibility
saved-evidence-v1/E01-INDEPENDENT-REVIEW.json preserves the initial strict FAIL.
saved-evidence-v2/E01-INDEPENDENT-REVIEW.json includes the diagnosis, all strict
failures, declared refinements and source/result hashes. verify_saved_e01.py
reads the frozen result package and invokes node_probe.mjs against the actual
browser module; check_portal_response.py supplies the independent circuit
solve. No production numerical engine is imported by the reviewer.

Earlier pre-freeze findings corrected by the implementation owners were an
invalid/zero Xmax browser input acceptance problem, undefined zero-signal
phase reporting, omitted CAD air-domain portions, a rotated acoustic-foot
rib and misleading final-window metrics on terminated impact cases. These
corrections predate the relevant final review or trajectory freeze. Mechanical
assembly fit and physical characterization remain separate gates.
