R07 independent numerical review

This review checks the recorded computation. It does not validate a contact law
against physical measurements and does not establish fracture performance.
Independent here means a separate computational implementation within this
project; it does not mean an outside laboratory or external replication.

Before the matrix

independent_preflight.py checks the exact frozen plan hash, all earlier input
and source hashes, the 144-row Cartesian product and ordering, the exact 18
selected trace IDs, and the full-resolution float64 storage arithmetic. It does
not import the new engine and does not execute a trajectory. Sixty-digit
Decimal Newton roots and Taylor trigonometric series independently evaluate
the cantilever basis for the passive-speed and timestep screen.

For contact exponent p and positive indentation d, the elastic force is K*d^p
and stored energy is K*d^(p+1)/(p+1). At the frozen reference indentation, the
force-matched p=1.5 model stores 80 percent of the linear-model contact energy;
the storage-matched model has 125 percent of its elastic force. Both nonlinear
models match the dissipative component at the reference point for every
approach velocity. These are different matching conventions, not simultaneous
equivalence or a material calibration.

For contact relative velocity r, the passive energy bound is
abs(r) <= sqrt(2*E0*(1/m_source + sum(phi_at_contact^2))). The stiffness and
velocity coefficients are bounded at d=0.001 m and approach speed=3 m/s over
the frozen rectangle. All planned cases pass this conservative finite-mode
screen. That screen is not a proof of nonlinear RK4 accuracy.

Portable checks

custom-poc/efficiency/iteration07_checks.py supplies the runner's preflight and
full saved-state trace checker. It imports the independently verified R06
modal helpers but does not import the R07 integrator. The checker reconstructs
every saved force, indentation, elastic storage, modal/source kinetic energy,
loss/work ledger, contact event and peak. It rejects missing samples, nonfinite
states, inconsistent initial gaps/budgets, wrong constitutive metadata and
summary disagreements. Contact events use the declared saved-step threshold
of force greater than 1e-9 N; no substep timing is inferred.

test_trace_checks.py verifies intact records for all three laws and deliberately
corrupted time, state, loss, law and reported-peak data. Seven tests passed before
matrix execution. These tests are separate from scientific case results.

After the matrix

audit_runs.py imports neither the R07 engine nor its portable checker. It reads
the saved results, hashes and traces without modifying them. Every completed
row is checked against its frozen parent command, specimen, timestep, protocol
energy and initial gap. All 48 linear-reference rows must reproduce the R06
target energy, final energy ledgers and legacy feasibility flags. The new
whole-span displacement bound and relative-speed bound are reconstructed.

All 18 selected full-resolution traces are loaded without pickle support. The
review independently reconstructs their physics using high-precision modal
references and also rebuilds each four-stage RK4 transition from its recorded
starting state. This is a consistency check of saved transitions, not a new
global trajectory. With all planned traces available, it checks 1,224,000
transitions. A 5e-12 componentwise tolerance with a unit-sized absolute floor
is declared for this additional recurrence diagnostic; the frozen stricter
energy/peak and monotonicity criteria still apply separately.

test_independent_audit.py verifies all three intact laws and rejects deliberate
interior modal-state and cumulative-work corruption. Three tests passed before
matrix execution. The short nonlinear smoke test's maximum normalized RK4
transition discrepancy was 1.1102230246251565e-16.

Coarse/fine and five/seven-mode comparisons are independently recomputed,
including the inherited ratio sensitivity rule and unresolved classifications.
The 24 fine-step nonlinear-versus-linear energy changes are independently
recomputed. Local acceptance, final mode-screen acceptance, row counters and
the complete artifact-manifest inventory, sizes and hashes are checked. Each
trace sidecar must agree with its report and frozen source/input references.
All run IDs, unfavorable results and failed statuses are retained. Data-integrity
acceptance and favorable numerical performance are different findings.

Figures

plot_verified.py accepts only a summary with a matching independent audit hash,
intact rows and verified retained traces. It creates a six-case contact-law
comparison and one force/energy-ledger figure for each of the three frozen
trace-focus cases. Raw traces remain untouched. Every saved sample is passed
to Matplotlib with path simplification disabled; force axes use a labeled
contact window and ledger axes retain the full 170 ms. Force maxima are marked
at their exact saved-step locations. Figure data, source/evidence hashes and
publication-ready image hashes are retained in a separate figure manifest.
Unresolved, invalid and incomplete screens remain visible in the comparison.

Reproduction

python independent_preflight.py --efficiency-root PATH_TO_EFFICIENCY
python test_trace_checks.py --efficiency-root PATH_TO_EFFICIENCY -v
python test_independent_audit.py --efficiency-root PATH_TO_EFFICIENCY -v
python audit_runs.py --efficiency-root PATH_TO_EFFICIENCY

The audit scripts write only their review outputs. They never change frozen
plans, model sources, original evidence, hardware or publication state.
