Continuous full-space analytical reference, kept separate from the official benchmark

The official benchmark report remains FAIL. These files do not replace its original
semi-analytical reference or change any acceptance criteria. No solver output has
been read, scored, shifted or scaled while defining this reference.

The physical inputs are the original benchmark DATA: homogeneous rho=2300kg/m³,
Vp=2800m/s, Vs=1500m/s, point double-couple Mxz=Mzx=1e16N*m at the origin, and the
six original receivers. The source half-duration remains2s (Gaussian width2/1.628).
The original angular factors and Cartesian rotation conventions are retained.

Two reference approximations are removed: the near-field convolution uses an
exact antiderivative at the physical P/S travel-time endpoints, and the Gaussian
moment-rate uses the analytic derivative of the erf source instead of differences.

How to reproduce from the workspace root:

PYTHONPATH=work/python-libs /Users/giuseppegambino/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/bin/python3 work/analytical-reference/generate_reference.py

The generator only reads original physical DATA and generator source metadata. It
writes inside work/analytical-reference. No solver, network or site operation runs.

accuracy-criteria.json contains requirements declared before numerical evaluation.
accuracy-verification.json records90 prescribed station/time comparisons against
independent adaptive quadrature,13 tests integrating the Gaussian derivative to
the erf CDF, and6 static-limit checks against the Kelvin Green tensor. It verifies
all18 exported files preserve their calculated float64 values on ASCII round-trip.
manifest.json records hashes of all18 traces, physical inputs and the verification.

Each trace has14000 rows, dt=.001s, time−4..+9.999s. Six columns are documented in
its header: time, Cartesian displacement, spherical displacement and spherical
near/intermediate/far terms. All displacement columns are in metres. The spherical
term columns correspond to R/THETA/PHI respectively for files X/Y/Z; they are not
individual Cartesian decompositions.

Accuracy verification concerns a homogeneous infinite medium and this fixed
parameter set. It is not Etna validation, boundary-condition validation or a new
benchmark PASS. Any later comparison with the solver must be reported separately
after this reference verification and must preserve the original benchmark record.
