rfx v1.7.0 — Fidelity, Progress, Mesh, and RF Workflow Hardening

Highlights

rfx-fdtd 1.7.0 is a broad fidelity and production-workflow release spanning
181+ commits since v1.6.7.

• Added Simulation.fidelity_report() and expanded preflight diagnostics so
  realized geometry, mesh registration, absorber placement, and extraction
  limits are visible before claims-bearing runs.
• Added deterministic long-run progress reporting with report_every.
• Generalized microstrip-line extraction beyond the previous fixed-axis
  assumptions and hardened whole-port wire S-parameter extraction.
• Added patterned surface-impedance sheets and strengthened thin-conductor,
  two-plane PEC, and cavity-stack fidelity handling.
• Added float64 and mixed-precision controls across supported solver paths.
• Made NTFF chunking exact and bounded for large near-to-far transforms.
• Added multi-band graded-z mesh support, including measured grading limits
  and preflight findings.
• Added rasterized geometry views and expanded public RF workflow guidance.
• Fixed multiple RF validation and cross-solver fidelity issues, including
  patch-cavity stack realization, PMC mirror-plane convention, MSL plane
  V/I extraction, and wire multiport transmission scaling.

See the full changelog
for detailed behavior changes, measured evidence, and issue references.

Install

pip install rfx-fdtd==1.7.0

Compatibility

• Python 3.10+
• Pure-Python universal wheel
• JAX remains the numerical backend; accelerator support follows the selected
  JAX installation.

Release verification

Release commit: 35f2aab3714670bd1ee3cc882f51ced233bb5555

• GitHub required checks: all passed, including fast-suite shards 1-4,
  guards, Ruff, public docs, API reference, packaging, and Linux/macOS/Windows
  wheel smoke.
• Final-SHA targeted release tests: 27 passed.
• Local release validation: 172 passed with 1 optional skip; high-memory lane
  8 passed; local shards recorded 1,840 + 899 + 813 passes.
• External cross-validation:
  VESSL run 369367257496
  completed at the release SHA with cv05_rc=0 and cv11_rc=0.
• GPU suite:
  VESSL run 369367257495
  ran at the release SHA on RTX 4090: 173 passed, 22 skipped, 2 expected
  xfailed, and 1 failed. The sole failure is the pre-existing
  test_recover_debye_reference_mode_public_entry CPU-pinned optimizer
  tolerance also observed on the preceding main GPU run. The fit itself
  reduced loss from 1.675e-4 to 9.959e-6 and recovered tau within 11.4%,
  but delta-epsilon error was 30.9% against a 20% CPU-measured gate. A
  float64 adjudication at the same SHA
  (run 369367257556)
  reproduced the backend-sensitive delta-epsilon tolerance miss; no new
  v1.7.0 regression was identified.
• python -m build and twine check passed for the attached wheel and sdist.

Artifact checksums

ddf3398c280090520558d4cb6b4be7612a4a772a41cf9d364cddd9f05ec0f424  rfx_fdtd-1.7.0-py3-none-any.whl
e43ef416e298a4cd898ea49f3c7a2f5cabcf564dd3bc197a04af4bb1e39e11e6  rfx_fdtd-1.7.0.tar.gz