Golden impedance matrix: eleven minimal decks and two CI jobs that gate on them (#15)

[stevenmburns]

The CI matrix offered on #15: eleven minimal decks, one per feature axis, with pinned input impedances, a stdlib compare script, and two jobs appended to ci.yml. Nothing else in the workflow changes (37 lines added, none removed); both jobs reuse the onec-linux-openblas artifact build-linux already uploads, so they add no build.

What is in it

tests/golden_tests/:

• decks/01_dipole_free.nec … 10_gx_symmetry.nec — free-space dipole, GN 1, GN 2, LD 0, LD 5, a three-wire junction, TL, NT, a three-point FR sweep, GX symmetry. 10–12 lines each; every deck's CM lines state its axis, the pin and the measured spread across engines.
• decks/11_gn2_low_height.nec — GN 2 with the dipole 0.02 λ up: the case All GN grounds (including GN 1 PEC) silently produce free-space results on the sequential path #14's fix did not reach (see below).
• expected.json — pins for decks 01–10 (twelve points), tolerance 1e-4 relative.
• expected_open.json — pins for decks tied to an open issue (deck 11, All GN grounds (including GN 1 PEC) silently produce free-space results on the sequential path #14). When the issue closes, move the entry into expected.json and the floor grows by one deck.
• run_golden.py — stdlib only, Python 3.8+. Runs onec per deck, parses every ANTENNA INPUT PARAMETERS block, fails on |Z − Z_expected| / |Z_expected| > 1e-4, on a nonzero exit, on a missing block, and — the All GN grounds (including GN 1 PEC) silently produce free-space results on the sequential path #14 guard — on an environment block that does not echo the deck's own GN εr and σ. Each failure names the deck.
• PROVENANCE.md — every engine's raw number for every pin.

Two jobs, both needs: build-linux:

job	decks	at acafb37
golden-impedance	01–10	green — twelve points pass, widest 4.0e-05 (05_ld5)
golden-open-issues	11	red by design until #14 closes — 79.228 +j4.256 against 91.758 +j13.222, 1.7e-01

The floor is the regression contract and stays green; the red lives in the second job and points at exactly one deck.

Where the pins come from

I had suggested a three-way agreement at 1e-4 on #15. Measured, that does not exist between NEC-2 transliterations: nec2c 1.3.1 and nec2++ (via PyNEC) land about 2e-4 apart on these decks, and an independent MoM code (momwire, NEC-2's sinusoidal basis) about 1 % away. So every pin is nec2c 1.3.1's printed value, corroborated by nec2++ to ~2e-4 and by momwire to ~1 %, and the 1e-4 gate is a regression tolerance against the pin, not a claim of absolute accuracy. onec itself sits 1e-6 to 4e-5 from nec2c on the floor decks. Deck 10 is two-way (nec2c + momwire) because PyNEC's gx_card path disagrees with its own explicit two-wire twin; deck 11 is three-way, and it is the tightest row in the matrix: nec2c, nec2++ and momwire inside 1e-3 of each other, onec 17 % away while echoing RELATIVE DIELECTRIC CONST.= 10.000 correctly — the Sommerfeld path runs and returns a wrong number, which is why the matrix pins impedances rather than only the echo.

Reproducing

python3 tests/golden_tests/run_golden.py --onec ./onec                                   # exit 0
python3 tests/golden_tests/run_golden.py --onec ./onec --expected tests/golden_tests/expected_open.json   # exit 1 until #14 closes

Parser notes for anyone extending the script: a TL/NT deck prints a STRUCTURE EXCITATION table with the same column shape as the input-parameters row, so the parser scopes to the ANTENNA INPUT PARAMETERS header; onec writes CRLF by default and packs negative fields without a separator, both handled.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KGphPpW5qvL51b5xXQ8m5i