Measure the EFIT domain and grid: neither recovers a slice (#459)

[HongSik-Yun-Fusion]

Runs #459's EFIT 2×2 over the three reference discharges: the routine box against a reduced one, at 129×129 against 129×257. Phase 1A's capability questions are answered from the EFIT source, and the Phase 2 matrix is measured. CHEASE and TES are not touched.

What EFIT's grid contract actually is

• nw and nh are runtime arguments, argv(1) and argv(2) (efit.F90:93-104), and nh defaults to nw. A rectangular grid needs no rebuild.
• The Green table name encodes the grid and only the grid (table_name_ch, tables.F90:34).
• EFIT has no notion of the box of its own. It reads rgrid/zgrid out of the table (tables.F90:158) and derives the cell size from them. The box is whatever EFUND's &in5 baked in.
• So two tables for the same grid and different boxes have the same name, and running against the wrong one is undetectable from EFIT's output. verify_table reads the EFUND manifest and refuses a mismatch before any case runs; that refusal is tested.

EFUNDConfig already carried the box. regenerate_legacy_table.py exposed only the grid and now takes --rleft/--rright/--zbotto/--ztop.

The result, over 82 plasma slices

case	box (m)	grid	cell (mm)	produced	accepted	collapsed	findax
A	R 0.05–1.2, Z ±1.5	129×129	9.0 × 23.4	46	30	29	2
B	R 0.05–1.0, Z ±1.35	129×129	7.4 × 21.1	36	23	29	13
C	R 0.05–1.2, Z ±1.5	129×257	9.0 × 11.7	44	28	29	4
D	R 0.05–1.0, Z ±1.35	129×257	7.4 × 10.5	36	23	29	13

• Neither lever recovers a slice. The reduced box costs ten equilibria, the finer grid two. B and D agree on every count on every shot.
• The reduced box fails exactly as the box predicts. findax rejects separatrix points within two cells of the edge (find_axis.F90:308-311); bringing the edge in takes those rejections from 2 to 13. The vacuum volume the proposal wanted to trim is not unused.
• The collapse block is untouched. The same 29 slices collapse in all four cases. [EFIT] Qualify first-slice equilibrium seed sensitivity and convergence basin for VEST #588 ruled out the seed; this rules out the domain and the grid.
• Resolution barely moves the reconstruction. Halving the vertical cell changes every global quantity by under 0.3 % (li is the largest at 0.26 %) and chi-square not at all. The axis height moves by 1.1718 cm, which is half the coarse Z cell to five digits, so that is quantisation rather than a different equilibrium.

All four cases use one acceptance envelope, case A's. aminor_min is a resolution floor and would otherwise move with the grid, leaving four cases judged against four bars.

A log-reading defect, fixed because it changed these numbers

bound can reject a slice before its first Picard iteration, which prints only the error line. parse_slices delimited slices on the iteration counter alone, so it dropped that slice and handed its error to the previous one. On 39915 a converged slice was carrying two failures that were not its own. The parser now opens a slice when a solver error names a new time, and a test pins the real log tail that exposed it.

This changes the earlier denominators: the reference set has 82 plasma slices, not 77. The baseline and seed studies' relative comparisons stand, since every row went through the same parser; their absolute counts predate the fix.

Open, and not resolved here

Case A is the routine box and grid, yet it produces 46 equilibria against the merged baseline's 31. The two runs differ in both the Green table and the mhdin.dat geometry (freshly generated vs packaged), so which one matters is not established. #194's A/B changed the table alone on four slices and found nothing; this is the first evidence that something in that pair matters on the marginal slices. The effect is larger than either lever this issue set out to test, and it needs its own controlled A/B.

Verification

• Full offline suite: 9664 passed, 1063 skipped; one failure, test_plasma_features.py::test_a_spike_on_the_window_edge_does_not_move_the_peak, is the known parallel-only flake and passes in isolation.
• Four tables generated with the controlled EFUND (129×129 in about 2 min, 129×257 in about 4–6 min; rv129257.ddd is 260 MB), each verified against its manifest, EFIT run over every plasma slice of 39915, 41524 and 41672.

Refs #459 (the umbrella stays open for CHEASE, TES and the multi-discharge stage).

🤖 Generated with Claude Code