[Umbrella][Equilibrium] Qualify VEST computational-domain and grid-resolution sensitivity across EFIT, CHEASE, and TES

[HongSik-Yun-Fusion]

Context

VEST equilibrium workflows currently rely on solver-specific computational grids and domain choices that have not been systematically qualified for the device geometry.

A representative existing VEST equilibrium box is approximately:

R = 0.0 .. 1.2 m
Z = -1.5 .. +1.5 m

with a common structured-grid baseline of:

129 x 129

The present domain appears to retain generous margin around the PF-coil set. A plausible reduced VEST domain is:

R = 0.0 .. 1.0 m
Z = -1.35 .. +1.35 m

which still covers the relevant PF-coil geometry while reducing unused vacuum volume.

At the same time, the present 129 x 129 grid produces substantially different physical spacing in R and Z because the domain is much taller than it is wide. A rectangular grid such as:

129 x 257

may therefore improve effective spatial resolution and reduce directional grid anisotropy without increasing both dimensions equally.

This issue does not pre-select a new default. Its purpose is to establish the solver capability, regenerate any grid/domain-dependent numerical artifacts correctly, and quantify whether the resulting equilibrium changes are significant.

Scope

This umbrella covers EFIT, CHEASE, and TES.

The work is split into two stages:

1. establish and, where needed, implement solver-specific support for alternate computational domains/resolutions and all dependent geometry/Green/response artifacts;
2. perform a controlled numerical-sensitivity study for each solver.

TokaMaker is intentionally out of scope here because its unstructured finite-element discretization is not directly comparable to the structured (R, Z) grid question studied in this issue.

---

Phase 1 — solver capability and artifact-generation paths

Create one child issue per code. Each child must first audit the native solver contract before changing the wrapper.

1A. EFIT

Determine how VEST EFIT defines and consumes:

• NR, NZ or their executable-equivalent controls;
• Rmin/Rmax/Zmin/Zmax;
• geometry / Green-function tables;
• magnetic response matrices and any other precomputed grid-dependent artifacts.

Questions to resolve:

• Does the installed VEST EFIT support NR != NZ?
• Can 129 x 257 be used directly?
• Which geometry/Green/response artifacts must be regenerated when the grid or domain changes?
• Is artifact identity/provenance sufficient to prevent reuse of a table generated for a different grid/domain?

If needed, extend the VAFT EFIT wrapper to expose the required grid/domain controls and reproducible regeneration path.

EFIT child deliverables

• document the actual EFIT grid/domain contract;
• verify rectangular-grid support;
• identify all dependent geometry/Green/response artifacts;
• regenerate alternate-resolution artifacts reproducibly;
• add wrapper support where required;
• retain grid/domain provenance with generated artifacts.

1B. CHEASE

VAFT currently ties the two CHEASE box dimensions together through one nw setting, emitting conceptually:

NRBOX = nw
NZBOX = nw

Audit whether upstream CHEASE supports independent NRBOX and NZBOX values. If supported, split the VAFT configuration so that rectangular refinement/output grids can be tested explicitly.

The key comparison requested here is:

513 x 513
vs
513 x 257

CHEASE is not a diagnostic-response reconstruction code like EFIT/TES, so do not force a common "response matrix" abstraction where none exists. Instead, distinguish clearly among:

• input GEQDSK grid;
• CHEASE numerical/refinement grid;
• boundary sampling/smoothing;
• output GEQDSK grid.

CHEASE child deliverables

• verify NRBOX != NZBOX upstream capability;
• document the meaning of the current single nw setting;
• split NRBOX/NZBOX in the wrapper if valid;
• verify 513 x 513 and 513 x 257 runs;
• record input/output/refinement-grid provenance separately.

1C. TES

TES already exposes independent structured-grid controls such as nr, nz, rmin, rmax, zmin, and zmax, so it should be the most direct reference implementation for this study.

Audit:

• geometry generation;
• Green/response-matrix generation;
• cache keys and invalidation;
• whether grid/domain changes can accidentally reuse incompatible artifacts.

TES child deliverables

• document geometry and response-artifact generation;
• verify cache/artifact identity includes resolution and domain;
• generate 129 x 129 and 129 x 257 variants;
• generate current-domain and reduced-domain variants;
• retain full numerical provenance.

---

Phase 2 — controlled numerical-sensitivity study

For EFIT and TES, use the same 2 x 2 experiment wherever the solver supports it.

Case	Domain	Grid
A	R=0..1.2 m, Z=+-1.5 m	129 x 129
B	R=0..1.0 m, Z=+-1.35 m	129 x 129
C	current domain	129 x 257
D	reduced domain	129 x 257

Interpretation:

B - A : domain effect
C - A : resolution / cell-aspect-ratio effect
D - A : combined effect

and, where useful, quantify the interaction term rather than assuming the effects are additive.

All other physics inputs and solver settings must remain fixed within a comparison. In particular, do not mix this study with EFIT termination/configuration tuning tracked in #171.

EFIT sensitivity

Compare at minimum:

• convergence state, iteration count, runtime;
• magnetic-probe and flux-loop residuals;
• chi-square / fit-quality metrics;
• magnetic-axis position;
• LCFS and X-point geometry where available;
• elongation, triangularity, area, volume and wall gaps;
• q0, q95, qmin;
• li, beta_p, stored energy;
• pressure, F, FF', and toroidal-current-density profiles.

The objective is not merely to obtain a visually smoother contour, but to determine whether the reconstruction changes by more than the intrinsic numerical/reconstruction uncertainty.

TES sensitivity

Run the same A-D matrix while holding fixed:

• PF currents;
• plasma-current target;
• pressure/current source model;
• limiter/wall geometry;
• passive-current model;
• solver tolerances.

Compare the same geometry/profile/global metrics plus synthetic diagnostic responses and solver cost.

CHEASE sensitivity

CHEASE requires two separate tests.

C1. Propagation of upstream EFIT resolution/domain changes

Compare refinement starting from alternate EFIT equilibria, e.g.:

EFIT baseline -> CHEASE
EFIT alternate grid/domain -> CHEASE

Question:

Does CHEASE fixed-boundary refinement suppress, preserve, or amplify numerical differences introduced upstream by EFIT?

C2. CHEASE-native resolution sensitivity

Hold the same EFIT input fixed and compare CHEASE itself at:

513 x 513
vs
513 x 257

If needed, add one higher-resolution point only to establish a convergence direction.

---

Relationship to the near-axis q0 / GEQDSK pathology

This study is also relevant to #317, which documented that the kinetic-EFIT reference becomes unreliable near the magnetic axis (psi_N <~ 0.05). In that case, the stored q profile has a strong first-point outlier (q[0] around 8 while the next values are around 2), and the inconsistency propagates into toroidal-flux / rho_tor-related quantities.

Do not assume that the q0 spike is caused by coarse resolution. Instead, use this study to separate possible sources:

solver internal resolution
vs
solver output-grid sampling
vs
GEQDSK interpolation / serialization
vs
GEQDSK -> ODS conversion
vs
derived-profile calculation

Track near-axis behavior explicitly for every alternate-resolution case.

Suggested report-only diagnostics include:

• q[0], q[1], q[2], q[3];
• qmin and its location;
• a first-point q-spike metric relative to the next few nodes;
• dV/dpsi near the axis;
• phi(psi) / rho_tor consistency;
• monotonicity and regularity of the first few flux surfaces.

The goal is to determine whether the pathology decreases systematically with improved numerical resolution before considering any smoothing, clipping, or import-time sanitization.

---

Common comparison outputs

For every solver/case, retain a machine-readable comparison record containing at least:

Category	Quantities
Grid	NR, NZ, dR, dZ, cell aspect ratio
Domain	Rmin, Rmax, Zmin, Zmax
Geometry	axis, X-points, LCFS displacement, kappa, delta, area, volume
Profiles	q, pressure, F, FF', j_phi
Globals	q0, q95, qmin, li, beta_p, stored energy
Near-axis	q-spike metric, dV/dpsi, rho_tor consistency
Solver	convergence state, iterations, residuals
Cost	runtime, artifact size
Provenance	solver version, grid/domain, geometry/response revision

For LCFS comparison, prefer quantitative contour metrics in addition to visual overlays, such as:

• mean/RMS boundary displacement;
• maximum boundary displacement;
• Hausdorff distance.

Because VEST is small, also report absolute contour displacement in millimetres.

---

Execution plan

Stage 1 — infrastructure

• create EFIT child issue;
• create CHEASE child issue;
• create TES child issue;
• audit native grid/domain semantics;
• implement missing wrapper controls only where required;
• establish reproducible alternate geometry/response/grid artifact generation;
• make numerical provenance explicit.

Stage 2 — one-reference-case study

Use one well-behaved VEST equilibrium slice first and complete the full matrix before broadening the scan.

• EFIT A-D;
• TES A-D;
• CHEASE upstream-EFIT propagation comparison;
• CHEASE 513 x 513 vs 513 x 257.

The priority is attribution of numerical effects, not database-scale statistics.

Stage 3 — representative regimes

Extend only after Stage 2 is understood. Candidate cases include:

• nominal limited plasma;
• high-current equilibrium;
• strongly elongated plasma;
• X-point / near-diverted topology where available;
• a case with a strong near-axis q pathology.

Stage 4 — recommendation

Produce a solver-specific recommendation containing:

recommended domain
recommended grid/resolution
required generated artifacts
supported rectangular-grid combinations
expected numerical uncertainty
runtime/cost trade-off
known pathological regimes

Any production-default change should be a separate follow-up PR after this qualification is complete.

---

Related issues

• The kineticEfit reference ODS is not trustworthy within psi_N < 0.05 #317 — near-axis kinetic-EFIT reference pathology / q0 spike;
• Characterize and explicitly define VEST's EFIT solver configuration #171 — EFIT solver-configuration and termination characterization; keep fixed during this study;
• Add CHEASE-based fixed-boundary equilibrium variation and sensitivity workflows #66 — CHEASE fixed-boundary sensitivity workflow; reuse its scan architecture where appropriate;
• [P2] Add a TES-to-EFIT synthetic-diagnostic round-trip benchmark #75 — TES -> synthetic diagnostics -> EFIT round-trip benchmark; this work can provide qualified numerical settings and provenance for that benchmark.

---

Proposed child issues

1. [EFIT] Support and validate alternate VEST computational grids and response matrices
2. [CHEASE] Decouple NRBOX/NZBOX and validate rectangular refinement grids
3. [TES] Generate and validate alternate-resolution VEST geometry and response artifacts
4. [Validation] Quantify VEST equilibrium domain and resolution sensitivity across EFIT, CHEASE, and TES

The first three are infrastructure/capability work. The fourth is the cross-solver numerical study and should begin only once the solver-specific paths are reproducible.

---

Non-goals

• introducing a new equilibrium physics model;
• changing EFIT diagnostic weights/profile basis;
• tuning EFIT stopping criteria (Characterize and explicitly define VEST's EFIT solver configuration #171);
• broad CHEASE profile-variation research beyond the numerical-grid question (Add CHEASE-based fixed-boundary equilibrium variation and sensitivity workflows #66);
• smoothing or clipping q0 spikes without numerical evidence;
• sanitizing GEQDSK values by default;
• immediately rerunning the full VEST database;
• declaring 129 x 257 or the reduced domain as the new default before convergence evidence exists.

Acceptance criteria

• EFIT, CHEASE, and TES Phase-1 child issues are created and completed.
• Each solver's computational-domain and resolution semantics are documented.
• Alternate grid/domain-dependent geometry/Green/response artifacts can be regenerated reproducibly where applicable.
• Numerical artifact provenance prevents incompatible reuse.
• EFIT rectangular-grid capability is established.
• CHEASE independent NRBOX/NZBOX capability is established and the wrapper is updated if valid.
• TES reproduces both domains at 129 x 129 and 129 x 257.
• Domain-only, resolution-only, and combined effects are quantified separately.
• CHEASE upstream-EFIT propagation and native 513 x 513 vs 513 x 257 sensitivity are quantified.
• Geometry, profiles, global quantities, convergence, fit quality, and runtime are compared quantitatively.
• Near-axis q0 / dV/dpsi / rho_tor behavior is compared across resolution changes and explicitly related back to The kineticEfit reference ODS is not trustworthy within psi_N < 0.05 #317.
• A justified VEST numerical-configuration recommendation is produced.
• Any default change is deferred to a separate follow-up implementation PR.

[HongSik-Yun-Fusion]

Scope update — use full-discharge trajectories as the qualification unit

This issue should no longer use a single equilibrium slice as the primary scientific basis for choosing a VEST grid/domain configuration.

The infrastructure/capability work in Phase 1 remains unchanged, but the numerical qualification stage is amended as follows:

Full-discharge requirement

For EFIT and TES, every candidate grid/domain case A–D must be evaluated on the same complete usable discharge time grid for the first reference discharge:

A -> E_A(t)
B -> E_B(t)
C -> E_C(t)
D -> E_D(t)

Single equilibrium slices may still be used for debugging, visualization, or inspecting a specific numerical pathology, but they are not sufficient evidence for selecting a production/default spatial configuration.

Primary comparisons should therefore be trajectory-level quantities such as:

Delta LCFS(t)
Delta area(t)
Delta volume(t)
Delta q95(t)
Delta li(t)
Delta beta_p(t)
convergence(t)
runtime(t)

with whole-discharge summaries and phase-resolved summaries over:

1. current ramp-up;
2. quasi-stationary / slow-evolution phase;
3. current ramp-down / termination.

Use the same broad phase semantics as #579 / #76 where practical rather than introducing another phase taxonomy.

Revised execution logic

Replace the previous scientific role of the "one-reference-case" stage with:

Stage 2 — one-reference-full-discharge study
    EFIT A-D over the complete usable discharge
    TES A-D over the corresponding complete usable trajectory where applicable

Stage 3 — limited multi-discharge confirmation
    repeat only the configurations needed to confirm the observed sensitivity
    on a small number of representative full discharges

CHEASE remains a special case because its fixed-boundary refinement cost/semantics differ from EFIT/TES. CHEASE-native resolution tests may use matched representative slices when needed, but propagation of upstream EFIT spatial sensitivity should sample the full-discharge result sufficiently to establish whether CHEASE suppresses, preserves, or amplifies the observed trajectory-level differences.

Decision evidence

A recommended VEST grid/domain should be justified from full-discharge behavior, including at least:

• convergence fraction and failure distribution;
• runtime/cost per discharge;
• distribution and maximum excursion of LCFS displacement;
• area/volume sensitivity over time;
• global/profile sensitivity over time;
• phase dependence;
• near-axis pathology behavior over time where relevant.

This amendment keeps the original orthogonality of the study intact: profile basis, termination settings, initialization policy, and temporal cadence/averaging remain fixed while spatial discretization/domain is varied.

[HongSik-Yun-Fusion]

Where the EFIT work stands, and what should run next

Posting this here because #459 is the equilibrium umbrella and the studies under it now share one set of prerequisites. Five pull requests landed on develop today, which changes the blocking map, and a magnetics-quality verification ran ahead of any further study. Nothing below closes an issue; it says what is now unblocked and in what order the remaining work is worth doing.

What landed on develop

PR	Issue	What it means for the studies
#568	#433	Constraint values are now a true box average over the samples the window contains, not an interpolation onto a 0.1 ms grid that discarded ~60 % of a 25 kHz record. This was the stated prerequisite for #468.
#555	#296	EFIT gets one channel decision per slice, formed upstream and only translated by the writer. Five states, recorded in the product under equilibrium.code.parameters.channel_decisions, so a zero weight in a k-file can be traced to its reason. The manual list is retired (#295 §4) and a test fails if one reappears.
#543	#295	The flux-loop evidence table, refreshed now that every packaged shot has a model witness.
#584	#194	The Green-table provenance chain: both toolchain roles under one EFITHOME, canonical geometry → EFUND projection, typed EFUNDConfig, table manifests, and every EFIT run recording the table it consumed.
#575	#468	The cadence/window harness, and a k-file that can express a sub-millisecond time.

The blocking map, after those merges

• [EFIT] Qualify VEST temporal cadence and constraint-averaging window #468 is unblocked. EFIT k-file constraint averages sample a 1e-4 grid over 4e-5 data #433 was its stated prerequisite and it has landed.
• Add physics-constrained Bayesian/GP magnetic-sensor recovery with uncertainty-aware EFIT propagation #297 is unblocked. It waited on Decouple magnetic-channel validation and recovery policy from EFIT constraint generation #296 for the EFIT-facing contract, which now exists.
• Local EFIT reconstruction diverges from legacy reference in magnetic-axis/boundary (deferred) #119 is no longer a Green-table question. The [Feature] Generate Versioned EFIT Green Tables from Canonical VEST Geometry and Add Reproducible EFUND Support #194 A/B reconstructed the same 39915 slices with a table generated from canonical geometry and with the bundled one: identical χ², plasma current and axis to every printed digit, and both failing identically. The remaining suspects are the constraint set and the termination settings, which is Characterize and explicitly define VEST's EFIT solver configuration #171.
• [Feature] Generate Versioned EFIT Green Tables from Canonical VEST Geometry and Add Reproducible EFUND Support #194 still waits on [Feature] Version the VEST Wall/Passive Model by Shot and Add 965-Loop Coupling and Effective Resistance Calibration #191 for era-aware table selection and packaged versioned tables. The A/B result did not justify doing them now, and they were deliberately left out.
• Build the VAFT EFIT baseline stage and finalized main ODS #91 still waits on [P2] Classify equilibrium reconstruction failures by VEST operating regime #76, the only open item in its declared dependency list.

The input-quality answer, ahead of any study

workflow/magnetics_quality/ now asks the question none of the studies asked: is a shot's magnetics fit to reconstruct from? It composes the three existing layers — signal quality (#189), the channel decisions (#296), the vacuum-model comparison (#190) — into one per-shot verdict over the window the routine pipeline would actually reconstruct, with every threshold in one policy the table records. Model disagreement is reported beside the verdict and never inside it, since #190 is explicit that acceptance bounds wait for the VEST distribution.

On the three packaged shots the verdict is degraded on all three, unfit on none:

shot	window [s]	slices	usable frac (min)	condemned probes	measured to [s]	model residual (median)
39915	0.306–0.331	26	0.97	H3-08	0.340	0.016
41524	0.315–0.336	22	0.95	H3-08, C3-01, C3-05	0.340	0.037
41672	0.312–0.352	41	0.91	H3-08, C1-01, C3-01, C3-02, C3-05, C4-02	0.360	0.026

Five findings the studies should carry:

1. No EFIT-facing channel carries a stored validity in the packaged pre-EFIT products — 0 of 75. The only 8 channels with a record are IMPA probes, which have no waveform at all. "All valid" and "never looked at" are the same bytes in a product, so the sweep assesses and gates into a copy rather than trusting what it is handed. Anything that reads projected validity from these products finds nothing.
2. H3-08 (b_field_pol_probe[25]) is condemned on all three shots, with an amplitude 40 times its family's median, above the physical ceiling, and no usable sample anywhere in the record. The condemned set grows with shot number and is entirely inboard and C-array probes.
3. One EFIT-facing channel is missing on every shot (b_field_pol_probe[0], H1-01), as a deterministic zero-weight placeholder rather than a silent gap.
4. The routine window ends before each record does. 39915's magnetics hold their last value from 0.340 s while its window ends at 0.331 s, so the hold does not touch the reconstructed slices — but a study that widened the window past 0.340 s would be fitting held values.
5. The flux loops agree with the vacuum model on all three shots: 11 of 11 usable, normalized residual 1.6–3.7 % median. Wall authority is low (0.05–0.06), so this is weak corroboration, not a strong test. Together with Replace manual EFIT flux-loop blacklist with validation-driven automatic channel selection #295's table this is the input-side answer there: nothing rejects a flux loop on any packaged shot, and the automatic layer instead rejects probes the manual list never named.

Recommended order

#171 first. It owns what the others need and cannot borrow: a typed numerics configuration, per-slice runtime instrumentation, and a curated reference shot set. Today every local fit exits on the χ² criterion at 11 iterations with a GS error around 8e-3 while the k-file asks for 1e-5, and the stored 2023 reconstruction reaches a residual ten times smaller with a condition number three orders of magnitude lower. Until that difference is named, every other study measures the configuration rather than its own variable.

Then #468, with the harness that just landed and the box average underneath it. Its first pass is already recorded as a baseline; the note is explicit that no cadence converges under the configuration it can currently hold fixed, so the study proper starts when #171 pins one.

Then #196, holding #171's configuration fixed. It must not be scanned together with #171 or #468 — three temporal controls varied at once measure nothing.

Then #459 and #579, the spatial and profile-model counterparts, on the same pinned configuration.

Alongside, independently: #386 (the packaged reconstruction's pressure is ~90× below force balance and the diamagnetic constraint is effectively weightless under legacy_weight) and #325 (the stored r0) are cheap, do not depend on the numerics work, and both currently corrupt derived quantities that the studies above would otherwise report.

Two statements on issues that are now stale

• Computed diamagnetic flux has the opposite sign to the measured loop: the k-file writer feeds EFIT |DFLUX| #385 says the k-file writer feeds EFIT |DFLUX|. The code fix landed in PR Write the diamagnetic flux to EFIT with its sign (#385) #387: the default is now "imas" and the sign is written. What remains is regenerating stored products, which is worth saying on the issue so nobody re-fixes the writer.
• Migrate validate_* out of vaft.process and unify the result model (#253 increment A) #337 and Move the stage-QA artifact contract to vaft.database (#253 increment B) #338 are still open although their implementing PRs (Give VAFT one validation boundary and one result model (#337) #344, Move the stage-QA artifact contract to vaft.database (#338) #369) merged. Either they have residual scope worth naming or they can be closed.

[HongSik-Yun-Fusion]

A second failure mode for this issue, from the #171 baseline

Distinct from the leading-block collapse I have reported to #196, and it looks like a domain question rather than an initialization one.

Scattered through the slices that otherwise fit normally — Grad-Shafranov error 0.007–0.010, the fit itself unremarkable — findax raises "1st separatrix point is off grid" and the slice writes no output. On 39915 that is 320–324 and 328; on 41524, 331, 332 and 335; on 41672, eight slices interleaved with successful ones. Roughly a third of the slices that survive the fit are lost this way.

find_axis.F90:308-311 raises it when the separatrix point lands within two cells of the grid edge. The current computational box is R 0.05–1.2 m, Z ±1.5 m at 129x129, which is the anisotropy this issue already names: dR = 9.0 mm against dZ = 23.4 mm, a factor of 2.6. A separatrix search that leaves the box vertically has 2.6 times coarser resolution to work with than it does radially, and the box extends to |Z| = 1.5 m where the limiter reaches only 1.185 m.

So this is a concrete, measurable target for the domain and grid study: does a box matched to the machine — the limiter reaches R 0.104–0.760 m and |Z| ≤ 1.185 m — or an isotropic cell recover those slices? The count of findax losses is a cleaner metric for the domain question than convergence is, because these slices already converged.

One caveat before anyone reruns: a table generated for a different grid is a different table, so the comparison has to hold the Green tables fixed or regenerate them deliberately. workflow/efit_tables/ does the second properly, and its README records what changed when the tables were last regenerated.

[HongSik-Yun-Fusion]

The collapse block is yours: measured, not inferred

I handed you the findax off-grid losses earlier. The #588 seed study has now ruled initialization out of the other population as well, so both belong here.

Of the 29 slices across the three reference discharges that collapse to a null solution and fail in bound, four are ever recovered by any seed, and every one of them sits at the trailing edge of its block, adjacent to where the fit starts working:

shot	collapse block	ever recovered by any seed
39915	307–315 (9 slices)	315
41524	315–327 (13)	327
41672	315–321 (7)	320, 321

The sweep covered a threefold range in seed minor radius (0.10–0.30 m), a factor of two in elongation (1.0–2.0), ±10 cm in the seed's radial centre and ±5 cm vertically. 86 % of the block is untouched by all of it. The seed can move the boundary between the failing and working phases by a slice or two; it cannot touch the block.

So the leading collapse is a shape, limiter or grid question, not an initialization one. Together with the 17 findax losses this issue already has, that is 46 of 77 plasma slices — 60 % — currently producing no equilibrium for reasons that live in this issue's scope.

Two concrete starting points from the solver, both quoted in my earlier comment: bound raises "first and last contour points are too far apart" (boundary.f90:433) after 60 bisections fail to close a contour, and findax raises "1st separatrix point is off grid" when the search lands within two cells of the grid edge (find_axis.F90:308-311). The box is R 0.05–1.2 m, Z ±1.5 m at 129×129, so dR = 9.0 mm against dZ = 23.4 mm, and it extends to |Z| = 1.5 m where the limiter reaches only 1.185 m.

The harness that produced all of this is workflow/efit_numerics/ on develop; the per-slice tables are in test/data/efit_seed_basin.json.

(This replaces an earlier version of this comment in which a shell quoting error stripped the code spans.)

[HongSik-Yun-Fusion]

EFIT: Phase 1A answered, and the 2×2 measured (#688)

Short answer: neither the domain nor the grid recovers a single slice, and the reduced domain makes things worse.

The EFIT grid contract (1A)

• Rectangular grids work with no rebuild. nw and nh are argv(1) and argv(2) (efit.F90:93-104); efit 129 257 runs, and EFUND generated the 129×257 table in about 4 minutes.
• Dependent artifacts: the Green tables (ec, ep, rv<nw><nh>.ddd) carry the grid and the box. EFIT reads rgrid/zgrid from the table (tables.F90:158) and has no notion of the domain of its own. lim.dat and the &incheck envelope are EFIT inputs in the same directory, not EFUND products.
• Provenance is not sufficient in EFIT itself. The file name encodes nw/nh only (tables.F90:34), so tables for one grid and two boxes are named identically. The EFUND manifest does record the box, and the study now refuses any table whose manifest disagrees with its case.
• Wrapper support: EFUNDConfig already had the box; the regeneration workflow now exposes it.
• Not exercised: npoint (boundary-point array size) is chosen from nw alone (efit.F90:114-120).

The routine box starts at R = 0.05 m, not 0.0 m as quoted above; the Green functions are singular on the axis.

The 2×2, over 82 plasma slices of 39915, 41524 and 41672

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

• B − A (domain): ten fewer equilibria. Bringing the edge in to |Z| = 1.35 m takes findax's "off grid" rejections from 2 to 13, as find_axis.F90:308-311 predicts. The vacuum margin is doing work.
• C − A (resolution): no slice recovered. Every global quantity moves by under 0.3 % (li the largest at 0.26 %); chi-square does not move at all. The magnetic axis height moves by 1.1718 cm, which is half the coarse Z cell, so that is quantisation.
• D − A: identical to B on every count on every shot. Once the box is reduced, the grid changes nothing.
• The collapse block I handed over here is untouched: 29 slices collapse in all four cases. With [EFIT] Qualify first-slice equilibrium seed sensitivity and convergence basin for VEST #588 having ruled out the seed, this block is not a spatial-discretisation problem. I am taking it back out of this issue's scope.

The shots disagree in detail (41672 gains two accepted slices from the reduced box; the other two lose), so none of this is a recommendation. Per the scope update above, a default should come from more full discharges.

Two corrections to what I reported earlier

1. The slice counts were short. The log parser dropped any slice bound rejects before its first iteration and attributed its error to the previous slice. Fixed in Measure the EFIT domain and grid: neither recovers a slice (#459) #688. The reference set has 82 plasma slices, not 77; earlier relative comparisons stand, absolute counts do not.
2. The "46 of 77 lost to this issue" figure no longer holds. In case A only 36 of 82 produce nothing, and of those, the 29 collapses have now been ruled out of [Umbrella][Equilibrium] Qualify VEST computational-domain and grid-resolution sensitivity across EFIT, CHEASE, and TES #459's scope.

Open question this raised

Case A (routine box and grid) produces 46 equilibria where the merged #171 baseline produced 31. The two differ in both the Green table and the mhdin.dat geometry (freshly generated vs packaged), so the cause is not isolated. That is a bigger effect than anything this 2×2 found, and it reopens #194's question with new evidence. It needs a controlled A/B that changes one of the two at a time.

Harness: workflow/efit_numerics/domain_grid.py; record test/data/efit_domain_grid.json.