[d-burg]

Release note

• Audience: users
• Numerical impact: SLAYER roots are now reported in the lab frame in both coupling modes. Uncoupled: each surface's Roots/Q_root and Roots/omega move by τ_k·n·Ω_E and n·Ω_E (on diiid_slayer_n1 ω goes −10.5 k → +31.0 k rad/s at the 2/1, −9.7 k → +21.0 k at the 3/1, −11.1 k → −2.8 k at the 4/1); γ is unchanged. Coupled (coupling_mode = "coupled"): the determinant is now Doppler-shifted per surface and rescaled in the right direction; on diiid_slayer_n1_coupled ω moves −10.5 k → +31.0 k rad/s and γ by −0.3 % (220.25 → 219.67 s⁻¹), inside the coupled root extraction's accuracy. Every layer input is bit-identical. (harness @ 546d9fb)
• Migration: Roots/Q_root and Roots/omega are lab-frame in both modes; an uncoupled root's previous, E×B-frame value is ω − n·Ω_E from PerSurface/omega_E and PerSurface/n, and Diagnostics/* stay in each surface's E×B frame. A kinetic file without omega_E gives Ω_E = 0 (lab = E×B frame), records NaN in PerSurface/omega_E, and warns. In coupled mode widen Q_re_range if the run warns that Doppler offsets fall outside it. Interface removals: slayer_parameters(omega=…), multi_surface_coupling_full(rotation=…, ntor=…), KineticProfiles(omega_e=…, omega_i=…) and build_slayer_inputs(compute_omega_star=…); drop those arguments. KineticProfiles.omega is renamed omega_E.

The coupled tearing determinant evaluated every non-reference surface at the wrong Q (off by (τ_k/τ_ref)²), so its roots depended on which surface was chosen as reference. It also ignored the E×B rotation that the kinetic file already supplies. Both are fixed, and the root extraction is now reproducible run to run.

Left: on develop the coupled root sits at the 2/1's own-frame frequency (−10.5 krad/s). With this PR it moves by n·Ω_E at the 2/1 (41.5 krad/s) into the lab frame, landing where shifting the uncoupled 2/1 root predicts (30,993 against 30,996 rad/s); γ changes by under 1 % (221.9 → 219.7 s⁻¹ in that run). Right: the E×B Doppler shift n·Ω_E now applied at each surface, from the kinetic file. Both from the harness run at c4c95c4. With uncoupled roots now also reported in the lab frame, that comparison is in the raw output: the uncoupled 2/1 is at 30,996 rad/s and the coupled root at 30,993.

What changed

change	where
Surface k's layer is evaluated at Q·τ_k/τ_ref, not Q·τ_ref/τ_k	Coupled.jl, CoupledFullMatch.jl (one line each)
Real per-surface Doppler offset q_shift = −τ_k·n·Ω_E, from the kinetic file's omega_E, applied in coupled mode only	SurfaceCoupling.jl, run_slayer.jl, TearingProblem.jl
The full 4m×4m form's + i·ntor·rotation[k] is removed: it shifted γ instead of ω and added s⁻¹ to a dimensionless Q	CoupledFullMatch.jl
The AMR Delaunay triangulation is seeded	GrowthRateExtraction.jl (3 lines)
PerSurface/omega_E and PerSurface/q_shift are written to the output; omega_E is NaN when the kinetic file has none	Result.jl, HDF5Output.jl
Uncoupled roots are reported in the lab frame, Q + τ_k·n·Ω_E; each layer is still scanned and solved in its own E×B frame	run_slayer.jl
Always-zero rotation and diamagnetic arguments are removed	KineticProfiles.jl, LayerInputs.jl, LayerParameters.jl and eight test files
New harness case diiid_slayer_n1_coupled	regression-harness/cases/

The source change is +169/−119 lines. Most of the 24 files are tests that only drop the removed arguments.

What moves

regress --refs 9a51446df,546d9fbec --force on feynman (SLURM, 8 threads), all 20 cases: develop @ 9a51446 against this branch @ 546d9fb, which has develop merged. Julia 1.11.9, manifest pinned.

case	result
diiid_slayer_n1 (uncoupled): ω	2/1 −10,526 → +30,996; 3/1 −9,676 → +21,048; 4/1 −11,097 → −2,768 rad/s, each by n·Ω_E
diiid_slayer_n1 (uncoupled): γ	unchanged (≤ 0.03 s⁻¹, develop's triangulation noise, below the harness threshold); layer inputs bit-identical
diiid_slayer_n1_coupled: ω	−10,522 → +30,993 rad/s
diiid_slayer_n1_coupled: γ	220.25 → 219.67 s⁻¹ (−0.3 %)
diiid_slayer_n1_coupled: q_shift	new output, absent on develop
the other 18 cases	unchanged (gal_resistive_pe extracts nothing on either ref, as on develop)

Since the previous report (e7ce73c on SDCC), this branch adds the sign-anchor test, the NaN omega_E and the lab-frame uncoupled roots. Only the last moves a number, and only uncoupled ω and Q_root. The coupled γ came out −1.0 % (feynman, c4c95c4), −0.3 % (SDCC, e7ce73c) and −0.3 % here; see How far to trust the coupled γ below.

For the reviewer

• Why τ_k/τ_ref. With Q = τ_k·(ω + iγ), one physical eigenvalue reaches surface k as Q·τ_k/τ_ref. With a diagonal Δ′ the coupled roots must then reproduce each surface's uncoupled root whichever surface is the reference; the old direction got only the reference surface right. Pinned by "Coupled roots do not depend on the reference surface" in runtests_dispersion_rotation.jl.
• Why the factor n. GPEC's omega_E is per unit n, so the mode sees n·Ω_E.
• Why this sign. The layer's Q is the mode frequency in the local E×B frame, so Q_layer = τ_k·ω_lab − τ_k·n·Ω_E. A test on two real SLAYER surfaces checks that the rotating determinant at Q + τ_ref·n·Ω_E equals the static one at Q to within 1e-4 (measured agreement 3e-6, the layer solve's rounding noise); the opposite sign is off by 27–394 %.
• How far to trust the coupled γ. The coupled root is the raw contour intersection, with no polishing. Re-extracting one set of scan samples under different triangulations gave γ = 190.6 and 191.9 s⁻¹, a 0.65 % spread (measured before develop's Sauter default). The seed makes it repeatable, but it still shifts between environments: this branch gives 219.67 s⁻¹ on feynman (Julia 1.11.9) and 220.05 s⁻¹ on SDCC (Julia 1.12.6). The ω shift is resolved; the γ shift (−1.0 % in one run, −0.3 % in the other) is not.
• Fortran has the same inversion on its slayer_growthrate branch, so the earlier Julia/Fortran coupled comparison did not validate the rescaling direction.
• Not in this PR. A per-surface override of the file rotation for rotation scans (omega_E_kHz) is split into Tearing - FEATURE - Add a per-surface omega_E_kHz override of the kinetic-file E×B rotation #519, stacked on this one.

Full regression report (SLAYER cases)

Regression Report: diiid_slayer_n1
============================================================================================
Ref 1: 9a51446df  @ 9a51446df (2026-10-22)
       env: julia 1.11.9, x86_64-linux-gnu, manifest f826ee26 (pinned), 8 threads/8 BLAS, jit build
Ref 2: 546d9fbec  @ 546d9fbec (2026-10-23)
       env: julia 1.11.9, x86_64-linux-gnu, manifest f826ee26 (pinned), 8 threads/8 BLAS, jit build
--------------------------------------------------------------------------------------------
Quantity                            9a51446df  546d9fbec  Diff                 Status
--------------------------------------------------------------------------------------------
SLAYER surface indices              [6 elem]   [6 elem]   0.0e+00              OK
SLAYER poloidal m                   [6 elem]   [6 elem]   0.0e+00              OK
SLAYER toroidal n                   [6 elem]   [6 elem]   0.0e+00              OK
SLAYER minor radius rs              [6 elem]   [6 elem]   0.0e+00              OK
SLAYER r-based shear                [6 elem]   [6 elem]   0.0e+00              OK
SLAYER Lundquist S                  [6 elem]   [6 elem]   0.0e+00              OK
SLAYER D_norm                       [6 elem]   [6 elem]   0.0e+00              OK
SLAYER P_perp                       [6 elem]   [6 elem]   0.0e+00              OK
SLAYER tauk                         [6 elem]   [6 elem]   0.0e+00              OK
SLAYER iota_e                       [6 elem]   [6 elem]   0.0e+00              OK
SLAYER Q_e                          [6 elem]   [6 elem]   0.0e+00              OK
SLAYER Q_i                          [6 elem]   [6 elem]   0.0e+00              OK
SLAYER Q_root [2/1,3/1,4/1]         [3 elem]   [3 elem]   2.463e+00 (394.38%)  ** CHANGED **
SLAYER ω_Hz [2/1,3/1,4/1]           [3 elem]   [3 elem]   4.152e+04 (374.18%)  ** CHANGED **
SLAYER γ_Hz [2/1,3/1,4/1]           [3 elem]   [3 elem]   2.5e-02              OK
SLAYER no_root flags [2/1,3/1,4/1]  [3 elem]   [3 elem]   0.0e+00              OK
SLAYER enabled flag                 1          1          0.0e+00              OK
Runtime (s)                         298.6s     300.2s                          --
============================================================================================
Summary: 2 changed, 15 unchanged

Regression Report: diiid_slayer_n1_coupled
=====================================================================================
Ref 1: 9a51446df  @ 9a51446df (2026-10-22)
       env: julia 1.11.9, x86_64-linux-gnu, manifest f826ee26 (pinned), 8 threads/8 BLAS, jit build
Ref 2: 546d9fbec  @ 546d9fbec (2026-10-23)
       env: julia 1.11.9, x86_64-linux-gnu, manifest f826ee26 (pinned), 8 threads/8 BLAS, jit build
-------------------------------------------------------------------------------------
Quantity                     9a51446df  546d9fbec  Diff                 Status
-------------------------------------------------------------------------------------
SLAYER tauk                  [6 elem]   [6 elem]   0.0e+00              OK
SLAYER q_shift               N/A        [6 elem]   N/A                  N/A
SLAYER coupled Q_root        [1 elem]   [1 elem]   2.463e+00 (394.48%)  ** CHANGED **
SLAYER coupled ω             [1 elem]   [1 elem]   4.151e+04 (394.56%)  ** CHANGED **
SLAYER coupled γ             [1 elem]   [1 elem]   5.753e-01 (0.26%)    ** CHANGED **
SLAYER coupled no_root flag  [1 elem]   [1 elem]   0.0e+00              OK
SLAYER enabled flag          1          1          0.0e+00              OK
Runtime (s)                  310.1s     311.8s                          --
=====================================================================================
Summary: 3 changed, 3 unchanged, 1 missing/N/A

The other 18 cases report every quantity unchanged.

The run-to-run scatter in the coupled γ, and its fix

• An earlier version of this PR saw the coupled γ take two values run to run with identical source.
• Δ′, the energies, all per-surface SLAYER inputs and the stored AMR scan samples were bit-identical in every run; only the extracted roots differed.
• The cause was the triangulation of the scan samples. DelaunayTriangulation inserts points in a random order drawn from Julia's global RNG, and an AMR grid is full of cocircular points, so the order picks the diagonals. Re-extracting the same stored samples under 10 seeds gave exactly the two observed values.
• The fix draws the order from a fixed-seed generator. A test asserts that extraction no longer touches the global RNG; it fails without the fix.

Other notes

• Rotation enters the determinant in coupled mode only. Each uncoupled layer is solved in its own plasma frame, and its root is then relabelled into the lab frame; PerSurface/q_shift stays 0 there. PerSurface/omega_E is written in both modes.
• GGJ surfaces are not rotated, because they carry no time normalization; the run warns.
• n = 0 with nonzero rotation is an error, since the shift would silently vanish.
• External callers that build KineticProfiles with omega_e=zeros(...), omega_i=zeros(...) must drop those two arguments.
• ForceFreeStates.ResistiveMatch(; rotation) is unaffected: it uses the Laplace-variable eigenvalue, where the imaginary sign is correct.

---

NO MERGE WITHOUT HUMAN REVIEW. This PR needs a third-party human reviewer's approval of the head that lands.

🤖 Generated with Claude Code

[matt-pharr]

@d-burg you requested I review this today so I will assume it is ready for review. I will review it and leave it to you to edit after that

[matt-pharr]

I pushed two small commits on top of the review:

The Doppler sign test now has a physical anchor (test/runtests_dispersion_rotation.jl)

• The "Real SLAYER surfaces" test checks that the rotating determinant at Q + τ_ref·n·Ω_E equals the static one at Q. With q_shift = −τ_k·n·Ω_E that holds identically. So it pins how _q_shift composes with the determinant, but it would also pass with the sign flipped throughout. Its comments now say so, and its inputs use GPEC-signed ω_* (ω_*e < 0, ω_*i > 0).
• New testset, "Scanned Re(Q) runs along the ζ axis of Ω_E". The Doppler sign is right only if the scanned Re(Q) lies along the same ζ axis as the kinetic file's Ω_E and GPEC's ω_*e. The layer Δ vanishes where the mode co-rotates with the electron fluid. The test checks that this zero sits at Re(Q) = +τ_k·ω_*e, not −τ_k·ω_*e: the |Δ| ratio is about 0.015 at Im Q = 0.01. Reversing either Q_e = −τ_k·ω_*e or the i·conj(Q) bridge in solve_inner moves the zero to −τ_k·ω_*e and fails the test.

omega_E is NaN when the kinetic file has none (run_slayer.jl, TearingProblem.jl, Result.jl, HDF5Output.jl)

• _load_profiles now reports whether the file carries omega_E, and the tearing solve passes no rotation when it doesn't. That still applies no shift (q_shift stays 0). But result.omega_E and PerSurface/omega_E now record NaN, the HDF5 not-computed sentinel, so a missing profile can be told apart from a measured zero rotation. The same holds for run_slayer_from_inputs called without omega_E. A new check in runtests_slayer_runner.jl covers it.

Harness case comment (regression-harness/cases/diiid_slayer_n1_coupled.toml)

• The comment above the coupled-root thresholds now explains why they are tight: the seeded triangulation repeats exactly in one environment. It also says the 0.65 % triangulation spread applies across environments. The unfinished "ω and Q need the same measurement" line is gone, and the thresholds are unchanged.

No harness number moves. The NaN only replaces zeros in an output that applies no shift; everything else is tests and comments.

[matt-pharr]

@d-burg the physics checks out. The τ_k/τ_ref direction is right, and so is the Doppler sign: the two sign flips (Q_e = −τ_k·ω_*e and the i·conj(Q) bridge in solve_inner) cancel, so Re(Q) runs along GPEC's ζ axis, the same axis as Ω_E. The new sign-anchor test pins that. The harness at e7ce73c moves only the coupled-SLAYER results, as described.

One design question before I approve: which frame should Roots/omega and Roots/Q_root report in?

Right now it depends on the mode: uncoupled roots are in each surface's own E×B frame, and the coupled root is in the lab frame. I'd prefer always lab frame:

• The coupled root has no single E×B frame, because each surface rotates at its own Ω_E. So the lab frame is the only frame both modes share. Reporting the coupled root in the reference surface's frame instead would bring back the reference-surface dependence this PR removes.
• Uncoupled and coupled results become directly comparable. The figure shifts the uncoupled 2/1 root by n·Ω_E by hand (30,996 against 30,993 rad/s); with lab-frame output, the raw data shows that comparison directly.
• It's what experiments measure. Mirnov coils see lab-frame frequencies.
• For uncoupled roots it is only a relabel. Report Q_root + τ_k·n·Ω_E,k and the matching ω, so ω = Re(Q)/τ_k still holds. The scan and root-finding stay in each layer's own frame, and γ doesn't change. The E×B-frame value stays recoverable as ω − n·Ω_E from PerSurface/omega_E and PerSurface/n. When the kinetic file has no omega_E, take Ω_E = 0, as the coupled path already does.

The cost is that uncoupled ω moves, by n·Ω_E at each surface (+41.5 krad/s at the 2/1 in diiid_slayer_n1). So the release note would need updating, plus a fresh harness run.

Could you either:

1. switch to always reporting in the lab frame, as above, or
2. tell me you feel strongly that uncoupled roots should stay in their own E×B frame, and why? Then I'll approve as is, and we'll add the frame to the Roots/Q_root description, which doesn't mention it yet.

[matt-pharr]

@d-burg sorry I messed up my worktrees and accidentally pushed something here, I force pushed back over it to undo it but please beware of if I messed anything up on this branch

[d-burg]

@matt-pharr Lab frame everywhere, as you suggested, in 546d9fbec:

• Uncoupled roots. Each layer is still scanned and solved in its own E×B frame. Its root is reported as Q + τ_k·n·Ω_E and ω + n·Ω_E, so γ is unchanged and the E×B-frame value is ω − n·Ω_E. Diagnostics/* (valid roots, poles, filtered roots) stay in each surface's E×B frame, and the docstrings and HDF5 descriptions now say which frame each output is in.
• No omega_E in the kinetic file. Ω_E is taken as 0, so lab and E×B frame coincide. PerSurface/omega_E stays NaN, as in your commit, and the tearing solve now warns that the roots assume Ω_E = 0.
• Harness (feynman, all 20 cases, develop 9a51446df vs 546d9fbec): diiid_slayer_n1 ω moves by n·Ω_E (2/1 −10,526 → +30,996, 3/1 −9,676 → +21,048, 4/1 −11,097 → −2,768 rad/s) with γ unchanged. The coupled case is as before (ω +30,993, γ −0.3 %), and the other 18 cases are unchanged. The uncoupled 2/1 and the coupled root now agree in the raw output, 30,996 against 30,993.
• New test. The uncoupled Q_root moves by exactly τ_k·n·Ω_E, while γ and the layer-frame extraction stay identical.

Release note and description are updated. Thanks for merging develop (cbfa58d5e); since that touches src/, I'm re-running the harness there and will update the stamp. Your accidental push didn't cost anything: the branch was back at 9e95d2395 when I built on it.