defect: pass the pipeline distortion coefficients to the forward model

build_forward_model negated the amplitudes of the converted distortion
coefficients before handing them to midas_diffract's HEDMForwardModel.
That was a first-order inverse written for a forward model that applied
the pipeline's correction (raw -> ideal) in the wrong direction. The
midas_diffract commit "FF tilt convention and forward distortion
direction" makes the model invert the correction itself, with a fixed
point, so the compensation now double-inverts: with a tilted and
distorted geometry (p0 = 1e-4) the forward model and the reference
predictor raster.predict_reflections disagreed by 0.078 px, against the
1e-2 px the test requires. CI caught it on the first run after that
change: test_forward_model_matches_reference_predictor for the
"tilted+distortion" and "tilted+distortion+wedge" cases.

The coefficients are now passed as md.v1_to_v2_coeffs returns them. This
also removes the O(delta^2) residual the negation left (about 2e-3 px
for realistic coefficients). The two predictors are independent
implementations (the reference path uses midas_transforms
apply_tilt_distortion), so their agreement without the negation is a
second check that the midas_diffract direction is right.

Needs midas-diffract >= 0.10.0: the floor is raised from 0.9.0 and
midas-defect goes from 0.9.0 to 0.9.1. With an older midas_diffract the
distortion would now be applied in the wrong direction.

Behaviour change: spot positions predicted by build_forward_model for a
geometry with distortion move by up to the size of the distortion
(about 0.08 px for the test geometry), toward the reference predictor.
Index-from-cloud results that used distortion on a calibrated geometry
should be re-checked.

Tests: tests/test_indexing_forward_conventions.py passes for all four
cases (flat, tilted, tilted+distortion, tilted+distortion+wedge).

Author: marinerhemant

packages/midas_defect/midas_defect/__init__.py

@@ -13,4 +13,4 @@
 # time (before either runtime initializes) so callers don't hit a hard crash.
 os.environ.setdefault("KMP_DUPLICATE_LIB_OK", "TRUE")
 
-__version__ = "0.9.0"
+__version__ = "0.9.1"

packages/midas_defect/midas_defect/indexing.py

@@ -113,19 +113,18 @@ def build_forward_model(crystal, geom: Geometry, *,
     #   multi_mode="panel"; the default "layered" mode applies the NF ray-plane intersection,
     #   which put spots a median 7.6 px off at tilts (0.3, 0.15, 0.37) deg.
     # * distortion: the 15-term p_coeffs were not passed at all. They are LEGACY paramstest
-    #   p0..p14 applied raw -> ideal by midas_transforms; HEDMForwardModel applies the
-    #   midas_distortion v2 layout ideal -> raw. The bridge is v1_to_v2_coeffs (a reindex) with
-    #   the AMPLITUDES negated and the phases kept: the first-order inverse of D = 1 + delta,
-    #   leaving an O(delta^2) residual (~2e-3 px for realistic coefficients).
+    #   p0..p14 applied raw -> ideal by midas_transforms. HEDMForwardModel (midas_diffract >= 0.10.0)
+    #   takes those same coefficients (v1_to_v2_coeffs is a reindex) and inverts the correction itself
+    #   (fixed point), so they are passed as they are. Before 0.10.0 the model applied the correction
+    #   in the wrong direction and this code compensated with negated AMPLITUDES (a first-order
+    #   inverse with an O(delta^2) residual); with 0.10.0 that compensation would double-invert.
     # With all three, the two predictors agree to <1e-2 px (tests/test_indexing_forward_conventions.py).
     omega_edge = geom.omega_first_deg - 0.5 * geom.omega_step_deg
     has_dist = any(abs(float(p)) > 0 for p in geom.p_coeffs)
     p_ideal_to_raw = None
     if has_dist:
         import midas_distortion as md
-        v2 = np.asarray(md.v1_to_v2_coeffs(np.asarray(geom.p_coeffs, float)), float)
-        amp = np.array([not n.startswith("phi") for n in md.P_COEF_NAMES])
-        p_ideal_to_raw = np.where(amp, -v2, v2).tolist()
+        p_ideal_to_raw = np.asarray(md.v1_to_v2_coeffs(np.asarray(geom.p_coeffs, float)), float).tolist()
     hgeom = HEDMGeometry(
         Lsd=geom.lsd_um, y_BC=geom.bcy_px, z_BC=geom.bcz_px, px=geom.px_um,
         omega_start=omega_edge, omega_step=geom.omega_step_deg,

packages/midas_defect/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "midas-defect"
-version = "0.9.0"
+version = "0.9.1"
 description = "Differentiable, phase-agnostic diffuse-scattering defect metrology on top of standard FF-HEDM: intensity-budget closure, per-grain dislocation density, fault-rod and selection-rule tests."
 readme = "README.md"
 license = "BSD-3-Clause"
@@ -41,7 +41,7 @@ dependencies = [
                                  # midas_hkls.asymmetry_sign_test, added there;
                                  # >=0.11 for centring_allowed, which rows.py now delegates to
     "midas-stress>=0.1",         # quat / rotation / misorientation primitives
-    "midas-diffract>=0.9.0",     # HEDMForwardModel for predicted spot positions; >=0.2.0 for the
+    "midas-diffract>=0.10.0",     # HEDMForwardModel for predicted spot positions; >=0.2.0 for the
                                  # distortion layer and FF-panel tilts build_forward_model now uses
     "midas-distortion>=0.2.0",   # v1 -> v2 coefficient bridge in indexing.build_forward_model
     "midas-index>=0.1",          # initial-pass indexer for seed_index.py