FieldDriver derived the shield ellipsoid from its own transform.localScale;
ShieldManager derived it independently from the Shield prefab instance's
transform, doing the same normalise-and-project by hand against a different
object. Add ShieldEnvelope as the one reader of a transform's centre, radii,
surface projection, and true outward normal (derived, not authored, per R6 --
no second copy of numbers the physical MeshCollider already owns), and switch
both consumers onto it, deleting their hand-rolled projections.
Radii read the attached collider mesh's own bounds.extents per axis rather
than assuming a flat 0.5 or 1.0: Shield.prefab's Icosphere collider measures
(0.0095106, 0.01, 0.01), not a uniform round number, and the cubesphere's
collider measures exactly (1,1,1) -- confirmed by a throwaway batchmode probe
against the actual mesh assets, not guessed.
SurfaceNormal is not normalize(p): a non-uniform ellipsoid's true normal is
the gradient of the implicit surface, transformed by rotation and the INVERSE
of lossyScale (a normal is a covector). The first implementation used
TransformDirection (rotation and forward scale) and the batchmode probe's
finite-difference gradient check caught it immediately -- exactly the
plausible-looking-wrong-answer failure mode the cut map called out by name.
Both original consumers keep normalize(p)-equivalent behaviour via the new
SurfaceDirection method, since they only ever wanted a direction.
FieldDriver falls back to its pre-cut formulas (logged once) when no
ShieldEnvelope is present, since it has no guaranteed carrier yet.
ShieldManager does not: every object that carries it gets an envelope in this
cut, so a missing one logs an error and degrades to a fixed direction rather
than recomputing a second projection.
No prefab or scene file is touched here (Q5: the operator rigs prefabs and
scenes by hand); ShieldPanelCut2Verify's batchmode probe instead builds
synthetic transforms carrying Longinus's and Djinni's actual override values
plus the real Icosphere mesh, and proves ProjectToSurface lands exactly on the
ellipsoid and SurfaceNormal matches a finite-difference gradient to 1e-3.
mutation_tests_shield_panel_cut2.py mutates SurfaceNormal three ways (back to
normalize(p), back to TransformDirection, and a flattened 0.5 extents factor)
and confirms the probe goes red for each, naming the reason.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
