Two related fixes flagged by an audit of vm_vector_2_matrix_norm callers:
ParticleEffectAction: locals.direction defaulted to vmd_zero_vector,
which would trip vm_vector_2_matrix_norm's unit-length assertion (and
divide by zero in release) when an action program emitted a particle
effect without first running a Set Direction action. Default to
vmd_z_vector so the unset-direction path always feeds a unit vector.
matrix_h / Lua orientation type: the element indexer (m[1..9] = x)
wrote arbitrary floats straight into mtx.a1d[idx] and the resulting
non-orthonormal matrix could flow into obj->orient via the Orientation
virt-vars on object/modelinstance/subsystem, breaking downstream engine
code that assumes the basis vectors are unit length. getInterpolated
had the same problem by design (pure linear interp, no normalization).
Adds a fourth MatrixState, NeedsOrthonormalize, which the indexer
setter and getInterpolated use to flag their output. ValidateMatrix
and ValidateAngles now call vm_orthogonalize_matrix when they observe
that state, so any consumer that reads through GetMatrix/GetAngles
gets a valid orientation without each call site needing its own guard.
getInterpolated is also marked ADE_FUNC_DEPRECATED at FSO 26.0 in
favor of rotationalInterpolate; mods targeting older versions keep
the legacy linear blend (now safely orthonormalized on use).
Goober5000