Error model: harden the fit against pathological inputs (code review)

Addresses code-review findings on RefineErrorModel:
- Floor the 1/dev^2 bin weight relative to the data scale (1e-3 of the median
  bin dev^2), not an absolute 1e-30: a near-zero-scatter bin could otherwise
  acquire a runaway weight and hijack the global (a,b) fit.
- Reject a near-collinear normal-equation system relatively (det > 1e-10*Ass*AII)
  instead of with an absolute threshold that an ill-conditioned fit can pass.
- Reset the model to identity at entry so any early return leaves it inactive
  rather than keeping a stale a/b alongside a freshly-cleared mean map (which
  would make CorrectedSigma fall back to the per-observation I).
- PixelRefine: correct the orient_prior comment - with the sweep on, the LSQ
  anchor is the swept orientation (intended), not the spot-centroid one.

Verified unchanged on the lyso test set (ISa 1.1, CC1/2 90.3%).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

image_analysis/pixel_refinement/PixelRefine.cpp

@@ -920,8 +920,11 @@ void PixelRefine::Run(const T *image,
         BuildParameterBlocks(data, beam, dist_mm, detector_rot,
                              latt_vec0, latt_vec1, latt_vec2);
 
-        // Anchor for orientation regularization = the spot-centroid orientation we
-        // started from (captured before any pixel-level refinement moved it).
+        // Anchor for orientation regularization = the orientation the LSQ starts from
+        // (captured before the predict<->refine iterations move it). When the global
+        // sweep ran first this is the swept orientation, not the original spot-centroid
+        // one - which is intended: the regularizer keeps the LSQ near its own starting
+        // point, it is not meant to pull a deliberate sweep back.
         if (iter == 0)
             for (int i = 0; i < 3; ++i)
                 orient_prior[i] = latt_vec0[i];