* claim: #7635
* Keep the drag cursor on the manifold of an ill-conditioned sketch (#7635)
A construction line collapsed to zero length makes its angle row's entries
on that line ~1e7 while the same row's entries on the measured line are ~1.
Dragging an endpoint of the midpoint-pinned line then failed two ways: the
null-space rank analysis, which scaled rows only, hid the small (physically
meaningful) entries below its pivot tolerance and let the measured line
rotate freely; and the cursor's refinement, which demanded a TOL-tight
correction the ill-conditioned system cannot deliver, gave up entirely and
let the drag leave an unsolvable step placed - breaking the perpendicular
and angle constraints.
Equilibrate the null-space Jacobian on both sides (columns by their largest
entry, then rows) so one row's entries share a scale, and let the cursor
keep working from a pose the solver itself calls solved (residual below
ACCEPT) when refinement stalls above TOL, accepting corrected trials at the
same tolerance the final validation uses.
---------
Co-authored-by: Dahak Builds <dahakbuilds@users.noreply.github.com>