Current finite nonempty square task-row/coordinate-column matrix.
Current finite task derivative, in the same row units/second.
Positive finite physical scale for each task row, unchanged during the call.
Fresh rates/diagnostics, or an explicit unsupported, pivot-refused or unrepresentable result.
Solve the current square differential kinematics J(q) qdot = goalRate. Rows and the matching goal rate use the same task frame and physical unit; each column uses its actual coordinate unit, so output is that unit/second. Positive rowScale divides both sides of each row, for example the current chain length for metre rows and one for radian rows. It changes neither the equality nor the coordinate rates and never weights a residual fit. Keep all supplied numeric rows and vectors unchanged until return.
The caller owns an actual analytic Jacobian and goal derivative on one smooth branch. Uncontrolled ancestor/root motion must be included in that Jacobian or subtracted from the goal rate in the same task frame. This numerical leaf neither samples a pose nor proves that a supplied matrix is its derivative. Active constraints, changing active sets, rectangular systems and unsupported charts need the caller's explicit derivative fallback; no clamped or damped answer is manufactured here.
The existing LU owner's pivot refusal is retained, not interpreted as a nonsingularity certificate. A solved result reports actual original-unit residuals, scaled backward errors, an inverse-based infinity condition estimate and the computed inverse defect. These are numerical diagnostics, not certified forward-error bounds or a physical acceptance tolerance.