[Tarzelf]

Learned tool-pose targets need a geometric bridge that does not read hidden task state. This adds an isolated PiPER model that accepts six measured encoders and computes FK, a world-frame TCP Jacobian and one bounded differential-IK joint-target update. It explicitly distinguishes the existing base-frame link6-origin feature from the authored tool point 0.12003 m along link6.

The API takes no live environment, object or contact input and executes no dynamics or actuators. It preserves the original authored DLS settings and target bounds, refuses encoder overshoot and malformed inputs, and leaves existing learning/installer behavior unchanged. It is a prerequisite for future learned pose/progress experiments, not a trained skill or a collision planner.

Validation: ten static tests passed, including 17 independent XML FK checks, 90 finite-difference Jacobian columns and 34 comparisons against the unchanged original controller. The existing native-assets CI step now runs these tests. A portable static CPU probe measured 0.0305 ms median / 0.0324 ms p95 per joint-target calculation on the disclosed shared host; both the earlier prototype and portable execution receipts are retained. No task rollout, training, rendering or full-loop deadline claim is made.

Peer review covered all nine files. The reference-action research report explains the approach-precision and progress failures motivating this component. All learned qualification results remain unchanged. Stacked on #36.