Add dynamic SE(3) kinematic consistency and correction

[Peter-cuhk]

What changed

• add full SE(3) FK-versus-logged-EEF residual evaluation with Pinocchio
• dynamically discover URDF chains, joint/state mappings, EEF pose fields, and rotation conventions
• fit reviewable base, TCP, joint-sign, and joint-zero correction proposals with held-out metrics and safe refusal states
• add proposal persistence, UI review/acceptance, filter previews, corrected exports, and provenance manifests
• share the same engine across the Qwen filtering CLI, API, and align-robot-kinematics Codex skill

Why

Qwen-RobotManip-style kinematic curation requires reconciling heterogeneous robot datasets with their URDFs without hard-coding vendor link names or silently rewriting source observations.

Impact

Corrections remain reviewable metadata until accepted. Accepted profiles affect derived previews and exports only, with source data preserved and correction evidence recorded in conversion reports.

Validation

• 49 focused Python tests passed
• Ruff checks passed
• 67 frontend tests and production build passed
• skill validator and API/skill parity tests passed
• real Astribot S1 read-only validation reduced episode-held-out position median from 97.0 mm to 2.76 mm and P95 from 100.3 mm to 11.1 mm