Summary

• Twist3.Ad() was return self.SE3().Ad() -- builds a full, validated SE3 object purely to immediately extract its array inside .Ad() and discard the object.
• Computes the same result directly: smb.tr2adjoint(smb.trexp(self.S, check=False)). Same underlying math (trexp then tr2adjoint), just without the intermediate pose-object construction.
• No behavior change: bit-identical output for the general case, only the computation path is different.
• No existing test covered Ad() at all. Added one first, with hand-verified ground truth (not derived from Ad() itself) for pure-rotation (block-diagonal [[R,0],[0,R]]) and pure-translation (skew(t) coupling) cases, plus a cross-check against SE3.Ad() over 20 random transforms -- confirmed it passes unchanged against both the old and new implementation.

Test plan

• New test_Ad passes against the old implementation (self.SE3().Ad()) -- establishes the correctness baseline
• Passes unchanged after the optimization -- confirms behavior-preserving
• pytest tests/ -- full suite, 343 passed, no regressions
• Benchmarked old vs new in the same process: 35.7us vs 25.6us per call, 1.40x faster, results bit-identical (np.testing.assert_almost_equal)