Summary

• trlog's general-case branch (SO(3)) divides by sin(theta), computed from trace(R) via acos. The only existing coverage (test_trlog) used R = np.eye(3) exactly — never a numerically near-identity matrix as produced by real computation.
• That gap is exactly what Fix trlog exceptional condition #63 reported: a near-identity R reaching the general case and dividing by (effectively) zero.
• Confirmed against the pre-a9fc08a code ("rework code to be more robust to nearly identity rotation matrix") that rotx(1e-9) — not caught by iseye(), and not exactly np.eye(3) — raises FloatingPointError: divide by zero encountered in divide there under strict numpy error settings. Current code's st == 0 guard (added in a9fc08a) plus the acos domain clamp (added in 5d1044a) handle it cleanly.
• Also adds a case at theta=1e-7 (small but not clamped) to confirm the general-case formula stays numerically accurate rather than needing a coarse fuzzy-tolerance fallback.

Test plan

• pytest tests/base/test_transforms3d.py -k trlog -v — both tests pass
• pytest tests/base/test_transforms3d.py — full file, 30 passed
• Verified the new case reproduces a real FloatingPointError against the pre-a9fc08a code path