three_parallel (UR arms) returns a spurious >1e-7 branch at near-singular poses

[siddhss5]

Repro (deterministic, on main)

```python
import numpy as np
from ssik.prebuilt import ur16e_ik as U
q = np.array([0.0, -2.4834242925436327, -2.446578501260409, -2.6545207067859944, -2.5361296670486584, 0.0])
T = U.fk(q)
sols = U.solve(T, respect_limits=False)

7 sols, worst FK = 7.44e-06 (> the 1e-7 fuzz ceiling; allow_refinement doesn't help)

```

At this near-singular pose (q0 = q5 = 0 → wrist/shoulder alignment) the ikgeo.three_parallel solver returns a branch that FK-closes to only ~7e-6 — above its own FK tolerance, so it should have been filtered. max_fk over the returned set exceeds the arm's 1e-7 ceiling.

Impact

• test_prebuilt_uniform_fuzz.py::test_prebuilt_6r_random_q_roundtrip[ur16e_ik] can fail when Hypothesis samples near this pose (found during an unrelated full-suite run; ur16e artifact unchanged). Latent flake — CI hasn't reliably hit it.
• Pre-existing; independent of Generalize seven_r.srs past ZYZ: arbitrary-axis spherical triples (Davenport) — unlocks Galaxea R1 Pro + every non-canonical SRS arm #354/Generalize native SRS solver to any concurrent-axis 7R (#354) #356/Onboard Galaxea R1 Pro (7-DOF) as a prebuilt #357/Bulletproof: wire the T-perturbation rescue into the thin-wrapper (SRS-family) arms #358. Likely affects other UR (three_parallel) arms at the same singularity.

Directions

• The FK-tolerance gate in three_parallel should drop the spurious branch (or the non_singular_q6r fuzz strategy should exclude q0=q5=0-type alignments).
• Investigate why LM refine (allow_refinement=True) can't pull the 7e-6 branch in.