[ZJLi2013]

Stacked on #35 (and #33, #34 below it); merge those first.

What

• Geometric candidates list every grasp in the base wrist orientation, then the same list with the hand turned 180° about the approach axis. The hand is symmetric, but the two orientations give IK different joint branches. With first-success selection, the base list is tried first, so results where the base orientation already works are unchanged.
• GripperSpec.usable_width replaces "full opening minus 1 cm" and defaults to the measured 7.5 cm.
• Pick trial records carry the candidate's orientation.

Results (empty table, MI300X)

• Usable width: boxes 3–7.5 cm wide, 5 cm tall, all lift top-down. In feat(grasp): geometric candidates from the object's oriented box, chosen by trying them in simulation #35, the wide candidates that failed were all on round or cylindrical objects, so the opening was not the cause.
• The 180° orientation does not avoid the IK discontinuity: apple_02 at (0.42, +0.27) fails motion generation for all 16 candidates, in both orientations. die_01, mug_02 and block_01 lift at both positions, mostly on the first candidate.
• apple_02 at (0.42, -0.27): a candidate lifted in trial but the recorded episode failed; not yet explained.

Test

• New: the flipped list repeats the base list with the finger axis reversed; usable_width bounds which axes get candidates.
• pytest: 182 passed.

AI disclosure

Written with Cursor (Claude) as the coding agent: the change, tests, the remote runs and this description. The human reviews before merging.