feat(juice_shot): v3 — gate at visible disengagement, longer travel, easier twist
Operator feedback on the first 19 latched takes: SUCCESS fired at 5 mm with the cap
visibly still screwed on (the skirt bottom only clears the thread ridges' top at
+10 mm), and the 8 mm slide stop physically truncated the unscrew — the takes maxed
at 5.4-8.1 mm with two parked exactly on the stop (decoded from every take's
_mujoco_state buffers).
v3 changes (joints + doc header only; meshes and the screw equality untouched):
- cap_slide range 0..0.008 -> 0..0.014 (cap can rise fully above the 0.082 neck lip)
- cap_twist range 0..6.2056 -> 0..10.8598 (the 14 mm travel expressed in twist)
- cap_twist frictionloss 0.035 -> 0.030 Nm (breakaway headroom vs the operator's
measured >=0.047 Nm: 1.34x -> 1.57x) - shot_cap_col grip friction 3.0 -> 4.0 (ribbed-interlock knurl; the sanctioned
"easier to twist" lever — +33% deliverable torque at the 14.5 mm radius, thread
yank resistance untouched) - juice_shot.urdf joint limits mirrored (slide 0.014, twist 10.8598)
The task-side gate moves 5 -> 10 mm in the same change (WUJI_SimBench PR: v0.6.3
fix(hand_track)). Yank guard after the friction trim: straight-pull threshold ~23 N
(was 27), but reaching the NEW gate by yanking costs 23 N x 10 mm ~ 0.23 J vs 0.135 J
before — the cheat got harder. Verified: asset regression suite 7/7 (incl. a new
geometric gate lock measured from shot_thread.obj), end-to-end task checker latches
at 10.5 mm and NO LONGER at 5 mm, friction holds the cap at any written height.
NOTE for existing data: all 19 old completed takes stop short of the new gate (8 mm
stop era) — they no longer score as successes and need re-recording under v3.