OmniSim v9.2.6

Light grippers now hold a part by friction, a URDF robot's own root collider is in the solver by
default, and the OmniLink bridges never start a turn without a fresh prompt. Found running robot
makers' published models (Robonine SO-ARM 102, Preferred Robotics Kachaka, WIRobotics ALLEX, MAB
HB50W). Each fix has a pin test and a value-parsed hatch; the engine, runtime and bridges all change,
so use the v9.2.6 installer. Arm grasps and multi-robot contact change: read the behaviour changes.

⚠️ Behaviour changes — read before upgrading

• Light grippers hold a part by friction: fixed-base position servos carry the rotor inertia their
  damping implies. Every position servo is built ke = effort*10, kd = effort*0.5 with no armature,
  whatever the joint weighs. MuJoCo's implicit integrator applies the damping as M + dt*kd, but the
  constraint solve (contacts, joint limits, a URDF <mimic> coupling) sizes its force against the bare
  M, so on a joint with a large dt*kd/M every constraint acted at M/(M+dt*kd) strength. Measured on
  Robonine's SO-ARM 102 (7 g jaws, dt*kd/M 13.6; wrist 857), OmniSim's own exported model replayed in
  plain MuJoCo 3.11: the jaws closed 21 mm into a 40 mm cube, ran past their own limits (the mimic
  follower 43 mm past with no part in the way) and the cube squirted out at lift-off at every jaw force
  from 3 to 100 N, at 1 and 2 ms. Now every fixed-base position servo over dt*kd/M = 1 gets exactly
  dt*kd - M of armature (at most a 2× change in the inertia the integrator already applied; a mimic
  follower gets its leader's), and MuJoCo's constraint weights for those joints are refreshed. Measured
  after, in-engine through the gated /prompt ("Pick up the cube, then go home."): the SO-ARM 102 carries
  the 40 mm cube 17.8 cm to home, 3.8 mm from the tool point, friction only — also at 1 ms, at 3 N, with
  a larger cube and with two driven jaws; with the hatch off it drops it, as before. Elsewhere: ladder0
  rung 8 with newtonContactKe 4000 / 6000 / 8000 now carries (carry_rel 2.0 mm, was ejected at up to
  2.33 m/s — the "ke trap" in the friction-grasp guide was this dilution); friction_grasp_minimal now
  carries the part between its pads (it rose 0.2025 m for 0.2023 m of wrist travel, where before it was
  squeezed out onto the wrist plate and read +0.2817 — the pin tests/test_friction_grasp_holds.py now
  asserts the band around the wrist's travel); omniarm6_real_pick_place drift 4.8 → 0.7 mm; the TRLC
  DK1 pilot still lifts its cube 21.1 cm; the UR5e chat arm's wave and home and the 2F-140 chat pick are
  unchanged; SO-101 pick-place still succeeds and ALOHA's battery is no longer knocked 27 cm. ⚠
  omniarm6_physics_pick_place still passes but its place distance went from 12 to 92 mm (limit 100):
  the demo releases the cube ~25 cm above the place height (a separate, pre-existing bug in that demo)
  and the dropped cube now slides further. A gain written after the first step (setControlPID, servo
  promotion) re-judges that joint and its mimic followers. Free-floating robots (legged, humanoid, mobile) are NOT changed by default — their gaits and
  policies were tuned on the old model, although they carry the same dilution (Go2 73, B2 74, a G1 wrist
  723): OMNISIM_NEWTON_SERVO_ARMATURE_SCOPE=all covers them. Skipped when an RL recipe sets
  OMNISIM_NEWTON_TARGET_KD / _FINGER_KD / _ARM_KD. Hatch: OMNISIM_NEWTON_SERVO_ARMATURE_BOUND=0
  (a number sets the bound). Pins: tests/test_newton_servo_armature_bound.py (9 engine-free tests).
• An OmniLink bridge never starts a turn without a fresh operator prompt. Observed on WIRobotics
  ALLEX after an engine restart: the relay was woken by a joint.limit_hit at startup (the relaxed
  fingers rest on their stops), short-term memory had restored the previous session's sentence, and the
  arm re-ran that whole task before anyone typed. Now event-triggered turns are OFF by default
  (OMNILINK_EVENT_WAKE=1 opts in, value-parsed); even opted in, no wake fires before this session's
  first operator turn, startup / settle events (first 5 s of sim or wall time) never wake, and restored
  history carries a session-boundary note so nothing from a previous session is resumed unless asked
  again. PROTOCOL.md wake section. Pins: tests/test_no_unprompted_turns.py, test_event_wake.py.
• A Robot's own root collider registers by default unless it would lift the robot off its wheels.
  A Robot wrapper with colliding links below it (every URDF mobile base, quadruped and humanoid) used to
  get a 1 mm placeholder instead of its own boundingObject unless WorldInfo.newtonRobotColliders TRUE,
  so a URDF robot's root-link <collision> did not exist in the solver and the log told the author to
  "declare a boundingObject" it already declared. Measured on Preferred Robotics' Kachaka (chassis box
  underside exactly on the tyre contact plane, no caster): by default the published file tipped onto its
  nose (pitch 79°) where plain MuJoCo drags the box. Now the root shapes register and stay colliding
  unless their lowest point is more than 2 mm below the level the robot stands on at the spawn pose
  (OMNISIM_NEWTON_ROBOT_ROOT_TOL), the one envelope that really holds every wheel off the floor. That
  level is the lowest one shared by at least two branches below the root (each child link and its
  subtree: a wheel, a leg, a caster), so an arm hanging below the tyres in its zero pose does not set it;
  the load log reports each verdict and the gap. Kachaka by default now drags at 0.464 of the commanded
  0.26 m/s (identical to the TRUE run), turns at 0.182 of 1 rad/s, and with the box raised 5 mm plus two
  casters drives at 0.998 / 0.983. Blast radius, open-loop drive A/B against the hatch (straight and pivot,
  speed / yaw-rate ratio, rest height, min height, max pitch / roll): Husky, Jackal, ROSbot, ROSbot XL,
  TurtleBot3 Burger and Waffle identical to 5 digits in all 12 cases — their chassis colliders now
  register but sit 8.5-132 mm above the tyre plane (Burger 8.5, Waffle 10, ROSbot 22.5, XL 30.6, Jackal
  63.5, Husky 132), so on flat ground they touch nothing; they will now hit an obstacle or a step with
  the body as the file draws it. Full census (all 219 catalogued worlds, hatch A/B, bitwise-deterministic
  replays): multi-robot worlds where robots meet (husky_fleet_arena, agent_fleet_collision) now
  collide chassis to chassis and diverge after the first contact; newton_smoke_test's robot no longer
  sinks through the floor (lowest z -0.154 m before, 0.006 m now); the 8-Husky swarm, lidar-walls,
  Jackal / TurtleBot3 drives and the X30 trot are identical. The warning for a wrapper that is left on the placeholder now says it
  declares a boundingObject and why it is not registered. newtonRobotColliders TRUE still registers it
  unconditionally. Hatch: OMNISIM_NEWTON_ROBOT_ROOT_COLLIDER=0 (value-parsed) restores the old skip.
  Pins: tests/test_newton_robot_root_collider.py (12 engine-free tests), tests/test_newton_urdf_fixed_links.py.

Fixed — OmniLink bridges

• A /prompt turn's budget honours timeout_s and counts simulated time. The arm bridge called the
  relay with a fixed 90 s wall budget and ignored the body; on WIRobotics ALLEX at 0.09-0.18x real time a
  correct three-hover turn was cancelled after two hovers (timeout_after_actions, no reply, no home).
  All four bridges (arm, mobile, quadruped, drone) now parse timeout_s (default 90, at most 600) and an
  optional wall_timeout_s with one helper, and the turn is cancelled only once both the sim and the wall
  clock have run timeout_s; wall_timeout_s (default 600) caps the wall clock alone for a sim that stops
  stepping. A cancelled turn says which clock ran out (budget.expired_on). PROTOCOL.md §5.7. Pins:
  packages/omnisim-bridges/tests/test_prompt_budget.py.
• set_tcp_target no longer reports a slow world as stalled. The arm's wait called the world stalled
  after 2 s of wall silence from the sim thread; one ~2.4 s engine tick on ALLEX ended a good hover as
  stalled, achieved: None, settled: False, and the arm had reached the point (1.4 mm). The threshold is
  10 s on every bridge (bridge_base.STALL_GRACE_S), and a wait that does give up now returns the joints'
  last measured reading (last_reading, with the tool point and its error for a TCP verb) while achieved
  stays null. The audit line names stalled=True. Pins: packages/omnisim-bridges/tests/test_arm_slow_world_wait.py.
• The mobile bridge faults a tipped base. It measured only a planar pose, so the MAB HB50W rolled to a
  48° tilt and every drive, turn and stop still reported success. It now measures the base tilt every tick
  (get_robot_state.tilt_deg); above 35° it declares fault: {code: "base_tilt"}, stops the wheels, ends
  the motion in flight with settled: false and the fault attached, refuses drive/turn until the base is
  upright again, and the parser's reply says the robot tipped instead of "Drove +1.98 m". Limit:
  OMNISIM_MOBILE_TILT_LIMIT_DEG or a robot config's tilt_limit_deg; hatch OMNISIM_MOBILE_TILT_GUARD=0
  (value-parsed). PROTOCOL.md §11. Pins: packages/omnisim-bridges/tests/test_mobile_tilt_guard.py.

Full Changelog: v9.2.5...v9.2.6