fix(entity): write root velocity via free-joint DoFs, not set_vel/set_ang

[KraHsu]

Summary

Fixes #242 — and the same dead idiom everywhere else it was copied.

Genesis's RigidEntity API is asymmetric: get_vel / get_ang exist, but set_vel / set_ang exist only on FEM / tool entities. Every write site built on getattr(handle, "set_vel", None) therefore silently no-opped on rigid articulations:

• mdp.push_by_setting_velocity — the entire event (the bug: push_by_setting_velocity 对 RigidEntity 静默失效（set_vel/set_ang 不存在） #242 report)
• mdp.reset_root_state_uniform — the velocity half
• ArticulationWriter.write_root_state — velocity halves (pose worked)
• examples/wuji reorient — cube reset velocity zeroing, the interval velocity kick (its docstring calls it "the important robustness lever"; it read current velocity via the existing getters, then wrote through the missing setters), and the goal-marker velocity zeroing
• examples/franka — cube reset velocity zeroing

Fix: new genelab.entity.write_root_velocity(handle, lin_vel_w, ang_vel_w, env_ids) routes the write through set_dofs_velocity on the base free joint's 6 DoFs (0–2 world-frame linear, 3–5 angular; free-joint discovery mirrors the binder's n_dofs >= 6 idiom), with a fallback to the direct setters for entity types that genuinely have them (FEM / tool entities, existing test fakes). All seven sites now call it. Fixed-base entities correctly write nothing (False return).

Impact note (in CHANGELOG): policies previously trained with push events effectively had those disturbances disabled; retraining may be needed where push robustness matters.

Test plan

• 11 new tests in tests/test_root_velocity.py: fakes mirror the Genesis 1.2 surface (deliberately without set_vel/set_ang) so a regression to the old idiom fails instead of no-opping — covers the helper, both event terms, the writer seam, fixed-base and fallback paths
• genesis_runtime-gated integration test: builds a real free box (RigidObjectCfg(fixed=False)), writes velocity, reads it back via get_vel/get_ang
• Existing test_articulation_state root-state round-trip still passes (its fake has 1-DoF joints only → takes the fallback path)
• Local: ruff check + ruff format --check clean; pyright adds no new errors over main; full suite 925 passed / 4 skipped (one pre-existing, environment-only Qt/plotter abort in full-process runs reproduces identically on clean main and is unrelated)

🤖 Generated with Claude Code

[KraHsu]

Follow-up: same-class audit (pushed as 10dad32)

The root cause of #242 isn't set_vel/set_ang specifically — it's the getattr(handle, "...", None) → silently skip idiom, which turns any upstream API mismatch into a silent no-op. Audited every getattr-guarded Genesis call in src/ + examples/ against the installed Genesis 1.2:

Verdict	Sites
✅ Exist upstream, working	all DR setters (set_dofs_kp/kv, set_dofs_frictionloss, set_dofs_armature, set_dofs_force_range, set_friction_ratio, set_mass_shift), sensor getters (get_contacts, get_links_net_contact_force, get_links_inertial_mass, get_dofs_force, get_dofs_control_force), RigidSolver.set_gravity, set_pos/set_quat
⚠️ Dead fallback arms, harmless	set_dofs_friction, set_dofs_position_target — only the second arm of or-chains whose first arm exists
❌ Live silent no-op, fixed	teleop HUD: outer Viewer forwards register_keybinds but not set_message_text → HUD text never displayed; now falls back to the inner _pyrender_viewer

Also added tests/test_genesis_api_contract.py: pins every getattr-guarded method name against the installed Genesis (import-only, runs on headless CI). A future Genesis bump that renames any of these seams now fails tests instead of silently disabling physics or UI.