Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@newton/_src/solvers/kamino/examples/rl/example_rl_bipedal.py`:
- Around line 169-171: The _make_balls_fn function currently slices
BALL_POSITIONS with num_balls without validation leading to negative or
too-large inputs producing unexpected results; before using num_balls in the
slice, validate and clamp it to a safe range (e.g., clamp num_balls = max(0,
min(num_balls, min(5, len(BALL_POSITIONS)))) or alternatively raise a ValueError
for out-of-range values), and update any doc/help text if you change behavior so
callers know the enforced max; refer to the _make_balls_fn function and the
BALL_POSITIONS constant when making the change.

In `@newton/_src/solvers/kamino/examples/rl/simulation.py`:
- Around line 82-89: The single-world path can leave negative shape_world (wid
== -1) entries in Newton contact data which later feed
convert_contacts_newton_to_kamino; to fix, when use_newton_collisions is true
and newton_model.world_count == 1, sanitize the Newton contact/shape_world data
by replacing negative shape_world/wid values with 0 on the model/state/contacts
you store (referencing newton_model, self._newton_state, self._newton_contacts
and the converter convert_contacts_newton_to_kamino) before any conversion or
use — i.e., after assigning self._newton_state/self._newton_contacts (and before
creating ContactsKamino or calling convert_contacts_newton_to_kamino) iterate
the contact shape_world/wid entries and set negative values to 0.
- Around line 82-89: The Newton-CD branch currently sizes ContactsKamino using
newton_model.rigid_contact_max and ignores the configured cap, so update the
allocation logic in the use_newton_collisions branch (where self._newton_model,
self._newton_state, self._newton_contacts are set and per_world/world_max are
computed) to clamp per_world by the configured max_contacts_per_world (e.g.,
config.collision_detector.max_contacts_per_world or the corresponding variable
passed into the module) before building world_max and calling
ContactsKamino(capacity=world_max, ...); ensure the clamp applies per-world (use
min(rigid_contact_max-derived value, max_contacts_per_world)) so the configured
cap is honored for heightfield/mesh scenes that force
use_newton_collisions=True.

---

Outside diff comments:
In `@newton/_src/solvers/kamino/_src/core/shapes.py`:
- Around line 626-699: The function that returns contact capacity for (type_a,
type_b) can return (0,0) for primitive vs HFIELD because only the final HFIELD
and mesh branches handle HFIELD; to fix, ensure any branch handling a primitive
also checks for type_b == GeoType.HFIELD and returns the conservative capacity
(e.g. _MESH_CONVEX_MAX, 0), or perform a canonical swap early so GeoType.HFIELD
is always handled by the dedicated HFIELD branch; update the primitive branches
(cases for GeoType.SPHERE/CAPSULE/ELLIPSOID/CYLINDER/BOX and the mesh/convex
branch) to explicitly handle type_b == GeoType.HFIELD (returning
_MESH_CONVEX_MAX, 0) so ModelKamino.from_newton and conversions.py sizing no
longer under-allocates.

---

Nitpick comments:
In `@newton/_src/solvers/kamino/_src/core/model.py`:
- Around line 1324-1331: No functional change required because Model.finalize()
populates model._shape_voxel_resolution before Kamino.from_newton() consumes it;
if you want a safety improvement, either add a short assertion at the start of
Kamino.from_newton() verifying model._shape_voxel_resolution is set and
device-backed (to fail fast with a clear message), or add a public accessor
property on Model (e.g., Model.shape_voxel_resolution) that returns the wp.array
and use that in from_newton() instead of the private attribute.

Verify each finding against the current code and only fix it if needed.

Outside diff comments:
In `@newton/_src/solvers/kamino/examples/rl/simulation.py`:
- Around line 862-863: The docstring for the q_j_ref property incorrectly says
"num_joint_coords" while the stored shape is (nw, njd) in DOF space; update the
q_j_ref docstring to state "(num_worlds, num_joint_dofs)" (or use
"num_joint_dofs" consistently) to match the actual reshape and the attribute
_q_j_ref so callers understand this is DOF-space joint position references.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@newton/_src/solvers/kamino/_src/geometry/unified.py`:
- Around line 524-527: The current detection packs heightfields with meshes into
_has_explicit, causing NarrowPhase to be created with has_meshes=True even when
only heightfields exist; change the detection to compute two booleans: one (e.g.
_has_meshes) that checks for GeoType.MESH and GeoType.CONVEX_MESH, and another
(e.g. _has_heightfields) that checks for GeoType.HFIELD, then pass both
explicitly to NarrowPhase as has_meshes=_has_meshes and
has_heightfields=_has_heightfields so the mesh-mesh contact kernel is
enabled/disabled correctly.