[TomCC7]

VoxelMapTransformer and VoxelGridMapper now consume and emit generated PointCloud2 values. The voxel stores accept finite coordinate arrays, while the existing voxel math, legacy local add_frame/reprojection callers and batching behavior remain intact. Output CDR retains the last source Header's exact nanoseconds and uses the configured map frame.

Parent: #4672. Project: https://linear.app/dimensional/project/message-encoding-refactor-05a3ddb4af3e/overview

Category	Files	Change
Voxel stores	3	Numeric array input shared with existing voxel math
Stream boundary	1	Generated values and exact source Header preservation
Tests	1	CPU stores, batching, stamped regression and actual voxel-to-costmap CDR chain
CI	1	Installed-wheel provenance and focused checks

Manual test after installing this commit (CPU only):

import numpy as np
from dimos.mapping.voxels.module import VoxelMapTransformer
from dimos.memory.type.observation import Observation
from dimos.messages.sensor_msgs.msg import PointCloud2
from dimos.messages.std_msgs.msg import Header
from dimos.msgs.pointcloud import pointcloud_from_xyz, pointcloud_xyz
from dimos.msgs.time import time_from_nanoseconds, to_nanoseconds

stamp = 1700000000123456789
cloud = pointcloud_from_xyz(
    np.array([[0.025, 0.025, 0.025]], dtype=np.float32),
    header=Header(frame_id="world", stamp=time_from_nanoseconds(stamp)),
)
mapper = VoxelMapTransformer(device="CPU:0", show_startup_log=False)
result = list(mapper(iter([Observation(ts=42, data_type=PointCloud2, _data=cloud)])))
assert len(pointcloud_xyz(result[0].data)) == 1
assert to_nanoseconds(result[0].data.header.stamp) == stamp

Validation: 23 source and 23 installed-wheel tests pass, including real MemoryRPC/CdrTransport memq Module ports and VoxelGridMapper → CostMapper. Four production files pass strict typing; applicable hooks pass; normal main wheel builds; full default runtime environment passes pip check (198 packages). The block above runs against the installed main wheel. GPU/hardware/performance and GitHub CI results are not claimed. Benchmark work remains paused; Numba's existing TBB fallback warning remains visible.