[TomCC7]

Collection's episode monitor and recorder now agree with WebXR/coordinator generated CDR ports: UInt32 buttons, String status, Image and JointState. The existing EpisodeStatus JSON document and state machine stay intact; the document's timestamp still defines episode boundaries. DataPrep projects canonical image row layout and named joint vectors, retaining sparse-command reconstruction and existing dataset selection behavior.

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

Parent: #4716 (a9b0d06211aae2d47e7b140834e6eccc8ef0cd61). This is one verified graph in an incomplete proposal. Isolated LeRobot host/worker ports, other remaining graphs, final dependency retirement and stack organization follow.

File change summary

Category	Files	Added	Deleted	Review focus
Collection/DataPrep runtime	4	84	50	Canonical ports/default CDR SQLite codec; padded image views; named joint reconstruction; explicit historical MCAP refusal
Behavior/regression tests	5	268	121	Actual local CDR monitor/recorder/reopen; canonical synthetic fixtures with retained assertions; real HDF5/LeRobot exports
CI and existing documentation	2	27	0	Installed-package acceptance and embedded offline test command

MCAP dataset fixtures now contain complete ros2msg schemas with plain CDR values and standard NONE/ZSTD chunk compression. DataPrep no longer imports historical Python message classes from recording metadata to load LCM codecs. Historical typed LCM/JPEG-in-LCM MCAP inputs fail with an explicit conversion instruction. Unrelated generic object/telemetry formats are outside this change.

Verification

• 152 source tests and the same 152 tests against production modules from a wheel installed in a separate venv, with python -I and site-packages assertions.
• Four production files pass strict typing. Scoped configured hooks pass except Nix lint, which is unavailable because nix is absent.
• Wheel SHA-256: af1f4a11f6e31d16db5f1533854f43593d059e7ae6c78f343f0e2717ae781fcb.
• Actual memq and discovery-disabled loopback Zenoh UInt32 input -> EpisodeMonitor -> CDR String -> CollectionRecorder -> SQLite close/reopen -> episode extraction and feature projection.
• Exact generated Header nanoseconds, padded RGB pixels, joint-name ordering, document/source-vs-reception episode time, original sparse-command/episode filtering assertions, HDF5 arrays and actual LeRobot video/Parquet/metadata assertions.
• Existing optional-MCAP absence test retains its assertions; its subprocess now uses isolated Python so checkout files cannot shadow an installed package.

Manual offline test

With the development and learning test dependencies installed:

DIMOS_TRANSPORT=lcm LCM_DEFAULT_URL=memq:// uv run pytest -o addopts='' --import-mode=importlib -q \
  dimos/imitation/collection dimos/imitation/dataprep \
  dimos/imitation/test_datacollection_e2e.py

No checkpoint/model, inference, training, camera, robot or external provider was used. This verifies the dataset format writer, not a full LeRobot policy runtime. No CI conclusion or human acceptance is claimed; CI polling and benchmark investigation remain paused.