[TomCC7]

Offline relocalization replay reads canonical PointCloud2/TFMessage recordings and .pc2.cdr premaps, then writes a canonical placed loaded_map once. Existing attempt spacing, sparse-frame filtering, first accepted fix, post-fix recording window and numeric error calculations are retained. Explicit geometry helpers feed the existing native RayTrace/Open3D algorithms. Viewer palette registration, fine premap size and fixed probe colors live outside generated values.

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

Parent: #4694. Tuning utilities receive encoding/geometry/display adapters only; no tuning study, search or benchmark was run or adjusted. The guide explicitly requires converted or regenerated canonical inputs. Renamed .pc2.cdr fixture paths are not a claim that historical LFS files have already been converted.

File category	Files	Review focus
Runtime/adapters	5	Replay and placed-map persistence; utility message boundaries; fine renderer and shared viewer helpers
Tests	2	Original replay interval/stop/first-fix assertions retained; actual native RayTrace; source ownership, display colors and palette IDs
Documentation	1	CDR premap reproduction and explicit historical-format break
CI	1	Installed production provenance, the same focused tests and five strict files

Validation: 23 focused source and the same 23 installed-wheel checks pass; production modules and the optional RayTrace extension asserted to come from site-packages. Five production files pass scoped strict mypy; applicable hooks pass; normal main wheel builds. SHA256: e5b20c5be1818a4e53f388390738675b33e381a11c494a99cdc7c501974c50a0.

Manual offline regression review requires the optional RayTrace Python math extension built in #4688 (it is not included as a precompiled native module in the main wheel):

python -c 'import dimos_voxel_ray_tracing'
OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 python -m pytest \
  --noconftest --import-mode=importlib -o addopts='' -q \
  dimos/mapping/relocalization/lidar/test_replay.py \
  dimos/mapping/cli/test_generated_cli.py \
  dimos/visualization/rerun/test_cdr.py::test_cloud_renderer_filters_nonfinite_positions_with_matching_colors \
  dimos/visualization/rerun/test_cdr.py::test_fixed_display_color_overrides_sensor_rgb_after_invalid_points_filter \
  dimos/visualization/rerun/test_cdr.py::test_height_palette_uses_all_class_ids_at_the_recording_root

The SQLite replay regression uses real generated cloud/TF payloads and the actual RayTrace math extension. The matcher decision is mocked at the external registration boundary so the test deterministically verifies refusal, retry, acceptance and stopping rather than running a tuning study. Placed-map tests check timestamp/frame/geometry, original input bytes and the existing no-overwrite behavior. No GPU/hardware, performance comparison, CI polling or human visual acceptance claim. Known GTSAM optional metadata mismatch, Nix and external gates remain separate; default cutover and legacy retirement are still incomplete.