feat: message-input point-lio rust module — WIP scaffold + raw sensor replay (#2821)

[kumarganduri]

Contribution path

• Linked issue or discussion: Rewrite point lio as Rust native module #2821 (DIM-1154)

Draft / work in progress. This PR implements option 1 from the issue
(message inputs), as proposed in
#2821 (comment) —
happy to change course if maintainers prefer another direction. The
estimator itself is intentionally not implemented yet (see "Remaining" section).

Problem

Point-LIO is welded to the Livox SDK: pointlio_native consumes raw network
packets, so it only works with one sensor family, and testing it without
hardware requires impersonating a lidar on the network (VirtualMid360:
sudo, loopback IP aliases, multicast routes). The issue asks for point cloud

• IMU message inputs instead.

Solution (so far)

Everything up to the estimator, wired end to end and testable from a file:

• pcap_to_raw_db.py (livox/scripts): offline converter, Mid-360 pcap ->
  memory2 db with lidar (PointCloud2, 10 Hz frames) + imu (Imu, 200 Hz)
  streams. Parses the Livox UDP payloads directly (layout from the public
  Livox-SDK2 header; conventions byte-for-byte with the live driver: mm->m,
  reflectivity/255, frame ts = first packet ts, accel g->m/s^2 by 9.80665).
  No SDK, no sudo, no network. Validated on the mid360_shake_stairs
  capture: 9.98 Hz frames at ~20k pts (spec: 200k pts/s), 200.0 Hz IMU, zero
  per-stream timestamp regressions, geometry verified visually.
• RawSensorReplay (livox/replay.py): publishes such a db as
  raw_lidar/imu topics on the shared replay clock — the message-boundary
  counterpart of VirtualMid360.
• PointLioRust (pointlio/rust): dimos-module crate, raw_lidar
  (PointCloud2) + imu (Imu) in -> lidar + odometry out. Output ports
  keep the existing PointLio names so downstream consumers wire identically.
  The estimator is an explicit slot (currently: passthrough clouds + identity
  odometry) so the wiring is provable before the math exists.
• pointlio-rust-replay blueprint: the whole chain from a file.

Verified live over zenoh: raw_lidar 10 Hz + imu 200 Hz in, lidar/
odometry out 1:1 with input frames, zero drops over 30 s of replay.

Remaining (in scope for this PR or follow-ups, per maintainer preference)

• The estimator. Blocked on direction (see the issue comment): the
  existing core (dimos-module-pointlio) is GPL-2.0, so a faithful Rust port
  can't live in this repo — options are an FFI wrap of the C++ core behind
  this module's message interface, or a port in a separate GPL repo consumed
  as a binary, or a clean-room implementation.
• Undistortion/TF details, tuning-parameter plumbing (the current config is
  the minimal frame-id pair), nix flake for the crate, trajectory-level
  validation harness against the C++ implementation.

Base-branch note

This branch is based on 29f35555 (mid-July) rather than current main:
current main cannot start any blueprint on macOS (first coordinator->worker
set_transport RPC times out — details and bisect in #3331). Will rebase
onto main once that is resolved; the diff here touches nothing near the RPC
layer. The conflict GitHub reports is the auto-generated
all_blueprints.py registry; it regenerates cleanly on rebase
(pytest dimos/robot/test_all_blueprints_generation.py).

How to Test

# one-time: fetch pcap + build the replay db (~3 min)
uv run python -m dimos.hardware.sensors.lidar.livox.scripts.pcap_to_raw_db \
    --pcap "$(python -c "from dimos.utils.data import get_data; print(get_data('mid360_shake_stairs/mid360_shake_stairs.pcap'))")"

# build + run the chain
(cd dimos/hardware/sensors/lidar/pointlio/rust && cargo build --release)
uv run dimos run pointlio-rust-replay

Expect lidar + odometry publishing at 10 Hz (identity poses; the rerun
view shows the replayed clouds). Replay-tested only; no hardware involved.

AI assistance

Done with the help of Claude Code (Fable 5)

Checklist

• This PR is scoped to one issue or clearly stated problem.
• I ran the relevant checks (pre-commit, cargo clippy/fmt, blueprint
  registry generation test, live replay verification) for the files I
  changed.
• I have reviewed and understood every line in this PR.
• I disclosed AI assistance above.
• I have read and approved the CLA.