Online position velocity and attitude evaluation

The deliverable is a hardware-free, reproducible public-log workflow for position, velocity and attitude quality, including scenarios, improvement decisions and runnable Docker/binary distribution artifacts.

The baseline is the received-event implementation merged in PR #559. Additive attitude/velocity export must preserve its position, status, queue and reset behavior. The existing batch attitude export is an offline comparator, with its future-input preprocessing clearly distinguished from received-event processing. Raw PPC inference never opens reference.csv.

Develop on Tokyo run1; the existing six PPC runs are development/regression data. Closed application holdouts and paused smartphone work remain closed. Record source, configuration, binary and input hashes and full raw argv. A bounded smoke establishes wiring; full runs establish development metrics.

Quaternion is body FLU to the filter's fixed local ENU frame. Published RPY is body FRD to world NED, aerospace 3-2-1, degrees; heading is clockwise from north. GPST and the original IMU axis mapping/lever arms are fixed. Evaluation converts the fixed local frame to each reference location explicitly. No fitted global heading, mounting offset, time shift, retrospective initialization or smoothing may improve the reported online result. Unavailable fields are NaN in CSV and null in JSON; initial heading is not qualified until its latch is observed.

Report position/velocity norm and per-axis errors, circular roll/pitch/heading, SO(3) rotation error, RMSE/P50/P95/maximum, availability, exact time matching, uninitialized states, first alignment delay, resets/recovery delay and latency. Use all available latched attitudes for primary errors, including unhealthy heading; report healthy-only subsets separately. Reference labels are offline only. Stops (<0.2 m/s), low speed (0.2–2 m/s), turns (>5 deg/s at >=1 m/s), reverse (reference forward velocity <-0.5 m/s) and ordinary motion may overlap. Scenario GNSS removal and IMU gaps use fixed elapsed-time windows; their coverage and post-event recovery are retained, including right censoring.

After baseline diagnosis, freeze one causal candidate and its acceptance contract before its comparison. Across six development runs require no more than 1% regression in position, velocity and full-rotation RMSE/P95, no more than 0.1 percentage point availability loss, no slower initial heading or scenario recovery, and at least one targeted attitude metric improvement. Runtime is reported with host contention. Failure keeps the original default and records No-Go; threshold tuning after seeing comparison results is not permitted in that experiment.

Ship a one-command scorer with JSON/CSV/plots, independent math/missing-data tests, native frame/snapshot tests, real prefix parity and outage checks. Re-run the applicable broad CLI/benchmark/binding/package/native checks. Docker and binary bundles must expose the same command and a tracked offline smoke; record platform/runtime requirements and actual distribution verification. Registry or release publication is a separate action.