fix(benchmark): start the dead-reckoning control from the true initial state (T2-4)

[Telschow]

Why

The synthetic truth starts away from the origin, and DeadReckoning started at the origin. The control's unaligned error therefore mixed a start offset with drift. In the benchmark record the unaligned ATE for this case was far above the start-aligned one.

What changed

• DeadReckoning.run takes an optional initial (attitude, position, velocity). The default is the same as before: the identity pose at rest at the origin. Biases are ignored, because the integrator has no bias state. The type is a small Protocol, because importing InitialState from eskf made an import cycle that tests/test_no_import_cycles.py catches.
• run_case passes the true first pose (velocity zero, as the synthetic scene documents) for the DeadReckoning case only. The ESKF cases are untouched.
• configs/benchmark.yaml: the dead_reckoning description now says it starts from the true state. The stale typed result in the vision comment (the one T1-5 left) is replaced by words.
• docs/scenarios.md mirrors the new description (a test requires it).
• tests/test_injection_path.py: the noiseless-IMU test pinned a hand-typed number for the old behaviour. It now asserts that the noiseless control is far below the noisy one.

The failing test (written first)

tests/test_estimators.py. On main the two new tests that pass initial= fail with TypeError: DeadReckoning.run() got an unexpected keyword argument 'initial'. The default-start test passes on main and still passes, so other callers are unaffected.

• first-epoch position and attitude error are zero with the true start, and the error stays bounded;
• an initial velocity is carried through;
• no initial state still starts at the origin.

Golden diff (case-confined)

Compared the old and new tests/golden/benchmark.json, all 15 cases:

• cases that changed: dead_reckoning only (55 values: ate, drift, headline, rpe_1s, stats, summary, and the description);
• other 14 cases: identical;
• top level: config_sha256 only. That is the file hash and moves because the config bytes changed. No case-level config_hash moved.

Before regenerating, the golden test without --update failed in dead_reckoning only (and, once the config was edited, in config_sha256 only).

Run comparison, navkit run --out on main and on this branch with runtime_s, wall_s, realtime_factor and environment stripped: the only differences are the dead_reckoning case and config_sha256.

Table rows, before and after

File	Row	Before	After
README.md, docs/results.md	Dead reckoning (no aiding), ATE RMSE (unaligned)	2.359	1.248

Read from the benchmark file. Other alignments of the same case are unchanged:

Alignment	Before	After
none	2.359	1.248
rigid	0.838	0.838
rigid_start	1.231	1.231
similarity	0.828	0.828

The unaligned figure now sits next to the start-aligned one, as it should. Other tables match (check_doc_tables.py). No figure uses this row.

The control still drifts

test_the_dead_reckoning_control_starts_from_the_true_state_and_still_drifts: the minimum position error is zero (the first epoch), the maximum exceeds the median, and the fitted growth slope and the final drift share are positive. In the record the growth slope was 0.0931 m/s before and 0.1279 m/s after, so the row still shows growing error.

Removed stale comment

In configs/benchmark.yaml: "and measures measurably worse (mean NEES 23 rather than 3)." is now "and measures measurably worse in mean NEES."

Not changed, and one thing to look at

Not changed: the ESKF initialisation, any default, the row (it stays), the 4x preset.

Unexpected, not fixed here: the ESKF cases also call run without an initial, so they also start at the identity pose at the origin with the declared one-metre prior, while the truth starts about two metres away. The comment above the scoring step in benchmark.py says the filter is "initialised in the reference frame". That reads wrong for the current start position. I left it, as asked. It needs its own issue.

Gates

ruff, format, mypy, check_doc_tables.py, metrics.py --check (count updated), check_links.py, the golden test, and the full suite (1562 passed, 29 skipped, 2 xfailed).

Closes #111

🤖 Generated with Claude Code