Wider real-front-end measurement (ADR-0023 follow-up); ADR-0024 draft

[Telschow]

What

• The wider measurement of the two-view front end: six windows of five EuRoC Machine Hall sequences (windows chosen by ground-truth speed, a second non-overlapping MH_01 window), the measured errors of each window replayed into the stochastic clone at assumed noise of 1, 2, 4 and 8 times the measured size, three seeds each, with the single anchor for comparison. The design (windows, scales, seeds, gaps) was committed before the run (b4e4f0b).
• navkit euroc vo --wide runs it. Per-window tables on docs/real_front_end.md, committed CSVs, and tests that check each statement against them.
• The first window is replayed again with the run aligned to its frames. The earlier run began 1.08 s after the frames (the ground truth starts at 1.08 s), so its last second had no vision. The figures moved slightly (NEES 27, 54% at the measured noise; 4.1, 92% at four times) and the conclusion did not.
• ADR-0024 (proposed): the decision section is for the maintainer.

Result

Replayed real errors, gate = NEES below 10 and 2-sigma coverage above 90%:

• at the measured noise: 2 of 6 windows meet it; at 2x: 2 of 6; at 4x: 6 of 6; at 8x: 6 of 6;
• the single anchor: 0 of 6.
  The tightest window at 4x has a coverage of 91%. Front-end errors are small in every window (0.07 to 0.16 deg and 5 to 20 mm at one frame) and correlated for 0.05 to 0.5 s.

Limits

One dataset, a front end that is not a production one, a translation with the true scale. Each window's replayed errors are one realisation, so a window has no spread of its own. The default is unchanged.

Full suite (1309 passed), lint, types, links, strict docs build, SRS trace and metrics pass locally. Coverage 94.3%. The camera frames are git-ignored and not in the repository.

🤖 Generated with Claude Code