[Telschow]

Re-measures the start-attitude-draw study on the corrected fixture as a committed CSV, and replaces the typed numbers on the pages by a generated, test-checked section. Development seeds 0 to 9 only.

Hook

run_case gets a keyword-only start_hook. It receives the StartError that draw_start_error drew and returns the one to use. It applies to ESKF cases that start from prior_draw. Like gnss_hook and imu_hook it is not in config_hash, and the docstring says so. The draw is made first on its own stream, so every variant of a seed sees the same draw.

Study

navkit sweep start-draw --csv PATH [--page PATH] [--compare PATH] [--seeds A-B] (src/navkit/start_draw_study.py). Cases outage_visual, vision_only, outage_visual_clone, outage_control. Variants full, no_attitude, attitude_only. One CSV row per (case, variant, seed): ATE RMSE (alignment none, full run) and mean NEES over the full run, no verdicts. Deterministic. Seeds from 100 are refused.

Result (medians over seeds 0 to 9)

Case	Full draw	Attitude zeroed	Attitude only	Seeds with larger ATE when the attitude is drawn
outage_visual	25.44 m	4.21 m	28.12 m	9 of 10
vision_only	54.23 m	7.55 m	52.18 m	10 of 10
outage_visual_clone	0.86 m	0.85 m	0.85 m	9 of 10
outage_control	9.74 m	9.43 m	7.66 m	6 of 10

The new numbers agree with the old claim. The clone is not exactly unchanged: per seed its ATE differs across variants by up to about 13% and its mean NEES by up to about 18%. The page therefore says "within a quarter of their mean in every seed" (tolerance chosen from those values), not "does not change". The mechanism stays a HYPOTHESIS; nothing here tests it.

Tests, with failing-on-main evidence

• tests/test_start_hook.py: identity hook gives an identical record (timing keys stripped); the hook sees the drawn error and is not in the hash; a hook that zeroes dtheta gives an exactly zero initial attitude error; every variant of a seed sees the same draw. All 4 failed on main (run_case had no start_hook).
• tests/test_start_draw_page.py: CSV shape and seeds, page section equals the generated text, every table cell recomputed from the rows, each sentence tested against the rows (several times, close to, factor of two, quarter for the clone, 2 degrees equals the filter's declared prior, HYPOTHESIS wording and no mechanism words), a fresh run of seed 0 equals the committed rows, held-out refusal.
• tests/test_quoted_numbers.py: table rows between the start-draw markers are exempt (they are checked cell by cell by the page test); sentences around them are still scanned. A new test pins that.
• Mutation checks: hook ignoring its return value fails the zero-attitude test. A hook-dependent draw seed fails the same-draw test and the identity test. Allowing seeds from 100 fails the refusal test.

Statements changed

File	Before	After
README.md, sweeps bullet on the attitude draw	words about the medians, "measured once ... at e94c107", "produced under the earlier draw ... being re-measured (issue #168)"	the same findings in words, no label, pointer to the generated table on the Results page and docs/data/start_draw_study.csv; HYPOTHESIS sentence kept
docs/results.md, paragraph "The anchor model does not recover ..."	words, old label	generated section between <!-- start-draw:start/end -->: table, sentences, HYPOTHESIS
CONSTRAINTS.md B1	"(MEASUREMENT, measured once ... at e94c107, being re-measured ...; HYPOTHESIS ...). The clone is unaffected."	"(MEASUREMENT, the generated table on the Results page; HYPOTHESIS ...). The clone does not follow the attitude draw."

No new typed numbers outside the generated table. Defaults and the golden file are unchanged.

CI

A step in the ensemble reproducibility job runs navkit sweep start-draw --compare docs/data/start_draw_study.csv (about 3.6 minutes locally). That job already took about 24 minutes against about 36 for the slowest test job, so the job should stay under it. The new page test is in the docs-only list.

Notes

• PR chore(metrics): quote the test count rounded down to a hundred, so PRs stop conflicting #174 was not merged when this was cut, so metrics.json and the count markers hold the exact count. Expect a count conflict when chore(metrics): quote the test count rounded down to a hundred, so PRs stop conflicting #174 merges; python scripts/metrics.py --write fixes it.
• Gates: ruff, format, mypy, check_doc_tables, metrics --check, check_links, mkdocs build --strict, framing and quoted-numbers tests, the golden test (unchanged), full suite (2307 passed, 31 skipped, 2 xfailed).

Closes #168

🤖 Generated with Claude Code