test_apply_comet compares the original 1/a against the CODE catalogue with
atol=1000 (units of 1e-6/au). Running the full 369-comet fixture: 325 of 334
reproduce the catalogue to a median of 0.009, against a median catalogue
uncertainty of 2.2. The test would pass with the integration badly wrong.
Tightening it first needs the nine that disagree (|delta| 1.6-29). Not step
size -- 10, 4 and 1 day steps agree to six significant figures. Non-gravitational
forces explain two, both the smallest-perihelion cases: C/2020 S4 needs
A1 = 7.6e-8 au/d^2, and JPL fits 6.34e-8 under the same distant-activity
scaling. The other seven carry osculating elements inconsistent with the
catalogue value stored beside them; substituting current JPL elements takes
C/2003 WT42 from 29.0 to 1.0.
Doing that substitution for the whole sample makes agreement much worse
(median 0.009 -> 3.9), so this is nine rows, not a stale fixture.
Suggest fixing or excluding those nine, then setting atol near 0.1.
test_apply_cometcompares the original 1/a against the CODE catalogue withatol=1000(units of 1e-6/au). Running the full 369-comet fixture: 325 of 334reproduce the catalogue to a median of 0.009, against a median catalogue
uncertainty of 2.2. The test would pass with the integration badly wrong.
Tightening it first needs the nine that disagree (|delta| 1.6-29). Not step
size -- 10, 4 and 1 day steps agree to six significant figures. Non-gravitational
forces explain two, both the smallest-perihelion cases: C/2020 S4 needs
A1 = 7.6e-8 au/d^2, and JPL fits 6.34e-8 under the same distant-activity
scaling. The other seven carry osculating elements inconsistent with the
catalogue value stored beside them; substituting current JPL elements takes
C/2003 WT42 from 29.0 to 1.0.
Doing that substitution for the whole sample makes agreement much worse
(median 0.009 -> 3.9), so this is nine rows, not a stale fixture.
Suggest fixing or excluding those nine, then setting
atolnear 0.1.