Model the annual aberration of light in the astrometric kernels

[Cybis320]

Stacked on #989 (base branch fix-nutation-frame); only the aberration change shows in the diff. Rebased from the study branch aberration-correction and simplified: the correction is always on, there is no .config option.

Summary

The star catalog gives barycentric directions; the camera sees them displaced by the annual aberration of light, up to 20.5″ towards the apex of the Earth's motion. Across a night's hour-angle range the alt/az projection of that displacement changes by up to ~0.6′. Without a model, the star fit absorbs it into RA_d, dec_d and pos_angle_ref: the fitted pointing of a fixed camera wanders through the night (the residual "chase" left in the calibration variation plots after the precession fix), and the platepar's reference pointing is not the true pointing.

• cyraDecToXY displaces catalog directions by the aberration before projecting; cyXYToRADec removes it. The platepar then describes the true pointing and stars still come out at their catalog positions.
• Earth's velocity from the Keplerian orbit (Meeus ch. 25, mean longitudes referred to the J2000 equinox): 0.02″ from erfa's epv00 + ab.

Stars and meteors are different frames

Light from a meteor is not aberrated: source and observer move together (the 0.3 ms light-time displacement and the aberration cancel exactly to first order). So:

conversion	stars (measurement=False, default)	image measurements (measurement=True)
xyToRaDecPP	aberration removed → catalog direction	left in → geometric direction in the Earth's frame, which is what a trajectory solver uses
raDecToXYPP	catalog direction aberrated, then projected	projected as is (new keyword)

The fit, star matching, photometry and the sky residuals all go through the star path. Every caller that converts meteor centroids or picks already passes measurement=True (applyPlateparToCentroids, applyAstrometryFTPdetectinfo, the ECSV writer, EventMonitor, SkyFit2 and Astra picks), so FTPdetectinfo gets Earth-frame directions with no further change. Callers that project Earth-frame objects through the star path (satellite positions, FOV footprints, EventMonitor's trajectory projection) are off by the 20″, i.e. 0.1 to 0.3 px at GMN scales, and were left alone.

Without this distinction (the study branch removed the aberration for everything), meteor directions would have stayed where they are today: biased by the aberration, because the uncorrected fit absorbs it and the removal then takes it out again.

Verification

Tests/test_aberration.py: shift magnitude and inverse, velocity against erfa (0.1″), forward/reverse round trips for both paths, that a measurement at a star's pixel comes out at the star's apparent (aberrated) direction to 0.01″, and the star fit as in real life on the distortion-free synthetic camera of #989's test: with the platepar holding the true pointing, stars project to their catalog directions (erfa velocity aberration inverted) within 0.5″ over the whole 82×46° frame through the night, with no roll (0.3″) or scale (0.1″) error. Today's kernel on the same test drifts 33″ at the centre over a night and leaves 7 to 15″ at the corners that no pointing or rotation refit can absorb.

44 nights from 17 GMN stations, recalibrated with the correction off and on (study branch): the fits move by the predicted amount (slope 1.01 in altitude, 0.99 in azimuth, expected 1.00, for every lens, pointing and season including two southern stations), and the settled-night altitude drift of the quiet cameras drops by about 40%. The residual after the correction has no aberration signature (slope 0.01 on the prediction) and does not repeat by sidereal time; it is thermal and mechanical. The nutation error fixed in #989 contributes only a constant offset within a night (its diurnal wander is 0.02 to 0.09′ p2p for the study dates), so those conclusions carry over.

USV001 2026-09-05 (217 FFs, 693 meteor centroids), this branch against the study's runs:

	off	study on	this branch
star-match residual	0.1587 px	0.1594 px	0.1586 px
settled pointing scatter alt / az	0.156′ / 0.076′	0.057′ / 0.068′	0.057′ / 0.061′
meteor centroid shift vs off, mean / max		6.4″ / 16.5″	16.9″ / 29.3″

The aberration at those centroids' directions and times is 16.8″ mean (20.2″ max); the measured shift divided by it has a median of 0.98. The maximum includes the constant 7 to 9″ from the nutation fix. The study run moved the meteors only by the part of the aberration that a rigid pointing shift cannot absorb across this 82° field.

What changes for users

• Meteor RA/Dec and alt/az in FTPdetectinfo move by the aberration at the meteor's direction, up to 20.5″ (direction-dependent, so not a common-mode shift between stations): ~10 m at 100 km range and ≲0.01° in radiants, systematic with respect to the apex. WMPL treats the observations as geometric Earth-frame directions, so this removes a bias rather than adding a correction to be undone downstream.
• Existing platepars absorbed the aberration of their fit date; applied without a refit they are off by the difference to today's aberration (up to ~0.6′ seasonally). The nightly recalibration refits them. In the calibration variation plots the corrected fits sit at a constant offset from an uncorrected reference platepar until the reference is refitted.
• The kernel keyword aberration (default True) remains on cyraDecToXY / cyXYToRADec for tests and for reproducing the old behaviour.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CzZHCZoftFjV712njkWB5v