Fix the ADS-B truth path for ADSBHub-fed sources

[Purple10101]

What

Three changes to the ADS-B truth path, all needed before a node can read truth from an ADSBHub-fed source such as adsb.retina.fm.

1. alt_baro fallback when alt_geom is absent.
2. Extrapolation window 5 s → 10 s, as a named constant.
3. baro_rate fallback when geom_rate is absent.

All three are the same shape: blah2 read only the GNSS-derived field, and ADSBHub's SBS feed carries none of them.

1. Altitude fallback

extrapolatePosition read only alt_geom, so an aircraft without one was placed at 0 ft. Zero is falsy, bistatic.js rejects it, and geometry.js turns that rejection into a delay of exactly 0: a truth target that looks present and is useless.

bistatic.js already had alt_geom ?? alt_baro, but it could never fire, because extrapolatePosition had already replaced the missing value with 0 and handed it over labelled alt_geom. The fallback belongs where the field is first read.

SBS/BaseStation carries no geometric altitude, so adsb.retina.fm serves alt_geom for 0 of 43 aircraft and alt_baro for 43 of 43.

What the approximation costs. alt_baro is pressure altitude, alt_geom is height above the WGS-84 ellipsoid. Measured across 202 aircraft carrying both: median difference +750 ft, p90 1350 ft, and systematically positive rather than scattered. Projected onto bistatic delay with this repo's bistatic.js and jn1's real geometry, in 150 m bins:

ground range	elevation	error at +500 ft	at +1500 ft
5 km	64.9°	287 m (1.9 bins)	862 m (5.7 bins)
25 km	23.1°	124 m (0.8 bins)	375 m (2.5 bins)
100 km	6.1°	34 m (0.2 bins)	104 m (0.7 bins)
400 km	1.5°	17 m (0.1 bins)	53 m (0.4 bins)

It scales with sin(elevation): sub-bin beyond ~50 km, a few bins for near-overhead targets.

2. Wider extrapolation window

The 5 s threshold was chosen when every source refreshed positions about once a second. adsb.retina.fm refreshes in whole-population batches every 8.6 s, so each aircraft's age sweeps a 1-9 s sawtooth and the whole population crosses 5 s together.

It is a confidence cutoff on a constant-velocity model, not a property of the data, so widening it costs projection error rather than correctness. Worst realistic case, a standard-rate turn held for the whole window: 131 m at 5 s, 520 m at 10 s, about three range bins here, and near zero for an airliner in cruise.

3. Vertical rate fallback

geom_rate reaches this code from neither source: SBS has no such field, and the tar1090 converter does not pass adsb.lol's through. vel_up was therefore always 0, so a climbing aircraft got the doppler of a level one and its projected altitude never moved. baro_rate is carried by both feeds and already passed by the converter.

This one is correctness, not capability. 122 of 193 live aircraft were climbing or descending by more than 100 ft/min, and including their vertical rate moved the doppler by a median of 0.19 Hz, at most 1.38 Hz. Against 2 Hz bins at a 0.5 s CPI that is under one bin even at worst, because these targets sit at low elevation. It would matter for a near-overhead climber; this sample had none.

Measured effect

Live data through this pipeline with jonathan-node-1's real rx/tx geometry (WGBY-TV), usable truth targets per poll over 10 polls:

	adsb.retina.fm	api.adsb.lol
before	0 every poll	38
altitude fallback only	mean 10.7, zero on 7/10	40.7
plus wider window	mean 53.2, zero on 0/10	60.0

adsb.lol gains 47% as a side effect, because its sparser aircraft were crossing the 5 s cutoff too. That was available at any point and is independent of adsb-service.

Testing

node api/test_alt_baro_fallback.js     # 14 checks, all pass

Assertion-based, exits non-zero on failure. Each group was verified to fail with its change reverted and the tests retained: 3 fail without the altitude fallback, 5 without the vertical-rate fallback.

test_bistatic, test_tar1090_extrapolation, test_timestamp_units and test_location still pass.

Deploying

This is an api/ change, so it needs the build-frontend workflow; build-blah2 does not build blah2-api.

Must land before or with offworldlabs/retina-node#44, or nodes pointed at adsb-service get an empty overlay. Pairs with offworldlabs/tar1090-node#20.

🤖 Generated with Claude Code