Radar model is monostatic; bistatic observations are silently mis-modelled

[matthewholman]

Observation carries one station, used for both the transmit and the receive leg. That is monostatic by design — ISSUE_146_RADAR_DESIGN.md scopes the first implementation to TX == RX and lists bistatic as a refinement — but a bistatic measurement passed in anyway is not rejected. The receive leg is simply evaluated at the transmitting antenna.

Measured on JPL radar for (6489) Golevka: observed/model Doppler is 1.039–1.044 on bistatic rows against 0.999934 on monostatic ones. Against a 0.11 Hz uncertainty that is roughly a hundred sigma per observation. Golevka has 18 such rows out of 40; including them gives flag 2 with a reduced chi-square of 1.5e7, excluding them gives flag 0 at 0.80.

Bistatic pairs are common in the archive — Goldstone DSS-14 transmitting with DSS-13 or DSS-38 receiving. Supporting them needs a second station on the observation; until then the caller has to filter, which #529 documents in _radar_observation.

[matthewholman]

A simpler fix than the one this issue was opened with, and it subsumes a second problem.

The up leg currently derives the transmit position by extrapolating the receive station backwards:

rtx = xe - vo*tau - 0.5*ao*tau^2

That is wrong for bistatic in the obvious way -- the transmitting antenna is not the receiving one moved back in time -- and it also carries a truncation error even when TX == RX. Over Golevka's ~100 s round trip the Earth's rotation takes the station about 170 m past the linear term, roughly 1.1 us, against delay uncertainties of 0.3 us.

Both go away if we simply evaluate the station position at each of the two epochs instead of extrapolating between them.

The transmit epoch is not something we have to solve for: the round-trip delay is the observable, so the transmit time is t_receive - delay_observed. The only error in that is the delay residual, of order microseconds, which moves a station by about half a millimetre.

_append_observer_acceleration already shifts et and re-evaluates obscodes_to_barycentric, so evaluating a station at an arbitrary epoch is machinery layup already has -- it is currently used to build a Taylor series where a direct evaluation would do:

shifted = data.copy()
shifted["et"] = data["et"] - delay_seconds   # transmit epoch, from the observable
shifted["stn"] = data["stnTx"]               # the transmitting antenna
tx = observatory.obscodes_to_barycentric(shifted)

Observation then carries a transmitter position (and velocity, for Doppler) alongside the receiver's, and compute_radar_residuals uses it directly for the up leg.

Net effect: bistatic becomes correct by construction, the ~1.1 us truncation disappears, and the code gets smaller -- the tau expansion goes, the ax/ay/az plumbing goes, and _append_observer_acceleration loses its only caller.

One implementation note: compute the transmit epoch from the observed delay once, not from the model delay each iteration. Early in a fit the model delay can be far off, and the station position should not move with it.