The non-grav weak-constraint guard measures the orbit's conditioning, not the amplitude's

[matthewholman]

WEAK_NONGRAV_RCOND rejects a joint state + non-grav fit when the reciprocal condition number of the whole npar×npar correlation matrix falls below 1e-8. That is not a statement about the non-grav column: a six-parameter orbit fit is itself strongly correlated, so the test is dominated by the orbit underneath. On (6489) Golevka the state block alone has rcond 2.6e-10 — twenty-five times below the threshold — while A2 correlates with the state at only 0.91–0.93 and its formal uncertainty, 1.6e-15 au/day², matches the published one. The fit is rejected for the conditioning of its own orbit, so A2 is reported as NaN even after #536.

I tried replacing it with the amplitude's variance inflation factor and measured that this does not work either. On a synthetic A2 arc (truth −5.0e-14):

arc	VIF	A2 ± σ
10 d	5.7e4	−7.5e−12 ± 2.6e−7
30 d	1.2e4	−3.0e−13 ± 1.8e−8
90 d	23	−5.2e−14 ± 1.3e−10
4 yr	19	−5.00e−14 ± 3.6e−13
6489 Golevka	3.2e5	−10.2e−15 ± 1.6e−15

Golevka is a 6σ detection with a higher VIF than the useless 10-day arc. Correlation with the state is not what separates them; whether the amplitude is determined is, and the covariance already reports that faithfully.

So the question is what the guard is for. With the solve fixed the state is no longer contaminated, and a large σ is an honest answer rather than a failure. Worth a decision before picking a new criterion — test_a2_weak_constraint_guard_on_short_arc encodes the current one.

[matthewholman]

Now that #536 and #538 are on main, this is what the guard costs. Reduced chi-square in brackets:

object	optical only	+ radar
2062 Aten	−14.96 ± 1.70 (0.44)	NaN (1.55)
3908 Nyx	—	NaN (0.28)
6489 Golevka	—	NaN (0.78)

(A2 in 1e-15 au/day².) All four optical-only fits report an uncertainty; all three radar fits are suppressed, at reduced chi-square 0.28–1.55.

Aten is the clean case: same object, same 980 optical observations, and adding seven radar rows turns −14.96 ± 1.70 into NaN. Radar tightens the state, which lowers the full correlation matrix's rcond, so better data makes the guard reject.

Worth considering as an option: drop the guard and report the amplitude with its uncertainty. It was added because the solve returned garbage, and the solve is now fixed — a large sigma is an honest answer, and it does not suppress a well-constrained null.