[davidhbernstein]

Pin vcov()'s diagonal fallback against a zero variance

A follow-up to #62, one test file and no code change.

When no Hessian can be inverted, vcov.sfareg() falls back to a diagonal built
by squaring the reported standard errors (R/sfareg_methods.R:209). A64's
clamp therefore reached it too: a parameter whose variance estimate came back
negative arrived here as 0 and contributed an exact zero variance — the same
claim of infinite precision, one level up. Since #62 it arrives as NA.

This is a characterization test, not a regression test, and the comment in it
says so. vcov() itself is unchanged, so given the NA the test constructs,
it passes on main as well; what #62 changed is which value arrives. It is
worth pinning because the fallback squares whatever it is handed, so a
zero-fill or a pmax() reintroduced there would restore the defect silently
with every other A64 test still green.

The object is constructed rather than fitted, so — unlike #62's end-to-end
case, which had to be gated on skip_if_not(any(v < 0), ...) after all five CI
platforms disagreed with this Mac about whether the design reaches a negative
variance — this one cannot depend on an optimizer landing point.

test-a64-negative-variance-se.R: 19 of 19 pass against the main build at
5163c44. No R/ file is touched.

🤖 Generated with Claude Code