NaN follow-ups after #45/#47: mcrg np.var pre-checks bypass autocorr; DBL_MAX spacing edge

[marcohost33-maker]

From the independent rounds on #45 (d46e768) and #47 (4436c1e), 2026-09-26:

• Callers pre-check np.var(x) > 0 before calling autocorr (mcrg_matrix.py:341, mcrg_multirg.py:263/338/480/591): np.var of a NaN series is NaN, nan > 0 is False, so the series never reaches the new _require_finite gate and tau silently becomes 0.5. Same orientation class as NaN passes two rejection guards (autocorr var0, clt sigma2_g): fail-open orientation #46. Fix: validate finiteness first, or call the autocorr entry and let it fail closed.
• rhat.py spacing floor at DBL_MAX: the docstring says the floor changes nothing for normal numbers; at folded_scale = DBL_MAX np.spacing overflows to inf (RuntimeWarning) and the verdict flips OK → DEGENERATE_FOLDED (fail-closed direction). Document or guard.
• '234/400' in the rhat_defined docstring (rhat = 1.0 under DEGENERATE_FOLDED) was not reproduced by the second reviewer (165/400 with another distribution) — state it as distribution-dependent.

🤖 Generated with Claude Code

[marcohost33-maker]

Additions from the delta round on #47 (496747c):

• jackknife_ratio: the den_terms finiteness guard (autocorr.py:346) has no test (mutant stays green over test_autocorr+test_clt+test_mcrg).
• c_window=+inf passes (autocorr.py:163): AR(1) phi=0.9 gives tau 1.40 instead of 9.26 (error bar ~2.6× too small); confidence_interval(n=+inf) returns (0, 0) (clt.py:199). Add np.isfinite.
• Finite input |x| ≈ 1e308: binning_error / obm_variance return NaN/inf silently; integrated_autocorr_time with empty x and given rho returns NaN (rho shape unchecked). jackknife_ratio passes a NaN from combine through (mcrg.py:463-464 uses that deliberately as a placeholder — decide).

🤖 Generated with Claude Code