fix(autocorr,clt): NaN-safe guards and finite inputs (#46)

[marcohost33-maker]

Closes #46.

Guards of the form if stat <= 0: raise rejected only when an ordered comparison was True, so NaN passed and produced silent NaN results. Found by the NaN-orientation lint (nan-gate-lint.py 1.0.1, Vero tools).

Correction (after the independent round on 4436c1e): the first version of this PR claimed that finiteness checked in autocorr_function_fft reached binning_error — that was wrong, binning_error never calls it, and further public entries still returned NaN silently (obm_variance, jackknife_ratio, integrated_autocorr_time(..., rho=...), ar1_clt_variance(sigma_eps=nan), confidence_interval(n=nan)). The PR now checks every public entry.

Changes

• autocorr._require_finite(values, name): fail closed on NaN/inf, called at each public entry: autocorr_function_fft, integrated_autocorr_time (samples always, and a precomputed rho), binning_error, jackknife_ratio (numerator and denominator terms), clt.obm_variance.
• positively formulated guards: var0, c_window, sigma2_g (finite and ≥ 0), mean finite, ar1_clt_variance sigma_eps finite and > 0, confidence_interval n NaN-safe.
• an overflowing series (|x| ≈ 1e308, finite samples) now reports "variance overflows float64" instead of the misleading "constant series".

Evidence (Windows, py3.12) at 496747c

• 20 new cases red on the base (10 in 6d445d5, 10 in 706bfbf), all by assertion
• test_autocorr + test_clt: 60 passed; indirect consumers (test_checkpoint, test_manifest, test_mcrg*, test_rbim_nishimori, test_snis, test_wolff2d + the two): 181 passed, exit 0
• python -m adaptiverg_qec.cli --selftest --json results/selftest-ci.json: 45/45 PASS, exit 0 (exit measured directly, not through a pipe)
• ruff check + ruff format --check clean
• diskriminierung.py: DK-20260926T102148-f647563dbe6a (first four guards) and DK-20260926T112239-2956ffa7d19a (8/8 new guards: binning, rho-path x, rho finite, jackknife, overflow message, obm, sigma_eps, n); positive control on a normal-path test collapsed as required (DK-20260926T112337-292c42f43469)

Pre-existing, not changed here: callers in mcrg_matrix.py/mcrg_multirg.py pre-check np.var(x) > 0, so a NaN series never reaches autocorr there and tau silently becomes 0.5 — tracked as follow-up.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

[marcohost33-maker]

Delta round (Equalita, same family) at 496747c: PASS WITH CONDITIONS (P3 only). NaN/+inf/-inf now fail loudly at all 15 array entries (autocorr.py:89/162/169/254/345-346, clt.py:136); integer arrays, lists, unmasked np.ma, np.int64 sizes and empty arrays still behave; no circular import (5 import orders). All body numbers reproduced by the reviewer: 60 / 181 passed, selftest 45/45 exit 0, 20 red on the bases. Discrimination DK-20260926T114219-9018dfa66a5c + positive control DK-20260926T114305-b34c91bbab56.

Precision of my body: in run DK-…112239 my first positive-control case was blind (it targeted a test that expects the raise) — replaced by a normal-path control (DK-…112337); 'jackknife' there discriminates the num guard only — the den_terms guard has no test (P3-1). Follow-ups moved to #48: den=NaN test, +inf for c_window and n (c_window=+inf → tau 1.40 instead of 9.26 at phi 0.9), non-finite results from finite |x|≈1e308 in binning_error/obm_variance, rho shape unchecked.

🤖 Generated with Claude Code