Add homogeneity_test(): chi2, likelihood-ratio, and Kullback tests for transition regimes

[dshkol]

Second tranche of the 0.2.0 feature work: the inference layer. This closes the biggest scoping caveat — griddy could show that lag-conditioned matrices differ but not test whether they differ beyond sampling noise.

New function

homogeneity_test(x) — generic with two methods:

• grd_spatial_markov: tests the Rey, Kang & Wolf (2016) question directly — do the transition matrices conditioned on spatial-lag class differ from the pooled matrix?
• list of count matrices: the same machinery for arbitrary regime stratifications (LISA Markov quadrant regimes, time regimes, etc.)

Returns a tidy tibble with one row per test: Pearson chi-squared Q and likelihood-ratio (both with the Bickenbach & Bald 2003 degrees-of-freedom adjustment), and the Kullback, Kupperman & Ku (1962) information test.

On the bundled usjoin panel: Q = 200.6 (df 60), LR = 170.7 — spatial conditioning decisively matters, which is the headline number for the release post.

Validation

• Exact agreement with PySAL giddy 2.3.9 Spatial_Markov homogeneity attributes (Q, LR, dof, p-values) and giddy.markov.kullback() on the usjoin spatial count matrices; fixture generated by the new tools/oracle/pysal_giddy_inference.py (kept separate from the existing oracle script to avoid cross-branch conflicts)
• The published Kullback (1962) Table 9.2 example ships as an in-package reference test (statistic 160.961, df 30)
• Shipped tests also cover identical-regime invariants (Q = LR = 0; the Kullback statistic is deliberately not zero there — verified against giddy) and error paths
• devtools::check(): 0 errors, 0 warnings, 0 notes

Note: branched off main independently of #2; both add a section under the same NEWS heading, so whichever merges second needs a trivial conflict resolution there.

🤖 Generated with Claude Code