Add sparse second-order grid evidence

docs/reference/sparse_second_order_grid.md

@@ -0,0 +1,16 @@
+# Sparse second-order grid evidence
+
+`pyrecest.filters.sparse_second_order_grid_evidence` computes exact evidence for
+finite-support second-order grid hidden Markov models.  The latent state is a
+sparse pair `(x[t-1], x[t])`; callers provide the initial pair lattice and sparse
+transition rows.
+
+The primitive is useful when a dense `n_states^2 x n_states` transition would be
+too large, but the model has finite local support.  It returns the normalized
+log evidence, terminal single-state posterior, optional fixed-interval smoothed
+single-state marginals, and support/cache diagnostics.
+
+The implementation is intentionally domain-neutral: replay, map matching,
+maneuvering target tracking, and other second-order grid models can supply their
+own transition-row builder while sharing the same forward/backward evidence
+calculation.

src/pyrecest/filters/__init__.py

@@ -61,6 +61,12 @@
     "sticky_mode_transition_matrix": ".discrete_state",
     "uniform_log_probabilities": ".discrete_state",
     "uniform_probabilities": ".discrete_state",
+    "SparsePairInitializer": ".sparse_second_order_grid",
+    "SparsePairTransitionCacheKeyBuilder": ".sparse_second_order_grid",
+    "SparsePairTransitionRowBuilder": ".sparse_second_order_grid",
+    "SparseSecondOrderGridResult": ".sparse_second_order_grid",
+    "SparseTransitionRows": ".sparse_second_order_grid",
+    "sparse_second_order_grid_evidence": ".sparse_second_order_grid",
     "DecorrelatedSCGPTracker": ".gprhm_tracker",
     "DecorrelatedScGpTracker": ".gprhm_tracker",
     "EKFSplineTracker": ".ekf_spline_tracker",