test(scripts): add 31 tests for sudoku/core/graph.py + dataset.py (cycle 81)

[jsboige]

Summary

Adds 31 tests for 2 Sudoku core modules:

**** (16 tests):

• (9): shape [2,1620], dtype long, no self-loops, 20 incoming edges per node, row/col/box neighbor verification, deterministic output
• (7): batch_size=1 matches base, correct offset for batch>1, shape scales, valid index ranges

**** (15 tests):

• (8): string-to-numpy conversion, zeros handling, dtype, edge cases
• (3): length, getitem, int8 casting
• Cross-invariants (4): edge count matches peer count, batch structure preserved, parse_81 output compatible with solver

LIVE verification

13+ callers across training/eval/diagnostic scripts (grep build_sudoku_edge_index / make_batch_edge_index).

Test results

31 passed in 2.26s
Full suite: 1429 passed, 0 failed, 0 regression

po-2024 test-coverage cumulative (cycles 69-81)

Cycle	Script	Tests	PR
69	validate_pr_notebooks	34	#1878
70	weekly_digest	19	#1880
71	validate_sc_notebooks	56	#1882
72	update_sc_navigation	42	#1885 (CLOSED dead-code)
73-78	genai-stack (7 modules)	341	#1886/#1892/#1894/#1897/#1898/#1903/#1905
80	sudoku/core/solvers.py	24	#1906
81	sudoku/core/graph.py + dataset.py	31	this PR
Total		579

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com

[clusterManager-Myia]

[Hermes] — APPROVED

Pure test file (251 LOC, single file). Tests for deterministic graph construction functions — no business logic changes, no I/O, no secrets. CI passes, security scan clean. Tests are well-structured with meaningful edge cases (self-loops, batch offsets, determinism). LGTM.