test(sudoku): add 24 tests for core/solvers.py (cycle 80)

[jsboige]

Summary

Adds 24 unit tests for scripts/sudoku/core/solvers.py — Norvig-style constraint propagation solver.

Test classes (24 tests)

Class	Tests	Target
TestIsValidPuzzle	9	Row/col/box duplicate detection, empty grid, partial grid
TestSolveSudoku	11	Easy/hard/empty puzzles, unsolvable, preserves givens, return type
TestCrossInvariants	4	Validity consistency, deterministic results

Coverage

• solve_sudoku(puzzle_81): full solver (constraint propagation + MRV backtracking)
• is_valid_puzzle(puzzle_81): duplicate check in rows/cols/boxes
• Edge cases: empty puzzle, complete grid, single empty cell, unsolvable (row/col/box conflicts)

Verification

24 passed in 0.20s
Full suite: 1422 passed, 0 regression

Context

• po-2024 test-coverage lane (T1: sudoku core)
• LIVE: 2 callers (eval_notebook.py, cpu_diagnostic.py)
• Pure computation: no I/O, no GPU, no network
• Cumulative: 524 + 24 = 548 tests across cycles 69-80

[clusterManager-Myia]

[NanoClaw]

Well-structured test suite. 24 tests across 3 classes with clear separation of concerns, good edge-case coverage, and correct assertion patterns. A few observations:

Nit (non-blocking):

• test_duplicate_in_row (L84-86): The comment on L85 says "row 0 already has 5 at index 0 — now duplicate at [0,0]" but the p[0] = 5 assignment is a no-op (the value was already 5). The comment on L86 correctly self-corrects, but the first comment is misleading. Consider removing L85 or simplifying to just: p[1] = 5 # duplicate 5 in row 0 with p[0].

Minor observations:

• sys.path.insert for import resolution (L17-19) works but is fragile relative to the execution context. If tests are ever run from a different working directory, Path(__file__).resolve() should still hold, so this is acceptable for a scripts-local test — just noting it as something a CI runner could trip on.
• TestCrossInvariants.test_solve_produces_valid_puzzle is functionally identical to TestSolveSudoku.test_easy_puzzle in terms of setup and assertions (same puzzle, same validity check). It adds semantic value as a cross-invariant test, but the duplication is worth being aware of if the easy puzzle string ever changes.
• pytest is imported but never used (no pytest.mark, pytest.raises, or fixtures). Harmless, but import pytest on L7 could be removed if no parametrized or fixture-based tests are planned.

Verified:

• All puzzle grids validated: solved grid is a correct complete Sudoku, easy/hard puzzle givens have no duplicates and easy givens match the solved grid.
• No security concerns (no secrets, credentials, or sensitive patterns).
• Test count matches PR title: 24 test_* methods.
• Assertion patterns are sound: is True/is None/is not None used correctly, numpy array equality and masking used appropriately.

[myia-ai-01]

[ai-01 forensic APPROVE] Tests-only PR (+296/-0, 1 fichier). Cible VERIFIEE LIVE : scripts/sudoku/core/solvers.py existe (3697 B), exporte solve_sudoku (L86) + is_valid_puzzle (L108) — les 2 symboles testes — et a un caller reel (scripts/sudoku/cpu_diagnostic.py:29). scripts/sudoku/ cataloge dans docs/scripts-reference.md. Pas de dead-code (lecon #1876/#1885). Aucune cellule notebook touchee.