[DRAFT] test: cover behavior gaps surfaced by external review

[speckhard]

Summary

Pure test additions — no production code changes. Each test pins or exercises a documented public-API contract that previously had no coverage. Surfaced by external code review.

What's tested

test_database.py (+3 tests, +1 parametrize value)

• test_database_open_mmap_populate — Database.open(path, mmap=True, populate=True). Smoke test only on macOS (memmap2 PopulateRead advise is best-effort there); pre-faulting performance is a benchmark concern, not a unit-test one. Documented in the test docstring.
• test_database_negative_indexing_raises_overflow_error — pins the actual current exception (OverflowError from PyO3's usize extraction) and its message. If wraparound semantics are ever added, this test will fail loudly.
• test_database_empty_molecule_roundtrip — n_atoms == 0 round-trip through the SOA encoder and parser. The 0-atom path has no upstream guard; this test exercises it end-to-end.
• lz4 added to four mmap-parametrized tests (test_database_single_item_reads_are_view_compatible, ..._custom_array_properties_roundtrip_all_numeric_tags, ..._single_item_mutation_is_copy_on_write, ..._single_item_pickle_materializes_owned_roundtrip). The basic lz4+mmap=True path was already covered by test_database_roundtrip[lz4], but the deeper mutation/pickle/view round-trips ran only ["none", "zstd"].

test_atom_molecule.py (+3)

• test_from_arrays_rejects_wrong_dtype_positions and ..._atomic_numbers — wrong numpy dtypes (float64 positions, int64 atomic numbers) must error cleanly across the PyO3 boundary, not silently truncate or panic.
• test_set_property_python_bool_pins_current_behavior — locks in that mol.set_property(key, True) stores Int(1) because Python bool is an int subclass and PyO3 extracts i64 before f64. If anyone wants real bool storage, this test will fail loudly so the design choice is explicit.

test_from_ase.py (+1)

• test_from_ase_expands_voigt6_stress_to_3x3 — covers ase_bridge._voigt6_to_mat3x3. ASE's get_stress(voigt=True) returns a (6,) Voigt-form vector and the bridge must expand it to a (3,3) symmetric tensor. The test asserts both the exact element mapping and matrix symmetry.

Test plan

• Full suite green; new tests counted:

$ uv run --extra dev --locked --with "maturin>=1.4,<2.0" maturin develop --release
🛠 Installed atompack-db-0.2.1

$ uv run --extra dev --locked pytest tests/ -q
154 passed, 1 skipped in 6.95s   # with ASE installed

(test_from_ase.py is gated by pytest.importorskip("ase") like the rest of the file; the Voigt test runs only when ASE is available. Without ASE, the suite still passes — the test_from_ase module is module-level skipped.)

• Each new test runs and passes individually (verified with -v).

What this is not

No production code changes, no API changes, no fixture changes. Risk surface is zero — this PR can only fail by running tests against unmerged behavior changes. All assertions are against current main semantics.