test(stage-2): add proptest + linalg edge cases (#393)

[Mec-iS]

Summary

Stage 2 of #391, tracked in #393. Adds proptest as a dev-dependency and writes property-based invariant tests + edge cases for core linalg, sort, and distance. +14 tests (474 → 488 unit), all gates green.

Changes

proptest = "1.5" added to [dev-dependencies]

linalg/basic/arrays.rs — 5 proptest invariants + 5 edge cases

• Transpose involution: (A^T)^T == A
• Matmul identity: A * I == A
• Matmul associativity: (AB)C ≈ A(BC) (approximate comparison — FP accumulation across three matmuls can exceed exact PartialEq)
• Transpose-matmul identity: (AB)^T == B^T * A^T
• Reshape preserves element count: reshaped matrix has the expected (rows, cols) shape
• Edge: 1×1 matmul, row×col matmul (→ 1×1), shape-mismatch panic, reshape-incompatible panic, 1×N transpose

algorithm/sort/quick_sort.rs — 1 proptest invariant

• Valid permutation: quick_argsort produces a permutation where all indices appear exactly once and values are non-decreasing in permutation order

metrics/distance/euclidian.rs — 3 proptest invariants

• Zero for identical points: d(a, a) == 0
• Symmetry: d(a, b) == d(b, a)
• Triangle inequality: d(a, c) ≤ d(a, b) + d(b, c)

Verification (run, not guessed)

Gate	Result
cargo fmt --all -- --check	exit 0
cargo clippy --all-features -- -Drust-2018-idioms -Drust-2024-compatibility -Dwarnings	exit 0
cargo test --all-features	exit 0 — 488 unit, 68+3 doctests

Notes

• Matmul associativity uses approximate comparison (1e-9 tolerance) because three-level FP accumulation can cause exact PartialEq to fail — discovered by running, not guessing.
• Distance proptests generate same-length vectors via a shared len parameter (the distance function panics on size mismatch).
• Each batch was verified by running before moving to the next.

Version bumped 0.6.3 → 0.6.4; CHANGELOG updated under [0.6.4].

Closes #393.

[Mec-iS]

Summary of Actionable Notes

Location	Severity	Note
arb_small_matrix	🟡 Minor	Input range 0.0..100.0 excludes negatives; make range a parameter or widen to -50.0..50.0 to avoid silently weak coverage for future callers
proptest_matmul_transpose_identity	🟠 Fragile	Uses exact prop_assert_eq! for a FP computation — switch to approximate comparison (1e-10) to survive future matmul implementation changes
proptest_matmul_associativity comment	ℹ️ Typo	"matmils" → "matmuls"
quick_argsort_is_valid_permutation	🟡 Minor	if n == 0 { return Ok(()) } guard is unreachable (min length = 1); remove to avoid misleading dead code
euclidean_distance_symmetric / triangle_inequality	🟡 Minor	Slice-truncation input generation is wasteful; prefer prop_flat_map length-first pattern consistent with arb_small_matrix

The proptest_matmul_transpose_identity fragility is the only item that could cause a spurious CI failure under future implementation changes. Everything else is minor polish. PR #410 is ready to merge, with the transpose-identity assertion mode worth fixing before or shortly after landing.

[Mec-iS]

Pushed 204e39b — fixes the wasm32 build failure and addresses all review feedback:

wasm32 build fix: proptest pulls in wait-timeout (Unix syscalls unavailable on wasm32). Moved proptest to [target.'cfg(not(target_arch = "wasm32"))'.dev-dependencies] and added #[cfg(not(target_arch = "wasm32"))] to all proptest test functions. Verified: cargo build --target wasm32-unknown-unknown --all-features = exit 0.

Review feedback:

• 🟠 proptest_matmul_transpose_identity: switched from exact prop_assert_eq! to approximate comparison (1e-10) — prevents spurious CI failures under future matmul impl changes.
• 🟡 arb_small_matrix: widened input range from 0.0..100.0 to -50.0..50.0 for negative-value coverage.
• ℹ️ Fixed typo matmils → matmuls.
• 🟡 quick_argsort_is_valid_permutation: removed unreachable n == 0 guard (min length is 1).

Verified: fmt=0, clippy=0 (with -Drust-2024-compatibility), test=0 (488 unit + 68+3 doctests), wasm32 build=0.