deps: upgrade wide/thiserror/criterion + once_cell→std; fast_max/min recovers ~2.2x AVX2

[isPANN]

Summary

Two connected changes — a dependency refresh, and the SIMD fix that the refresh turned out to need.

1. Dependency upgrades

• wide 0.7 → 1 (SIMD)
• thiserror 1.0 → 2.0
• criterion 0.5 → 0.8 (dev/bench only; criterion::black_box → std::hint::black_box)
• once_cell → std OnceLock (external dep removed from tropical-gemm-cuda)

2. fast_max/fast_min — recovering performance the wide bump silently cost

wide 1.x's f32xN::max/min are NaN-aware, and that handling is not free:

arch	.max() / .min() (wide 1.x default)	.fast_max() / .fast_min()
x86 AVX	max_m256 + is_nan() + vblendvps (3 ops + a max→isnan→blend dependency chain in the hot loop)	single max_m256
ARM NEON	NaN-aware vmaxnmq_f32 / vminnmq_f32	plain vmaxq_f32 / vminq_f32

Tropical GEMM never produces NaN — MaxPlus inputs are finite or −∞, MinPlus finite or +∞, MaxMul finite (zero = 0.0) — so the only NaN source (+∞ + −∞) cannot occur and the NaN handling is dead work. Switching all 7 microkernels (avx2 ×4, neon ×3) to fast_max/fast_min lowers each to a single maxps/minps/vmaxq/vminq.

Without this, the wide 0.7→1.x bump alone silently ~halves AVX2 throughput.

Benchmarks (A/B: fast_* vs default max/min, same wide 1.4.0, criterion, p < 0.05)

x86 AVX2 (HKUST-GZ HPC, f32):

size	default max/min	fast_*	speedup
128	~600 µs	~291 µs	2.07×
256	~4.59 ms	~2.09 ms	2.20×
512	~36.5 ms	~16.4 ms	2.22×
1024	~291 ms	~131 ms	2.23×

(MaxPlus / MinPlus / MaxMul all ≈ −54%.)

ARM NEON (Apple M-series, f32): −3% to −4% on MinPlus/MaxMul; MaxPlus ≈ flat.

Correctness

cargo test -p tropical-gemm with fast_* in place: 279 unit + 24 doctests, 0 failures on both x86 AVX2 (HPC) and ARM NEON (local). Includes the assertion-based AVX2 kernel unit tests. cargo audit clean. CUDA crate (thiserror 2.0 + once_cell→std + cudarc) builds and its 56 GPU tests pass on A40/CUDA 12.8.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.29%. Comparing base (35e589c) to head (4fca49c).

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #56   +/-   ##
=======================================
  Coverage   96.29%   96.29%           
=======================================
  Files          19       19           
  Lines         918      918           
=======================================
  Hits          884      884           
  Misses         34       34           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.