Fix CI: pin to Rust 1.94.0, fix clippy warnings, fix feature gates

AGENTS.md

@@ -0,0 +1,26 @@
+# AGENTS.md
+
+## Cursor Cloud specific instructions
+
+This is a Rust library crate (ndarray fork with HPC extensions). No external services (databases, APIs) are needed.
+
+### Quick reference
+
+| Action | Command |
+|--------|---------|
+| Build | `cargo build` |
+| Lint | `cargo clippy -- -D warnings` |
+| Test (lib) | `cargo test --lib -p ndarray` |
+| Test (workspace) | `cargo test` |
+| Test (HPC subset) | `cargo test --lib -p ndarray -- hpc::` |
+| Run example | `cargo run --example life` |
+| Format check | `cargo fmt -- --check` |
+
+### Environment notes
+
+- **Rust 1.94.0** is pinned via `rust-toolchain.toml`; rustup auto-selects it in `/workspace`.
+- **No AVX-512 hardware** in Cloud Agent VMs — SIMD kernel tests using `#[target_feature(enable = "avx512f")]` are compile-gated and will be skipped at runtime. This is expected behavior.
+- **Feature gates**: `intel-mkl` and `openblas` are mutually exclusive and require system libraries not installed by default. The default build uses `native` (pure Rust SIMD) which needs no extra libs.
+- **Build time**: ~18s cold, <1s incremental. Tests (~1819) take ~70s.
+- The workspace has sub-crates under `crates/` and `ndarray-rand/`. Default members exclude `blas-tests` and `blas-mock-tests` (they activate the `blas` feature which needs cblas-sys linking).
+- `libssl-dev` is needed as a build dependency for some transitive crates.

Cargo.toml

@@ -34,6 +34,9 @@ name = "ndarray"
 bench = false
 test = true
 
+[[example]]
+name = "ocr_benchmark"
+required-features = ["std"]
 
 [dependencies]
 num-integer = { workspace = true }

src/hpc/renderer.rs

@@ -425,7 +425,7 @@ mod tests {
         let _ = &*GLOBAL_RENDERER;
         // First-touch: tick count is 0; capacity is at least 4096
         // (could be greater if PREFERRED_F32_LANES > 16 at some future tier).
-        assert!(GLOBAL_RENDERER.tick_count() >= 0);
+        assert_eq!(GLOBAL_RENDERER.tick_count(), 0);
         let f = GLOBAL_RENDERER.read_front();
         assert!(f.capacity >= 4096);
     }

src/hpc/vml.rs

@@ -475,7 +475,7 @@ mod tests {
 
         // Generate "weight" data (deterministic, mimics Gaussian distribution)
         let weights: Vec<f64> = (0..d)
-            .map(|i| ((i as f64 * 0.7 + 13.0).sin() * 2.5))
+            .map(|i| (i as f64 * 0.7 + 13.0).sin() * 2.5)
             .collect();
 
         // ── ENCODING: f64[4096] → f64[17] (golden-step projection) ──

src/hpc/vnni_gemm.rs

@@ -315,7 +315,7 @@ mod tests {
         let n = 5;
         let k = 8;
         let a: Vec<u8> = (0..m * k).map(|i| (i % 200) as u8).collect();
-        let b: Vec<i8> = (0..k * n).map(|i| ((i % 100) as i8 - 50)).collect();
+        let b: Vec<i8> = (0..k * n).map(|i| (i % 100) as i8 - 50).collect();
         let expected = scalar_gemm(&a, &b, m, n, k);
         let mut c = vec![0i32; m * n];
         int8_gemm_vnni(&a, &b, &mut c, m, n, k);

src/simd_int_ops.rs

@@ -284,8 +284,8 @@ mod tests {
     #[test]
     fn add_i16_matches_scalar_for_tail_lengths() {
         for &len in &[0usize, 1, 31, 32, 33, 64, 65, 100] {
-            let a_init: Vec<i16> = (0..len).map(|i| (i as i16 * 7 - 1000)).collect();
-            let b: Vec<i16> = (0..len).map(|i| (i as i16 * -3 + 500)).collect();
+            let a_init: Vec<i16> = (0..len).map(|i| i as i16 * 7 - 1000).collect();
+            let b: Vec<i16> = (0..len).map(|i| i as i16 * -3 + 500).collect();
 
             let mut a_simd = a_init.clone();
             add_i16(&mut a_simd, &b);