diff --git a/Cargo.lock b/Cargo.lock
index 54548a61..a68c345e 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1152,6 +1152,7 @@ version = "0.7.0"
 dependencies = [
  "numpy",
  "pyo3",
+ "pyo3-build-config",
  "turbovec",
 ]
 
diff --git a/turbovec-python/Cargo.toml b/turbovec-python/Cargo.toml
index 9cc4a980..8cfc72cd 100644
--- a/turbovec-python/Cargo.toml
+++ b/turbovec-python/Cargo.toml
@@ -12,3 +12,6 @@ crate-type = ["cdylib"]
 turbovec-core = { package = "turbovec", path = "../turbovec" }
 pyo3 = { version = "0.27.0", features = ["extension-module", "abi3-py39"] }
 numpy = "0.27.0"
+
+[build-dependencies]
+pyo3-build-config = "0.27.0"
diff --git a/turbovec-python/build.rs b/turbovec-python/build.rs
new file mode 100644
index 00000000..88c0fea0
--- /dev/null
+++ b/turbovec-python/build.rs
@@ -0,0 +1,10 @@
+fn main() {
+    // Emit the platform-correct linker arguments for a Python extension
+    // module. On macOS this passes `-undefined dynamic_lookup` so symbols
+    // from the Python interpreter (e.g. `Py_True`) resolve at load time
+    // instead of failing the link step. Without it, a plain `cargo build`
+    // on macOS fails with "symbol(s) not found for architecture arm64"
+    // (issue #92). Building via maturin already injects these args; this
+    // makes a bare `cargo build` work too.
+    pyo3_build_config::add_extension_module_link_args();
+}
diff --git a/turbovec/src/search.rs b/turbovec/src/search.rs
index 4fda9433..078580ee 100644
--- a/turbovec/src/search.rs
+++ b/turbovec/src/search.rs
@@ -61,7 +61,6 @@ unsafe fn score_4bit_block_neon(
     for batch in 0..n_batches {
         let g_start = batch * FLUSH_EVERY;
         let g_end = (g_start + FLUSH_EVERY).min(n_byte_groups);
-        let n_groups = g_end - g_start;
 
         let mut accum = [vdupq_n_u16(0); 4];
 
@@ -1097,7 +1096,6 @@ unsafe fn score_4query_block_neon(
     for batch in 0..n_batches {
         let g_start = batch * FLUSH_EVERY;
         let g_end = (g_start + FLUSH_EVERY).min(n_byte_groups);
-        let n_groups = g_end - g_start;
 
         let mut acc: [[uint16x8_t; 4]; 4] = [[vdupq_n_u16(0); 4]; 4];
 
@@ -1343,6 +1341,10 @@ pub(crate) fn block_pair_has_allowed(mask: Option<&[u64]>, base_vec_pair: usize)
 /// VM / emulator that doesn't expose AVX2 to userspace). Without this
 /// fallback, pre-AVX2 x86_64 silently returned empty top-k results
 /// instead of falling back to a slower-but-correct kernel.
+///
+/// Not compiled on aarch64, where the NEON kernel is always available and
+/// this scalar path is never reached (it would warn as dead code).
+#[cfg(not(target_arch = "aarch64"))]
 #[allow(clippy::too_many_arguments)]
 fn score_query_into_heap(
     qlut_uint8: &[u8],