diff --git a/crates/ml-core/build.rs b/crates/ml-core/build.rs
index 22c156cd3..ad5a0afaa 100644
--- a/crates/ml-core/build.rs
+++ b/crates/ml-core/build.rs
@@ -29,6 +29,7 @@ fn main() {
     println!("cargo:rerun-if-changed={}", common_header_path.display());
 
     // Detect GPU architecture from env or default to sm_80
+    println!("cargo:rerun-if-env-changed=CUDA_COMPUTE_CAP");
     let cuda_compute_cap = std::env::var("CUDA_COMPUTE_CAP").unwrap_or_else(|_| "80".to_string());
     let arch = format!("sm_{cuda_compute_cap}");
 
diff --git a/crates/ml-dqn/build.rs b/crates/ml-dqn/build.rs
index 74486d572..853c774a7 100644
--- a/crates/ml-dqn/build.rs
+++ b/crates/ml-dqn/build.rs
@@ -21,7 +21,13 @@ fn main() {
 
     println!("cargo:rerun-if-changed={}", common_header_path.display());
 
-    // Detect GPU architecture from env or default to sm_80
+    // Detect GPU architecture from env or default to sm_80.
+    // CRITICAL: pair env::var() with rerun-if-env-changed per
+    // pearl_build_rs_rerun_if_env_changed, otherwise sccache + cubin caches
+    // hit-on-SHA across arch changes (e.g., SM 89 L40S → SM 90 H100), and
+    // training fails at runtime with CUDA_ERROR_NO_BINARY_FOR_GPU on rmsnorm
+    // cubin load (commit 0371d6a76 on H100, post-Class-B chain).
+    println!("cargo:rerun-if-env-changed=CUDA_COMPUTE_CAP");
     let cuda_compute_cap = std::env::var("CUDA_COMPUTE_CAP").unwrap_or_else(|_| "80".to_string());
     let arch = format!("sm_{cuda_compute_cap}");
 
diff --git a/crates/ml-ensemble/build.rs b/crates/ml-ensemble/build.rs
index 654739587..90a92d514 100644
--- a/crates/ml-ensemble/build.rs
+++ b/crates/ml-ensemble/build.rs
@@ -13,6 +13,7 @@ fn main() {
     let kernel_dir = Path::new("src");
 
     // Detect GPU architecture from env or default to sm_80
+    println!("cargo:rerun-if-env-changed=CUDA_COMPUTE_CAP");
     let cuda_compute_cap = std::env::var("CUDA_COMPUTE_CAP").unwrap_or_else(|_| "80".to_string());
     let arch = format!("sm_{cuda_compute_cap}");
 
diff --git a/crates/ml-explainability/build.rs b/crates/ml-explainability/build.rs
index 51c1765d0..799dbf628 100644
--- a/crates/ml-explainability/build.rs
+++ b/crates/ml-explainability/build.rs
@@ -13,6 +13,7 @@ fn main() {
     let kernel_dir = Path::new("src");
 
     // Detect GPU architecture from env or default to sm_80 (matches crates/ml/build.rs).
+    println!("cargo:rerun-if-env-changed=CUDA_COMPUTE_CAP");
     let cuda_compute_cap = std::env::var("CUDA_COMPUTE_CAP").unwrap_or_else(|_| "80".to_string());
     let arch = format!("sm_{cuda_compute_cap}");
 
diff --git a/crates/ml-ppo/build.rs b/crates/ml-ppo/build.rs
index 1da83cbe8..6c044ccb2 100644
--- a/crates/ml-ppo/build.rs
+++ b/crates/ml-ppo/build.rs
@@ -29,6 +29,7 @@ fn main() {
     println!("cargo:rerun-if-changed={}", common_header_path.display());
 
     // Detect GPU architecture from env or default to sm_80
+    println!("cargo:rerun-if-env-changed=CUDA_COMPUTE_CAP");
     let cuda_compute_cap = std::env::var("CUDA_COMPUTE_CAP").unwrap_or_else(|_| "80".to_string());
     let arch = format!("sm_{cuda_compute_cap}");
 
diff --git a/crates/ml-supervised/build.rs b/crates/ml-supervised/build.rs
index dd1f05508..ef2abb27e 100644
--- a/crates/ml-supervised/build.rs
+++ b/crates/ml-supervised/build.rs
@@ -25,6 +25,7 @@ fn main() {
     let manifest_dir = PathBuf::from(std::env::var("CARGO_MANIFEST_DIR").unwrap());
 
     // Detect GPU architecture from env or default to sm_80 (L40S / A100).
+    println!("cargo:rerun-if-env-changed=CUDA_COMPUTE_CAP");
     let cuda_compute_cap = std::env::var("CUDA_COMPUTE_CAP").unwrap_or_else(|_| "80".to_string());
     let arch = format!("sm_{cuda_compute_cap}");