fix(ci): make DeviceOffloadPolicy Default derivable (clippy -D warnings)

.squad/decisions/inbox/cohaagen-cuda-clippy-hotfix.md

@@ -0,0 +1,39 @@
+# CI hotfix — make `DeviceOffloadPolicy` `Default` derivable (clippy `-D warnings`)
+
+- Date: 2026-08-01
+- By: Cohaagen
+- Branch: `squad/fix-cuda-clippy-derivable-default`
+- Lineage: unblocks origin/main "CUDA compile (Linux+Windows)"; regression introduced by #572
+
+## Root cause
+
+#572 flipped `DeviceOffloadPolicy::async_pagein`'s default to `false`. That made
+the hand-written `impl Default` fully equal to a derivable default (all three
+fields = their type default: `bool` → `false`, `Option` → `None`), so
+`clippy::derivable_impls` fires under the CI's `-D warnings`, turning every PR's
+CUDA-compile job RED (weight_paging.rs:117).
+
+## Fix (behavior-identical)
+
+Replaced the hand-written `impl Default for DeviceOffloadPolicy` with
+`#[derive(Default)]` on the struct. Verified byte-identical: the three field
+defaults are exactly `enabled: false`, `device_budget_bytes: None`,
+`async_pagein: false` — which is precisely what the derive produces. Doc comments
+on fields preserved. `from_env()` and all runtime behavior unchanged. Net diff:
+1 insertion, 11 deletions (one file).
+
+## Green confirmation
+
+- `cargo clippy -p onnx-runtime-ep-cuda --features cuda -- -D warnings` → **exit 0**
+  (was the failing gate; no other `derivable_impls`/`-D warnings` hits in the crate).
+- CI first step `cargo check --locked -p onnx-runtime-ep-cuda -p onnx-runtime-python
+  --features onnx-runtime-python/cuda` → **exit 0**.
+- `cargo fmt --all --check` → clean.
+- Guard test `weight_paging::tests::device_policy_defaults_to_disabled` (already
+  present, :678) asserts `enabled=false`, `device_budget_bytes=None`,
+  `async_pagein=false` → **passes**, locking the derived defaults against field
+  default-drift. No new test needed.
+
+## Verdict
+
+Trivial, provably behavior-identical lint hotfix. Unblocks main (and #574).

crates/onnx-runtime-ep-cuda/src/weight_paging.rs

@@ -102,7 +102,7 @@ pub(crate) fn async_pagein_from_env_value(value: Option<&str>) -> bool {
 /// Whether/how the CUDA EP should page offloaded weights into a bounded VRAM
 /// residency cache. Disabled by default so the resident fast path is untouched
 /// and byte-identical.
-#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Default)]
 pub struct DeviceOffloadPolicy {
     pub enabled: bool,
     /// Explicit VRAM budget in bytes, if the operator pinned one.
@@ -114,16 +114,6 @@ pub struct DeviceOffloadPolicy {
     pub async_pagein: bool,
 }
 
-impl Default for DeviceOffloadPolicy {
-    fn default() -> Self {
-        Self {
-            enabled: false,
-            device_budget_bytes: None,
-            async_pagein: false,
-        }
-    }
-}
-
 impl DeviceOffloadPolicy {
     /// Read the policy from the process environment.
     pub fn from_env() -> Self {