diff --git a/changelog.d/9777-build-cache-concat-site-cache.md b/changelog.d/9777-build-cache-concat-site-cache.md
new file mode 100644
index 0000000000..fa6c1dedf7
--- /dev/null
+++ b/changelog.d/9777-build-cache-concat-site-cache.md
@@ -0,0 +1,10 @@
+**`PERRY_CONCAT_SITE_CACHE` is now a build-cache input.** #9514's per-site
+concat cache reads the variable as a build-time kill switch — setting it to
+`0` removes the lowering lane entirely — but it was registered neither in
+`BUILD_CACHE_ENV_VARS` nor as a justified exclusion. A build with the switch
+flipped could therefore be served a cached object produced with it in the
+other state.
+
+`codegen_env_vars_are_build_cache_inputs` caught this by scanning
+`crates/perry-codegen/src` for every `env::var("PERRY_…")`, which is why the
+check scans the source instead of trusting a hand-maintained list.
diff --git a/crates/perry/src/commands/compile/build_cache.rs b/crates/perry/src/commands/compile/build_cache.rs
index 285918a5f9..340efb9a06 100644
--- a/crates/perry/src/commands/compile/build_cache.rs
+++ b/crates/perry/src/commands/compile/build_cache.rs
@@ -94,6 +94,11 @@ const BUILD_CACHE_ENV_VARS: &[&str] = &[
     // #9026: gates the once-per-closure-entry resolution of read-only boxed
     // capture cells — flipping it changes every closure body that qualifies.
     "PERRY_BOX_CAPTURE_ENTRY_CELLS",
+    // #9514: gates the per-site concat cache. `PERRY_CONCAT_SITE_CACHE=0`
+    // removes the lane at build time, so a qualifying `"literal" + value`
+    // site lowers to a different sequence with it on and off and a cached
+    // object from one setting must not serve the other.
+    "PERRY_CONCAT_SITE_CACHE",
     // The guarded-preinline IR-size ceiling: functions on either side of the
     // budget inline differently, so a run with a raised ceiling must not be
     // served objects a default run produced (same rule as the RS4GC budget