diff --git a/changelog.d/8274-next-gate-native-backend.md b/changelog.d/8274-next-gate-native-backend.md
new file mode 100644
index 0000000000..02117b1882
--- /dev/null
+++ b/changelog.d/8274-next-gate-native-backend.md
@@ -0,0 +1,17 @@
+### Fixed
+
+- The production Next.js App Route dylib gate now exercises the **default**
+  codegen backend. `tests/test_next_app_route_dylib.sh` pinned
+  `PERRY_LLVM_INPROCESS="${PERRY_LLVM_INPROCESS:-0}"`, which selected the text
+  transport and — because `${VAR:-0}` cannot express "unset" — left no way for
+  the gate to run the configuration users actually get. The pin was reasonable
+  while #8228 stopped the native in-process path compiling five of this
+  fixture's modules; #8241 fixed that. The variable is now forwarded only when
+  the caller sets it, so an explicit backend can still be selected for a
+  bisection while the unset default reaches the compiler unchanged.
+
+  Measured on `183d30c53a` before landing this: with the native backend the
+  fixture compiles 104/104 modules and the gate passes 100/100 verifier
+  repetitions twice, with `freeze/LLVM pipeline started` appearing 5× per
+  compile (the five split modules #8228 could not build) versus 0× on the text
+  path — so the gate now demonstrably runs the backend it claims to.
diff --git a/tests/test_next_app_route_dylib.sh b/tests/test_next_app_route_dylib.sh
index f3751b626f..3e8eecc6ab 100755
--- a/tests/test_next_app_route_dylib.sh
+++ b/tests/test_next_app_route_dylib.sh
@@ -170,10 +170,18 @@ app="$providers/next-app-route.$library_extension"
 compile_log="$scratch/compile.log"
 (
     cd "$fixture"
+    # The backend is deliberately NOT pinned: this gate exists to exercise the
+    # configuration users get, which is the native in-process path (default ON
+    # wherever the runtime can walk the frames). It was pinned to the text
+    # transport while #8228 made the native path unable to compile five of this
+    # fixture's modules; #8241 fixed that, and a pin of `${VAR:-0}` cannot
+    # express "unset", so the pin left this gate structurally unable to test the
+    # default. `PERRY_LLVM_INPROCESS` is forwarded only when the caller sets it,
+    # so a bisection can still select a backend explicitly.
     env \
         PERRY_NO_AUTO_OPTIMIZE=1 \
         PERRY_DISABLE_WELL_KNOWN=1 \
-        PERRY_LLVM_INPROCESS="${PERRY_LLVM_INPROCESS:-0}" \
+        ${PERRY_LLVM_INPROCESS:+PERRY_LLVM_INPROCESS="$PERRY_LLVM_INPROCESS"} \
         PERRY_LLVM_CLANG="$llvm_clang" \
         PERRY_LLVM_OPT="$llvm_opt" \
         PERRY_LL_SIZE_OPT="${PERRY_LL_SIZE_OPT:-0}" \