diff --git a/benchmarks/nemotron_3.5_super/README.md b/benchmarks/nemotron_3.5_super/README.md
index a6cc07b2a5..0ebea8f2e7 100644
--- a/benchmarks/nemotron_3.5_super/README.md
+++ b/benchmarks/nemotron_3.5_super/README.md
@@ -40,6 +40,24 @@ worker every 10 cycles regardless of load.
 - bug: https://github.com/vllm-project/router/issues/197
 - fix: https://github.com/vllm-project/router/pull/216 (unmerged upstream)
 
+#216 on its own is not enough, and for prefill-heavy benchmarks it is worse than not
+applying it. It makes the worker load counters honest, which switches on a second
+latent bug: `cache_aware` decides whether to use prefix affinity from the
+*fleet-wide* load spread, so one hot worker discards affinity for every request --
+including requests whose own worker is idle. Under P/D that gate is open almost
+permanently, because prefill worker load counts queued requests as well as running
+ones. Routing degenerates to shortest-queue, already-cached prompts get recomputed,
+prefill saturates and decode starves behind it.
+
+The pin therefore points at a branch carrying #216 plus a fix that applies the same
+load check per request, against the worker that request wants:
+
+- prefill fix: https://github.com/vllm-project/router/pull/238
+
+Both are plain commit SHAs fetched from `vllm-project/router`; a PR head is a ref
+there even when the branch lives on a contributor's fork. Repin to a released commit
+once these land upstream.
+
 Build the wheel once with `build_vllm_router_wheel.sh`, then pass it to the container
 build. The wheel is built inside the eval base image, so its extension module matches
 the Python that runs `vllm-router` at eval time:
diff --git a/benchmarks/nemotron_3.5_super/build_vllm_router_wheel.sh b/benchmarks/nemotron_3.5_super/build_vllm_router_wheel.sh
index d19f7d0c70..7067be2f55 100755
--- a/benchmarks/nemotron_3.5_super/build_vllm_router_wheel.sh
+++ b/benchmarks/nemotron_3.5_super/build_vllm_router_wheel.sh
@@ -30,10 +30,11 @@ set -euo pipefail
 CONTAINER=$CONTAINER
 OUTPUT_DIR=${OUTPUT_DIR:-$(pwd)/results/vllm_router}
 VLLM_ROUTER_GIT_URL=${VLLM_ROUTER_GIT_URL:-https://github.com/vllm-project/router}
-# Head of vllm-project/router#216 ("[Bugfix]: Fix health checker to preserve
-# worker load during checks"). Fetching a bare SHA works because GitHub serves
-# any commit reachable from a ref, and PR heads are refs.
-VLLM_ROUTER_COMMIT=${VLLM_ROUTER_COMMIT:-9e6fce282a877c65185468692c6ba8a483409d9b}
+# Head of vllm-project/router#238, which carries #216 ("[Bugfix]: Fix health checker
+# to preserve worker load during checks") unchanged plus a fix for the prefill-side
+# regression #216 exposes; see the README. Fetching a bare SHA works because GitHub
+# serves any commit reachable from a ref, and PR heads are refs.
+VLLM_ROUTER_COMMIT=${VLLM_ROUTER_COMMIT:-5b7ce08ea8cd27123cf6c23313cd92d3ba0da70e}
 RUST_TOOLCHAIN=${RUST_TOOLCHAIN:-1.95.0}
 # Load-accounting unit tests from the fix. Compiling the test binary roughly
 # doubles the job, so it is opt-out.