diff --git a/.github/workflows/cuda-nightly.yml b/.github/workflows/cuda-nightly.yml
index 53dafd909d..c33e72b56d 100644
--- a/.github/workflows/cuda-nightly.yml
+++ b/.github/workflows/cuda-nightly.yml
@@ -707,6 +707,11 @@ jobs:
         if: ${{ !cancelled() && steps.decide.outputs.proceed == 'true' }}
         run: |
           set -uo pipefail
+          # -e OFF, as in the perf041 step: the default shell is `bash -e {0}`, so the
+          # RED control's expected exit 101 aborted the step at `; r=$?` before its
+          # assertions ran (run 37161229337: GREEN passed, step failed 101, no ::error::).
+          # Every failure path below is guarded explicitly.
+          set +e
           t=api::openai_handlers::qwen35_serve_tests::gpu_every_serve_route_prefills_through_the_batched_prefill
           model="$HOME/models/Qwen3.5-0.8B-Q4_K_M.gguf"
           [ -f "$model" ] || { echo "::error::#4250 UNMEASURABLE on ${{ matrix.name }} (provisioning): $model is absent"; exit 1; }
diff --git a/docs/roadmaps/entries/PMAT-4715.yaml b/docs/roadmaps/entries/PMAT-4715.yaml
new file mode 100644
index 0000000000..398c99140b
--- /dev/null
+++ b/docs/roadmaps/entries/PMAT-4715.yaml
@@ -0,0 +1,19 @@
+- id: PMAT-4715
+  github_issue: 4715
+  item_type: task
+  title: 'B2: root-cause fixes for red nightlies on main'
+  status: inprogress
+  priority: critical
+  assigned_to: aprender-wmainred
+  created: 2026-10-04T00:00:00Z
+  updated: 2026-10-04T00:00:00Z
+  acceptance_criteria:
+  - Each red or cancelled scheduled workflow on main has a root cause named in its fix PR, one PR per defect, listed on issue #4715
+  - Each fix PR changes no gate's acceptance; a threshold or acceptance change goes to the train lead's sign-off batch instead
+  - After each fix merges, the next scheduled or dispatched run of that workflow on main gets past the step that failed
+  phases: []
+  subtasks: []
+  estimated_effort: null
+  labels:
+  - kind:ci
+  notes: 'Parent #4678 (0.70.2), bar B2: no check red on main for more than 24 h. Proposed by aprender-a7, minted by cop aprender-14.'
diff --git a/docs/roadmaps/roadmap.yaml b/docs/roadmaps/roadmap.yaml
index f65016f6ce..6d12cec68b 100644
--- a/docs/roadmaps/roadmap.yaml
+++ b/docs/roadmaps/roadmap.yaml
@@ -24060,3 +24060,22 @@ Not done, pending the ruling: moving the 31 fixture files to an obviously synthe
   labels:
   - kind:docs
   notes: 'Operator C197, 2026-10-01 12:40Z: re-date via the spec''s own procedure; check unchanged; not a waiver.'
+- id: PMAT-4715
+  github_issue: 4715
+  item_type: task
+  title: 'B2: root-cause fixes for red nightlies on main'
+  status: inprogress
+  priority: critical
+  assigned_to: aprender-wmainred
+  created: 2026-10-04T00:00:00Z
+  updated: 2026-10-04T00:00:00Z
+  acceptance_criteria:
+  - Each red or cancelled scheduled workflow on main has a root cause named in its fix PR, one PR per defect, listed on issue #4715
+  - Each fix PR changes no gate's acceptance; a threshold or acceptance change goes to the train lead's sign-off batch instead
+  - After each fix merges, the next scheduled or dispatched run of that workflow on main gets past the step that failed
+  phases: []
+  subtasks: []
+  estimated_effort: null
+  labels:
+  - kind:ci
+  notes: 'Parent #4678 (0.70.2), bar B2: no check red on main for more than 24 h. Proposed by aprender-a7, minted by cop aprender-14.'