diff --git a/.github/workflows/coverage.yml b/.github/workflows/coverage.yml
index f1b8eb9f3a2..d24a7293669 100644
--- a/.github/workflows/coverage.yml
+++ b/.github/workflows/coverage.yml
@@ -8,6 +8,12 @@ on:
   schedule:
     - cron: '0 0 * * *' # This runs the workflow every day at 12:00 AM UTC
   workflow_dispatch: {}
+  # TEMPORARY (do not merge): runs this workflow against a PR through the
+  # copy-pr-bot mirror branch, so the wheel repair below can be verified
+  # before it is proposed on its own. Same trigger as ci.yml.
+  push:
+    branches:
+      - "pull-request/[0-9]+"
 
 env:
   PY_VER: "3.14"
@@ -32,6 +38,9 @@ jobs:
   coverage-linux:
     name: Coverage (Linux)
     needs: [coverage-vars]
+    # TEMPORARY (do not merge): only Windows is under test here, and the Linux
+    # leg holds an A100 for half an hour.
+    if: ${{ !startsWith(github.ref, 'refs/heads/pull-request/') }}
     runs-on: "linux-amd64-gpu-a100-latest-1"
     permissions:
       id-token: write
@@ -235,13 +244,64 @@ jobs:
           cd cuda_bindings
           ../.venv/Scripts/pip wheel -v --no-deps . -w ../wheels/
 
+      # PIP_PRE is set so pip will consider the cuda-bindings wheel built one
+      # step ago: it carries a .devN version, which counts as a pre-release.
+      # The side effect is that PyPI's pre-releases become eligible too, and
+      # since cuda-bindings 13.4.0b1 was published on 2026-07-29 it outranks
+      # the local build for `cuda-bindings==13.*`.  Its cydriver.pxd is
+      # generated from CTK 13.4 headers, where CUmemLocation has a `localized`
+      # field, while cuda.core here compiles against the 13.3.0 mini-CTK, where
+      # it does not -- and Cython generates a struct converter over every
+      # field, so the build fails on `error C2039`.  Constraining the version
+      # keeps what PIP_PRE is for and drops what it lets in.
+      #
+      # The local segment is stripped because PEP 440 has a public `==`
+      # specifier ignore a candidate's local label: `==13.3.2.dev152` still
+      # selects 13.3.2.dev152+g0123456, and constraints files stay free of
+      # local versions.
       - name: Build cuda.core wheel
         run: |
           export PIP_FIND_LINKS="$(pwd)/wheels"
           export PIP_PRE=1
+          bindings_whl="$(ls ./wheels/cuda_bindings-*.whl | head -1)"
+          bindings_ver="$(basename "$bindings_whl" | cut -d- -f2)"
+          echo "cuda-bindings==${bindings_ver%%+*}" > "$GITHUB_WORKSPACE/constraints.txt"
+          cat "$GITHUB_WORKSPACE/constraints.txt"
+          export PIP_CONSTRAINT="$GITHUB_WORKSPACE/constraints.txt"
           cd cuda_core
           ../.venv/Scripts/pip wheel -v --no-deps . -w ../wheels/
 
+      # Every other Windows build repairs its wheels and this one never did.
+      # cibuildwheel runs each package's repair-wheel-command from
+      # pyproject.toml, which on Windows is delvewheel: it copies dependent
+      # DLLs into <package>.libs under content-hashed names, rewrites the .pyd
+      # import tables to match, and registers the directory from the package
+      # __init__.  A plain `pip wheel` never reads [tool.cibuildwheel], so
+      # these wheels imported a bare "MSVCP140.dll" and resolved it against
+      # whatever the test machine had in System32 -- 14.00.24215.1, from 2015,
+      # on the runner this job uses.
+      #
+      # That is fatal rather than cosmetic.  _resource_handles.pyd, compiled by
+      # MSVC 14.44, imports only _Mtx_lock and _Mtx_unlock and never
+      # _Mtx_init_in_situ, because std::mutex has had a constexpr constructor
+      # since VS 2022 17.10.  The 2015 runtime still expects that
+      # initialisation and dereferences a null handle on the first lock, which
+      # _stream.pyx takes while cuda.core is still importing: every Windows
+      # coverage run since 2026-03-17 died there, before collecting a test.
+      #
+      # --namespace-pkg cuda is required; `cuda` is a namespace package and
+      # delvewheel otherwise patches the wrong __init__.
+      - name: Repair the Windows wheels
+        run: |
+          .venv/Scripts/pip install delvewheel
+          mkdir -p wheels-repaired
+          for whl in ./wheels/cuda_bindings-*.whl ./wheels/cuda_core-*.whl; do
+            .venv/Scripts/delvewheel repair --namespace-pkg cuda \
+              --exclude "torch_cpu.dll;torch_python.dll" \
+              -w ./wheels-repaired "$whl"
+          done
+          mv -f ./wheels-repaired/*.whl ./wheels/
+
       - name: List wheel artifacts
         run: |
           echo "=== Windows wheel artifacts ==="
@@ -393,7 +453,10 @@ jobs:
     name: Combine Coverage and Deploy
     needs: [coverage-linux, coverage-windows]
     runs-on: ubuntu-latest
-    if: always()
+    # TEMPORARY (do not merge): a PR run has no Linux half, so this would both
+    # fail on the missing .coverage.linux and publish a partial report to
+    # gh-pages.
+    if: ${{ always() && !startsWith(github.ref, 'refs/heads/pull-request/') }}
     permissions:
       id-token: write
       contents: write