ci(miri): shard the run across eight jobs

.github/workflows/ci.yaml

@@ -393,9 +393,16 @@ jobs:
     # Only runs on merge queue / push to main — never blocks PR checks.
     if: github.event_name == 'merge_group' || github.event_name == 'push'
     runs-on: ubuntu-latest
-    name: miri
-    # A healthy run takes well under this; the per-test limit is in the
+    # Sharded: under Miri every test is its own interpreted process (about
+    # 30 s minimum each), so the ~1400 selected tests take 4-5 hours serial.
+    # The 2026-11-05 master run reached 590 tests in the 2-hour cap. Eight
+    # shards keep each well under the cap; the per-test limit is in the
     # `miri` nextest profile (.config/nextest.toml).
+    strategy:
+      fail-fast: false
+      matrix:
+        shard: [1, 2, 3, 4, 5, 6, 7, 8]
+    name: miri/${{ matrix.shard }}
     timeout-minutes: 120
     steps:
       - uses: actions/checkout@v4
@@ -405,6 +412,8 @@ jobs:
       - uses: Swatinem/rust-cache@v2
       - uses: taiki-e/install-action@nextest
       - run: ./scripts/miri-tests.sh
+        env:
+          MIRI_PARTITION: count:${{ matrix.shard }}/8
 
   cross_test:
     # Gated on merge_group only — cross-compile via docker (cross-rs) for

scripts/miri-tests.sh

@@ -56,8 +56,12 @@ export RUSTFLAGS="-Zrandomize-layout"
 # The polyfill is no longer the bottleneck — the missing piece is a
 # `cfg(miri)` switch in `src/simd.rs` that re-exports from `simd_nightly`
 # instead of `simd_avx*` under Miri.
+# CI shards the run (MIRI_PARTITION=count:K/N, see ci.yaml § miri): under Miri
+# each test is its own interpreted process, so the whole suite takes about
+# 4-5 hours on one runner. Unset, the full suite runs in one go.
 cargo +nightly miri nextest run -v --profile miri \
     --no-fail-fast \
+    ${MIRI_PARTITION:+--partition "$MIRI_PARTITION"} \
     -p ndarray -p ndarray-rand \
     --features approx,serde,nightly-simd \
     -E '!(