Add on_retry_command to clean build state between retry attempts

When a build is killed mid-way (e.g., SIGTERM), the build directory is
left in a corrupted state. Subsequent retry attempts fail immediately
because set -e exits on the first error from the dirty build tree.
Running ./mfc.sh clean between retries restores a fresh build state.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: sbryngelson

.github/workflows/bench.yml

@@ -110,6 +110,10 @@ jobs:
             (cd pr     && ${{ matrix.build_script }}) &
             (cd master && ${{ matrix.build_script }}) &
             wait %1 && wait %2
+          on_retry_command: |
+            (cd pr     && ./mfc.sh clean) &
+            (cd master && ./mfc.sh clean) &
+            wait
 
       - name: Bench (Master v. PR)
         run: bash pr/.github/scripts/run_parallel_benchmarks.sh ${{ matrix.device }} ${{ matrix.interface }} ${{ matrix.cluster }}

.github/workflows/test.yml

@@ -257,6 +257,7 @@ jobs:
           retry_wait_seconds: 60
           timeout_minutes: 480
           command: bash .github/workflows/${{ matrix.cluster }}/build.sh ${{ matrix.device }} ${{ matrix.interface }}
+          on_retry_command: ./mfc.sh clean
 
       - name: Test
         run:  bash .github/workflows/${{ matrix.cluster }}/submit.sh .github/workflows/${{ matrix.cluster }}/test.sh ${{ matrix.device }} ${{ matrix.interface }} ${{ matrix.shard }}