Fix race conditions and cleanup in build cache

- Only remove build/staging (not build/install) on retry, so concurrent
  test jobs reading installed binaries are not disrupted
- Remove stale symlink in lock-timeout fallback path to prevent writing
  into the shared cache without holding the lock
- Remove redundant flock --unlock (closing fd is sufficient)

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

Author: sbryngelson

.github/scripts/setup-build-cache.sh

@@ -48,6 +48,11 @@ if flock --timeout 3600 9; then
 else
     echo "  WARNING: Cache lock timeout (1h), building locally without cache"
     exec 9>&-
+    # Remove any existing symlink to the shared cache so we don't write
+    # into it without the lock. Then create a real local directory.
+    if [ -L "build" ]; then
+        rm -f "build"
+    fi
     mkdir -p "build"
     echo "========================="
     return 0 2>/dev/null || true

.github/workflows/frontier/build.sh

@@ -49,7 +49,7 @@ while [ $attempt -le $max_attempts ]; do
 
     if [ $attempt -lt $max_attempts ]; then
         echo "Build failed on attempt $attempt. Clearing staging/install and retrying in 30s..."
-        rm -rf build/staging build/install build/lock.yaml
+        rm -rf build/staging build/lock.yaml
         sleep 30
     fi
     attempt=$((attempt + 1))

.github/workflows/frontier_amd/build.sh

@@ -49,7 +49,7 @@ while [ $attempt -le $max_attempts ]; do
 
     if [ $attempt -lt $max_attempts ]; then
         echo "Build failed on attempt $attempt. Clearing staging/install and retrying in 30s..."
-        rm -rf build/staging build/install build/lock.yaml
+        rm -rf build/staging build/lock.yaml
         sleep 30
     fi
     attempt=$((attempt + 1))

.github/workflows/phoenix/test.sh

@@ -24,7 +24,7 @@ while [ $attempt -le $max_attempts ]; do
 
     if [ $attempt -lt $max_attempts ]; then
         echo "Build failed on attempt $attempt. Clearing staging/install and retrying in 30s..."
-        rm -rf build/staging build/install build/lock.yaml
+        rm -rf build/staging build/lock.yaml
         sleep 30
     else
         echo "Build failed after $max_attempts attempts."
@@ -36,7 +36,6 @@ done
 # Release the cache lock before running tests. Tests only read installed
 # binaries and can take hours — no need to block other builds.
 if [ "${_cache_locked:-false}" = true ]; then
-    flock --unlock 9
     exec 9>&-
     echo "Released build cache lock before tests"
 fi