infra: run frozen V121 from seeded transcript cache

Author: heathsanchez

.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml

@@ -8,6 +8,11 @@ on:
       - '.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml'
   workflow_dispatch:
 
+# Infrastructure-only serialization. Scientific protocol is unchanged.
+concurrency:
+  group: trace-ace-frozen-transcripts-v1
+  cancel-in-progress: false
+
 jobs:
   semantic:
     runs-on: ubuntu-24.04
@@ -21,25 +26,44 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn gdown fastembed==0.8.0
-      - name: Download frozen metadata and transcripts
+
+      - name: Restore frozen transcript archive from GitHub cache
+        uses: actions/cache/restore@v4
+        with:
+          path: transcripts.zip
+          key: trace-ace-transcripts-v1-603547640
+          fail-on-cache-miss: true
+
+      - name: Validate frozen transcript archive
+        shell: bash
         run: |
-          set -e
+          set -euo pipefail
+          test -f transcripts.zip
+          test "$(stat -c%s transcripts.zip)" = "603547640"
+          unzip -tq transcripts.zip >/dev/null
+          ACTUAL=$(sha256sum transcripts.zip | cut -d' ' -f1)
+          test "$ACTUAL" = "e685b85b04694e130c25b17d09cdd1892fbda5e9fa685e98b2300114b915aa2d"
+          echo "Frozen transcript cache verified: $ACTUAL"
+
+      - name: Download frozen metadata and extract data
+        shell: bash
+        run: |
+          set -euo pipefail
           gdown 1EpqoamY0vFI2qE57R6wdqU5HwuoVk3Zz -O metadata.zip
-          if ! gdown 1nOjremWhpZ_QKSLvZfGcNkS_C3kMMBUI -O transcripts.zip; then
-            echo 'primary frozen transcript transport failed; trying frozen mirror'
-            gdown 1UWuHKIJ86yINhQOBGx2fFJ890hiEa6Md -O transcripts.zip
-          fi
-          sha256sum metadata.zip transcripts.zip
           mkdir -p data/meta data/transcripts
           unzip -q metadata.zip -d data/meta
           unzip -q transcripts.zip -d data/transcripts
           FEATURES=$(find data/meta -type f -name 'train_features*.csv' -print -quit)
           LABELS=$(find data/meta -type f -name 'train_labels*.csv' -print -quit)
           FIRST=$(find data/transcripts -type f -name '*.csv' -print -quit)
+          test -n "$FEATURES"
+          test -n "$LABELS"
+          test -n "$FIRST"
           TRANSCRIPTS=$(dirname "$FIRST")
           echo "FEATURES=$FEATURES" >> "$GITHUB_ENV"
           echo "LABELS=$LABELS" >> "$GITHUB_ENV"
           echo "TRANSCRIPTS=$TRANSCRIPTS" >> "$GITHUB_ENV"
+
       - name: Preflight
         run: |
           python -m py_compile competitions/trace_the_ace/v121_pretrained_semantic_residual.py
@@ -58,7 +82,8 @@ jobs:
             --rows 2500 \
             --out ../../v121_pretrained_semantic_residual.json
       - name: Show decision
-        run: cat v121_pretrained_semantic_residual.json
+        if: always()
+        run: test -f v121_pretrained_semantic_residual.json && cat v121_pretrained_semantic_residual.json || true
       - uses: actions/upload-artifact@v4
         if: always()
         with: