fix(release): keep large snapshots attached and low impact

.gitea/workflows/release-gates.yml

@@ -11,7 +11,10 @@ permissions:
 
 concurrency:
   group: geointel-release-${{ gitea.ref }}
-  cancel-in-progress: true
+  # A cancelled HTTP caller does not terminate the allowlisted controller
+  # process that already owns the production lock. Queue a newer revision
+  # instead of orphaning an in-flight backup or deploy.
+  cancel-in-progress: false
 
 jobs:
   quality:
@@ -162,7 +165,10 @@ jobs:
     needs: [quality, dependency-audit, container]
     if: ${{ gitea.event_name == 'push' && gitea.ref == 'refs/heads/main' }}
     runs-on: unraid-deploy
-    timeout-minutes: 180
+    # The first byte-complete storage snapshot can exceed 100 GiB. Keep the
+    # gated caller attached for the full conservative backup/build window;
+    # the controller and deploy script still serialize every mutation.
+    timeout-minutes: 720
     steps:
       - name: Deploy only after every release gate is green
         run: |

backend/tests/test_rc5_release_deployment.py

@@ -53,6 +53,20 @@ def test_release_creates_verified_backup_before_candidate_migrations() -> None:
     assert "preflight_backup_capacity" in script
     assert "select_verified_link_dest" in script
     assert "release_backup_snapshot.py" in script
+    assert "run_low_impact()" in script
+    assert "ionice -c 2 -n 7" in script
+    assert "nice -n 10" in script
+    assert "run_low_impact bash scripts/backup_release_state.sh" in script
+    assert "run_low_impact bash scripts/verify_release_backup.sh" in script
+
+
+def test_gitea_deploy_waits_for_the_mandatory_large_snapshot() -> None:
+    workflow = (ROOT / ".gitea" / "workflows" / "release-gates.yml").read_text(encoding="utf-8")
+
+    assert "cancel-in-progress: false" in workflow
+    deploy_job = workflow.split("\n  deploy:\n", maxsplit=1)[1]
+    assert "timeout-minutes: 720" in deploy_job
+    assert "docker exec gitea-deploy-control" in deploy_job
 
 
 def test_release_starts_only_the_locally_attested_ai_image() -> None:

deploy/unraid/deploy-release.sh

@@ -355,14 +355,34 @@ print(
 PY
 }
 
+run_low_impact() {
+  local priority_command=()
+
+  # Backups are mandatory, but their first byte-complete copy and SHA-256
+  # verification must not starve the live Unraid services. BusyBox hosts do
+  # not always provide both tools, so use every available scheduler without
+  # weakening the backup when one is absent.
+  if command -v ionice >/dev/null 2>&1; then
+    priority_command+=(ionice -c 2 -n 7)
+  fi
+  if command -v nice >/dev/null 2>&1; then
+    priority_command+=(nice -n 10)
+  fi
+  if [ "${#priority_command[@]}" -eq 0 ]; then
+    "$@"
+    return
+  fi
+  "${priority_command[@]}" "$@"
+}
+
 select_verified_link_dest() {
   local candidate=""
   GEOINTEL_BACKUP_LINK_DEST=""
   while IFS= read -r candidate; do
     if (
       cd "$candidate" \
-        && sha256sum -c CHECKSUMS.sha256 >/dev/null \
-        && python3 "$ROOT/scripts/release_backup_snapshot.py" verify-backup --backup-dir "$candidate"
+        && run_low_impact sha256sum -c CHECKSUMS.sha256 >/dev/null \
+        && run_low_impact python3 "$ROOT/scripts/release_backup_snapshot.py" verify-backup --backup-dir "$candidate"
     ); then
       GEOINTEL_BACKUP_LINK_DEST="$candidate"
       echo "Using verified prior byte snapshot as link-dest: ${candidate}"
@@ -482,7 +502,7 @@ PY
 
   GEOINTEL_PREDEPLOY_BACKUP_DIR="${GEOINTEL_BACKUPS_PATH%/}/${release_id}"
   echo "Creating mandatory pre-deploy backup ${release_id}..."
-  if ! bash scripts/backup_release_state.sh \
+  if ! run_low_impact bash scripts/backup_release_state.sh \
       --container geointel \
       --output-root "$GEOINTEL_BACKUPS_PATH" \
       --release-id "$release_id" \
@@ -491,7 +511,7 @@ PY
       --inventory-mode sha256 \
       --rollback-image-tag "$GEOINTEL_PREDEPLOY_ROLLBACK_TAG" \
       "${backup_link_args[@]}" \
-    || ! bash scripts/verify_release_backup.sh \
+    || ! run_low_impact bash scripts/verify_release_backup.sh \
       --container geointel \
       --backup-dir "$GEOINTEL_PREDEPLOY_BACKUP_DIR"; then
     echo "Pre-deploy backup failed; restarting the unchanged current release." >&2

docs/RELEASE_RUNBOOK.md

@@ -80,6 +80,15 @@ Create an immutable byte-complete backup. Initial storage/model copy and
 verification can be I/O-heavy; subsequent backups deduplicate unchanged bytes
 against the newest verified prior snapshot:
 
+The gated Unraid job keeps its controller request attached for up to twelve
+hours because an initial snapshot can exceed 100 GiB. Release runs for the same
+branch queue instead of cancelling an in-flight deploy. Snapshot copying and
+checksum verification use the lowest available best-effort CPU and I/O
+priority (`nice` and `ionice`) while the deploy lock prevents overlap. If an
+operator manually cancels the caller, inspect the controller state and
+`/mnt/user/appdata/gitea-deploy-runners/deploy.lock` before retrying; never
+assume that cancelling the HTTP client terminated the server-side operation.
+
 ```bash
 bash scripts/backup_release_state.sh \
   --container geointel \