fix(ci): set PGDATABASE=te_mgmt_dev, add Tier E/X evals, fix step order

.github/workflows/quality-gate.yml

@@ -77,7 +77,8 @@ jobs:
       PGPORT: "5432"
       PGUSER: postgres
       PGPASSWORD: postgres
-      PGDATABASE: postgres
+      PGDATABASE: te_mgmt_dev
+      API_ALLOW_DESTRUCTIVE: "true"
     steps:
       - name: Checkout
         uses: actions/checkout@v7
@@ -138,15 +139,18 @@ jobs:
             fi
           done
 
-      - name: Evals — Tiers P, I, S
-        run: python3 evals/runner.py --tiers p,i,s --verbose
-
       - name: Deploy all environments (for cross-env parity)
         run: |
           for env in dev test staging prod; do
             bash build/deploy_all.sh "$env"
           done
 
+      - name: Evals — Tiers X, E (post-deploy, before test artifacts)
+        run: python3 evals/runner.py --tiers x,e --verbose
+
+      - name: Evals — Tiers P, I, S
+        run: python3 evals/runner.py --tiers p,i,s --verbose
+
       # Prints a final block accounting for every test: PASSED / FAILED /
       # ERROR / SKIPPED (with reasons) / NOT RUN (deselected). --strict fails
       # the build on any skip.

evals/datasets/tier_e/01_all_envs_same_tables/NOTES.txt

@@ -0,0 +1,10 @@
+Tier E — Cross-environment structural parity.
+
+After all four environments (dev, test, staging, prod) have been deployed,
+their te_core_schema tables must be structurally identical: same table names,
+same column names, same column types, same column order.
+
+This scenario queries information_schema.columns for each environment and
+asserts the structural fingerprints match.
+
+Requires: PostgreSQL reachable, all four environment databases deployed.

evals/datasets/tier_e/02_empty_schema_is_detected/NOTES.txt

@@ -0,0 +1,3 @@
+Negative control: verify that _get_schema_fingerprint returns zero tables
+for an empty schema, so 01_all_envs_same_tables would correctly fail if a
+deploy silently wrote nothing.

evals/datasets/tier_x/01_csv_round_trip_postgresql/NOTES.txt

@@ -0,0 +1,11 @@
+Tier X — CSV round-trip through PostgreSQL.
+
+Loads each sample CSV (build/csv/samples/*.csv) into the dev database via
+csv_loader.sh, exports it back via csv_utilise.sh export, and diffs the
+data columns against the original.  Marker columns (_csv_row_id, _loaded_at)
+are excluded from the diff.
+
+Proves that the loader → DB → export pipeline preserves data fidelity for
+arbitrary CSV shapes.
+
+Requires: PostgreSQL reachable via psql, config.local.env present.

evals/expected/tier_e/01_all_envs_same_tables.json

@@ -0,0 +1,9 @@
+{
+  "scenario": "01_all_envs_same_tables",
+  "description": "All four environments must have identical table structure (names, columns, types).",
+  "expected": {
+    "all_envs_match": true,
+    "min_envs_compared": 4,
+    "min_tables_checked": 12
+  }
+}

evals/expected/tier_e/02_empty_schema_is_detected.json

@@ -0,0 +1,7 @@
+{
+  "scenario": "02_empty_schema_is_detected",
+  "description": "Negative control: an empty schema must report 0 tables so the parity eval cannot false-positive on missing deploys.",
+  "expected": {
+    "empty_schema_tables": 0
+  }
+}

evals/expected/tier_x/01_csv_round_trip_postgresql.json

@@ -0,0 +1,8 @@
+{
+  "scenario": "01_csv_round_trip_postgresql",
+  "description": "Load sample CSVs into PostgreSQL, export back, and diff data columns against originals.",
+  "expected": {
+    "all_round_trips_match": true,
+    "min_csvs_tested": 3
+  }
+}