fix(dab): restart dab-postgres on-failure so a mid-run crash doesn't lose the DB (#12)

The dab-postgres service had a pg_isready healthcheck and a service_healthy
startup gate but no restart policy (unlike dab-mongo). When postgres:17 dies
mid-run (crash/OOM) it is never restarted, the container leaves dab-net, and
`dab-postgres` stops resolving for the rest of the trial ('could not translate
host name dab-postgres'). For hybrid pg+duckdb datasets like PANCANCER_ATLAS the
clinical data then becomes unreachable and the agent can only abstain — observed
on PANCANCER_ATLAS q2/q3 (both 'UNABLE TO DETERMINE', reward 0), which alone
account for ~0.056 of stratified pass@1.

Add 'restart: on-failure' to dab-postgres, mirroring the dab-mongo fix, so the
populated data dir comes straight back up and main's healthcheck recovers.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: kentwelcome

packages/razorback-plugin-dab/src/razorback_plugin_dab/generate/compose.py

@@ -116,6 +116,12 @@ def generate_compose(
                 "POSTGRES_PASSWORD": POSTGRES_PASSWORD,
                 "POSTGRES_DB": pg_dbs[0],
             },
+            # postgres:17 can die mid-run (crash/OOM). Without a restart policy the
+            # container leaves dab-net and `dab-postgres` stops resolving for the rest
+            # of the trial — clinical data (e.g. PANCANCER_ATLAS) becomes unreachable
+            # and the agent can only abstain. Restart on failure so the populated data
+            # dir comes straight back up and the healthcheck recovers (mirrors dab-mongo).
+            "restart": "on-failure",
             "healthcheck": {
                 "test": ["CMD-SHELL", f"pg_isready -U {POSTGRES_USER} -d {pg_dbs[0]}"],
                 "interval": "5s",

packages/razorback-plugin-dab/tests/unit/test_compose_postgres.py

@@ -40,6 +40,20 @@ def test_postgres_service_emitted(tmp_path: Path):
     assert "dab-net" in pg["networks"]
 
 
+def test_postgres_has_restart_policy(tmp_path: Path):
+    # postgres:17 can die mid-run (crash/OOM). Without restart-on-failure the
+    # container leaves dab-net and `dab-postgres` stops resolving for the rest of
+    # the trial, so clinical data becomes unreachable and the agent can only
+    # abstain (observed on PANCANCER_ATLAS q2/q3). Mirror the dab-mongo fix.
+    compose_text = generate_compose(
+        db_config=_BOOKREVIEW_LIKE,
+        dataset_name="bookreview",
+        data_root=tmp_path,
+    )
+    pg = yaml.safe_load(compose_text)["services"]["dab-postgres"]
+    assert pg["restart"] == "on-failure"
+
+
 def test_main_service_depends_on_postgres(tmp_path: Path):
     compose_text = generate_compose(
         db_config=_BOOKREVIEW_LIKE,