fix(dev): switch Keycloak to dev-file + persistent volume

KC_DB was set to dev-mem (H2 in-memory) — every time the keycloak
container went idle or restarted, the realm + the M7.1 test users
dissolved and the next smoke run failed with token_not_found until
realm-export was re-imported manually. Bit twice during the M7.1
testing session.

dev-file points H2 at /opt/keycloak/data/h2/ and the new
keycloak-data named volume keeps it across container lifecycles. The
realm-export.json bind-mount is still read-only and re-imported on
fresh starts via `--import-realm`, so a `docker compose down -v` is
still the canonical "rebuild from scratch" lever.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

docker-compose.yml

@@ -5,7 +5,11 @@ services:
     environment:
       KC_BOOTSTRAP_ADMIN_USERNAME: admin
       KC_BOOTSTRAP_ADMIN_PASSWORD: admin
-      KC_DB: dev-mem
+      # dev-file persists Keycloak's H2 DB under /opt/keycloak/data/h2/
+      # so the realm + test users survive container restarts. With the
+      # previous dev-mem the realm dissolved every time KC went idle,
+      # breaking local M7.1 smoke testing.
+      KC_DB: dev-file
       KC_HEALTH_ENABLED: "true"
     ports:
       - "8080:8080"
@@ -15,6 +19,7 @@ services:
     volumes:
       - ./keycloak/realm-export.json:/opt/keycloak/data/import/realm-export.json:ro
       - ./keycloak/themes/certifai:/opt/keycloak/themes/certifai:ro
+      - keycloak-data:/opt/keycloak/data/h2
     healthcheck:
       test: ["CMD-SHELL", "exec 3<>/dev/tcp/localhost/8080 && echo -e 'GET /realms/master HTTP/1.1\\r\\nHost: localhost\\r\\nConnection: close\\r\\n\\r\\n' >&3 && head -1 <&3 | grep -q '200 OK'"]
       interval: 10s
@@ -256,3 +261,4 @@ volumes:
   langfuse-db-data:
   langfuse-clickhouse-data:
   langfuse-clickhouse-logs:
+  keycloak-data: