fix(selfhost): attach a pg.Pool error listener to prevent uncaught-exception crashes

GITTENSORY-1R/1S ("uncaughtException: terminating connection due to
administrator command") and GITTENSORY-1T ("boot: Postgres not ready after
30000ms") are one incident, traced on the same container within a 29-minute
window: Postgres itself was restarted while the app held live pooled
connections. node-postgres crashes the WHOLE process with an uncaught
exception if a Pool has no "error" listener and an idle client's connection
drops -- Node's EventEmitter throws on an unhandled "error" event. The crash
then triggered a boot-retry crash loop (waitForPostgres's 30s wait, repeatedly
failing since Postgres was mid-restart) until Postgres was fully back up.

Attach pool.on("error", ...) right after the pool is constructed. The pool
already removes a broken client and opens a fresh one on the next checkout on
its own -- the only thing missing was a listener so Node stops treating an
idle client's connection-level error as unhandled. Logs via the same
console.error JSON convention used throughout this codebase so the failure
still reaches Sentry/Loki, just without taking the process down.

Author: JSONbored

src/server.ts

@@ -218,6 +218,22 @@ async function buildPostgresBackend(
   const pg = (await import("pg")).default;
   pg.types.setTypeParser(20, (v: string) => Number.parseInt(v, 10)); // int8 (COUNT) → number, like D1
   const pool = new pg.Pool({ connectionString: url, max: resolvePostgresPoolMax() });
+  // node-postgres crashes the WHOLE process with an uncaught exception if the pool has no "error" listener and
+  // an IDLE client's connection drops (Node's EventEmitter throws on an unhandled "error" event) -- confirmed
+  // live (GITTENSORY-1R/1S): Postgres itself being restarted ("terminating connection due to administrator
+  // command") took the whole app down, which then crash-looped for ~29 minutes hitting waitForPostgres's 30s
+  // boot timeout (GITTENSORY-1T) until Postgres was fully back up. The pool already removes a broken client and
+  // opens a fresh one on the next checkout on its own -- the only thing missing was a listener so Node stops
+  // treating an idle client's connection-level error as unhandled.
+  pool.on("error", (error) => {
+    console.error(
+      JSON.stringify({
+        level: "error",
+        event: "selfhost_pg_pool_error",
+        message: error instanceof Error ? error.message : "unknown error",
+      }),
+    );
+  });
   const db = createPgAdapter(pool);
   const queue = createPgQueue(pool, consume);
   await queue.init();