testing: real-socket scope-close TCP teardown test (comms-4lz5)

packages/testing/realm-hooks.ts

@@ -9,14 +9,17 @@ import { afterEach, beforeEach } from 'bun:test'
  * wrapping (comms-xwqm). 30s gives that contention wide headroom while still
  * failing if a hook genuinely wedges.
  *
- * This is headroom, not a cure-all: the long-poll teardown tests (the
- * gap-replay / scope-close-interrupt cases that drive an infinite server
- * handler and rely on AbortSignal unwinding a forked drain fiber) can still
- * exceed even this under aggressive contention. That residual is a fiber/scope
- * lifecycle problem, not a too-small-number problem, and is tracked for the
- * Effect-Scope-based realm rework (comms-4lz5 / comms-30hq) — running the whole
- * test, realm acquisition included, inside one scope with guaranteed finalizers
- * removes the leftover-fiber-starves-teardown failure mode structurally.
+ * The infinite-long-poll cases that used to strain even this headroom (the
+ * gap-replay / scope-close-interrupt drains whose forked fiber's AbortSignal
+ * unwinding could starve teardown under aggressive contention) no longer run
+ * here: comms-e5vm.2 moved that LOGIC onto the stub HttpClient + TestClock
+ * (deterministic, no socket, no forked drain), and the one genuinely-socket
+ * teardown assertion that remains is self-contained in
+ * `packages/zulip/scope-teardown.test.ts` with its own `Effect.acquireRelease`
+ * (comms-4lz5). So every fixture still wired through these hooks is a plain
+ * request/response realm that starts and stops in a few ms — the 30s is wide
+ * headroom for contention, not a cure for a teardown-starvation mode that is
+ * now structurally gone.
  */
 export const REALM_HOOK_TIMEOUT_MS = 30_000
 

packages/zulip/scope-teardown.test.ts

@@ -0,0 +1,108 @@
+/**
+ * Tier-3 residue (comms-4lz5): the **one** assertion in the suite that
+ * genuinely cannot leave the socket — proving that interrupting an in-flight
+ * `FetchHttpClient` long-poll on scope close actually tears down the underlying
+ * TCP connection (`AbortSignal → fetch → socket`).
+ *
+ * Everything else the event pump does — gap-replay, 429 retry,
+ * BAD_EVENT_QUEUE_ID reconnect, the Effect fiber-interrupt LOGIC of scope close
+ * — moved onto the owned-fake stub HttpClient + TestClock in `adapter-events.test.ts`
+ * (comms-e5vm.2): deterministic, no socket. The stub proves the fiber unwinds.
+ * It cannot prove the *socket* unwinds, because there is no socket. That gap is
+ * this test, and it is the **only surviving `Bun.serve` long-poll** in the suite.
+ *
+ * The honest assertion is **server-side**: a real `Bun.serve` whose long-poll
+ * handler parks forever but listens on the request's `req.signal`. When the
+ * client tears the connection down, Bun fires that abort. Observing it is proof
+ * the TCP teardown reached the peer — not merely that the Effect fiber
+ * unwound on our side. The Effect platform client wires this end to end:
+ * `httpClient.make` aborts the request's `AbortController` `onInterrupt`, and
+ * `FetchHttpClient` passes that `signal` into `fetch`.
+ *
+ * Lifecycle is `Effect.acquireRelease`/`Scope` throughout: the server is
+ * acquired with `server.stop(true)` as its release, and the long-poll is forked
+ * into an inner `Effect.scoped` whose close is the interruption under test — so
+ * release is guaranteed even on the interruption path.
+ */
+
+import { effectTest } from '@commy/testing/effect-test'
+import { FetchHttpClient, HttpClient } from '@effect/platform'
+import { Data, Duration, Effect } from 'effect'
+
+class ConnectionNotTornDown extends Data.TaggedError('ConnectionNotTornDown')<{
+  readonly message: string
+}> {}
+
+interface LongPollServer {
+  readonly url: string
+  /** Resolves once the server has the long-poll request in flight. */
+  readonly pollInFlight: Promise<void>
+  /** Resolves when the client tears the in-flight connection down (`req.signal`). */
+  readonly clientDisconnected: Promise<void>
+  readonly stop: () => Promise<void>
+}
+
+// A real Bun.serve whose /events handler parks forever and reports, via
+// req.signal, when the client disconnects mid-poll — the server-side proof of
+// TCP teardown.
+const startLongPollServer = (): LongPollServer => {
+  const inFlight = Promise.withResolvers<void>()
+  const disconnected = Promise.withResolvers<void>()
+  const server = Bun.serve({
+    port: 0,
+    fetch: (req) => {
+      inFlight.resolve()
+      return new Promise<Response>((resolve) => {
+        req.signal.addEventListener('abort', () => {
+          disconnected.resolve()
+          resolve(new Response(null, { status: 499 }))
+        })
+      })
+    },
+  })
+  if (typeof server.port !== 'number') {
+    throw new Error('long-poll server failed to bind a TCP port')
+  }
+  return {
+    url: `http://localhost:${server.port}/events`,
+    pollInFlight: inFlight.promise,
+    clientDisconnected: disconnected.promise,
+    stop: () => server.stop(true),
+  }
+}
+
+effectTest(
+  'closing the scope of an in-flight FetchHttpClient long-poll tears down the TCP connection (comms-4lz5)',
+  () =>
+    Effect.gen(function* () {
+      const server = yield* Effect.acquireRelease(Effect.sync(startLongPollServer), (s) =>
+        Effect.promise(() => s.stop()),
+      )
+      const client = yield* HttpClient.HttpClient
+
+      // Fork the real long-poll into an inner scope, hold until the server has
+      // it in flight on the socket, then close the scope — the interruption is
+      // what must propagate to the socket.
+      yield* Effect.scoped(
+        Effect.gen(function* () {
+          yield* Effect.forkScoped(client.get(server.url))
+          yield* Effect.promise(() => server.pollInFlight)
+        }),
+      )
+
+      // The scope is closed; the forked fiber is interrupted. Proof of teardown
+      // is the server observing the client's disconnect. A FetchHttpClient that
+      // failed to abort the socket would leave this parked, and the timeout
+      // fails the test loudly.
+      yield* Effect.promise(() => server.clientDisconnected).pipe(
+        Effect.timeoutFail({
+          duration: Duration.seconds(3),
+          onTimeout: () =>
+            new ConnectionNotTornDown({
+              message: 'scope close did not tear down the in-flight long-poll connection',
+            }),
+        }),
+      )
+    }),
+  { layer: FetchHttpClient.layer, timeout: 10_000 },
+)