e2e: fix real intermittent IPC drop in agent-tool-approval tests

The diagnostics added last time confirmed the chat request always reaches
the fake Ollama server (request count incremented correctly on every
failure) — the response chunk carrying the tool call was the thing going
missing, not the request. This is the same class of bug already documented
in fake-ollama.ts (Electron can silently coalesce/drop a webContents.send()
that lands in the same event-loop tick as another IPC send), just triggered
by a different neighbor: enabling agent mode kicks off its own background
IPC (sessions:update, agentTools.detectProjectScripts()), and the tool-call
turn is a single chunk carrying the *entire* response, so losing it leaves
nothing behind to recover — unlike token streaming, where losing one chunk
mid-stream still leaves partial text.

Fix: let agent-mode side effects settle (300ms) before sending, and widen
the tool-call turn's own response delay (250ms, up from the fixture's
15ms default) for more headroom. Passed 3x repeated locally plus the full
suite; previously flaked deterministically on CI even at a 30s wait.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Author: voidstackloop

e2e/tests/agent-tool-approval.spec.ts

@@ -55,7 +55,25 @@ async function enableAgentModeAndSendToolCallingMessage(instance_: LaunchedApp):
     await window.getByRole("button", { name: "Agent", exact: true }).click();
     await expect(window.getByRole("button", { name: /Agent/ })).toHaveAttribute("aria-pressed", "true");
 
-    fakeOllama.setNextChatTurn({ toolCall: { name: "run_command", arguments: { command: "echo hello-from-agent" } } });
+    // Enabling agent mode kicks off its own IPC round trips in the
+    // background (sessions:update, agentTools.detectProjectScripts() via the
+    // agentWorkspace effect) — sending the chat message immediately risks
+    // its single-chunk tool-call response's webContents.send() landing in
+    // the same event-loop tick as one of those, which Electron can coalesce
+    // away silently (this is the same class of bug worked around in
+    // fake-ollama.ts's DEFAULT_DELAY_MS, just triggered by a different
+    // neighbor now that agent mode is in play). Letting things settle first
+    // is cheaper and more direct than only padding the response delay.
+    await window.waitForTimeout(300);
+
+    fakeOllama.setNextChatTurn({
+        toolCall: { name: "run_command", arguments: { command: "echo hello-from-agent" } },
+        // Wider than the fixture's own default — this response is a single
+        // chunk carrying the *entire* tool call, so if Electron drops it,
+        // there's no partial content left behind to recover; CI has shown
+        // this needs more headroom than plain streaming does.
+        delayMs: 250,
+    });
 
     const input = window.getByPlaceholder("Send a message...");
     await input.fill("run a command for me");