test(eval): ondemand simulate uses matchGolden committed goldens (runtime-handler pattern)

Switch the golden assertions from bun toMatchSnapshot to matchGolden + committed
__fixtures__/*.golden.json, matching the runtime handlers' golden style. Driven by
TestCoreClient: fixtureFactories record/replay can't key this command because its inputs
carry random per-session UUIDs + now-based CloudWatch windows (fixturePath hashes the
input), so replay never matches — the same reason batch simulate uses TestCoreClient.

Author: jariy17

src/handlers/eval/ondemand/simulate/__fixtures__/ondemand-simulate-groundtruth.golden.json

@@ -0,0 +1,19 @@
+[
+  {
+    "context": {
+      "spanContext": {
+        "sessionId": "s1"
+      }
+    },
+    "assertions": [
+      {
+        "text": "polite"
+      }
+    ],
+    "expectedTrajectory": {
+      "toolNames": [
+        "lookup"
+      ]
+    }
+  }
+]

src/handlers/eval/ondemand/simulate/__fixtures__/ondemand-simulate-output.golden.json

@@ -0,0 +1,7 @@
+{
+  "sessionsRequested": 2,
+  "sessionsEvaluated": 2,
+  "results": [],
+  "examplesInvoked": 2,
+  "examplesFailed": 0
+}

src/handlers/eval/ondemand/simulate/__snapshots__/simulate.test.tsx.snap

@@ -1,13 +1,17 @@
 import { test, expect, describe } from "bun:test";
+import { join } from "node:path";
 import { createRootHandler } from "../../../index";
 import {
   createSilentLogger,
+  matchGolden,
   TestCoreClient,
   testIO,
   TestGlobalConfigAccessor,
 } from "../../../../testing";
 import type { EvaluateResult, InvokeDatasetResult, SessionTrace } from "../../types";
 
+const FIXTURES = join(import.meta.dir, "__fixtures__");
+
 // Two invoked sessions: e1 carries session-level ground truth (assertions + trajectory),
 // e2 carries none. The handler adapts these to EvaluationReferenceInput[] for evaluate.
 const INVOKE_RESULT: InvokeDatasetResult = {
@@ -163,18 +167,16 @@ describe("eval ondemand simulate", () => {
     const evaluate = core.eval.calls.find((c) => c.method === "evaluate");
     expect(evaluate).toBeDefined();
     const input = evaluate!.args[0] as { groundTruth: unknown };
-    expect(input.groundTruth).toMatchSnapshot();
+    matchGolden(
+      FIXTURES,
+      "ondemand-simulate-groundtruth.golden.json",
+      JSON.stringify(input.groundTruth, null, 2),
+    );
   });
 
-  test("renders scores inline with invoked/failed counts", async () => {
+  test("renders scores inline with invoked/failed counts [golden]", async () => {
     const { stdout } = await run(BASE);
-    expect(JSON.parse(stdout)).toEqual({
-      sessionsRequested: 2,
-      sessionsEvaluated: 2,
-      results: [],
-      examplesInvoked: 2,
-      examplesFailed: 0,
-    });
+    matchGolden(FIXTURES, "ondemand-simulate-output.golden.json", stdout);
   });
 
   test("refuses to evaluate when nothing was invoked", async () => {