fix(app): drop patchModel, let default model resolve for review seeds

The dynamic GPT model search was fragile across CI environments.
The e2eURL override routes any model to the mock server, so there's
no need to pick a specific model.

packages/app/e2e/session/session-review.spec.ts

@@ -40,26 +40,12 @@ function edit(file: string, prev: string, next: string) {
   )
 }
 
-async function patchModel(sdk: Parameters<typeof withSession>[0]) {
-  const all = await sdk.provider.list().then((res) => res.data?.all ?? [])
-  for (const provider of all) {
-    for (const model of Object.values(provider.models)) {
-      if (!model.id.includes("gpt-")) continue
-      if (model.id.includes("gpt-4")) continue
-      if (model.id.includes("oss")) continue
-      return { providerID: provider.id, modelID: model.id }
-    }
-  }
-  throw new Error("No apply_patch-capable GPT model found for review seed")
-}
-
 async function patchWithMock(
   llm: Parameters<typeof test>[0]["llm"],
   sdk: Parameters<typeof withSession>[0],
   sessionID: string,
   patchText: string,
 ) {
-  const model = await patchModel(sdk)
   const callsBefore = await llm.calls()
   await llm.toolMatch(
     (hit) => bodyText(hit).includes("Your only valid response is one apply_patch tool call."),
@@ -69,7 +55,6 @@ async function patchWithMock(
   await sdk.session.prompt({
     sessionID,
     agent: "build",
-    model,
     system: [
       "You are seeding deterministic e2e UI state.",
       "Your only valid response is one apply_patch tool call.",