fix(studio): recover transient AI failures and report provider diagnostics

test/studio-ai.test.ts

@@ -3,6 +3,7 @@ import { CORE_DESIGN_SYSTEMS } from "@wrnexus/design-system";
 import { designRevision, DEFAULT_DESIGN_BRIEF } from "@wrnexus/design-studio";
 import {
   AI_TIMEOUT_MS,
+  AI_OUTPUT_TOKENS,
   apimartProvider,
   studioAI,
 } from "../app/lib/studio-ai.ts";
@@ -57,7 +58,7 @@ test("APIMart uses the specified model, bearer authentication, non-streaming JSO
       const body = JSON.parse(init?.body as string);
       expect(body.model).toBe("gpt-6-luna");
       expect(body.stream).toBe(false);
-      expect(body.max_tokens).toBe(4096);
+      expect(body.max_tokens).toBe(AI_OUTPUT_TOKENS);
       expect(body.messages).toEqual([
         { role: "system", content: "rules" },
         { role: "user", content: "brief" },
@@ -200,3 +201,121 @@ test("a public AI endpoint enforces a bounded generation rate", async () => {
     expect((await handler(await request()))!.status).toBe(200);
   expect((await handler(await request()))!.status).toBe(429);
 });
+
+test("temporary APIMart failures retry once and return three validated designs", async () => {
+  let calls = 0;
+  const handler = studioAI(
+    "test-secret",
+    mockFetch(async (_url, init) => {
+      calls++;
+      expect(JSON.parse(init!.body as string).max_tokens).toBe(8192);
+      return calls === 1
+        ? new Response("sensitive upstream context test-secret", {
+            status: 503,
+          })
+        : completion();
+    }),
+  );
+  const batch = await (await handler(await request()))!.json();
+  expect(calls).toBe(2);
+  expect(batch.alternatives).toHaveLength(3);
+  expect(batch.provenance.attempts).toBe(2);
+  expect(batch.provenance.maxOutputTokens).toBe(8192);
+});
+
+test("authentication and quota failures provide actionable redacted diagnostics without retrying", async () => {
+  for (const [status, code, hint] of [
+    [401, "invalid_api_key", "server API key"],
+    [402, "insufficient_quota", "credit or quota"],
+    [404, "model_not_found", "gpt-6-luna"],
+  ] as const) {
+    let calls = 0;
+    const handler = studioAI(
+      "test-secret",
+      mockFetch(async () => {
+        calls++;
+        return Response.json(
+          { error: { code, message: "test-secret private brief" } },
+          { status },
+        );
+      }),
+    );
+    const body = await (await handler(await request()))!.text();
+    expect(calls).toBe(1);
+    expect(body).toContain(hint);
+    expect(body).not.toContain("test-secret");
+    expect(body).not.toContain("private brief");
+  }
+});
+
+test("truncated and empty completions retry but never exceed two total upstream calls", async () => {
+  for (const kind of ["length", "empty", "invalid-then-outage"]) {
+    let calls = 0;
+    const handler = studioAI(
+      "test-secret",
+      mockFetch(async () => {
+        calls++;
+        if (kind === "invalid-then-outage")
+          return calls === 1
+            ? Response.json({
+                choices: [{ message: { content: "invalid JSON" } }],
+              })
+            : new Response("private upstream body", { status: 502 });
+        return calls === 1
+          ? Response.json({
+              choices: [
+                {
+                  finish_reason: kind === "length" ? "length" : "stop",
+                  message: { content: "" },
+                },
+              ],
+            })
+          : completion();
+      }),
+    );
+    const batch = await (await handler(await request()))!.json();
+    expect(calls).toBe(2);
+    expect(batch.provenance.attempts).toBe(2);
+    expect(batch.alternatives).toHaveLength(
+      kind === "invalid-then-outage" ? 0 : 3,
+    );
+  }
+});
+
+test("network and repeated rate failures remain bounded and describe the actual failure category", async () => {
+  for (const network of [true, false]) {
+    let calls = 0;
+    const handler = studioAI(
+      "test-secret",
+      mockFetch(async () => {
+        calls++;
+        if (network) throw Error("private test-secret TLS details");
+        return new Response("private test-secret", { status: 429 });
+      }),
+    );
+    const body = await (await handler(await request()))!.text();
+    expect(calls).toBe(2);
+    expect(body).toContain(
+      network ? "could not reach APIMart" : "rate limiting",
+    );
+    expect(body).not.toContain("test-secret");
+  }
+});
+
+test("a retry followed by an invalid design does not spend a third request or mislabel a schema error", async () => {
+  let calls = 0;
+  const handler = studioAI(
+    "test-secret",
+    mockFetch(async () => {
+      calls++;
+      return calls === 1
+        ? new Response("unavailable", { status: 503 })
+        : Response.json({ choices: [{ message: { content: "not JSON" } }] });
+    }),
+  );
+  const batch = await (await handler(await request()))!.json();
+  expect(calls).toBe(2);
+  expect(batch.alternatives).toHaveLength(0);
+  expect(batch.diagnostics[0].code).toBe("WRN-STUDIO-PROPOSAL");
+  expect(batch.diagnostics[0].message).toContain("schema");
+});