fix(coach): graceful fallback when sentinel-only output strips empty

When the model emits only a `---KEYVALUES---` block with no leading
prose (or a malformed block missing the closing marker), the
post-strip prose is empty. The previous fallback chained back to
`rawReply`, surfacing raw sentinel markers to the user. Replace with
a structured `coach.provider.empty` error frame so the UI renders
the localised provider-error copy instead of leaking the markers.

Add a regression integration test pinning that the streamed body
never contains `---KEYVALUES---` for a sentinel-only reply, and that
no assistant message is persisted in this branch.

Closes Code-H1.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/app/api/insights/chat/route.ts

@@ -272,7 +272,26 @@ Reply now as the assistant, in ${locale === "de" ? "German" : "English"}.`;
   // provenance envelope so the UI can render the collapsible
   // "Worauf bezieht sich das?" disclosure.
   const sentinel = parseKeyValuesSentinel(rawReply);
-  const replyText = sentinel.prose.trim() || rawReply;
+  const proseAfterStrip = sentinel.prose.trim();
+  // v1.4.22 W5 reconcile (Code-H1) — when the model emits a
+  // sentinel-only / malformed reply, `sentinel.prose` is empty after
+  // stripping. The previous fallback `sentinel.prose.trim() || rawReply`
+  // surfaced raw `---KEYVALUES---` markers to the user. The empty-prose
+  // condition signals an unusable provider response: short-circuit to
+  // the structured `coach.provider.empty` error frame instead of
+  // streaming the raw sentinel body.
+  if (!proseAfterStrip) {
+    annotate({
+      action: { name: "coach.keyvalues.parse_failed" },
+      meta: {
+        kept: sentinel.keyValues.length,
+        reason: "empty_prose_after_strip",
+        promptVersion: PROMPT_VERSION,
+      },
+    });
+    return streamProviderError({ code: "coach.provider.empty" });
+  }
+  const replyText = proseAfterStrip;
   const enrichedProvenance =
     sentinel.keyValues.length > 0
       ? { ...snapshot.provenance, keyValues: sentinel.keyValues }

tests/integration/coach-chat.test.ts

@@ -398,4 +398,54 @@ describe("POST /api/insights/chat — integration", () => {
     };
     expect(persisted.keyValues).toBeUndefined();
   });
+
+  it("never leaks raw ---KEYVALUES--- markers when the provider returns a sentinel-only reply", async () => {
+    // v1.4.22 W5 reconcile (Code-H1) — regression test. The model
+    // emits ONLY the sentinel block with no leading prose AND a
+    // missing close marker, so `parseKeyValuesSentinel` returns
+    // `{ prose: "", malformed: true }`. The route must surface the
+    // structured `coach.provider.empty` error frame instead of
+    // streaming the raw sentinel body to the client.
+    await seedUserWithSession();
+    runProviderMock.mockResolvedValue({
+      result: {
+        content: [
+          "---KEYVALUES---",
+          "avg7 systolic: 138 [mmHg] (last7days)",
+          // Note: no closing ---END--- and no prose before the marker.
+        ].join("\n"),
+        tokensUsed: 64,
+        model: "mock",
+        providerType: "openai",
+      },
+      workingProvider: { providerType: "openai", instance: {} },
+      fallbackHops: [],
+    });
+    const { POST } = await import("@/app/api/insights/chat/route");
+
+    const res = await (POST as (req: Request) => Promise<Response>)(
+      new Request("http://localhost/api/insights/chat", {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({ message: "How is my BP this week?" }),
+      }),
+    );
+    expect(res.status).toBe(200);
+    const text = await readStream(res);
+
+    // Streamed body must NOT contain the raw sentinel markers and must
+    // surface the structured provider-empty error.
+    expect(text).not.toMatch(/---KEYVALUES---/);
+    expect(text).not.toMatch(/---END---/);
+    expect(text).toMatch(/"type":"error"/);
+    expect(text).toMatch(/coach\.provider\.empty/);
+
+    // No assistant message persisted — the empty-prose branch
+    // short-circuits before `appendMessage`.
+    const prisma = getPrismaClient();
+    const messages = await prisma.coachMessage.findMany({
+      where: { role: "assistant" },
+    });
+    expect(messages.length).toBe(0);
+  });
 });