fix(auto-reply): honor structured FailoverError reasons in terminal catch classification

context_overflow and transient (timeout/server_error/overloaded) FailoverError
reasons now classify structurally like billing/rate_limit already do, instead
of relying on message text that may lack overflow or HTTP-status tokens. A
typed transient failure without a leading status token now takes the single
transient retry; typed context overflow suppresses the generic failure reply.
Part of #104219 (row 12).

src/auto-reply/reply/agent-runner-execution.test.ts

@@ -1401,6 +1401,8 @@ describe("runAgentTurnWithFallback", () => {
   });
 
   afterEach(() => {
+    // Fake-timer tests in this describe must not leak into --isolate=false peers.
+    vi.useRealTimers();
     vi.clearAllMocks();
   });
 
@@ -7856,6 +7858,81 @@ describe("runAgentTurnWithFallback", () => {
     expect(state.updateSessionStoreMock).not.toHaveBeenCalled();
   });
 
+  it.each([
+    {
+      reason: "server_error" as const,
+      message: "upstream provider failed briefly",
+    },
+    {
+      reason: "timeout" as const,
+      message: "provider request timed out without status token",
+    },
+    {
+      reason: "overloaded" as const,
+      message: "provider capacity exhausted briefly",
+    },
+  ])(
+    "retries once for structured FailoverError $reason without leading HTTP status text",
+    async ({ reason, message }) => {
+      vi.useFakeTimers();
+      state.runEmbeddedAgentMock
+        .mockRejectedValueOnce(
+          new FailoverError(message, {
+            reason,
+            provider: "openai",
+            model: "gpt-5.5",
+          }),
+        )
+        .mockResolvedValueOnce({
+          payloads: [{ text: "recovered after transient failover" }],
+          meta: {},
+        });
+
+      const runAgentTurnWithFallback = await getRunAgentTurnWithFallback();
+      const resultPromise = runAgentTurnWithFallback(createMinimalRunAgentTurnParams());
+      await vi.advanceTimersByTimeAsync(2_500);
+      const result = await resultPromise;
+
+      expect(state.runEmbeddedAgentMock).toHaveBeenCalledTimes(2);
+      expect(result.kind).toBe("success");
+      if (result.kind === "success") {
+        expect(result.runResult.payloads?.[0]?.text).toBe("recovered after transient failover");
+      }
+    },
+  );
+
+  it("uses structured FailoverError context_overflow over non-overflow message text", async () => {
+    state.isLikelyContextOverflowErrorMock.mockReturnValue(false);
+    state.runEmbeddedAgentMock.mockRejectedValueOnce(
+      new FailoverError("provider rejected the request payload", {
+        reason: "context_overflow",
+        provider: "anthropic",
+        model: "claude",
+      }),
+    );
+
+    const runAgentTurnWithFallback = await getRunAgentTurnWithFallback();
+    const result = await runAgentTurnWithFallback(
+      createMinimalRunAgentTurnParams({
+        sessionCtx: {
+          Provider: "telegram",
+          Surface: "telegram",
+          ChatType: "direct",
+          MessageSid: "msg",
+        } as unknown as TemplateContext,
+      }),
+    );
+
+    expect(result.kind).toBe("final");
+    if (result.kind === "final") {
+      expect(result.payload.text).toBe(
+        "⚠️ Context overflow — prompt too large for this model. Try a shorter message or a larger-context model.",
+      );
+      expect(result.payload.text).not.toBe(GENERIC_RUN_FAILURE_TEXT);
+      expect(result.payload.text).not.toContain("provider rejected the request payload");
+    }
+  });
+
   it("uses the throwing fallback candidate model for compaction failure hints", async () => {
     state.isCompactionFailureErrorMock.mockReturnValue(true);
     state.runWithModelFallbackMock.mockImplementationOnce(async (params: FallbackRunnerParams) => {

src/auto-reply/reply/agent-runner-execution.ts

@@ -3063,7 +3063,13 @@ async function runAgentTurnWithFallbackInternal(
         : isFailoverError(err)
           ? err.reason === "billing"
           : isBillingErrorMessage(message);
-      const isContextOverflow = !isBilling && isLikelyContextOverflowError(message);
+      // Prefer structured FailoverError reasons over message-text heuristics so
+      // typed context-overflow/transient failures are not misclassified when the
+      // error string lacks overflow/HTTP status tokens.
+      const isContextOverflow =
+        !isBilling &&
+        ((isFailoverError(err) && err.reason === "context_overflow") ||
+          isLikelyContextOverflowError(message));
       const isCompactionFailure = !isBilling && isCompactionFailureError(message);
       // OAuth/auth-profile failures must reach buildExternalRunFailureReply so
       // the targeted re-auth/failover copy is surfaced instead of the generic
@@ -3078,7 +3084,12 @@ async function runAgentTurnWithFallbackInternal(
         !shouldSurfaceToControlUi
           ? classifyProviderRequestError(err)
           : undefined;
-      const isTransientHttp = isTransientHttpError(message);
+      const isTransientHttp =
+        isTransientHttpError(message) ||
+        (isFailoverError(err) &&
+          (err.reason === "timeout" ||
+            err.reason === "server_error" ||
+            err.reason === "overloaded"));
 
       // Drain/restart aborts stay silent and defer to post-restart
       // main-session recovery, which resumes the interrupted turn (or emits its