Do not retry an HTTP 408; lock retry cap and timeout values

The OpenAI and Anthropic adapters now report an HTTP 408 as
llm.TimeoutStatusError, which still matches llm.ErrTimeout but carries
its status. Retrying no longer retries any error carrying a 4xx status,
so a vendor's 408 is returned at once while a transport timeout, which
has no status, is still retried.

Tests: 408 (bare and wrapped) is not retried and a transport timeout is;
FromSettings integration cases for an OpenAI and Anthropic 408 and an
OpenAI transport timeout; the cap and default-timeout tests assert the
literal 20 s and 60 s instead of the constants they lock.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

internal/ai/llm/retry_test.go

@@ -77,7 +77,7 @@ func TestRetryingRateLimitedWaitIsCapped(t *testing.T) {
 	if _, err := r.Generate(context.Background(), retryReq); err != nil {
 		t.Fatal(err)
 	}
-	if want := []time.Duration{MaxRetryAfter}; !slices.Equal(c.sleeps, want) {
+	if want := []time.Duration{20 * time.Second}; !slices.Equal(c.sleeps, want) {
 		t.Errorf("sleeps = %v, want %v", c.sleeps, want)
 	}
 	assertConsumed(t, f, 2)
@@ -149,11 +149,13 @@ func TestRetrying5xxThreeTimesFails(t *testing.T) {
 
 func TestRetryingDoesNotRetryPermanentErrors(t *testing.T) {
 	for name, err := range map[string]error{
-		"401 auth":   ErrAuth,
-		"400":        NewProviderError(400, []byte("bad request"), ""),
-		"404":        NewProviderError(404, []byte("no such model"), ""),
-		"bad output": ErrBadOutput,
-		"other":      errors.New("llm: request has no messages"),
+		"401 auth":    ErrAuth,
+		"400":         NewProviderError(400, []byte("bad request"), ""),
+		"404":         NewProviderError(404, []byte("no such model"), ""),
+		"408":         NewTimeoutStatusError(408, ""),
+		"wrapped 408": fmt.Errorf("openai: generate: %w", NewTimeoutStatusError(408, "")),
+		"bad output":  ErrBadOutput,
+		"other":       errors.New("llm: request has no messages"),
 	} {
 		t.Run(name, func(t *testing.T) {
 			r, f, c := newTestRetrying()
@@ -170,6 +172,18 @@ func TestRetryingDoesNotRetryPermanentErrors(t *testing.T) {
 	}
 }
 
+// A transport timeout matches ErrTimeout like an HTTP 408, but carries no
+// status, so it is retried.
+func TestRetryingTransportTimeoutRetried(t *testing.T) {
+	r, f, _ := newTestRetrying()
+	f.QueueGenerate(fail(fmt.Errorf("openai: generate: %w", ErrTimeout)), ok("dinner"))
+
+	if _, err := r.Generate(context.Background(), retryReq); err != nil {
+		t.Fatal(err)
+	}
+	assertConsumed(t, f, 2)
+}
+
 func TestRetryingDeadlineShorterThanRetryAfter(t *testing.T) {
 	r, f, c := newTestRetrying()
 	rl := NewRateLimitedError(10*time.Second, "")
@@ -250,7 +264,7 @@ func TestRetryingTimeoutBoundsWholeCall(t *testing.T) {
 		timeout time.Duration
 		want    time.Duration
 	}{
-		{"default", 0, DefaultTimeout},
+		{"default", 0, 60 * time.Second},
 		{"custom", 10 * time.Second, 10 * time.Second},
 	} {
 		t.Run(tc.name, func(t *testing.T) {