Context refusal in llama-server's exceed_context_size_error shape; README for v2.3

Implemented-By: OpenCode session (model recorded in docs/implementer-log.md)

internal/proxy/ctxguard_router_test.go

@@ -95,11 +95,16 @@ func TestRouterUnloadedModelIsNotACandidate(t *testing.T) {
 	if big.hits.Load() != 0 {
 		t.Errorf("big served %d requests for a model it does not have loaded", big.hits.Load())
 	}
-	var e map[string]any
+	// v2.3: the refusal is llama-server's exceed_context_size_error shape; n_ctx is what "max" was.
+	var e struct {
+		Error struct {
+			NCtx float64 `json:"n_ctx"`
+		} `json:"error"`
+	}
 	if err := json.Unmarshal([]byte(body), &e); err != nil {
 		t.Fatalf("body %q is not JSON: %v", body, err)
 	}
-	if max, _ := e["max"].(float64); max != 4096 {
-		t.Errorf("max = %v, want 4096: the largest per-slot context among hosts that have shared loaded", e["max"])
+	if e.Error.NCtx != 4096 {
+		t.Errorf("error.n_ctx = %v, want 4096: the largest per-slot context among hosts that have shared loaded", e.Error.NCtx)
 	}
 }