diff --git a/include/pybind11/internal/builtins.h b/include/pybind11/internal/builtins.h
index aa616ca9..084376cb 100644
--- a/include/pybind11/internal/builtins.h
+++ b/include/pybind11/internal/builtins.h
@@ -81,9 +81,13 @@ inline bool isinstance(const handle& obj, const handle& type) {
 
 inline int64_t hash(const handle& obj) { return vm->py_hash(obj.ptr()); }
 
+namespace impl {
+
 template <typename T, typename SFINAE = void>
 struct type_caster;
 
+}
+
 template <typename T>
 handle cast(T&& value, return_value_policy policy, handle parent) {
     // decay_t can resolve c-array type, but remove_cv_ref_t can't.
@@ -107,7 +111,7 @@ handle cast(T&& value, return_value_policy policy, handle parent) {
                                                         : return_value_policy::move;
         }
 
-        return type_caster<underlying_type>::cast(std::forward<T>(value), policy, parent);
+        return impl::type_caster<underlying_type>::cast(std::forward<T>(value), policy, parent);
     }
 }
 
@@ -115,7 +119,7 @@ template <typename T>
 T cast(const handle& obj, bool convert) {
     assert(obj.ptr() != nullptr);
 
-    type_caster<T> caster = {};
+    impl::type_caster<T> caster = {};
 
     if(caster.load(obj, convert)) {
         return caster.value;
diff --git a/include/pybind11/internal/cast.h b/include/pybind11/internal/cast.h
index acc0897a..526c5a26 100644
--- a/include/pybind11/internal/cast.h
+++ b/include/pybind11/internal/cast.h
@@ -2,7 +2,7 @@
 #include "instance.h"
 #include "accessor.h"
 
-namespace pybind11 {
+namespace pybind11::impl {
 
 using pkpy::is_floating_point_v;
 using pkpy::is_integral_v;
diff --git a/include/pybind11/internal/class.h b/include/pybind11/internal/class.h
index 70d850d4..dd4e01e8 100644
--- a/include/pybind11/internal/class.h
+++ b/include/pybind11/internal/class.h
@@ -169,15 +169,15 @@ public:
     }
 
     enum_& value(const char* name, T value) {
-        handle var = type_caster<T>::cast(value, return_value_policy::copy);
-        this->m_ptr->attr().set(name, var.ptr());
+        handle var = pybind11::cast(value, return_value_policy::copy);
+        setattr(*this, name, var);
         m_values.emplace_back(name, var);
         return *this;
     }
 
     enum_& export_values() {
         for(auto& [name, value]: m_values) {
-            Base::m_scope.ptr()->attr().set(name, value.ptr());
+            setattr(Base::m_scope, name, value);
         }
         return *this;
     }
diff --git a/include/pybind11/internal/cpp_function.h b/include/pybind11/internal/cpp_function.h
index fa7a4364..f12e5b02 100644
--- a/include/pybind11/internal/cpp_function.h
+++ b/include/pybind11/internal/cpp_function.h
@@ -176,7 +176,7 @@ public:
 template <typename Fn, std::size_t... Is, typename... Args>
 handle invoke(Fn&& fn,
               std::index_sequence<Is...>,
-              std::tuple<type_caster<Args>...>& casters,
+              std::tuple<impl::type_caster<Args>...>& casters,
               return_value_policy policy,
               handle parent) {
     using underlying_type = std::decay_t<Fn>;
@@ -361,10 +361,10 @@ struct template_parser<Callable, std::tuple<Extras...>, std::tuple<Args...>, std
 
         // resolve keyword arguments
         const auto n = vm->s_data._sp - view.end();
-        std::size_t index = 0;
+        int index = 0;
 
         if constexpr(named_argc > 0) {
-            std::size_t arg_index = 0;
+            int arg_index = 0;
             auto& arguments = *record.arguments;
 
             while(arg_index < named_argc && index < n) {
@@ -403,7 +403,7 @@ struct template_parser<Callable, std::tuple<Extras...>, std::tuple<Args...>, std
         }
 
         // ok, all the arguments are valid, call the function
-        std::tuple<type_caster<Args>...> casters;
+        std::tuple<impl::type_caster<Args>...> casters;
 
         // check type compatibility
         if(((std::get<Is>(casters).load(stack[Is], convert)) && ...)) {
@@ -489,14 +489,14 @@ pkpy::PyVar setter_wrapper(pkpy::VM* vm, pkpy::ArgsView view) {
 
         if constexpr(std::is_member_object_pointer_v<Setter>) {
             // specialize for pointer to data member
-            type_caster<member_type_t<Setter>> caster;
+            impl::type_caster<member_type_t<Setter>> caster;
             if(caster.load(view[1], true)) {
                 self.*setter = caster.value;
                 return vm->None;
             }
         } else {
             // specialize for pointer to member function
-            type_caster<std::tuple_element_t<1, callable_args_t<Setter>>> caster;
+            impl::type_caster<std::tuple_element_t<1, callable_args_t<Setter>>> caster;
             if(caster.load(view[1], true)) {
                 (self.*setter)(caster.value);
                 return vm->None;
@@ -507,7 +507,7 @@ pkpy::PyVar setter_wrapper(pkpy::VM* vm, pkpy::ArgsView view) {
         using Self = remove_cvref_t<std::tuple_element_t<0, callable_args_t<Setter>>>;
         auto& self = handle(view[0])._as<instance>()._as<Self>();
 
-        type_caster<std::tuple_element_t<1, callable_args_t<Setter>>> caster;
+        impl::type_caster<std::tuple_element_t<1, callable_args_t<Setter>>> caster;
         if(caster.load(view[1], true)) {
             setter(self, caster.value);
             return vm->None;
diff --git a/include/pybind11/internal/kernel.h b/include/pybind11/internal/kernel.h
index bb117712..55d33f45 100644
--- a/include/pybind11/internal/kernel.h
+++ b/include/pybind11/internal/kernel.h
@@ -37,7 +37,7 @@ class handle;
 class object;
 class iterator;
 class str;
-class arg;
+struct arg;
 struct args_proxy;
 struct kwargs_proxy;
 
@@ -210,12 +210,7 @@ constexpr inline bool is_pyobject_v = std::is_base_of_v<object, T>;
 #if PK_VERSION_MAJOR == 2
 using error_already_set = pkpy::TopLevelException;
 #else
-class error_already_set : std::exception {
-public:
-    error_already_set() = default;
-
-    const char* what() const noexcept override { return "An error occurred while calling a Python function."; }
-};
+using error_already_set = pkpy::Exception;
 #endif
 
 inline void setattr(const handle& obj, const handle& name, const handle& value);
diff --git a/include/pybind11/internal/module.h b/include/pybind11/internal/module.h
index 420b11fc..6eef55e2 100644
--- a/include/pybind11/internal/module.h
+++ b/include/pybind11/internal/module.h
@@ -19,8 +19,10 @@ public:
     }
 
     module_ def_submodule(const char* name, const char* doc = nullptr) {
-        auto package = this->package()._as<pkpy::Str>() + "." + this->name()._as<pkpy::Str>();
-        auto m = vm->new_module(name, package);
+        // TODO: resolve package
+        //auto package = this->package()._as<pkpy::Str>() + "." + this->name()._as<pkpy::Str>();
+        auto fname = this->name()._as<pkpy::Str>() + "." + name;
+        auto m = vm->new_module(fname, "");
         setattr(*this, name, m);
         return m;
     }
diff --git a/include/pybind11/internal/object.h b/include/pybind11/internal/object.h
index ef1ffe15..007ccf9b 100644
--- a/include/pybind11/internal/object.h
+++ b/include/pybind11/internal/object.h
@@ -69,15 +69,19 @@ public:
     template <typename T>
     decltype(auto) _as() const {
         static_assert(!std::is_reference_v<T>, "T must not be a reference type.");
-#if PK_VERSION_MAJOR == 2
-        if constexpr(pkpy::is_sso_v<T>) {
-            return pkpy::_py_cast<T>(vm, ptr());
+        if constexpr(std::is_same_v<T, empty>) {
+            return empty();
         } else {
-            return ptr().template obj_get<T>();
-        }
+#if PK_VERSION_MAJOR == 2
+            if constexpr(pkpy::is_sso_v<T>) {
+                return pkpy::_py_cast<T>(vm, ptr());
+            } else {
+                return ptr().template obj_get<T>();
+            }
 #else
-        return (((pkpy::Py_<T>*)ptr())->_value);
+            return (((pkpy::Py_<T>*)ptr())->_value);
 #endif
+        }
     }
 };
 
diff --git a/include/pybind11/internal/type_traits.h b/include/pybind11/internal/type_traits.h
index 1bb17172..d24374bb 100644
--- a/include/pybind11/internal/type_traits.h
+++ b/include/pybind11/internal/type_traits.h
@@ -25,9 +25,9 @@ struct function_traits {
     static_assert(dependent_false<Fn>, "unsupported function type");
 };
 
-#define PYBIND11_FUNCTION_TRAITS_SPECIALIZE(qualifiers)                                                                \
+#define PYBIND11_FUNCTION_TRAITS_SPECIALIZE(...)                                                                       \
     template <typename R, typename... Args>                                                                            \
-    struct function_traits<R(Args...) qualifiers> {                                                                    \
+    struct function_traits<R(Args...) __VA_ARGS__> {                                                                   \
         using return_type = R;                                                                                         \
         using args_type = std::tuple<Args...>;                                                                         \
         constexpr static std::size_t args_count = sizeof...(Args);                                                     \
@@ -167,4 +167,32 @@ constexpr auto type_name() {
 #endif
 }
 
+template <typename... Args>
+struct overload_cast_t {
+    template <typename Return>
+    constexpr auto operator() (Return (*pf)(Args...)) const noexcept -> decltype(pf) {
+        return pf;
+    }
+
+    template <typename Return, typename Class>
+    constexpr auto operator() (Return (Class::*pmf)(Args...), std::false_type = {}) const noexcept -> decltype(pmf) {
+        return pmf;
+    }
+
+    template <typename Return, typename Class>
+    constexpr auto operator() (Return (Class::*pmf)(Args...) const, std::true_type) const noexcept -> decltype(pmf) {
+        return pmf;
+    }
+};
+
+/// Syntax sugar for resolving overloaded function pointers:
+///  - regular: static_cast<Return (Class::*)(Arg0, Arg1, Arg2)>(&Class::func)
+///  - sweet:   overload_cast<Arg0, Arg1, Arg2>(&Class::func)
+template <typename... Args>
+constexpr inline overload_cast_t<Args...> overload_cast;
+
+/// Const member function selector for overload_cast
+///  - regular: static_cast<Return (Class::*)(Arg) const>(&Class::func)
+///  - sweet:   overload_cast<Arg>(&Class::func, const_)
+constexpr static auto const_ = std::true_type{};
 }  // namespace pybind11
diff --git a/include/pybind11/internal/types.h b/include/pybind11/internal/types.h
index 1c079ca4..92a2dd37 100644
--- a/include/pybind11/internal/types.h
+++ b/include/pybind11/internal/types.h
@@ -385,11 +385,11 @@ public:
 };
 
 class args : public tuple {
-    PYBIND11_TYPE_IMPLEMENT(tuple, struct empty, vm->tp_tuple);
+    PYBIND11_TYPE_IMPLEMENT(tuple, pybind11::empty, vm->tp_tuple);
 };
 
 class kwargs : public dict {
-    PYBIND11_TYPE_IMPLEMENT(dict, struct empty, vm->tp_dict);
+    PYBIND11_TYPE_IMPLEMENT(dict, pybind11::empty, vm->tp_dict);
 };
 
 }  // namespace pybind11
diff --git a/include/pybind11/stl.h b/include/pybind11/stl.h
index 10e8e976..92c6667f 100644
--- a/include/pybind11/stl.h
+++ b/include/pybind11/stl.h
@@ -9,7 +9,7 @@
 #include <map>
 #include <unordered_map>
 
-namespace pybind11 {
+namespace pybind11::impl {
 
 template <typename T, std::size_t N>
 struct type_caster<std::array<T, N>> {
@@ -140,5 +140,5 @@ struct type_caster<T, std::enable_if_t<is_py_map_like_v<T>>> {
     }
 };
 
-}  // namespace pybind11
+}  // namespace pybind11::impl