[api] Name the null default invariant once and thread needs_decode through the pointer buffer base

The dead size parameter goes, the string type reads the inherited flag, the forced short
string path asserts against it, and the generated declarations say why the pointer may be
null.

script/api_protobuf/api_protobuf.py

@@ -566,11 +566,11 @@ def create_field_type_info(
         # For messages that decode (SOURCE_CLIENT or SOURCE_BOTH), use pointer
         # for zero-copy access to the receive buffer
         if needs_decode:
-            return PointerToBytesBufferType(field, None)
+            return PointerToBytesBufferType(field)
 
         # For SOURCE_SERVER (encode only), explicit annotation is still needed
         if get_field_opt(field, pb.pointer_to_buffer, False):
-            return PointerToBytesBufferType(field, None)
+            return PointerToBytesBufferType(field)
 
         return BytesType(field, needs_decode, needs_encode)
 
@@ -1135,9 +1135,9 @@ class PointerToBufferTypeBase(TypeInfo):
         return False
 
     def __init__(
-        self, field: descriptor.FieldDescriptorProto, size: int | None = None
+        self, field: descriptor.FieldDescriptorProto, needs_decode: bool = True
     ) -> None:
-        super().__init__(field)
+        super().__init__(field, needs_decode)
         self.array_size = 0
 
     @property
@@ -1215,28 +1215,32 @@ class PointerToStringBufferType(PointerToBufferTypeBase):
     reference_type = "StringRef &"
     const_reference_type = "const StringRef &"
 
-    def __init__(
-        self, field: descriptor.FieldDescriptorProto, needs_decode: bool
-    ) -> None:
-        super().__init__(field, None)
-        self._needs_decode = needs_decode
-
     @classmethod
     def can_use_dump_field(cls) -> bool:
         return True
 
+    @property
+    def _starts_null(self) -> bool:
+        """A field that is only encoded, and skipped when empty, never has its pointer read
+        before it is set, so it can default to a null StringRef and the message constructs as
+        one zero fill. Any encode path that copies unconditionally must check this."""
+        return not self._needs_decode and not self.force
+
     @property
     def public_content(self) -> list[str]:
-        # A field that is only ever encoded and skipped when empty never has its pointer read, so a
-        # null default lets the whole message construct as one zero fill
-        if not self._needs_decode and not self.force:
-            return [f"StringRef {self.field_name}{{nullptr, 0}};"]
+        if self._starts_null:
+            return [
+                f"StringRef {self.field_name}{{nullptr, 0}};  // null until set, encode only"
+            ]
         return [f"StringRef {self.field_name}{{}};"]
 
     @property
     def encode_content(self) -> str:
         max_len = self.max_data_length
         if max_len is not None and max_len < 128 and self.force:
+            assert not self._starts_null, (
+                "unconditional copy of a field that may start null"
+            )
             tag = self.calculate_tag()
             if tag < 128:
                 return _encode_call(