diff --git a/backend/domain/model.py b/backend/domain/model.py
index 78bdb1ce..9abf61da 100644
--- a/backend/domain/model.py
+++ b/backend/domain/model.py
@@ -155,13 +155,25 @@ class SceneNode:
     y: float
     title: str
     kind: str = "placeholder"
-    # R3.1 (doc/QT_REMOVAL_PLAN.md): the chat node's real persisted shape -
-    # graphlink_session/serializers.py's raw_content/is_user/is_collapsed,
-    # minus everything Qt-only (paint state, scroll position, docked-child
-    # widgets). Unused (default) for every other kind.
-    # R3.17: also reused verbatim as the html node's raw HTML source string -
-    # no separate field, same reuse pattern as R3.5's code text and R3.13's
-    # thinking text living in this same field.
+    # The node's single free-text body. Originally R3.1
+    # (doc/QT_REMOVAL_PLAN.md), the chat node's persisted shape -
+    # graphlink_session/serializers.py's raw_content, minus everything
+    # Qt-only (paint state, scroll position, docked-child widgets).
+    #
+    # It is NOT chat-only and has not been for a long time. This comment used
+    # to end "Unused (default) for every other kind" while a later line in the
+    # same block already contradicted it, and the list has kept growing since.
+    # Verified 2026-09-04 against the SceneNode(...) constructions in
+    # backend/domain/graph.py and backend/session_load.py, TEN kinds populate
+    # it: artifact, chat, document, harness, html, image, note, plan,
+    # thinking, web_research. Pinned by tests/test_shared_node_field_docs.py,
+    # so this list fails the build rather than rotting again.
+    #
+    # That shared use is deliberate, and it is why `content` did not move to a
+    # per-kind class in the ADR-002 stage 2.5 migration: a field ten kinds
+    # write would have to be duplicated across ten state classes to live
+    # there, which is worse than one core field. Treat it as core, like
+    # title - not as a kind-specific leftover.
     content: str = ""
     is_collapsed: bool = False
     # R3.13 (doc/QT_REMOVAL_PLAN.md): the ThinkingNode/docked-child increment -
@@ -170,14 +182,24 @@ class SceneNode:
     # time from this flag (scan nodes whose parent edge points at it), never
     # stored on the parent itself. Unused (default) for every other kind.
     is_docked: bool = False
-    # R3.25 (doc/QT_REMOVAL_PLAN.md): the ConversationNode's real persisted
-    # shape - graphlink_conversation_node.py's conversation_history, a
-    # growing list of {"role": "user"|"assistant", "content": text} dicts
-    # rendered as multiple bubbles inside one node card. This is the one R3
-    # kind whose OWN field is a LIST rather than a scalar - every prior kind
-    # (chat/code/document/thinking/html/image) stores one scalar value per
-    # node; a conversation node instead owns a whole message history. Unused
-    # (default empty list) for every other kind.
+    # A node's own message history: a growing list of
+    # {"role": "user"|"assistant", "content": text} dicts. Originally R3.25
+    # (doc/QT_REMOVAL_PLAN.md), the ConversationNode's persisted shape, ported
+    # from graphlink_conversation_node.py's conversation_history.
+    #
+    # It is NOT conversation-only and has not been since the plugin kinds
+    # landed. This comment used to say "Unused (default empty list) for every
+    # other kind"; verified 2026-09-04 against the restorers in
+    # backend/domain/graph.py and backend/session_load.py, SEVEN kinds
+    # populate it: artifact, chat, code_sandbox, conversation, gitlink, html,
+    # web_research. Any kind that holds a back-and-forth with a model keeps it
+    # here. Pinned by tests/test_shared_node_field_docs.py.
+    #
+    # Shared for the same reason `content` is, and left on SceneNode by the
+    # ADR-002 stage 2.5 migration for the same reason - see that field's own
+    # note above. backend/domain/node_states.py's docstring is right to list
+    # `conversation` among the kinds with no state class: its history is not a
+    # kind-specific field, it is this shared one.
     history: list[dict[str, str]] = field(default_factory=list)
     # R4.3 (doc/QT_REMOVAL_PLAN.md): transient per-node in-flight-request
     # marker - the id of the AgentDispatcher request currently generating a
diff --git a/tests/test_shared_node_field_docs.py b/tests/test_shared_node_field_docs.py
new file mode 100644
index 00000000..b27b0a96
--- /dev/null
+++ b/tests/test_shared_node_field_docs.py
@@ -0,0 +1,83 @@
+"""`content` and `history` are SHARED fields; their comments have to say so.
+
+Both live on SceneNode rather than in a per-kind class in
+backend/domain/node_states.py, and both were documented as belonging to one
+kind - `content` to chat, `history` to conversation - with an explicit
+"Unused for every other kind" line. Neither was true any more. A 2026-09-04
+audit measured ten kinds writing `content` and seven writing `history`,
+and in `content`'s case a later line in the SAME comment block already
+contradicted the earlier one.
+
+That is a recurring shape in this codebase rather than a one-off: a comment
+asserts a closed set, the set grows, and nothing fails. The same audit found
+backend/api/_shared.py naming its two callers by hand and being wrong within
+one commit of being written. This gate is the cheap answer for the two cases
+where the claim is load-bearing - it is the reason these fields were left on
+SceneNode, so a reader deciding whether to "finish the migration" needs the
+list to be right.
+
+Derived from the kinds that actually construct a SceneNode with each field,
+across the two modules that create nodes: backend/domain/graph.py (live
+creation) and backend/session_load.py (restore from a saved chat).
+
+Same posture as tests/undo_classification.py: a hand-authored expectation
+plus a gate that fails when the code and the expectation diverge. If this
+fails because a new kind legitimately started using the field, update BOTH
+the set below and the field's comment in backend/domain/model.py - that
+pairing is the whole point.
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+NODE_CREATING_MODULES = (
+    REPO_ROOT / "backend" / "domain" / "graph.py",
+    REPO_ROOT / "backend" / "session_load.py",
+)
+
+# Hand-authored, and deliberately duplicated in backend/domain/model.py's own
+# field comments. The gate exists to keep the two copies honest.
+CONTENT_KINDS = {
+    "artifact", "chat", "document", "harness", "html",
+    "image", "note", "plan", "thinking", "web_research",
+}
+HISTORY_KINDS = {"artifact", "chat", "code_sandbox", "conversation", "gitlink", "html", "web_research"}
+
+
+def _kinds_constructing_with(field_name: str) -> set[str]:
+    """Kinds passed to a SceneNode(...) call that also passes `field_name`."""
+    kinds: set[str] = set()
+    for path in NODE_CREATING_MODULES:
+        tree = ast.parse(path.read_text(encoding="utf-8"))
+        for node in ast.walk(tree):
+            if not isinstance(node, ast.Call):
+                continue
+            func = node.func
+            name = func.id if isinstance(func, ast.Name) else getattr(func, "attr", None)
+            if name != "SceneNode":
+                continue
+            passed = {kw.arg for kw in node.keywords if kw.arg}
+            if field_name not in passed:
+                continue
+            for kw in node.keywords:
+                if kw.arg == "kind" and isinstance(kw.value, ast.Constant):
+                    kinds.add(str(kw.value.value))
+    return kinds
+
+
+def test_content_is_written_by_the_kinds_its_comment_names():
+    assert _kinds_constructing_with("content") == CONTENT_KINDS
+
+
+def test_history_is_written_by_the_kinds_its_comment_names():
+    assert _kinds_constructing_with("history") == HISTORY_KINDS
+
+
+def test_the_scan_finds_a_sane_population():
+    """Guards the guard: a renamed class or a changed construction style would
+    make both checks above compare two empty sets and pass."""
+    assert len(_kinds_constructing_with("content")) >= 5
+    assert len(_kinds_constructing_with("history")) >= 5