fix(node): the roster, not the key alone, decides who gets a session

The handshake opened a session for anyone holding the group key with a hub
token naming the group; the roster was consulted only when wrapping the key
in a join. A member revoked or unpinned on the node but still a member on
the hub kept a full session with the key they already held — and was handed
the chat epoch their removal had just opened, since chat keys go to any
session. An honest client never met this (it asks for the key through
join_request every time); one that kept the key did not have to.

- After the proof, the node asks the roster and refuses with
  `not_authorized_for_group` unless the account is an active member of the
  group or the node's operator.
- A removal from any door — MNP, the node page, the CLI — now opens a new
  chat epoch in each group the person could read, broadcasts it, and closes
  every connection they hold (`ops.members._after_removal`). The CLI and the
  node page did neither.
- Design §5.2, protocol §6.1, §6.3, §14.2.

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

3 files changed, 91 insertions, 18 deletions

diff --git a/packages/meshbay-node/src/meshbay_node/ops/members.py b/packages/meshbay-node/src/meshbay_node/ops/members.py
index b5da9c0..2562dd6 100644
--- a/packages/meshbay-node/src/meshbay_node/ops/members.py
+++ b/packages/meshbay-node/src/meshbay_node/ops/members.py
@@ -262,6 +262,52 @@ async def cancel_link_invitation(state: dict, group_id: str, invite_id: str) ->
             "node": node_cancelled, "hub": hub_cancelled}
 
 
+async def _after_removal(state: dict, user_id: str, group_ids: list[str]) -> None:
+    """
+    What a removal has to do beyond the roster, whichever door it came through.
+
+    - **A new chat epoch in every group the person could read.** They keep the
+      key of the epoch they were in; the next one is what they must not get.
+    - **Every live connection of theirs closed.** The handshake now refuses them
+      (it consults the roster), so this is what makes the removal immediate
+      rather than "at their next reconnection".
+
+    It used to live in the MNP handlers only, so a removal from the CLI or the
+    node page — the doors an operator at their own machine uses — left the
+    epoch where it was and the person connected. Best effort: a failure here
+    must not turn a done removal into a refused one, and is logged loudly.
+    """
+    from meshbay_common import MNP_VERSION
+    from meshbay_common.protocol import MNP
+
+    from meshbay_node.ops.chat import open_chat_epoch
+
+    groups_ctx = state.get("groups_ctx") or {}
+    for gid in sorted({g for g in group_ids if g}):
+        try:
+            result = await open_chat_epoch(state, gid)
+        except Exception as e:
+            log.error("Removal of %s: no new chat epoch for group %s (%s) — "
+                      "they still hold the current chat key", user_id[:8], gid[:8], e)
+            continue
+        notice = {"type": MNP.CHAT_EPOCH_ACK, "v": MNP_VERSION, "epoch": result["epoch"]}
+        for session in list((groups_ctx.get(gid) or {}).get("_peers", {}).values()):
+            try:
+                session._send(notice)
+            except Exception:
+                pass
+
+    webrtc = state.get("webrtc")
+    if webrtc is not None:
+        for session in list(getattr(webrtc, "_sessions", {}).values()):
+            if session._user_id == user_id and (
+                    not group_ids or session._group_id in group_ids):
+                try:
+                    await session.close()
+                except Exception:
+                    pass
+
+
 async def revoke_member(state: dict, user_id: str, group_id: str) -> dict:
     """
     Stop serving the group key to someone.
@@ -285,6 +331,7 @@ async def revoke_member(state: dict, user_id: str, group_id: str) -> dict:
         raise OpError("No such member in that group", status=404)
     log.info("Member revoked: user=%s group=%s member=%s invites_dropped=%d",
              user_id[:8], group_id[:8], revoked, dropped)
+    await _after_removal(state, user_id, [group_id] if revoked else [])
     return {"status": "revoked", "user_id": user_id, "group_id": group_id,
             "was_member": revoked, "invites_dropped": dropped,
             # Only what is true: somebody who never redeemed a code never held
@@ -297,6 +344,8 @@ async def revoke_member(state: dict, user_id: str, group_id: str) -> dict:
 async def unpin_member(state: dict, user_id: str) -> dict:
     """Forget a pinned identity, so the person can pair again with a new key."""
     roster = _roster(state)
+    groups = [m["group_id"] for m in await roster.list_members()
+              if m.get("user_id") == user_id and m.get("group_id")]
     if not await roster.unpin(user_id):
         raise OpError("No such pinned identity", status=404)
     # Drop the stored keypair bundle too. Left behind, it is served to the next
@@ -310,4 +359,5 @@ async def unpin_member(state: dict, user_id: str) -> dict:
         except Exception:
             log.warning("unpin: could not drop keypair bundle for %s", user_id[:8])
     log.info("Identity unpinned: user=%s", user_id[:8])
+    await _after_removal(state, user_id, groups)
     return {"status": "unpinned", "user_id": user_id}
diff --git a/packages/meshbay-node/src/meshbay_node/transport/webrtc/group_ops.py b/packages/meshbay-node/src/meshbay_node/transport/webrtc/group_ops.py
index 5f5f9ed..3756eac 100644
--- a/packages/meshbay-node/src/meshbay_node/transport/webrtc/group_ops.py
+++ b/packages/meshbay-node/src/meshbay_node/transport/webrtc/group_ops.py
@@ -113,8 +113,9 @@ class GroupOpsMixin:
             self._audit("admin_auth_failed", f"member_unpin:{user_id[:8]}")
             return
         try:
+            # The new chat epochs and the closed sessions are the op's own
+            # (`ops.members._after_removal`), for every door alike.
             await self._run_op(ops.unpin_member, user_id)
-            await self._new_chat_epoch(self._group_id or "", "member_unpin")
         except ops.OpError as e:
             self._send({"type": "error", "detail": e.message})
             return
@@ -331,26 +332,17 @@ class GroupOpsMixin:
             return
 
         try:
+            # The op opens the new chat epoch and closes every connection the
+            # account holds in this group (`ops.members._after_removal`), for
+            # the loopback API and the CLI as much as for this door. They keep
+            # the key they already unwrapped; rotating it is the operator's
+            # call, and the ack says so.
             result = await self._run_op(
                 ops.revoke_member, user_id, self._group_id or "")
-            await self._new_chat_epoch(self._group_id or "", "member_revoke")
         except ops.OpError as e:
             self._send({"type": "error", "detail": e.message})
             return
 
-        # Anyone connected right now keeps the key they already unwrapped; what
-        # they lose is the next one. Rotating it is the operator's call, and the
-        # ack says so rather than implying this undid anything already read.
-        # Every connection that account holds, not "the" one: with device
-        # linking a person may be connected from several at once, and the
-        # registry is keyed per connection precisely because it cannot hold
-        # only one of them.
-        for peer in self._sessions_of(user_id):
-            try:
-                await peer.close()
-            except Exception:
-                pass
-
         self._audit("member_revoke", user_id)
         self._send({
             "type": MNP.MEMBER_REVOKE_ACK, "v": MNP_VERSION,
diff --git a/packages/meshbay-node/src/meshbay_node/transport/webrtc/handshake.py b/packages/meshbay-node/src/meshbay_node/transport/webrtc/handshake.py
index f701072..7822186 100644
--- a/packages/meshbay-node/src/meshbay_node/transport/webrtc/handshake.py
+++ b/packages/meshbay-node/src/meshbay_node/transport/webrtc/handshake.py
@@ -137,7 +137,8 @@ class HandshakeMixin:
         return {"sig": base64.b64encode(self._ctx["sk_node"].sign(transcript)).decode()}
 
     def _do_handshake_response(self, msg: dict) -> None:
-        if not self._gek_challenge or not hasattr(self, "_pending_sub"):
+        if not self._gek_challenge or not hasattr(self, "_pending_sub") \
+                or getattr(self, "_admitting", False):
             self._send({"type": "error", "detail": "No pending handshake challenge"})
             return
 
@@ -170,8 +171,38 @@ class HandshakeMixin:
             self._audit_auth_failed(group_id, "GEK HMAC mismatch")
             return
 
-        self._complete_handshake(gek, binding)
-        self._gek_challenge = None
+        # One answer at a time: the roster is asked asynchronously, and a
+        # second response arriving meanwhile must not be taken as a new one.
+        self._admitting = True
+        self._spawn(self._admit(gek, binding, group_id))
+
+    async def _admit(self, gek: bytes, binding: bytes, group_id: str) -> None:
+        """
+        Open the session only for someone this node's roster admits.
+
+        The group-key proof says the peer holds the key, and the token says the
+        hub counts the account a member. Neither is the node's own answer —
+        and the roster is supposed to be the authority (§6.1). Without this, a
+        member revoked here, or unpinned, but still a member on the hub, kept
+        full access with the key they already held: files, uploads, and — since
+        chat keys are handed to any session — the very epoch their removal had
+        just opened. An honest client never met the gap, because it asks for
+        the key through `join_request`, which does consult the roster; a
+        client that kept the key did not have to.
+        """
+        try:
+            roster = self._ctx.get("roster")
+            if roster is not None and not await roster.is_authorized(
+                    group_id, self._pending_sub):
+                self._send({"type": "error",
+                            "detail": "This node has not admitted you to this group",
+                            "code": "not_authorized_for_group"})
+                self._audit_auth_failed(group_id, "not admitted by the roster")
+                return
+            self._complete_handshake(gek, binding)
+        finally:
+            self._gek_challenge = None
+            self._admitting = False
 
     def _complete_handshake(self, gek: bytes, binding: bytes) -> None:
         # Authenticated peers may send large frames (file uploads); unauthenticated