test(C6): pin CLI start refused while a web session is grace-detached

Council item C6 (SIGNOFF-M2/ARBITRATION.md), matrix gap #1 of 3 (the
other two -- C1's interleaving, C4's restart-with-live-child -- are
closed by those items' own tests). No code change: claim_blocks_cli_start
never consults _grace state, only pid liveness, so this cell was already
correct by construction. The test exists to pin that the two mechanisms
compose correctly, not to fix a defect.

Test: TestWebThenCli::
test_cli_start_is_refused_while_a_web_session_is_detached_within_grace --
schedules a real grace-release timer for a web claim with a live pid,
confirms has_pending_release() is true, then asserts a CLI start still
refuses with the existing message and logs no reclaim.

Evidence: reviews/2026-09-02-full-repo-review/evidence/M2/step-9/C6/.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Author: NetDevAutomate

docs/architecture/session-authority.md

@@ -98,6 +98,16 @@ empty) and cell 2/4's stale-claim branch (CLI start, either owner shape):
 the recorded owner is provably dead — reclaimed, not blocked, in both
 directions.
 
+**C6 (council, matrix gap, closed):** cell 4's live-pid branch is also
+exercised while the web session is merely DETACHED within its grace
+window (`_grace.has_pending_release()` true — client's tab closed, agent
+still running, not yet released) — `TestWebThenCli::
+test_cli_start_is_refused_while_a_web_session_is_detached_within_grace`.
+`claim_blocks_cli_start` does not consult grace state at all (only pid
+liveness), which is exactly why this cell was worth pinning explicitly:
+proof the two mechanisms compose correctly, not merely that each is
+individually correct.
+
 ## 4. End matrix
 
 | Ends | Effect |

packages/studyloop/tests/test_session_authority_matrix.py

@@ -435,6 +435,40 @@ def test_a_dead_web_claim_is_reclaimed_by_a_cli_start(self, caplog) -> None:
         assert len(warnings) == 1
         assert "web-sess-crashed" in warnings[0].message
 
+    def test_cli_start_is_refused_while_a_web_session_is_detached_within_grace(
+        self, caplog
+    ) -> None:
+        """C6 (matrix gap the council named): a web session merely DETACHED
+        within its grace window (client's tab closed, agent still running,
+        not yet released) must still refuse a CLI start with the existing
+        message -- the grace window is not license to reclaim.
+        claim_blocks_cli_start does not consult grace state at all (only
+        pid liveness), which is exactly why this cell was worth pinning:
+        proof the two mechanisms compose correctly rather than merely each
+        being individually tested."""
+        from studyloop.web.routes.session import _grace
+
+        state = _fixture("web-live")
+        state["pid"] = os.getpid()
+
+        async def _schedule() -> None:
+            _grace.schedule_release("web-sess-1", grace=60.0)
+
+        run_async(_schedule())
+        assert _grace.has_pending_release("web-sess-1") is True
+
+        with (
+            patch("studyloop.tmux.is_tmux_available", return_value=True),
+            patch("studyloop.agent_launcher.shutil.which", return_value="/usr/bin/claude"),
+            patch("studyloop.session_state.read_session_state", return_value=state),
+            patch("studyloop.session.cleanup.auto_clean_zombies"),
+            pytest.raises(SessionStartError) as exc_info,
+        ):
+            start_session("Async IO", "claude", "study", "elapsed", 5, False)
+
+        assert "already active" in exc_info.value.message
+        assert not any("reclaim" in rec.message.lower() for rec in caplog.records)
+
     def test_a_web_claim_without_a_pid_blocks_conservatively(self) -> None:
         """A web-owned claim with no `pid` key (written by a build before
         this fix) blocks rather than silently reclaiming a claim whose