fix(benchmark): allow adjudicated runner recovery

loopx/capabilities/benchmark_toolkit/README.md

@@ -1200,6 +1200,11 @@ For a later comparison, call `experiment-board-show` with the same Goal/project
 route and relevant benchmark/study filters. Reuse a current board already read
 when it answers the decision; avoid rebuilding and returning it after every
 single-row update. Validation and invalid-transition rejection still apply.
+An initially fail-closed `runner_invalid` row may advance to `completed` when a
+later deterministic terminal adjudication supplies a schema-valid completed row
+for the same run identity. This promotion is one-way: completed and cancelled
+rows cannot be reopened or demoted. Provider-ledger reconciliation continues to
+reject conflicting terminal shards because it has no single adjudication owner.
 
 The default ledger is locked, atomically updated, and keyed by benchmark, study,
 case, and run identity. A compact row carries arm role, exact and comparison

loopx/capabilities/benchmark_toolkit/experiment_board.py

@@ -28,7 +28,10 @@
     "planned": _RUN_STATUSES,
     "running": {"running", "completed", "runner_invalid", "cancelled"},
     "completed": {"completed"},
-    "runner_invalid": {"runner_invalid"},
+    # A runner may fail closed before post-run evidence is fully reconciled.
+    # Permit the same stable run to advance once to a validated completed row;
+    # completed and cancelled rows remain immutable.
+    "runner_invalid": {"runner_invalid", "completed"},
     "cancelled": {"cancelled"},
 }
 _TERMINAL_RUN_STATUSES = {"completed", "runner_invalid", "cancelled"}

tests/capabilities/test_benchmark_experiment_board.py

@@ -792,6 +792,35 @@ def test_upsert_rejects_stale_transition_after_terminal_state(tmp_path: Path) ->
     assert rows[0]["status"] == "completed"
 
 
+def test_upsert_promotes_runner_invalid_after_terminal_adjudication(
+    tmp_path: Path,
+) -> None:
+    ledger = tmp_path / "experiment-board.jsonl"
+    runner_invalid = _running_baseline(
+        status="runner_invalid",
+        observed_at="2026-08-18T00:10:00+00:00",
+    )
+    completed = _baseline(observed_at="2026-08-18T00:20:00+00:00")
+
+    upsert_benchmark_experiment_board_row(ledger, runner_invalid)
+    receipt = upsert_benchmark_experiment_board_row(ledger, completed)
+    rows = read_benchmark_experiment_board_rows(ledger)
+
+    assert receipt["write"]["status"] == "updated"
+    assert rows == [normalize_benchmark_experiment_board_row(completed)]
+
+
+def test_upsert_does_not_demote_completed_to_runner_invalid(tmp_path: Path) -> None:
+    ledger = tmp_path / "experiment-board.jsonl"
+    upsert_benchmark_experiment_board_row(ledger, _baseline())
+    runner_invalid = _running_baseline(status="runner_invalid")
+
+    with pytest.raises(
+        ValueError, match="cannot move from completed to runner_invalid"
+    ):
+        upsert_benchmark_experiment_board_row(ledger, runner_invalid)
+
+
 def test_reconcile_promotes_terminal_and_skips_late_nonterminal_source() -> None:
     canonical = _running_baseline(observed_at="2026-08-18T00:10:00+00:00")
     source_running = _running_baseline(