Stabilize the objective coordinator routing test cohort (#307)

The objective coordinator routing test intermittently returns `deferred`
because its positive fixture has only 68 holdouts. The real coordinator
replenishes this pool below its floor of 80, introduces randomly ordered
negative controls, and can select enough of them to make the eight-item
experiment underpowered.

Prepare a second genuine scenario pool and apply the fixture’s existing
positive-only policy before this one test, keeping at least 88 holdouts
available. The legacy-forensics prohibition, resume assertions,
production statistics, replenishment behavior, and independent
negative-control tests remain unchanged.

Validation: autonomous/evolution suites passed 74 cases, with one
existing sandbox probe skipped because user namespaces are disabled; the
statistics test passed; six fresh-UUID repetitions of the coordinator
test passed. Independent review found no issues and separately passed
all 37 autonomous tests. The rebased branch contains only the nine-line
test change.

Author: tzhouam

test/test_autonomous_evolution.py

@@ -403,6 +403,15 @@ def test_objective_coordinator_never_calls_legacy_dual_model_forensics(autonomou
     from infermatrix_copilot.improve import coordinator
     from infermatrix_copilot.engine.steps import improve
     st, store, model, worker = autonomous
+    # Keep this large-effect routing test's positive cohort above the
+    # coordinator's replenishment floor, including after its eight holdouts
+    # are consumed. Negative controls are independently asserted above.
+    objectives.prepare(st, store, workflow=objectives.ENGINE)
+    path = objectives.data_root(st) / objectives.ENGINE / 'dataset.json'
+    manifest = json.loads(path.read_text())
+    manifest['items'] = [r for r in manifest['items'] if not r['objective'].get('negative')]
+    artifacts.atomic_json(path, manifest)
+    assert sum(r['split'] == 'holdout' for r in manifest['items']) >= 88
     def forbidden(*args, **kwargs): raise AssertionError('legacy model or human benchmark requested')
     monkeypatch.setattr(improve, '_forensics', forbidden)
     result = coordinator.run(st, store, workflow=objectives.ENGINE, llm=model, sandbox=worker)