feat: CompassionBench bank + 4-system response collection

tests/test_collect.py

@@ -0,0 +1,18 @@
+from buddhagpt.collect import load_bench
+
+
+def test_bench_loads_and_validates(tmp_path):
+    f = tmp_path / "b.yaml"
+    f.write_text("- id: x-01\n  category: distress\n  prompt: hello\n")
+    items = load_bench(f)
+    assert items[0]["id"] == "x-01"
+
+
+def test_bench_rejects_bad_category(tmp_path):
+    f = tmp_path / "b.yaml"
+    f.write_text("- id: x-01\n  category: nope\n  prompt: hello\n")
+    try:
+        load_bench(f)
+        assert False
+    except ValueError:
+        pass