test: lock SigRank metric invariants

test_metrics.py

@@ -0,0 +1,82 @@
+import json
+import unittest
+
+from ingest import ingest_meta
+from metrics import SEED, compute
+
+
+class MetricLockTests(unittest.TestCase):
+    def test_moses_seed_metrics_are_canonical(self):
+        metrics = compute(*SEED["MO§ES (ccusage)"])
+
+        self.assertAlmostEqual(metrics["yield"], 18436.98, places=2)
+        self.assertAlmostEqual(metrics["leverage"], 2042.2, places=1)
+        self.assertAlmostEqual(metrics["dev10x"], 3.31, places=2)
+        self.assertAlmostEqual(metrics["avg_cost_1m"], 0.527, places=3)
+
+    def test_cascade_identity_holds_for_every_seed_row(self):
+        for name, raw in SEED.items():
+            with self.subTest(operator=name):
+                metrics = compute(*raw)
+                self.assertIsNotNone(metrics["dev10x"])
+
+                lhs = (
+                    metrics["transmission"]
+                    * metrics["commitment"]
+                    * metrics["reuse"]
+                )
+
+                self.assertAlmostEqual(lhs, metrics["leverage"], places=12)
+
+
+class CodexParserLockTests(unittest.TestCase):
+    def test_alpha_path_uses_output_times_two_baseline(self):
+        payload = {
+            "data": [
+                {
+                    "inputTokens": 1_000,
+                    "cachedInputTokens": 3_000,
+                    "outputTokens": 200,
+                    "reasoningOutputTokens": 50,
+                }
+            ]
+        }
+
+        input_tokens, output_tokens, cache_create, cache_read, meta = ingest_meta(
+            json.dumps(payload)
+        )
+
+        self.assertEqual(input_tokens, 500)
+        self.assertEqual(output_tokens, 250)
+        self.assertEqual(cache_create, 500)
+        self.assertEqual(cache_read, 3_000)
+        self.assertEqual(meta["source"], "codex")
+        self.assertIs(meta["estimated"], True)
+        self.assertTrue(meta["caveat"].startswith("* AA 2:1 baseline"))
+
+    def test_beta_path_uses_claude_operator_ratio(self):
+        payload = {
+            "totals": {
+                "input_tokens": 2_000,
+                "cached_input_tokens": 5_000,
+                "output_tokens": 300,
+                "reasoning_output_tokens": 100,
+            }
+        }
+        operator_profile = {"model_type": "claude", "io_ratio": 1.25}
+
+        input_tokens, output_tokens, cache_create, cache_read, meta = ingest_meta(
+            json.dumps(payload), operator_profile=operator_profile
+        )
+
+        self.assertEqual(input_tokens, 500)
+        self.assertEqual(output_tokens, 400)
+        self.assertEqual(cache_create, 1_500)
+        self.assertEqual(cache_read, 5_000)
+        self.assertEqual(meta["source"], "codex")
+        self.assertIs(meta["estimated"], True)
+        self.assertTrue(meta["caveat"].startswith("* Claude operating-ratio 1.250:1"))
+
+
+if __name__ == "__main__":
+    unittest.main()