From f827df11c6cdfb180d44d8349e44ff4ed43a50ee Mon Sep 17 00:00:00 2001
From: karl <karl@users.noreply.tea.karlcom.de>
Date: Tue, 13 Oct 2026 13:44:06 +0200
Subject: [PATCH] Verify VAD allocation limits and decoder overshoot contracts

---
 docs/SPEECH-VAD.md              |  6 ++++++
 tools/speech-vad/test_native.py | 26 ++++++++++++++++++++++++++
 2 files changed, 32 insertions(+)

diff --git a/docs/SPEECH-VAD.md b/docs/SPEECH-VAD.md
index e41cab5..81096d6 100644
--- a/docs/SPEECH-VAD.md
+++ b/docs/SPEECH-VAD.md
@@ -211,6 +211,12 @@ was scored. All three native contracts passed again with the corrected build;
 four mapping regressions cover centiseconds, source origin, padded tails and
 invalid intervals. Segmentation parameters and model remained frozen.
 
+The expanded native contract suite then passed six tests in 0.471 seconds:
+missing-model failures emit no report, oversized PCM is rejected before model
+loading, and finite decoder overshoot is accepted by actual CPU inference
+without input clipping. The synthetic DC overshoot input tests the input
+contract only and is not included in the media quality pool.
+
 This small VAD-only result supports testing gated ASR next. It does not establish
 speech recall, transcript correctness, timestamp accuracy, retrieval improvement
 or a general music policy. No production gate was added and original Google
diff --git a/tools/speech-vad/test_native.py b/tools/speech-vad/test_native.py
index d225504..9217caf 100644
--- a/tools/speech-vad/test_native.py
+++ b/tools/speech-vad/test_native.py
@@ -48,6 +48,32 @@ class NativeContract(unittest.TestCase):
         self.assertEqual(result.returncode, 2)
         self.assertEqual(result.stdout, b"")
 
+    def test_missing_model_is_failure_without_report(self):
+        with tempfile.TemporaryDirectory(prefix="searchi-vad-missing-") as scratch:
+            pcm = Path(scratch) / "silence.f32"
+            pcm.write_bytes(b"\0" * 64000)
+            result = subprocess.run([str(self.worker), str(Path(scratch) / "missing.bin"), str(pcm)],
+                                    capture_output=True, timeout=5)
+            self.assertEqual(result.returncode, 3)
+            self.assertEqual(result.stdout, b"")
+
+    def test_oversized_pcm_rejected_before_model_loading(self):
+        with tempfile.TemporaryDirectory(prefix="searchi-vad-large-") as scratch:
+            pcm = Path(scratch) / "oversized.f32"
+            with pcm.open("wb") as stream:
+                stream.truncate(16000 * 600 * 4 + 4)
+            result = subprocess.run([str(self.worker), "missing.bin", str(pcm)],
+                                    capture_output=True, timeout=5)
+            self.assertEqual(result.returncode, 2)
+            self.assertEqual(result.stdout, b"")
+
+    def test_finite_decoder_overshoot_is_preserved(self):
+        # A DC input is not speech evidence. This checks only the input contract:
+        # valid floating-point decoder output must not be clipped or rejected.
+        result = self.run_pcm(struct.pack("<f", 1.01) * 16000)
+        self.assertEqual(result.returncode, 0, result.stderr.decode("utf-8", errors="replace"))
+        self.assertEqual(json.loads(result.stdout)["samples"], 16000)
+
 
 if __name__ == "__main__":
     unittest.main()