Fixing sinter post-selection feature

glue/sample/src/sinter/_collection/_collection_test.py

@@ -79,6 +79,51 @@ def test_collect():
     assert d[0.03].errors <= 70
     assert 1 <= d[0.04].errors <= 100
 
+def test_collect_postselection():
+    tasks = []
+    for p in [0.01, 0.02, 0.03, 0.04]:
+        circuit = stim.Circuit.generated(
+            'repetition_code:memory',
+            rounds=3,
+            distance=3,
+            after_clifford_depolarization=p,
+        )
+        mask = sinter._collection.post_selection_mask_from_predicate(
+            circuit_or_dem=circuit,
+            metadata={},
+            postselected_detectors_predicate=lambda _a, _b, _c: True,
+        )
+        tasks.append(sinter.Task(
+            circuit=circuit,
+            decoder='pymatching',
+            postselection_mask=mask,
+            json_metadata={'p': p},
+            collection_options=sinter.CollectionOptions(
+                max_shots=1000,
+                max_errors=100,
+                start_batch_size=100,
+                max_batch_size=1000,
+            ),
+        ))
+
+    results = sinter.collect(
+        num_workers=2,
+        tasks=tasks
+    )
+    probabilities = [e.json_metadata['p'] for e in results]
+    assert len(probabilities) == len(set(probabilities))
+    d = {e.json_metadata['p']: e for e in results}
+    print(d)
+    assert len(d) == 4
+    for k, v in d.items():
+        assert v.shots >= 1000
+        assert v.errors <= 1 # there is some small probability for undetected logical error
+    assert d[0.01].discards <= 200
+    assert d[0.02].discards <= 300
+    assert d[0.03].discards <= 500
+    assert 100 <= d[0.04].discards <= 1000
+
+
 
 def test_collect_from_paths():
     with tempfile.TemporaryDirectory() as d:

glue/sample/src/sinter/_decoding/_stim_then_decode_sampler.py

@@ -197,8 +197,8 @@ def sample(self, max_shots: int) -> AnonTaskStats:
             raise ValueError("predictions.dtype != np.uint8")
         if len(predictions.shape) != 2:
             raise ValueError("len(predictions.shape) != 2")
-        if predictions.shape[0] != num_shots:
-            raise ValueError("predictions.shape[0] != num_shots")
+        if predictions.shape[0] != num_shots - num_discards_1:
+            raise ValueError("predictions.shape[0] != num_shots - num_discards_1")
         if predictions.shape[1] < actual_obs.shape[1]:
             raise ValueError("predictions.shape[1] < actual_obs.shape[1]")
         if predictions.shape[1] > actual_obs.shape[1] + 1: