fix(vrf): replace mutable default RNG with per-call generator

backend/consensus/vrf.py

@@ -1,17 +1,27 @@
-from datetime import datetime
 import numpy as np
+from numpy.random import Generator as RNG
 from backend.consensus.base import DEFAULT_VALIDATORS_COUNT
 
 
 def get_validators_for_transaction(
     nodes: list[dict],
     num_validators: int | None = None,
-    rng=np.random.default_rng(seed=int(datetime.now().timestamp())),
+    rng: RNG | None = None,
 ) -> list[dict]:
     """
     Returns subset of validators for a transaction.
     The selelction and order is given by a random sampling based on the stake of the validators.
+
+    Args:
+        nodes: List of validator dicts, each with a "stake" key.
+        num_validators: How many validators to select. Defaults to DEFAULT_VALIDATORS_COUNT.
+        rng: Optional numpy Generator instance. A fresh default_rng() is created per call
+             when not provided — avoids the mutable-default-argument pitfall where a single
+             shared Generator would make the selection sequence predictable across calls.
     """
+    if rng is None:
+        rng = np.random.default_rng()
+
     if num_validators is None:
         num_validators = DEFAULT_VALIDATORS_COUNT
 

tests/unit/test_vrf.py

@@ -66,3 +66,26 @@ def choice_mock(a, p, size, replace):
 
     rng.choice.assert_called_once()
     assert validators == [{"stake": 3}, {"stake": 2}, {"stake": 1}]
+
+
+def test_get_validators_for_transaction_independent_rng_per_call() -> None:
+    """
+    Regression test for #1733: mutable default argument caused a single RNG instance
+    to be shared across all calls, making the selection sequence predictable.
+
+    Each call without an explicit rng must use an independent generator so that
+    the selections are not correlated across invocations.
+    """
+    nodes = [{"stake": i + 1} for i in range(10)]
+
+    # Collect first-element selections over many independent calls.
+    # With a shared RNG the sequence would be perfectly deterministic; with
+    # independent generators the distribution should spread across all nodes.
+    first_selections = [get_validators_for_transaction(nodes, 1)[0] for _ in range(200)]
+    unique_first = {tuple(v.items()) for v in first_selections}
+
+    # All 10 nodes should appear as first choice at least once over 200 trials.
+    assert len(unique_first) > 1, (
+        "Only one unique first-validator returned across 200 calls — "
+        "the RNG is likely being shared (mutable default argument bug)."
+    )