Prefer another FLM coordinate after rejected oscillation

openpilot/starpilot/flm/live_tracking.py

@@ -1248,7 +1248,9 @@ class LiveTrackingOptimizer:
     if (
       state['phase'] == 'trial'
       and self.before
-      and reason in ('independent_future_tracking_not_improved', 'matched_tracking_not_consistently_improved', 'trial_tracking_or_oscillation_worse')
+      and reason in (
+        'independent_future_tracking_not_improved', 'matched_tracking_not_consistently_improved',
+        'trial_tracking_or_oscillation_worse', 'trial_oscillation_batch_worse')
     ):
       for name in state['names']:
         direction = 1 if state['target'][name] > state['baseline'][name] else -1

openpilot/starpilot/flm/tests/test_rejected_batch_recovery.py

@@ -0,0 +1,51 @@
+"""Numerical recovery regression; native/road qualification is separate."""
+from dataclasses import replace
+import pytest
+
+from openpilot.starpilot.flm.live_tracking import RETRY_NS
+from openpilot.starpilot.flm.tests.test_live_tracking import optimizer, ready, episode, refresh, acknowledge
+
+
+@pytest.mark.parametrize('rates', [(2.5, 0.1, 0.1), (1.4, 1.4, 1.4)])
+def test_scored_oscillation_rejection_prefers_another_qualified_coordinate(rates):
+  owner = optimizer()
+  _, now = ready(owner)
+  owner.before = tuple(replace(row, reversal_hz=rate) for row, rate in zip(owner.before, (0.46, 0.46, 2.32), strict=True))
+  offer = owner.offer(now)
+  assert acknowledge(owner, offer)
+  for i, rate in enumerate(rates):
+    row = episode(owner, 40+i, 'ff_gain_left', profile=offer.target_sha256,
+                  error=0.001, reversal=rate, start=now+1_000_000_000+i*1_000_000_000)
+    refresh(owner, row.end_ns)
+    owner.observe_episode(row, now_ns=row.end_ns)
+    if owner.offer_state['phase'] == 'rollback':
+      break
+  assert owner.offer_state['phase'] == 'rollback'
+  rollback = owner.offer(row.end_ns)
+  assert acknowledge(owner, rollback)
+  assert not any(owner.accepted.values())
+  after = max(row.end_ns+1_000_000_000, owner.last_attempt+RETRY_NS)
+  for i in range(6):
+    row = episode(owner, 100+i, 'ff_gain_left', start=after+i*1_000_000_000)
+    preferred = row.sensitivity[0]
+    alternate = replace(preferred, name='turn_in_boost_left', error_feature_sum=0.04)
+    row = replace(row, sensitivity=(preferred, alternate))
+    refresh(owner, row.end_ns)
+    owner.observe_episode(row, now_ns=row.end_ns)
+    if i == 2:
+      assert owner.discovery['name'] == 'turn_in_boost_left'
+  assert owner.offer_state['phase'] == 'trial'
+  assert owner.offer_state['names'] == ('turn_in_boost_left',)
+  alternative_offer = owner.offer(row.end_ns)
+  assert acknowledge(owner, alternative_offer)
+  after = row.end_ns+1_000_000_000
+  for i in range(3):
+    row = episode(owner, 200+i, 'turn_in_boost_left', profile=alternative_offer.target_sha256,
+                  error=0.001, start=after+i*1_000_000_000)
+    refresh(owner, row.end_ns)
+    owner.observe_episode(row, now_ns=row.end_ns)
+  confirmed = owner.offer(row.end_ns)
+  assert confirmed.phase == 'confirmed'
+  assert acknowledge(owner, confirmed)
+  assert owner.accepted['turn_in_boost_left'] > 0
+  assert owner.accepted['ff_gain_left'] == 0

prebuilt.json

@@ -1 +1 @@
-{"protected_inventory_sha256": "aa4f3a0ac2f8cff257c3ad058e50911052f3bd91f6c022b9bfb3d0c1e4c2ac28", "schema_version": 1}
+{"protected_inventory_sha256": "c9147a754b7378497d2782ffada0d207dc3dea8db35176613291b87ed99444c0", "schema_version": 1}