From 9fc77a256b05db0417d0f090bd3b6e41f94ce186 Mon Sep 17 00:00:00 2001
From: firestar5683 <168790843+firestar5683@users.noreply.github.com>
Date: Sun, 25 Oct 2026 14:59:31 -0500
Subject: [PATCH] Try feasible Auto Learn directions at coordinate bounds

---
 openpilot/starpilot/flm/live_tracking.py      |  8 +++
 .../flm/tests/test_tracking_source_context.py | 57 +++++++++++++++++++
 prebuilt.json                                 |  2 +-
 3 files changed, 66 insertions(+), 1 deletion(-)

diff --git a/openpilot/starpilot/flm/live_tracking.py b/openpilot/starpilot/flm/live_tracking.py
index ad84caf1b0..bc1539bf76 100644
--- a/openpilot/starpilot/flm/live_tracking.py
+++ b/openpilot/starpilot/flm/live_tracking.py
@@ -1041,6 +1041,14 @@ class LiveTrackingOptimizer:
               for old_name, old_direction, old_profile, demand, old_end in self.closed_loop_worse
             )
 
+          # Without a fitted plant, the secant sign is only an experiment
+          # preference. At a hard bound it must not veto the feasible direction.
+          # Actual worse/inconclusive directions below remain disqualified;
+          # independent baseline, later feedback and slew are unchanged.
+          low, high = coordinate_bounds(name)
+          if self.rmse is None and min(high, max(low, self.accepted[name] + sign * min(MAX_STEP, 0.1 * confidence))) == self.accepted[name]:
+            sign = -sign
+
           if observed_worse(sign):
             if observed_worse(-sign):
               continue  # Both directions failed; do not oscillate guesses.
diff --git a/openpilot/starpilot/flm/tests/test_tracking_source_context.py b/openpilot/starpilot/flm/tests/test_tracking_source_context.py
index c3b1c49fa6..e44c58cdba 100644
--- a/openpilot/starpilot/flm/tests/test_tracking_source_context.py
+++ b/openpilot/starpilot/flm/tests/test_tracking_source_context.py
@@ -387,3 +387,60 @@ def test_small_step_can_improve_mean_and_median_with_normal_baseline_variation()
   assert _feedback_result([1.0] * 9, [0.1] + [1.01] * 8) == 0
   # Repeated new errors outside the original envelope are not kept.
   assert _feedback_result([1.0] * 9, [0.9] * 7 + [1.1] * 2) == 0
+
+
+@pytest.mark.parametrize('upper', [False, True])
+def test_source_boundary_tries_feasible_direction_without_claiming_plant_sign(upper):
+  from openpilot.starpilot.flm.canonical_surface_adapter import coordinate_bounds
+
+  o = owner()
+  name = 'center_deadband_highway_deg'
+  low, high = coordinate_bounds(name)
+  o.accepted[name] = high if upper else low
+  discovery_ids = set()
+  for i in range(5):
+    e = episode(o, i, name, kind='large', error=0.079)
+    # Magnitudes retained from c0's original native probe: one of four
+    # endpoints separates. This fixture tests selection, not native admission.
+    f = replace(e.sensitivity[0], error_feature_sum=0.015966462570723775 if upper else -0.015966462570723775,
+                feature_energy=0.2391186088079348, native_separated=1, native_endpoints=4, max_step_physical_rms=0.0012224938875305645)
+    e = replace(e, goal_rms=0.77, normalization_accel=0.77, sensitivity=(f,))
+    refresh(o, e.end_ns)
+    o.observe_episode(e, now_ns=e.end_ns)
+    if i == 1:
+      assert o.discovery is not None and o.offer_state is None
+      discovery_ids.update(o.selector_hint['source_episode_ids'])
+  selected = o.offer(e.end_ns)
+  assert selected is not None and selected.names == (name,)
+  assert dict(selected.target)[name] == pytest.approx(o.accepted[name] + (-0.005 if upper else 0.005))
+  assert len(o.before) == 3 and discovery_ids.isdisjoint(row.episode_id for row in o.before)
+  assert not o.offer_state['accepted'] and o.offer_state['alpha'] == 0.0
+
+
+@pytest.mark.parametrize('outcome', ['worse', 'inconclusive'])
+def test_source_boundary_fallback_does_not_repeat_rejected_feasible_direction(outcome):
+  from openpilot.starpilot.flm.live_tracking import _demand
+
+  o = owner()
+  name = 'center_deadband_highway_deg'
+  for i in range(4):
+    e = replace(episode(o, i, name, kind='large'), normalization_accel=0.5)
+    e = replace(e, sensitivity=tuple(replace(f, error_feature_sum=-f.error_feature_sum) for f in e.sensitivity))
+    if i == 0:
+      history = o.closed_loop_worse if outcome == 'worse' else o.inconclusive_trials
+      history.append((name, 1, e.profile_sha256, _demand(e), e.start_ns))
+    refresh(o, e.end_ns)
+    o.observe_episode(e, now_ns=e.end_ns)
+  assert o.discovery is None and o.offer_state is None
+
+
+def test_fitted_plant_does_not_reverse_identified_direction_at_boundary():
+  from openpilot.starpilot.flm.tests.test_live_tracking import optimizer, refresh as fitted_refresh
+
+  o = optimizer()
+  for i in range(20):
+    e = episode(o, i, 'center_deadband_highway_deg', kind='subtle')
+    e = replace(e, sensitivity=tuple(replace(f, error_feature_sum=-f.error_feature_sum) for f in e.sensitivity))
+    fitted_refresh(o, e.end_ns)
+    o.observe_episode(e, now_ns=e.end_ns)
+  assert o.discovery is None and o.offer_state is None
diff --git a/prebuilt.json b/prebuilt.json
index 4f9fce351f..daa13ae6cb 100644
--- a/prebuilt.json
+++ b/prebuilt.json
@@ -1 +1 @@
-{"protected_inventory_sha256": "fb8358c0f039c9b61c8725ff2bcd0cf9b2e44c9075b85555e0c156240896aba3", "schema_version": 1}
+{"protected_inventory_sha256": "da94b10aa940d202e229d8fb1e36321191a7aa2de68541fd8b2f19c6816bd876", "schema_version": 1}