Balance independent Ford lifecycle cases across CI shards

prebuilt.json

@@ -1 +1 @@
-{"protected_inventory_sha256": "d271bc09145447f9fc6d5d41d68a0d7e3fbc38be1f6765c21662dd79d8071b1e", "schema_version": 1}
+{"protected_inventory_sha256": "079e92a64c345ad09f03d5a7ede05b13deea46058bbdcc4a29011799cf17a50d", "schema_version": 1}

tools/ci/custom_pytest_shards.py

@@ -15,6 +15,7 @@ from tools.ci.lifecycle_shards import build_identity, same_build_inputs
 
 ROOT = Path(__file__).resolve().parents[2]
 HEAVY_MODULES = {
+  'openpilot/starpilot/tests/test_ford_bluepilot_library_lifecycle.py',
   'openpilot/starpilot/car/toyota/tests/test_camera_pedal_extension_callers.py',
   'openpilot/starpilot/car/toyota/tests/test_eps8_sdsu_callers.py',
   'openpilot/starpilot/car/toyota/tests/test_legacy_advert_callers.py',

tools/ci/custom_pytest_workloads.json

@@ -165,6 +165,10 @@
     "openpilot/starpilot/car/toyota/tests/test_prius_sdsu_callers.py": {
       "basis": "90 actual DEBUG/RELEASE caller cases passed in four bounded phases (186.885s combined wall time including startup). Scheduling estimate 3s/case; fresh per-case Params, raw CAN, native configuration and C++ host reset; no shared outcome dependence.",
       "seconds_per_case": 3.0
+    },
+    "openpilot/starpilot/tests/test_ford_bluepilot_library_lifecycle.py": {
+      "basis": "736580721a hosted shard7: 248 completed Ford lifecycle cases used573.067107459 phase seconds; slowest case4.394150353s. Round up to5s. Function-scoped library_settings and ford_joined create fresh Params, host/native safety per case and tear down each case; full in-case timelines remain unchanged.",
+      "seconds_per_case": 5.0
     }
   },
   "independent_max_cases": 60,
@@ -1982,12 +1986,6 @@
       "cases": 936,
       "seconds": 1.740757
     },
-    "openpilot/starpilot/system/starport/tests/test_starport.py": {
-      "basis": "Exact 65660a4 hosted main40623875380: all 72 nodes passed, 22.037476s setup/call/teardown; 25% estimate margin, retaining any higher same-source prior estimate. Scheduling only; no test deadline change.",
-      "blob": "c4c3d21bea7d749438afeb20886652bca06208bb",
-      "cases": 72,
-      "seconds": 27.546846
-    },
     "openpilot/starpilot/system/starport/tests/test_bugsink_report.py": {
       "basis": "Exact 65660a4 hosted main40623875380: all 33 nodes passed, 0.737380s setup/call/teardown; 25% estimate margin, retaining any higher same-source prior estimate. Scheduling only; no test deadline change.",
       "blob": "b759b5e8dc1c52623809ff957bc282ca31c7e539",
@@ -2006,6 +2004,12 @@
       "cases": 30,
       "seconds": 0.161266
     },
+    "openpilot/starpilot/system/starport/tests/test_starport.py": {
+      "basis": "Exact 65660a4 hosted main40623875380: all 72 nodes passed, 22.037476s setup/call/teardown; 25% estimate margin, retaining any higher same-source prior estimate. Scheduling only; no test deadline change.",
+      "blob": "c4c3d21bea7d749438afeb20886652bca06208bb",
+      "cases": 72,
+      "seconds": 27.546846
+    },
     "openpilot/starpilot/tests/test_captured_adrv_diagnostic_startup.py": {
       "basis": "Exact 65660a4 hosted main40623875380: all 64 nodes passed, 12.454431s setup/call/teardown; 25% estimate margin, retaining any higher same-source prior estimate. Scheduling only; no test deadline change.",
       "blob": "7849a690cd1ac2f3064f60d72622ea7150015cb8",

tools/ci/lifecycle_shards.py

@@ -92,6 +92,7 @@ def workload_policy():
   policy = json.loads(Path(__file__).with_name('lifecycle_workloads.json').read_text())
   if policy['version'] != 1 or set(policy['function_groups']) != {
       'openpilot/starpilot/car/toyota/tests/test_legacy_advert_callers.py',
+      'openpilot/starpilot/tests/test_ford_bluepilot_library_lifecycle.py',
       'openpilot/starpilot/tests/test_ford_configuration_parity.py',
       'openpilot/starpilot/aol/tests/test_stock_location_recovery.py',
       'openpilot/starpilot/car/toyota/tests/test_camera_2aa_callers.py',

tools/ci/lifecycle_workloads.json

@@ -29,6 +29,11 @@
       "basis": "Scheduling estimate2s/case: actual188 shard3 ordinal577..648 (all inside140case stock-location module) took91.261s; incomplete module timing, not a full measurement. Fresh function-scoped fixture creates CP/state/feedback and private Params per case; no shared outcome dependence; full in-case timelines retained.",
       "estimated_seconds_per_case": 2.0,
       "independent_parameters": true
+    },
+    "openpilot/starpilot/tests/test_ford_bluepilot_library_lifecycle.py": {
+      "basis": "736580721a hosted shard7: 248 completed Ford lifecycle cases used573.067107459 phase seconds; slowest case4.394150353s. Round up to5s. Function-scoped library_settings and ford_joined create fresh Params, host/native safety per case and tear down each case; full in-case timelines remain unchanged.",
+      "estimated_seconds_per_case": 5.0,
+      "independent_parameters": true
     }
   },
   "independent_function_max_cases": 60,

tools/ci/tests/test_custom_pytest_shards.py

@@ -621,3 +621,33 @@ def test_measured_slow_module_exact_function_and_leaf_groups(module, counts, exp
 def test_invalid_shard_count_still_denied(count):
   with pytest.raises(ValueError):
     shards.partition(['a.py::test_a'], count)
+
+
+@pytest.mark.parametrize("count", [24, 32])
+def test_ford_library_cases_keep_complete_timelines_in_bounded_groups(count):
+  module = 'openpilot/starpilot/tests/test_ford_bluepilot_library_lifecycle.py'
+  counts = {
+      "test_library_angle_curvature_aol_native_lifecycle": 96,
+      "test_model_loss_still_withdraws_library_output_through_actual_callers": 3,
+      "test_gas_held_first_engagement": 8,
+      "test_sustained_real_loop_lag_still_disables_ford": 1,
+      "test_cancel_at_every_native_receipt_phase": 20,
+      "test_manual_eps_refusal_keeps_aol_intent_without_steering_through_fault": 2,
+      "test_cruise_resume_preserves_deliberate_lateral_pause": 2,
+      "test_manual_steering_retires_held_turn_limit_alert": 2,
+      "test_fresh_cruise_withdrawal_after_previous_output_expires": 4,
+      "test_reengaging_in_a_manual_curve_respects_native_convergence": 32,
+      "test_brake_retains_independent_lateral_intent": 48,
+      "test_starpilot_tuning_uses_current_native_aol_path": 32,
+      "test_starpilot_low_speed_assistance_survives_brake_with_native_limits": 8
+  }
+  catalog = [f"{module}::{name}[{index}]" for name, cases in counts.items() for index in range(cases)]
+  assignment, loads = shards.partition(catalog, count)
+  assert len(catalog) == 258
+  assert sorted(node for group in assignment for node in group) == sorted(catalog)
+  assert max(map(len, assignment)) <= 48
+  assert max(loads) <= 240
+  assert sum(loads) == 258 * 5
+  reverse, reverse_loads = shards.partition(list(reversed(catalog)), count)
+  assert [set(group) for group in reverse] == [set(group) for group in assignment]
+  assert reverse_loads == loads

tools/ci/tests/test_lifecycle_shards.py

@@ -73,23 +73,20 @@ class LifecycleShardTests(unittest.TestCase):
     self.assertCountEqual(selected, identities)
     self.assertEqual(shards, helper.partition(list(reversed(targets)), list(reversed(identities)), 2))
 
-  def test_measured_bluepilot_lifecycle_is_not_packed_as_unknown_work(self):
+  def test_measured_bluepilot_lifecycle_cases_split_without_omission(self):
     module = 'openpilot/starpilot/tests/test_ford_bluepilot_library_lifecycle.py'
     policy = helper.workload_policy()
-    measured = policy['modules'][module]
-    self.assertNotIn(module, policy['function_groups'])
-    self.assertEqual(measured['cases'], 169)
-    self.assertGreater(measured['seconds'], 300 * policy['unknown_seconds_per_case'])
-    self.assertEqual(len(measured['source_sha256']), 64)
-    self.assertEqual(measured['reference']['artifact_id'], 13071624813)
+    self.assertTrue(policy['function_groups'][module]['independent_parameters'])
+    self.assertEqual(policy['function_groups'][module]['estimated_seconds_per_case'], 5.)
     prefix = module.removesuffix('.py').replace('/', '.') + '.'
-    identities = [prefix + f'test_case[{index}]' for index in range(169)]
-    identities += [f'other.test_case[{index}]' for index in range(300)]
-    shards = helper.partition([module, 'other.py'], identities, 2)
-    self.assertEqual(shards, [[module], ['other.py']])
+    identities = [prefix + f'test_case[{index}]' for index in range(258)]
+    shards = helper.partition([module], identities, 16)
     selected = [identity for shard in shards for identity in identities
                 if any(helper.matches(identity, target) for target in shard)]
     self.assertCountEqual(selected, identities)
+    self.assertEqual(len(selected), 258)
+    self.assertLessEqual(max(map(len, shards)), 17)
+    self.assertEqual(shards, helper.partition([module], list(reversed(identities)), 16))
 
   def test_only_large_reviewed_heavy_functions_split_into_exact_parameter_leaves(self):
     heavy = 'openpilot/starpilot/car/toyota/tests/test_legacy_advert_callers.py'