refactor(BA-6057): remove legacy scheduler package superseded by Sokovan

- Delete src/ai/backend/manager/scheduler/ (FIFO/LIFO/DRF schedulers,
  agent selectors, predicates, utils) — all functionality has been ported
  to manager/sokovan/scheduler/.
- Move ScheduleType enum into manager/sokovan/scheduler/types.py and
  update its six importers.
- Remove dead settle_agent_alloc() and KernelAgentBinding usage from
  AgentRegistry (no callers since the Phase 3 JSONB write removal).
- Drop the orphaned backendai_scheduler_v10 and backendai_agentselector_v10
  entrypoint groups from BUILD and plugin entrypoint metadata.
- Update tests/unit/plugin/test_entrypoint.py fixtures, model row
  comments, and manager/CLAUDE.md Scheduler guardrail to point at the
  Sokovan location.

Author: fregataa

src/ai/backend/manager/BUILD

@@ -40,17 +40,6 @@ python_distribution(
             "mgr": "ai.backend.manager.cli.__main__:main",
             "mgr.start-server": "ai.backend.manager.cli.start_server:main",
         },
-        "backendai_scheduler_v10": {
-            "fifo": "ai.backend.manager.scheduler.fifo:FIFOSlotScheduler",
-            "lifo": "ai.backend.manager.scheduler.fifo:LIFOSlotScheduler",
-            "drf": "ai.backend.manager.scheduler.drf:DRFScheduler",
-        },
-        "backendai_agentselector_v10": {
-            "legacy": "ai.backend.manager.scheduler.agent_selector:LegacyAgentSelector",
-            "roundrobin": "ai.backend.manager.scheduler.agent_selector:RoundRobinAgentSelector",
-            "concentrated": "ai.backend.manager.scheduler.agent_selector:ConcentratedAgentSelector",
-            "dispersed": "ai.backend.manager.scheduler.agent_selector:DispersedAgentSelector",
-        },
         "backendai_error_monitor_v20": {
             "intrinsic": "ai.backend.manager.plugin.error_monitor:ErrorMonitor",
         },

src/ai/backend/manager/CLAUDE.md

@@ -31,5 +31,5 @@ Each layer has its own `CLAUDE.md`. Read it before modifying code in that direct
 
 ## Scheduler
 
-- Scheduling logic belongs in `manager/scheduler/` — do not add scheduling decisions inside
-  API handlers or service methods.
+- Scheduling logic belongs in `manager/sokovan/scheduler/` — do not add scheduling decisions
+  inside API handlers or service methods.

src/ai/backend/manager/event_dispatcher/handlers/schedule.py

@@ -18,12 +18,12 @@
 from ai.backend.common.types import AgentId
 from ai.backend.logging.utils import BraceStyleAdapter
 from ai.backend.manager.data.deployment.types import DeploymentLifecycleSubStep
-from ai.backend.manager.scheduler.types import ScheduleType
 from ai.backend.manager.sokovan.deployment.coordinator import DeploymentCoordinator
 from ai.backend.manager.sokovan.deployment.route.coordinator import RouteCoordinator
 from ai.backend.manager.sokovan.deployment.route.types import RouteLifecycleType
 from ai.backend.manager.sokovan.deployment.types import DeploymentLifecycleType
 from ai.backend.manager.sokovan.scheduler.coordinator import ScheduleCoordinator
+from ai.backend.manager.sokovan.scheduler.types import ScheduleType
 from ai.backend.manager.sokovan.scheduling_controller import SchedulingController
 
 log = BraceStyleAdapter(logging.getLogger(__spec__.name))

src/ai/backend/manager/models/kernel/row.py

@@ -358,7 +358,7 @@ class KernelRow(Base):  # type: ignore[misc]
     #         // an ISO 8601 formatted timestamp of the last attempt
     #     "failed_predicates": [
     #       { "name": "concurrency", "msg": "You cannot run more than 30 concurrent sessions." },
-    #           // see the manager.scheduler.predicates module for possible messages
+    #           // see the sokovan scheduler validator modules for possible messages
     #       ...
     #     ],
     #     "passed_predicates": [ {"name": "reserved_time"}, ... ],  // names only

src/ai/backend/manager/models/session/row.py

@@ -553,7 +553,7 @@ class SessionRow(Base):  # type: ignore[misc]
     #         // an ISO 8601 formatted timestamp of the last attempt
     #     "failed_predicates": [
     #       { "name": "concurrency", "msg": "You cannot run more than 30 concurrent sessions." },
-    #           // see the manager.scheduler.predicates module for possible messages
+    #           // see the sokovan scheduler validator modules for possible messages
     #       ...
     #     ],
     #     "passed_predicates": [ {"name": "reserved_time"}, ... ],  // names only

src/ai/backend/manager/registry.py

@@ -184,7 +184,6 @@
 from .models.vfolder import (
     verify_vfolder_name,
 )
-from .scheduler.types import KernelAgentBinding
 from .types import UserScope
 
 type MSetType = Mapping[str | bytes, bytes | float | int | str]
@@ -1383,47 +1382,6 @@ async def update_scaling_group(self, agent_id: AgentId, scaling_group: str) -> N
         async with self._agent_client_pool.acquire(verified_agent_id) as client:
             await client.update_scaling_group(scaling_group)
 
-    async def settle_agent_alloc(
-        self,
-        kernel_agent_bindings: Sequence[KernelAgentBinding],
-    ) -> None:
-        """
-        Tries to settle down agent row's occupied_slots with real value. This must be called
-        after kernel creation is completed, to prevent fraction of resource dropped by agent scheduler
-        during kernel creation still being reported as used.
-        """
-
-        keyfunc = lambda item: item.agent_alloc_ctx.agent_id
-        for agent_id, group_iterator in itertools.groupby(
-            sorted(kernel_agent_bindings, key=keyfunc),
-            key=keyfunc,
-        ):
-            actual_allocated_slots = ResourceSlot()
-            requested_slots = ResourceSlot()
-
-            for kernel_agent_binding in group_iterator:
-                # this value must be set while running _post_create_kernel
-                actual_allocated_slot = self._kernel_actual_allocated_resources.get(
-                    kernel_agent_binding.kernel.id
-                )
-                requested_slots += kernel_agent_binding.kernel.requested_slots
-                if actual_allocated_slot is not None:
-                    actual_allocated_slots += ResourceSlot.from_json(actual_allocated_slot)
-                    del self._kernel_actual_allocated_resources[kernel_agent_binding.kernel.id]
-                else:  # something's wrong; just fall back to requested slot value
-                    actual_allocated_slots += kernel_agent_binding.kernel.requested_slots
-
-            # Phase 3 (BA-4308): Legacy JSONB write to agents.occupied_slots removed.
-            # Agent occupied slots are now solely managed by the normalized
-            # agent_resources table.  The agents.occupied_slots JSONB column is
-            # retained for historical audit but no longer written to.
-            if actual_allocated_slots != requested_slots:
-                log.debug(
-                    "agent {} has slot calibration diff (requested != actual); "
-                    "agent_resources table is the source of truth",
-                    agent_id,
-                )
-
     async def recalc_resource_usage(self, do_fullscan: bool = False) -> None:
         async def _recalc() -> Mapping[AccessKey, ConcurrencyUsed]:
             access_key_to_concurrency_used: dict[AccessKey, ConcurrencyUsed] = {}