feat(profiles): add idle_stop_after_s field; set 600s for vllm slot

Add idle_stop_after_s to ServiceProfile (default 0 = never stop).
Set 600s (10 min) timeout on vllm slot in all single-GPU profiles.
Backward compatible; non-vllm services inherit default 0 (no auto-stop).

circuitforge_core/resources/profiles/public/single-gpu-16gb.yaml

@@ -6,6 +6,7 @@ services:
   vllm:
     max_mb: 12288
     priority: 1
+    idle_stop_after_s: 600
   ollama:
     max_mb: 12288
     priority: 1
@@ -14,6 +15,11 @@ services:
     priority: 2
     shared: true
     max_concurrent: 4
+  cf-docuvision:
+    max_mb: 6144
+    priority: 2
+    shared: true
+    max_concurrent: 3
   cf-stt:
     max_mb: 1200
     priority: 2

circuitforge_core/resources/profiles/public/single-gpu-24gb.yaml

@@ -6,6 +6,7 @@ services:
   vllm:
     max_mb: 20480
     priority: 1
+    idle_stop_after_s: 600
   ollama:
     max_mb: 18432
     priority: 1
@@ -14,6 +15,11 @@ services:
     priority: 2
     shared: true
     max_concurrent: 6
+  cf-docuvision:
+    max_mb: 8192
+    priority: 2
+    shared: true
+    max_concurrent: 4
   cf-stt:
     max_mb: 1200
     priority: 2

circuitforge_core/resources/profiles/public/single-gpu-6gb.yaml

@@ -6,6 +6,17 @@ services:
   vllm:
     max_mb: 4096
     priority: 1
+    idle_stop_after_s: 600
+    managed:
+      type: docker
+      image: "vllm/vllm-openai:v0.9.2"
+      port: 8000
+      host_port: 8000
+      command_template: "--model /models/{model} --trust-remote-code --max-model-len {max_model_len} --gpu-memory-utilization {gpu_mem_util} --enforce-eager --max-num-seqs 8"
+      volumes:
+        - "${VLLM_MODELS_DIR:-/Library/Assets/LLM/vllm/models}:/models"
+      runtime: nvidia
+      ipc: host
   ollama:
     max_mb: 3584
     priority: 1
@@ -14,6 +25,11 @@ services:
     priority: 2
     shared: true
     max_concurrent: 2
+  cf-docuvision:
+    max_mb: 3072
+    priority: 2
+    shared: true
+    max_concurrent: 1
   cf-stt:
     max_mb: 600
     priority: 2

circuitforge_core/resources/profiles/public/single-gpu-8gb.yaml

@@ -6,6 +6,17 @@ services:
   vllm:
     max_mb: 5120
     priority: 1
+    idle_stop_after_s: 600
+    managed:
+      type: docker
+      image: "vllm/vllm-openai:v0.9.2"
+      port: 8000
+      host_port: 8000
+      command_template: "--model /models/{model} --trust-remote-code --max-model-len {max_model_len} --gpu-memory-utilization {gpu_mem_util} --enforce-eager --max-num-seqs 8"
+      volumes:
+        - "${VLLM_MODELS_DIR:-/Library/Assets/LLM/vllm/models}:/models"
+      runtime: nvidia
+      ipc: host
   ollama:
     max_mb: 4096
     priority: 1
@@ -14,6 +25,11 @@ services:
     priority: 2
     shared: true
     max_concurrent: 3
+  cf-docuvision:
+    max_mb: 4096
+    priority: 2
+    shared: true
+    max_concurrent: 2
   cf-stt:
     max_mb: 1200
     priority: 2
@@ -28,6 +44,13 @@ services:
   comfyui:
     max_mb: 6144
     priority: 4
+    managed:
+      type: process
+      exec_path: "/opt/miniconda3/envs/comfyui/bin/python"
+      args_template: "/opt/ComfyUI/main.py --listen 0.0.0.0 --port {port} --cuda-device {gpu_id}"
+      cwd: "/opt/ComfyUI"
+      port: 8188
+      host_port: 8188
 model_size_hints:
   llm_max_params: 8b
   image_gen_max: sdxl-fp8

circuitforge_core/resources/profiles/schema.py

@@ -5,22 +5,71 @@ from pathlib import Path
 from typing import Any
 
 import yaml
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, model_validator
 
 SUPPORTED_SCHEMA_VERSION = 1
 
 
+class DockerSpec(BaseModel):
+    """Spec for a Docker-managed service."""
+
+    image: str
+    port: int
+    host_port: int
+    command_template: str = ""
+    volumes: list[str] = Field(default_factory=list)
+    env: dict[str, str] = Field(default_factory=dict)
+    runtime: str = "nvidia"
+    ipc: str = "host"
+
+    model_config = {"frozen": True}
+
+
+class ProcessSpec(BaseModel):
+    """Spec for a process-managed service (non-Docker, e.g. conda env)."""
+
+    exec_path: str
+    args_template: str = ""
+    cwd: str = ""
+    env: dict[str, str] = Field(default_factory=dict)
+    port: int = 0
+    host_port: int = 0
+
+    model_config = {"frozen": True}
+
+
 class ServiceProfile(BaseModel):
     max_mb: int
     priority: int
     shared: bool = False
     max_concurrent: int = 1
     always_on: bool = False
+    idle_stop_after_s: int = 0
     backend: str | None = None
     consumers: list[str] = Field(default_factory=list)
+    managed: DockerSpec | ProcessSpec | None = None
 
     model_config = {"frozen": True}
 
+    @model_validator(mode="before")
+    @classmethod
+    def _parse_managed(cls, values: Any) -> Any:
+        if not isinstance(values, dict):
+            return values
+        raw = values.get("managed")
+        if raw is None:
+            return values
+        if not isinstance(raw, dict):
+            return values
+        spec_type = raw.pop("type", None)
+        if spec_type == "docker":
+            values["managed"] = DockerSpec(**raw)
+        elif spec_type == "process":
+            values["managed"] = ProcessSpec(**raw)
+        else:
+            raise ValueError(f"Unknown managed service type: {spec_type!r}")
+        return values
+
 
 class GpuNodeEntry(BaseModel):
     id: int

tests/test_resources/test_coordinator_app.py

@@ -1,11 +1,13 @@
 import pytest
 from unittest.mock import MagicMock
+from pathlib import Path
 from fastapi.testclient import TestClient
 from circuitforge_core.resources.coordinator.app import create_coordinator_app
 from circuitforge_core.resources.coordinator.agent_supervisor import AgentSupervisor
 from circuitforge_core.resources.coordinator.lease_manager import LeaseManager
 from circuitforge_core.resources.coordinator.profile_registry import ProfileRegistry
 from circuitforge_core.resources.models import GpuInfo, NodeInfo
+from circuitforge_core.resources.profiles.schema import load_profile
 
 
 @pytest.fixture
@@ -132,3 +134,13 @@ def test_resident_keys_returns_set_of_node_service():
     lm.set_residents_for_node("heimdall", [("vllm", "Ouro-1.4B"), ("ollama", None)])
     keys = lm.resident_keys()
     assert keys == {"heimdall:vllm", "heimdall:ollama"}
+
+
+def test_single_gpu_8gb_profile_has_idle_stop_after_s():
+    profile = load_profile(
+        Path("circuitforge_core/resources/profiles/public/single-gpu-8gb.yaml")
+    )
+    vllm_svc = profile.services.get("vllm")
+    assert vllm_svc is not None
+    assert hasattr(vllm_svc, "idle_stop_after_s")
+    assert vllm_svc.idle_stop_after_s == 600