sites: Inventar und Zielbild an echte Knoten (pve44–pve61) angepasst; ADR-0001/0002/0006 ergänzt

Aufbauplan Task 0.4:
- inventory.yaml: pve44/46/48/52/61 und GODE-PC-05/06 (Ada, Windows, cpu_roles: false)
- proxmox.yaml: eine GPU-VM pro Knoten mit allen Karten (D2), local-lvm/tank (D1 B),
  192.168.100.0/24 ohne VLAN, Gast-IPs .220–.239 (ANNAHME bis Task 0.2)
- registry.yaml/prometheus.yml: ops-Labels auf gpu-pve44/46/48, Ada ohne pve_node/vmid
- tofu: dynamische hostpci, Speicher je Gast; Ansible: pve_gpu_hosts/pve_kiosk,
  Replikation auf mehrere Ziele
- Placement: cpu_roles=false schließt Arbeitsplätze/Kiosk von CPU-Rollen aus
- PLAN §1.3/§1.4/§9 und ADR-Nachträge (D1 entschieden, D2/D3 Bestätigung offen)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

core/tests/test_artifacts.py

@@ -24,7 +24,7 @@ def test_healer_labels_available_on_alerts():
     """Labels, die Healer-Regeln brauchen, müssen von Metrik oder Scrape-Target kommen."""
     prom = yaml.safe_load((DEPLOY / "prometheus/prometheus.yml").read_text())
     jobs = {j["job_name"]: j for j in prom["scrape_configs"]}
-    for job in ("node", "gpu", "vllm"):
+    for job in ("node", "gpu", "llamacpp"):
         for sc in jobs[job]["static_configs"]:
             assert {"slot", "pve_node", "vmid"} <= set(sc["labels"]), (job, sc)
     from knaster.canary import OPS_LABELS
@@ -46,13 +46,17 @@ def test_compose_files_parse(role):
 
 
 def test_ops_labels_consistent_between_registry_and_prometheus():
+    """Das erste Scrape-Ziel eines Slots trägt dieselben ops-Labels wie die Registry (Healer-Ziel)."""
     reg = yaml.safe_load((REPO_ROOT / "sites/knaster/registry.yaml").read_text())
     prom = yaml.safe_load((DEPLOY / "prometheus/prometheus.yml").read_text())
-    vllm = {sc["labels"]["slot"]: sc["labels"] for sc in
-            next(j for j in prom["scrape_configs"] if j["job_name"] == "vllm")["static_configs"]}
-    for slot, labels in vllm.items():
+    first: dict[str, dict] = {}
+    for sc in next(j for j in prom["scrape_configs"] if j["job_name"] == "llamacpp")["static_configs"]:
+        first.setdefault(sc["labels"]["slot"], sc["labels"])
+    assert first
+    for slot, labels in first.items():
         ops = reg["slots"][slot]["ops"]
-        assert ops["vmid"] == labels["vmid"] and ops["container"] == labels["container"]
+        assert {k: ops[k] for k in ("docker_host", "container", "pve_node", "vmid")} == \
+            {k: labels[k] for k in ("docker_host", "container", "pve_node", "vmid")}, slot
 
 
 WORKFLOWS = sorted((REPO_ROOT / "n8n/workflows").glob("*.json"))
@@ -160,19 +164,27 @@ def test_proxmox_site_consistent_with_registry_and_prometheus():
     reg = yaml.safe_load((REPO_ROOT / "sites/knaster/registry.yaml").read_text())
     vms = site["vms"]
     for vm in vms.values():
-        assert vm["gpu"] in site["pci_mappings"]
-        assert site["pci_mappings"][vm["gpu"]]["node"] == vm["node"]  # Mapping liegt auf dem Host der VM
+        for gpu in vm["gpus"]:
+            assert gpu in site["pci_mappings"]
+            assert site["pci_mappings"][gpu]["node"] == vm["node"]  # Mapping liegt auf dem Host der VM
         assert vm["driver_generation"] in {"legacy", "modern"}
-    gpus = [vm["gpu"] for vm in vms.values()]
+        assert vm["node"] in site["nodes"] and site["nodes"][vm["node"]]["gpu_host"]
+    gpus = [g for vm in vms.values() for g in vm["gpus"]]
     assert len(gpus) == len(set(gpus))  # keine GPU doppelt durchgereicht
+    assert set(gpus) == set(site["pci_mappings"])  # jede gemappte Karte steckt in einer VM
     by_vmid = {str(vm["vmid"]): (name, vm) for name, vm in {**vms, **site["containers"]}.items()}
     for slot, spec in reg["slots"].items():
-        ops = spec.get("ops")
-        if not ops:
+        ops = spec.get("ops") or {}
+        if "vmid" not in ops:  # z. B. Ada-Worker unter Windows (D3), kein Proxmox-Gast
+            assert "pve_node" not in ops, slot
             continue
         name, vm = by_vmid[ops["vmid"]]
         assert vm["node"] == ops["pve_node"], slot
-        assert ops["docker_host"].startswith(name.split("-")[0]), slot
+        assert ops["docker_host"] == name, slot
+    for c in site["containers"].values():
+        if c.get("ha"):
+            assert site["nodes"][c["node"]]["zfs_tank"] and c["storage"] == "tank"
+            assert all(site["nodes"][t]["zfs_tank"] for t in c["replicate_to"])
     vmids = [str(v["vmid"]) for v in vms.values()] + [str(c["vmid"]) for c in site["containers"].values()]
     assert len(vmids) == len(set(vmids))
     # vLLM nur auf modernen Karten

core/tests/test_inventory_validate.py

@@ -28,29 +28,70 @@ def test_parsers():
     assert disks[0].size_gb == 256.1 and not disks[0].rotational
 
 
+PVE_GPU_HOSTS = {"pve44", "pve46", "pve48", "pve52"}
+
+
+def test_site_knaster_matches_real_nodes():
+    """Aufbauplan Task 0.4: Site-Dateien beschreiben die echten Knoten, nicht mehr die Annahmen pve-ada/pve1–4."""
+    import ipaddress
+
+    from knaster.util import load_yaml
+
+    nodes = {n.name: n for n in load_nodes(REPO_ROOT / "sites/knaster/inventory.yaml")}
+    assert set(nodes) == {"pve44", "pve46", "pve48", "pve52", "pve61", "gode-pc-05", "gode-pc-06"}
+    pve = load_yaml(REPO_ROOT / "sites/knaster/proxmox.yaml")
+    old_net = ipaddress.ip_network("10.20.0.0/24")
+    ips = []
+    for section in ("vms", "containers"):
+        ips += [ipaddress.ip_interface(g["ip"]).ip for g in pve[section].values()]
+    ips += [ipaddress.ip_address(pve["cluster"][k]) for k in ("gateway", "dns")]
+    assert not [ip for ip in ips if ip in old_net]
+    assert {m["node"] for m in pve["pci_mappings"].values()} <= PVE_GPU_HOSTS
+    assert {v["node"] for v in pve["vms"].values()} <= PVE_GPU_HOSTS
+    # eine GPU-VM pro GPU-Knoten mit allen Karten des Knotens (D2)
+    assert sorted(v["node"] for v in pve["vms"].values()) == sorted(PVE_GPU_HOSTS)
+    for vm in pve["vms"].values():
+        mapped = {name for name, m in pve["pci_mappings"].items() if m["node"] == vm["node"]}
+        assert set(vm["gpus"]) == mapped
+        gpus_inv = [g for g in nodes[vm["node"]].gpus]
+        assert len(vm["gpus"]) == len(gpus_inv)
+    # Ada-Rechner bleiben Windows-Arbeitsplätze (D3): keine Infra-/CPU-Rollen
+    assert not nodes["gode-pc-05"].cpu_roles and not nodes["pve61"].cpu_roles
+
+
 def test_placement_knaster():
     roles, models = load_catalog(REPO_ROOT / "catalog")
     p = plan(load_nodes(REPO_ROOT / "sites/knaster/inventory.yaml"), roles, models)
     by_role = {}
     for a in p.assignments:
         by_role.setdefault(a.role, []).append(a)
+    assert {by_role["gpu-main"][0].node, by_role["gpu-flex"][0].node} == {"gode-pc-05", "gode-pc-06"}
     assert by_role["gpu-main"][0].gpu_name.startswith("NVIDIA RTX 4000 Ada")
     assert "qwen3-14b-awq" in by_role["gpu-main"][0].models_fitting
     embed_nodes = {a.node for a in by_role["gpu-embed"]}
     assert len(embed_nodes) == 2  # Anti-Affinität
     assert all("P2000" in a.gpu_name for a in by_role["gpu-embed"] + by_role["gpu-small"])
-    assert by_role["infra"][0].mode == "cpu"
+    assert by_role["infra"][0].mode == "cpu" and by_role["infra"][0].node in PVE_GPU_HOSTS
     assert not p.unplaced
     # M4000 bewusst außer Betrieb (ADR-0006): weder frei noch belegt
     assert p.free_gpus == []
     assert len(p.disabled_gpus) == 4 and all("M4000" in g for g in p.disabled_gpus)
     tts = {a.role: a for a in p.assignments if a.role.startswith("tts")}
-    assert tts["tts-hq"].mode == "shared" and tts["tts-hq"].node == "pve-ada"
+    assert tts["tts-hq"].mode == "shared" and tts["tts-hq"].node == by_role["gpu-flex"][0].node
     assert "qwen3-tts-1.7b" in tts["tts-hq"].models_fitting
     assert tts["tts-cpu"].node == by_role["infra"][0].node
     assert tts["tts-cpu"].models_fitting == ["piper-de-thorsten-medium"]
 
 
+def test_nodes_without_cpu_roles_get_no_infra():
+    """Arbeitsplätze und Kiosk nehmen keine CPU-Rollen, auch wenn sie den meisten RAM haben."""
+    roles, models = load_catalog(REPO_ROOT / "catalog")
+    nodes = [NodeProfile(name="workstation", memory_gb=64, cpu_roles=False),
+             NodeProfile(name="server", memory_gb=32)]
+    p = plan(nodes, roles, models)
+    assert {a.node for a in p.assignments if a.mode in ("cpu", "cpu_fallback")} == {"server"}
+
+
 def test_placement_other_cluster():
     roles, models = load_catalog(REPO_ROOT / "catalog")
     p = plan(load_nodes(REPO_ROOT / "sites/example-single-gpu/inventory.yaml"), roles, models)