Fix: "Modelle finden" bevorzugt faehigste passende Modelle (kein Downgrade) + Rollen-Farben zentral

- discover.rank_runnable: rankt jetzt fit-level > params_b(desc) > downloads -> fuer die
  128GB-Box wird das faehigste passende (MoE-)Modell empfohlen statt kleiner Populaer-Modelle.
  Top-4-Anzeige nutzt dasselbe Ranking.
- maintenance.model_upgrades: schlaegt KEIN Downgrade mehr vor (rec.params_b >= installiert*0.95).
  Behebt: fast 35B-A3B -> 4B wurde faelschlich als Upgrade angeboten.
- Frontend: Rollen-Farb-Mapping zentralisiert in ModelBadges (ROLE_TONE/roleTone);
  Cockpit/RolesCard/ActiveModelsCard nutzen es statt eigener Duplikate.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/services/discover.py

@@ -47,12 +47,18 @@ def _fetch_author_models(author: str) -> list:
 
 
 def rank_runnable(models: list[dict]) -> list[dict]:
-    """EINE Quelle der Wahrheit fürs Ranking lauffähiger Modelle:
+    """EINE Quelle der Wahrheit fürs Ranking lauffähiger Modelle für DIESE Hardware:
-    bestes Fit-Level zuerst (perfect < marginal), bei Gleichstand meistgeladen.
+    1) nur was komfortabel passt (perfect vor marginal, too_tight fliegt raus),
-    Zu große Modelle (too_tight) fliegen raus."""
+    2) das FÄHIGSTE zuerst — mehr Parameter = mehr Können (bei MoE bleibt es dank
+       aktiver-Param-Schätzung schnell),
+    3) bei Gleichstand das meistgeladene.
+    So bevorzugt die 128-GB-Box große (MoE-)Modelle statt kleiner Populär-Modelle —
+    und „Modelle finden" schlägt nie ein Downgrade vor (z.B. 35B-A3B → 4B)."""
     return sorted(
         [m for m in models if m["fit"]["level"] != "too_tight"],
-        key=lambda m: (_FIT_ORDER[m["fit"]["level"]], -int(m.get("downloads") or 0)),
+        key=lambda m: (_FIT_ORDER[m["fit"]["level"]],
+                       -float(m.get("params_b") or 0.0),
+                       -int(m.get("downloads") or 0)),
     )
 
 
@@ -97,9 +103,8 @@ def refresh_discover(ram_gb: float) -> dict:
     for c in CATEGORIES:
         items = by_cat[c["role"]]
         ranked = rank_runnable(items)
-        # Top 4 je Kategorie (für die Anzeige) — gerankt, dann nach Downloads aufgefüllt.
+        # Top 4 je Kategorie = die best-gerankten (fähigstes-was-passt zuerst).
-        items.sort(key=lambda x: (x["fit"]["level"] != "too_tight", x["downloads"]), reverse=True)
+        top = ranked[:4]
-        top = items[:4]
         if top:
             cats.append({
                 "role": c["role"], "title": c["title"], "icon": c["icon"],

backend/services/maintenance.py

@@ -164,6 +164,12 @@ def model_upgrades() -> list[dict]:
 
     installed = llamaswap.list_models()
     active_roles = {m["role"] for m in installed if m.get("role")}
+    # Größtes installiertes Modell je Rolle (params_b) — Basis für „kein Downgrade".
+    inst_params: dict[str, float] = {}
+    for m in installed:
+        r, pb = m.get("role"), (m.get("capabilities") or {}).get("params_b")
+        if r and pb:
+            inst_params[r] = max(inst_params.get(r, 0.0), float(pb))
 
     cmds = " ".join(str(s.get("cmd", "")).lower()
                     for s in (llamaswap.read_config().get("models") or {}).values())
@@ -178,6 +184,14 @@ def model_upgrades() -> list[dict]:
         rec = c.get("recommended")
         if not rec:
             continue
+        # KEIN Downgrade: Empfehlung nur, wenn sie mind. so „groß"/fähig ist wie das
+        # installierte Modell der Rolle (verhindert z.B. fast 35B-A3B → 4B).
+        rec_model = next((x for x in c.get("models", []) if x.get("repo") == rec), None)
+        rec_params = float((rec_model or {}).get("params_b") or 0.0)
+        cur = inst_params.get(role, 0.0)
+        if cur and rec_params and rec_params < cur * 0.95:
+            continue
+
         base = rec.split("/")[-1].lower()
         stem = base[:-5] if base.endswith("-gguf") else base
         if base in cmds or (stem and stem in cmds):

frontend/dist/index.html

@@ -7,8 +7,8 @@
     <link rel="manifest" href="/manifest.webmanifest" />
     <link rel="icon" type="image/svg+xml" href="/favicon.svg" />
     <title>Mission Control 2.0</title>
-    <script type="module" crossorigin src="/assets/index-CrPS-3Fn.js"></script>
+    <script type="module" crossorigin src="/assets/index-oijhNb41.js"></script>
-    <link rel="stylesheet" crossorigin href="/assets/index-CjZVMvhL.css">
+    <link rel="stylesheet" crossorigin href="/assets/index-Cx3ZKkHL.css">
   </head>
   <body>
     <div id="root"></div>

frontend/src/components/dashboard/ActiveModelsCard.tsx

@@ -3,15 +3,7 @@ import { Activity, Zap, Moon } from "lucide-react"
 import { useModels, useTokenStats } from "@/lib/queries"
 import { fmtSize } from "@/lib/format"
 import { cn } from "@/lib/utils"
-
+import { roleTone } from "@/components/models/ModelBadges"
-const ROLE_TONE: Record<string, string> = {
-  fast: "bg-cyan-500/15 text-cyan-400 border-cyan-500/25",
-  heavy: "bg-amber-500/15 text-amber-400 border-amber-500/25",
-  coder: "bg-violet-500/15 text-violet-400 border-violet-500/25",
-  vision: "bg-pink-500/15 text-pink-400 border-pink-500/25",
-  scout: "bg-teal-500/15 text-teal-400 border-teal-500/25",
-  hermes: "bg-indigo-500/15 text-indigo-400 border-indigo-500/25",
-}
 
 export function ActiveModelsCard() {
   const { data } = useModels(2_000)
@@ -65,7 +57,7 @@ export function ActiveModelsCard() {
                 <div className="flex items-center gap-1.5">
                   {m.role && (
                     <span className={cn("text-[8px] font-semibold uppercase px-1.5 py-0.5 rounded font-mono border tracking-wider shrink-0",
-                      ROLE_TONE[m.role] || "bg-slate-500/15 text-slate-300 border-slate-500/25")}>
+                      roleTone(m.role))}>
                       {m.role}
                     </span>
                   )}

frontend/src/components/dashboard/RolesCard.tsx

@@ -1,8 +1,7 @@
 import { Layers } from "lucide-react"
 import { useModels } from "@/lib/queries"
 import { cn } from "@/lib/utils"
-
+import { ROLES, roleTone } from "@/components/models/ModelBadges"
-const ROLES = ["fast", "heavy", "coder", "vision", "scout"]
 
 export function RolesCard() {
   const { data } = useModels(3_000)
@@ -37,12 +36,7 @@ export function RolesCard() {
                 <div className="min-w-0 flex-1 mr-2">
                   <div className="flex items-center gap-2 min-w-0">
                     <span className={cn("text-[9px] font-semibold uppercase px-1.5 py-0.5 rounded font-mono border tracking-wider shrink-0",
-                      role === "fast" ? "bg-cyan-500/15 text-cyan-400 border-cyan-500/25" :
+                      roleTone(role))}>
-                      role === "heavy" ? "bg-amber-500/15 text-amber-400 border-amber-500/25" :
-                      role === "coder" ? "bg-violet-500/15 text-violet-400 border-violet-500/25" :
-                      role === "vision" ? "bg-pink-500/15 text-pink-400 border-pink-500/25" :
-                      "bg-teal-500/15 text-teal-400 border-teal-500/25"
-                    )}>
                       {role}
                     </span>
 

frontend/src/components/models/ModelBadges.tsx

@@ -4,6 +4,20 @@ import { cn } from "@/lib/utils"
 // Die 5 kanonischen Serving-Rollen (identisch zu Backend sources.py/llamaswap.ROLE_IDS).
 export const ROLES = ["fast", "heavy", "coder", "vision", "scout"]
 
+// Rolle → Farbton. EINE Quelle der Wahrheit für alle Rollen-Badges (Cockpit, RolesCard,
+// ActiveModelsCard, …). `hermes` = Agent-Hirn (taucht als geladenes Modell auf).
+export const ROLE_TONE: Record<string, string> = {
+  fast: "bg-cyan-500/15 text-cyan-400 border-cyan-500/25",
+  heavy: "bg-amber-500/15 text-amber-400 border-amber-500/25",
+  coder: "bg-violet-500/15 text-violet-400 border-violet-500/25",
+  vision: "bg-pink-500/15 text-pink-400 border-pink-500/25",
+  scout: "bg-teal-500/15 text-teal-400 border-teal-500/25",
+  hermes: "bg-indigo-500/15 text-indigo-400 border-indigo-500/25",
+}
+
+export const roleTone = (role?: string | null) =>
+  (role && ROLE_TONE[role]) || "bg-slate-500/15 text-slate-300 border-slate-500/25"
+
 export function FitBadge({ fit }: { fit: Fit }) {
   const tone = {
     perfect: "bg-emerald-500/15 text-emerald-400 border border-emerald-500/20",

frontend/src/views/models/Cockpit.tsx

@@ -6,7 +6,7 @@ import { useDialog } from "@/lib/useDialog"
 import { CapsChips } from "@/components/CapsChips"
 import { cn } from "@/lib/utils"
 import { fmtSize, fmtCtx } from "@/lib/format"
-import { getBrandInfo, ROLES } from "@/components/models/ModelBadges"
+import { getBrandInfo, ROLES, roleTone } from "@/components/models/ModelBadges"
 import { SpecDraftModal } from "@/components/models/SpecDraftModal"
 
 export function Cockpit() {
@@ -611,12 +611,7 @@ export function Cockpit() {
               >
                 <div className="flex items-center justify-between">
                   <span className={cn("text-[9px] font-bold uppercase tracking-wider font-mono px-1.5 py-0.5 rounded border",
-                    role === "fast" ? "bg-cyan-500/10 text-cyan-400 border-cyan-500/20" :
+                    roleTone(role))}>{role}</span>
-                    role === "heavy" ? "bg-amber-500/10 text-amber-400 border-amber-500/20" :
-                    role === "coder" ? "bg-violet-500/10 text-violet-400 border-violet-500/20" :
-                    role === "vision" ? "bg-pink-500/10 text-pink-400 border-pink-500/20" :
-                    "bg-teal-500/10 text-teal-400 border-teal-500/20"
-                  )}>{role}</span>
                   {isWarm && <span className="h-1.5 w-1.5 rounded-full bg-emerald-400 animate-pulse" />}
                 </div>