feat(v9): Phase 10 Schritt 1 — Modell-Capabilities aus GGUF-Metadaten

- model_caps.py: dependency-freier GGUF-Header-Reader (architecture,
  expert_count, context_length, parameter_count) — offline/authoritativ,
  bricht vor dem Tokenizer-Array ab. Leitet Tag-Set ab: MoE(+aktive B),
  Tools (yes via --jinja/Template | likely via Familie | no), Vision,
  Coder, Reasoning, Embedding, native ctx/params. Familien-Liste nur Fallback.
- models.py status(): meta.capabilities ergänzt.
- ModelsPanel: "KANN"-Spalte zeigt jetzt Capability-Chips statt Text/Code/Bild.

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

routers/models.py

@@ -20,6 +20,7 @@ from routers.cookbook import hf_file_size
 from llamaswap import (_swap_get, read_config, write_config,
                        model_id_from_path, set_role_alias, ROLE_IDS)
 from hw_math import extract_params_b, max_ctx_for, estimate_memory_gb
+from model_caps import capabilities
 import re
 import os
 import shutil
@@ -90,6 +91,7 @@ def status():
         size_bytes = None
         quant = ""
         filename = ""
+        path = ""
         m_path = re.search(r'-(?:m|-model)\s+([^\s]+)', cmd)
         if m_path:
             path = m_path.group(1).replace("'", "").replace('"', '')
@@ -137,6 +139,10 @@ def status():
                 "optimal_ctx": (_oc := (max_ctx_for(_pb, quant or "Q4_K_M", ram_gb) if ram_gb else None)),
                 "peak_ram_gb": round(estimate_memory_gb(_pb, quant or "Q4_K_M", ctx), 1),
                 "peak_ram_optimal_gb": (round(estimate_memory_gb(_pb, quant or "Q4_K_M", _oc), 1) if _oc else None),
+                "capabilities": capabilities(
+                    name=filename or name, cmd=cmd,
+                    gguf_path=(path if (path and os.path.exists(path)) else ""),
+                ),
             }
         }
     # Laufende Download-Jobs erkennnen: Modell bekommt state "downloading" + Fortschritt.