fix(hwfit): filter non-GGUF models on Windows (#2530)

Odysseus only supports llama.cpp on Windows (vLLM/SGLang are
explicitly blocked). llama.cpp requires GGUF, so AWQ/GPTQ/FP8
safetensors models without a GGUF alternate should not be
recommended in the Cookbook on Windows hosts.

Changes:
- hardware.py: add 'platform': 'windows' to _detect_windows()
  so downstream logic can identify Windows hosts.
- fit.py: include is_windows in the existing GGUF-only filter
  alongside apple_silicon and consumer_amd.
- tests: add test_hwfit_windows.py with regression tests.

Fixes #122, #614 (root cause: unservable models recommended).

services/hwfit/fit.py

@@ -576,6 +576,7 @@ def rank_models(system, use_case=None, limit=50, search=None, sort="score", quan
     system_backend = (system.get("backend") or "").lower()
     apple_silicon = system_backend in ("mps", "metal", "apple")
     rocm = system_backend == "rocm"
+    is_windows = system.get("platform") == "windows"
 
     # Consumer AMD Radeon (RDNA, gfx10/11/12): the practical local serving path
     # is GGUF via llama.cpp. vLLM/SGLang on ROCm are validated for datacenter
@@ -615,7 +616,11 @@ def rank_models(system, use_case=None, limit=50, search=None, sort="score", quan
         # servable path, so a model needs a real GGUF to be recommended.
         # Otherwise the Cookbook rates vLLM-only AWQ/GPTQ builds "GOOD" on a
         # Radeon that can't actually serve them.
-        if (apple_silicon or consumer_amd) and not (m.get("is_gguf") or m.get("gguf_sources")):
+        #
+        # Windows is the same: Odysseus only supports llama.cpp on Windows,
+        # which requires GGUF. vLLM/SGLang are explicitly blocked, so AWQ/GPTQ
+        # models without a GGUF source are unservable there.
+        if (apple_silicon or consumer_amd or is_windows) and not (m.get("is_gguf") or m.get("gguf_sources")):
             continue
 
         # Format filter: AWQ tab -> only AWQ models, FP4 tab -> FP4-family models, etc.

services/hwfit/hardware.py

@@ -539,6 +539,7 @@ def _detect_windows():
             "backend": d.get("gpu_backend", "cpu_x86"),
             "homogeneous": True,
             "gpu_error": None,
+            "platform": "windows",
         }
         # PowerShell only reports aggregate GPU info, not per-card detail, so we
         # can't tell a mixed box from a uniform one here — assume one homogeneous

tests/test_hwfit_windows.py

@@ -0,0 +1,74 @@
+"""Windows support for Cookbook hardware-fit.
+
+Odysseus only supports llama.cpp on Windows (vLLM/SGLang are explicitly
+blocked). llama.cpp requires GGUF, so non-GGUF models — including AWQ/GPTQ/
+FP8 safetensors repos — must be filtered out on Windows so the Cookbook does
+not recommend models the user cannot actually serve.
+"""
+
+from services.hwfit.fit import rank_models
+from services.hwfit.models import get_models
+
+
+def _windows_system(ram_gb=32.0, vram_gb=16.0):
+    return {
+        "has_gpu": True,
+        "backend": "cuda",
+        "gpu_name": "NVIDIA RTX 4060",
+        "gpu_vram_gb": vram_gb,
+        "gpu_count": 1,
+        "available_ram_gb": ram_gb * 0.7,
+        "total_ram_gb": ram_gb,
+        "platform": "windows",
+    }
+
+
+def _cuda_system():
+    return {
+        "has_gpu": True,
+        "backend": "cuda",
+        "gpu_name": "NVIDIA RTX 4090",
+        "gpu_vram_gb": 24.0,
+        "gpu_count": 1,
+        "available_ram_gb": 32.0,
+        "total_ram_gb": 64.0,
+    }
+
+
+def test_only_gguf_models_recommended_on_windows():
+    """llama.cpp (GGUF) is the only servable path on Windows, so every model
+    recommended there must ship a real GGUF — no vLLM-only AWQ/GPTQ/FP8."""
+    catalog = {m["name"]: m for m in get_models()}
+    unservable = [
+        r["name"] for r in rank_models(_windows_system(), limit=900)
+        if not (catalog.get(r["name"], {}).get("is_gguf")
+                or catalog.get(r["name"], {}).get("gguf_sources"))
+    ]
+    assert unservable == [], f"{len(unservable)} non-GGUF models on Windows, e.g. {unservable[:3]}"
+
+
+def test_safetensors_models_still_recommended_on_cuda():
+    """Regression guard: the GGUF-only rule must not leak onto CUDA."""
+    names = {r["name"] for r in rank_models(_cuda_system(), limit=900)}
+    assert "microsoft/Phi-mini-MoE-instruct" in names
+
+
+def test_awq_model_hidden_on_windows():
+    """The user's reported issue: Qwen2.5-3B-Instruct-AWQ is AWQ-only and must
+    not be recommended on Windows where it cannot be served."""
+    names = {r["name"] for r in rank_models(_windows_system(), limit=900)}
+    assert "Qwen/Qwen2.5-3B-Instruct-AWQ" not in names
+
+
+def test_awq_model_visible_on_cuda():
+    """The same AWQ model should still be visible on CUDA where vLLM can
+    serve it."""
+    names = {r["name"] for r in rank_models(_cuda_system(), limit=900)}
+    assert "Qwen/Qwen2.5-3B-Instruct-AWQ" in names
+
+
+def test_gguf_alternate_still_recommended_on_windows():
+    """Qwen2.5-3B-Instruct (the base model) has a GGUF source, so it should
+    still appear on Windows even though the AWQ variant is hidden."""
+    names = {r["name"] for r in rank_models(_windows_system(), limit=900)}
+    assert "Qwen/Qwen2.5-3B-Instruct" in names