enh: 模型排序加入上下文长度权重(15%)

新增 CONTEXT_LENGTHS 字典(20模型)
新增 _context_score() 评分函数
排名公式: probe(30%) ctx(15%) param(20%) family(15%) stab(10%) speed(10%)
输出: context_k + context_score 字段

scripts/model-health.py

@@ -37,6 +37,28 @@ CANDIDATE_POOL = [
     "nvidia/nemotron-mini-4b-instruct",  # 4B 兜底
 ]
 
+# 已知上下文长度（K=tokens）
+CONTEXT_LENGTHS = {
+    "openai/gpt-oss-120b": 128,
+    "nvidia/nemotron-3-super-120b-a12b": 128,
+    "mistralai/mistral-nemotron": 128,
+    "nvidia/nvidia-nemotron-nano-9b-v2": 128,
+    "meta/llama-3.1-8b-instruct": 128,
+    "nvidia/nemotron-mini-4b-instruct": 128,
+    "nvidia/llama-3.3-nemotron-super-49b-v1": 128,
+    "nvidia/llama-3.3-nemotron-super-49b-v1.5": 128,
+    "mistralai/mistral-medium-3.5-128b": 128,
+    "minimaxai/minimax-m2.7": 256,
+    "minimaxai/minimax-m3": 256,
+    "stepfun-ai/step-3.5-flash": 8,
+    "qwen/qwen3.5-122b-a10b": 128,
+    "qwen/qwen3-next-80b-a3b-thinking": 128,
+    "moonshotai/kimi-k2-instruct": 128,
+    "mistralai/devstral-2-123b-instruct-2512": 128,
+    "deepseek-v4-flash": 256,
+    "deepseek-ai/deepseek-v4-pro": 256,
+}
+
 # 已知忽略的模型（系统/不支持/垃圾，永远不测也不自动加入）
 KNOWN_IGNORE = {
     "gpt-4o", "gpt-4o-mini", "gpt-4o-audio-preview", "gpt-4o-mini-audio-preview",
@@ -165,6 +187,20 @@ def _speed_score(latency_ms: int, fastest_latency: int) -> float:
     return min(round(ratio * 100), 100)
 
 
+def _context_score(model: str) -> float:
+    """上下文长度分：越长越高 256K→100, 128K→80, 64K→60, 32K→40, 8K→10"""
+    ctx = CONTEXT_LENGTHS.get(model, 128)  # 未知默认128
+    if ctx >= 256:
+        return 100
+    if ctx >= 128:
+        return 80
+    if ctx >= 64:
+        return 60
+    if ctx >= 32:
+        return 40
+    return max(round(ctx / 8 * 10), 5)
+
+
 def _run_quality_probe(model: str) -> dict:
     """运行质量探针，返回探针分和详细结果"""
     probe_results = []
@@ -332,9 +368,10 @@ def test_model(model: str, fastest_latency: int = None) -> dict:
     ss = _speed_score(avg_latency, fastest_latency) if fastest_latency and avg_latency > 0 else 50
     stab_s = 100 if stability == "stable" else (50 if stability == "unstable" else 0)
     probe_s = probe["probe_score"]
+    cs = _context_score(model)
 
     rank_score = round(
-        probe_s * 0.35 + ps * 0.25 + fs * 0.20 + stab_s * 0.10 + ss * 0.10
+        probe_s * 0.30 + cs * 0.15 + ps * 0.20 + fs * 0.15 + stab_s * 0.10 + ss * 0.10
     )
 
     return {
@@ -351,6 +388,8 @@ def test_model(model: str, fastest_latency: int = None) -> dict:
         "probe_detail": probe["probe_detail"],
         "rank_score": rank_score,
         "param_b": param_b,
+        "context_k": CONTEXT_LENGTHS.get(model, 128),
+        "context_score": cs,
         "family_score": fs,
         "param_score": ps,
     }