fix(backend): unify coalesce skip condition at both sites

Inner spill loop clear now checks vregs>120 (matching outer site),
ensuring coalescing is consistently skipped for large functions.

src/mir/RegAlloc.cpp

@@ -1380,8 +1380,6 @@ namespace mir
         }
 
         // 大函数丢弃 move 偏好以保持分配稳定性
-        // 条件1: move偏好数量 >100 直接跳过
-        // 条件2: vreg数 * move偏好数 > 600 (conv2d: 71*15=1065)
         int mv = static_cast<int>(move_preferences.size());
         if (mv > 100 || static_cast<int>(function.GetNumVRegs()) * mv > 600)
           move_preferences.clear();
@@ -1704,7 +1702,8 @@ namespace mir
       }
 
       int mv = static_cast<int>(move_preferences.size());
-      if (mv > 100 || static_cast<int>(function.GetNumVRegs()) * mv > 600)
+      if (function.GetNumVRegs() > 120 || mv > 100 ||
+          static_cast<int>(function.GetNumVRegs()) * mv > 600)
         move_preferences.clear();
 
       std::vector<int> gp_alloc(GP_ALLOCATABLE, GP_ALLOCATABLE + GP_NUM_ALLOCATABLE);