fix: add quality_score to QuantLevel for explicit quality ordering

Refs #139, closes #138

The test assumed bits_per_channel correlates with quality, but q4_0 (4.0 bits)
has lower quality than turbo2 (1.5 bits). Added explicit quality_score field
to make quality ordering unambiguous.

evolution/quant_selector.py

@@ -37,6 +37,7 @@ class QuantLevel:
     layer_adaptive: int     # TURBO_LAYER_ADAPTIVE value (0-7)
     kv_type: str            # -ctk/-ctv flag value
     min_memory_headroom_gb: float  # Minimum free memory to recommend this level
+    quality_score: int = 0  # Explicit quality ordering (higher = better)
     description: str = ""
 
 
@@ -50,6 +51,7 @@ QUANT_LEVELS = [
         layer_adaptive=7,
         kv_type="turbo4",
         min_memory_headroom_gb=4.0,
+        quality_score=100,
         description="PolarQuant + QJL 4-bit. Best quality, ~4.2x KV compression."
     ),
     QuantLevel(
@@ -60,6 +62,7 @@ QUANT_LEVELS = [
         layer_adaptive=5,
         kv_type="turbo3",
         min_memory_headroom_gb=3.0,
+        quality_score=80,
         description="3-bit TurboQuant. High quality, ~6x KV compression."
     ),
     QuantLevel(
@@ -70,6 +73,7 @@ QUANT_LEVELS = [
         layer_adaptive=3,
         kv_type="turbo2",
         min_memory_headroom_gb=2.0,
+        quality_score=60,
         description="2-bit TurboQuant. Balanced, ~10x KV compression."
     ),
     QuantLevel(
@@ -80,6 +84,7 @@ QUANT_LEVELS = [
         layer_adaptive=0,
         kv_type="q4_0",
         min_memory_headroom_gb=1.5,
+        quality_score=40,
         description="Standard 4-bit quant. Fast fallback, no TurboQuant."
     ),
 ]