fix: eps=1.0 (19 clusters from 1486 items), add col_vector() for FixedSizeListArray reading

- eps: 0.1 → 1.0 (p50=1.011, meaningful semantic clustering)
- col_vector(): parse FixedSizeListArray and Float32Array for LanceDB vector column
- search() and scan_all() now read actual vectors instead of vec![]
- client.go: remove duplicate Epsilon field, eps=1.0

go/internal/consolidate/client.go

@@ -74,15 +74,10 @@ func Run(dataDir, sqlitePath, mode string) (*Result, error) {
 		LLMEndpoint:  os.Getenv("LLM_ENDPOINT"),
 		LLMModel:     os.Getenv("LLM_MODEL"),
 		LLMApiKey:    os.Getenv("LLM_API_KEY"),
-		LLMBudget:    20,
-		// eps: 1024-dim BGE-M3 单位向量
-		// euclidean² = 2(1-cosine), cosine = 1 - euclidean²/2
-		// eps=0.5 → cosine>0.875 (very strict, high similarity)
-		// eps=1.0 → cosine>0.5  (moderate, semantic related)
-		// eps=1.5 → cosine>0     (trivial, almost all cluster → 1 group)
-		// 从 1 cluster 调整到有意义的语义聚类，从 0.5 开始试
-		Epsilon:   0.1,
-		MinPoints: 3,
+		Epsilon:   1.0,
+				// 距离分布: p50=1.029, p95=1.135, max=1.184
+				// eps=0.1 → 1435 noise (太严), eps=1.0 → 合理语义聚类, eps=1.5 → 1个巨大cluster
+				MinPoints: 3,
 		ModelDir:  os.Getenv("BGE_MODEL_DIR"),
 	})
 }

rust/src/lancedb_ops.rs

@@ -106,7 +106,7 @@ impl LanceDBOps {
                     namespace: col_str(&batch, i, "namespace"),
                     content: col_str(&batch, i, "content"),
                     category: col_str(&batch, i, "category"),
-                    vector: vec![],
+                    vector: col_vector(&batch, i, "vector"),
                     tier: col_str(&batch, i, "tier"),
                     importance: col_f64(&batch, i, "importance"),
                     quality_score: col_f64(&batch, i, "quality_score"),
@@ -323,7 +323,7 @@ impl LanceDBOps {
                     namespace: col_str(&batch, i, "namespace"),
                     content: col_str(&batch, i, "content"),
                     category: col_str(&batch, i, "category"),
-                    vector: vec![],
+                    vector: col_vector(&batch, i, "vector"),
                     tier: col_str(&batch, i, "tier"),
                     importance: col_f64(&batch, i, "importance"),
                     quality_score: col_f64(&batch, i, "quality_score"),
@@ -404,3 +404,27 @@ fn col_bool(b: &RecordBatch, r: usize, c: &str) -> bool {
         .map(|a| a.value(r))
         .unwrap_or(false)
 }
+
+/// 读取 vector 列 (兼容 FixedSizeListArray 和 Float32Array)
+fn col_vector(b: &RecordBatch, r: usize, c: &str) -> Vec<f32> {
+    let col = match b.column_by_name(c) {
+        Some(col) => col,
+        None => return vec![0.0_f32; 1024],
+    };
+    // 优先尝试 FixedSizeListArray (当前使用的格式)
+    if let Some(list_arr) = col.as_any().downcast_ref::<FixedSizeListArray>() {
+        let item_arr = list_arr.value(r);
+        if let Some(float_arr) = item_arr.as_any().downcast_ref::<Float32Array>() {
+            return float_arr.values().to_vec();
+        }
+    }
+    // 回退: Float32Array (展平格式)
+    if let Some(float_arr) = col.as_any().downcast_ref::<Float32Array>() {
+        let stride = 1024;
+        let start = r * stride;
+        if start + stride <= float_arr.len() {
+            return float_arr.values()[start..start + stride].to_vec();
+        }
+    }
+    vec![0.0_f32; 1024]
+}