test(data): per-file vs global front-month filter — documents quarterly roll bug

The test proves that global filter_front_month on merged trades drops
entire quarters (Q1 lost when Q2 has higher volume). Per-file filtering
keeps both quarters. This is the regression test for the bug that
produced "Need at least 18 months of data" on H100.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml-features/src/trades_loader.rs

@@ -322,4 +322,46 @@ mod tests {
         let filtered = super::filter_front_month(&[]);
         assert!(filtered.is_empty());
     }
+
+    #[test]
+    fn test_filter_front_month_per_file_pattern() {
+        // Simulates quarterly roll: Q1 has front-month id=100, Q2 has id=200.
+        // If you filter globally, only one quarter survives (the one with more volume).
+        // Correct behavior: filter PER FILE, then concatenate both quarters.
+        let q1_trades = vec![
+            DbnTrade { timestamp: 1000, price: 5100.0, volume: 200, is_buy: true, instrument_id: 100 },
+            DbnTrade { timestamp: 2000, price: 5101.0, volume: 100, is_buy: false, instrument_id: 100 },
+            DbnTrade { timestamp: 1500, price: 5160.0, volume: 30, is_buy: true, instrument_id: 101 }, // back month
+        ];
+        let q2_trades = vec![
+            DbnTrade { timestamp: 5000, price: 5200.0, volume: 300, is_buy: true, instrument_id: 200 },
+            DbnTrade { timestamp: 6000, price: 5201.0, volume: 150, is_buy: false, instrument_id: 200 },
+            DbnTrade { timestamp: 5500, price: 5260.0, volume: 20, is_buy: true, instrument_id: 201 }, // back month
+        ];
+
+        // Per-file filtering (CORRECT): each file keeps its own front month
+        let q1_filtered = super::filter_front_month(&q1_trades);
+        let q2_filtered = super::filter_front_month(&q2_trades);
+        assert_eq!(q1_filtered.len(), 2, "Q1 should keep id=100 (2 trades)");
+        assert_eq!(q2_filtered.len(), 2, "Q2 should keep id=200 (2 trades)");
+        assert!(q1_filtered.iter().all(|t| t.instrument_id == 100));
+        assert!(q2_filtered.iter().all(|t| t.instrument_id == 200));
+        let combined_len = q1_filtered.len() + q2_filtered.len();
+        assert_eq!(combined_len, 4, "Per-file: 4 trades from both quarters");
+
+        // Global filtering (WRONG): merges all, picks single highest-volume id
+        let mut all_trades = q1_trades.clone();
+        all_trades.extend(q2_trades.clone());
+        let global_filtered = super::filter_front_month(&all_trades);
+        // id=200 has 450 volume vs id=100's 300 → global keeps only Q2
+        assert!(global_filtered.iter().all(|t| t.instrument_id == 200),
+            "Global filter incorrectly drops Q1 entirely");
+        assert_eq!(global_filtered.len(), 2, "Global filter loses Q1 data");
+
+        // This test documents WHY per-file filtering is required:
+        // global loses 50% of the data (one quarter).
+        assert!(combined_len > global_filtered.len(),
+            "Per-file ({combined_len}) must produce more data than global ({})",
+            global_filtered.len());
+    }
 }