From 15821a1c988f6b06301b2c2c1061d81814c4e5a9 Mon Sep 17 00:00:00 2001
From: DCCONSTRUCTIONS <dcconstructions@MacBook-Pro-DCCONSTRUCTIONS-2.local>
Date: Sat, 25 Jul 2026 21:20:13 +0300
Subject: [PATCH] fix: honor official RELLIS train sequence split

---
 src/k1link/datasets/rellis_qualification.py | 10 ++++--
 tests/test_rellis_qualification.py          | 40 +++++++++++++++++++++
 2 files changed, 48 insertions(+), 2 deletions(-)

diff --git a/src/k1link/datasets/rellis_qualification.py b/src/k1link/datasets/rellis_qualification.py
index b7bffea..fd12473 100644
--- a/src/k1link/datasets/rellis_qualification.py
+++ b/src/k1link/datasets/rellis_qualification.py
@@ -437,8 +437,14 @@ def _stratified_pairs(
     groups: dict[str, list[tuple[str, str]]] = {}
     for pair in pairs:
         groups.setdefault(pair[0].split("/", 1)[0], []).append(pair)
-    if sorted(groups) != ["00000", "00001", "00002", "00003", "00004"]:
-        raise RellisAdmissionError("RELLIS train split does not cover every sequence")
+    if (
+        len(groups) < 2
+        or any(
+            sequence not in {"00000", "00001", "00002", "00003", "00004"}
+            for sequence in groups
+        )
+    ):
+        raise RellisAdmissionError("RELLIS train split sequence coverage is incompatible")
     selected: list[tuple[str, str]] = []
     remaining = frame_count
     sequence_ids = sorted(groups)
diff --git a/tests/test_rellis_qualification.py b/tests/test_rellis_qualification.py
index 5e51e0b..5baeca7 100644
--- a/tests/test_rellis_qualification.py
+++ b/tests/test_rellis_qualification.py
@@ -57,6 +57,46 @@ def test_rellis_height_calibration_uses_only_stratified_train_frames(
     assert len(result["identity_sha256"]) == 64
 
 
+def test_rellis_calibration_accepts_the_official_train_sequence_subset(
+    tmp_path: Path,
+) -> None:
+    scan_path = tmp_path / "scans.zip"
+    label_path = tmp_path / "labels.zip"
+    pairs: list[tuple[str, str]] = []
+    train_sequences = ("00000", "00002", "00003", "00004")
+    with (
+        zipfile.ZipFile(scan_path, "w") as scans,
+        zipfile.ZipFile(label_path, "w") as labels,
+    ):
+        for sequence_id in train_sequences:
+            for frame_number in range(2):
+                point_member = (
+                    f"{sequence_id}/os1_cloud_node_kitti_bin/{frame_number:06d}.bin"
+                )
+                label_member = (
+                    f"{sequence_id}/os1_cloud_node_semantickitti_label_id/"
+                    f"{frame_number:06d}.label"
+                )
+                point_bytes, label_bytes = _frame(1.3)
+                scans.writestr(f"Rellis-3D/{point_member}", point_bytes)
+                labels.writestr(f"Rellis-3D/{label_member}", label_bytes)
+                pairs.append((point_member, label_member))
+
+    result = calibrate_rellis_sensor_height(
+        scan_path,
+        label_path,
+        tuple(pairs),
+        archive_identity="b" * 64,
+        frame_count=8,
+    )
+
+    calibrated_sequences = {
+        item["frame_id"].split("-")[1] for item in result["frame_estimates"]
+    }
+    assert calibrated_sequences == set(train_sequences)
+    assert "00001" not in calibrated_sequences
+
+
 def test_rellis_archive_index_rejects_path_traversal(tmp_path: Path) -> None:
     archive = tmp_path / "unsafe.zip"
     with zipfile.ZipFile(archive, "w") as target: