fix(arctic): address review findings and reconcile LFS rules

Regression-tested tools and evidence safeguards; preserve data bytes and pending human review. Reconcile current main LFS rules without changing unrelated dataset content.

ARCTIC/shared-v1.1/doc/FULL_MIGRATION_SUMMARY.json

@@ -254,5 +254,15 @@
     "note": "CONTRACT 6.5 keeps human visual acceptance separate from this record; every check above is mechanical. full_output_human_review in the view metainfo still reads pending and was not rewritten."
   },
   "publication_authorized": false,
-  "scope": "339 sequences, 3051 cameras, 2193687 frames; human post-full review and HF publication pending"
+  "scope": "339 sequences, 3051 cameras, 2193687 frames; human post-full review and HF publication pending",
+  "maintenance": {
+    "date": "2026-10-17",
+    "reviewed_parent": "5299fddd40102b598844da5cbe1cfcbe6dd6f815",
+    "scope": "tools, regression tests and documentation only; data unchanged",
+    "evidence": "doc/REVIEW_FIXES.md",
+    "historical_claims": "All original full-validation and publication-manifest claims describe the reviewed parent, not a fresh full run with these maintenance tools.",
+    "current_external_sha256_manifest_refreshed": true,
+    "human_review": "pending",
+    "manifest_method": "Ordinary Git blob bytes hashed; LFS payload SHA256/size from immutable stored pointers. See REVIEW_FIXES.md; no fresh full-data semantic validation."
+  }
 }

ARCTIC/shared-v1.1/doc/REVIEW_FIXES.md

@@ -0,0 +1,41 @@
+# PR #5 review fixes — 2026-10-17
+
+Reviewed parent: `5299fddd40102b598844da5cbe1cfcbe6dd6f815` in `ParityLab/Hand-Datasets-Raw`.
+This maintenance change updates tools, tests, documentation and evidence bookkeeping only.
+The `data/` subtree and raw data are unchanged from that immutable parent.
+The merge replaces the old release paths atomically and preserves other datasets.
+
+## Evidence scope
+
+No full real-media decode, full numerical comparison or human visual acceptance was rerun
+for this maintenance change. Historical validation results remain historical, not new
+sign-off. Human review is still pending; this document does not grant approval.
+Git tree/blob identities establish unchanged tracked content, not annotation correctness.
+The maintenance audit records exact before/after Git trees and path changes outside the
+release; Git blob SHA1 identities must not be confused with file-content SHA256 hashes.
+
+## Fix and tests
+
+`verify.py` requires existing source/target data directories and nonempty source views.
+Unselected `--complete` additionally requires 3051 source views and identical relative
+target view paths. Partial/incremental checks cannot replace formal acceptance or rebuild
+the publication manifest. Both verify and validate reject a manifest as a report output,
+including normalized path aliases. Invalid/empty selections fail before evidence writes.
+
+`python -B script/arctic_v1_1/test_review_fixes.py`: 10 tests pass. The combined
+ARCTIC/ReInterHand isolated CLI regression suite passes all 19 scenarios, including
+missing source, partial evidence protection and ReInterHand media-valid failures.
+Those CLI fixtures use cached numeric data; ARCTIC media/assets are marked placeholders,
+and ReInterHand declares unavailable video. They test control flow, not media correctness.
+Earlier independent numeric sampling covered 18 ARCTIC and 27 ReInterHand views; this
+maintenance does not expand that sample into a full comparison.
+
+`FULL_MIGRATION_SUMMARY.json` keeps its original validation claims as historical fields;
+its maintenance section narrows their provenance. The author's external SHA256 publication
+manifest was not available locally. A fresh external PUBLISH_MANIFEST.json is generated
+from the final fixed Git tree: SHA256 is computed over ordinary file bytes, and LFS
+content SHA256/size is taken from the exact stored pointer. This verifies published
+content identities without downloading or decoding the LFS payloads. A separate
+Git-identity ledger records paths and blob IDs for merge auditing.
+
+The existing object-keypoint registry exception is retained, not silently relabeled.

ARCTIC/shared-v1.1/script/arctic_v1_1/README.md

@@ -253,4 +253,12 @@ traceback。checkpoint 也记录同一份工具摘要，所以**改动 `migrate.
 
 **画外点怎么判定、怎么披露。** 页面只画发布自己的 camera 系几何，所以「物体大部分落在画外」是会真实发生的情形，而一个几乎空的面板光看是分不出「物体真的出画」和「投影坏了」的。抽样出的 54 个视图里只有 `s05_microwave_use_01/0` 命中：第 195 帧 32 个有效关键点只有 1 个落在画内，而**同一帧**的 8 路 exo 视图是 32/32——该视图是戴在头上的 ego 相机，物体正好在视野上缘之上，与发布声明的 `coordinate_frame: camera` 自洽，不是投影错误。处置是披露而不是改数：页面在标题下点名命中的帧并说明这是 ego 视图的视野问题，面板标题以「物体大部分落在画外（该帧 N 个有效关键点只有 M 个落在画内）」开头，规则一并写出（`inside_points / valid_points < 0.5` 即标记）。规则之外还逐帧记数：**每个抽样帧**的 `annotated`、`valid_points`、`inside_points`、`inside_ratio`、`mostly_out_of_frame` 都写进该视图报告的 `projection.frames`（源未标注的帧记 `annotated: false`，不省略），规则原文写进 `projection.out_of_frame_rule`，视图级计数写进 `geometry_panels.frames_mostly_out_of_frame`。门限只用于披露，不改动任何发布数据。
 
-这套页面是**人工验收要看的材料**，不是训练输入。CONTRACT §6.5 区分「重打包在数值上逐字节同一」与「人工目视验收」，页面给的是后者；渲出页面不等于有人看过，所以发布状态里的 `full_output_human_review` 与汇总里的 `human_full_review` 都仍是 `"pending"`。ARCTIC v1.0 自己的复核产物照旧随发布保留在 `doc/history-v1.0/validation/`。
+这套页面是**人工验收要看的材料**，不是训练输入。CONTRACT §6.5 区分「重打包在数值上逐字节同一」与「人工目视验收」，页面给的是后者；渲出页面不等于有人看过，所以发布状态里的 `full_output_human_review` 与汇总里的 `human_full_review` 都仍是 `"pending"`。ARCTIC v1.0 自己的复核产物照旧随发布保留在 `doc/history-v1.0/validation/`。
+## 2026-10-17 审查修复
+
+`verify.py` 在写任何验收产物前拒绝缺失/空源树；完整验收必须有完整源视图集合，且与目标集合一致。
+仅下载新发布树时，必须先从固定旧提交准备 `--v10`，不能把缺失的旧树当成已通过验收。
+子集/增量运行不再重建发布清单，也不能覆盖正式数值验收记录。
+`verify.py` / `validate.py` 均禁止将报告写到发布清单；选择数量必须为正。
+运行 `python -B test_review_fixes.py` 执行这些边界的离线回归。
+本次工具修改与旧验收记录的范围见 `../../doc/REVIEW_FIXES.md`。

ARCTIC/shared-v1.1/script/arctic_v1_1/test_review_fixes.py

@@ -0,0 +1,119 @@
+"""Offline regression checks for input completeness and protected evidence paths."""
+import contextlib
+import io
+import json
+from pathlib import Path
+import tempfile
+import unittest
+from unittest.mock import patch
+
+import validate
+import verify
+
+
+class ReviewFixes(unittest.TestCase):
+    def setUp(self):
+        self.temp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.temp.cleanup)
+        self.root = Path(self.temp.name)
+        self.source, self.target = self.root / 'v10', self.root / 'v11'
+        self.schema = verify.RELEASE_ROOT / 'doc/schema.json'
+        self.manifest = self.root / 'PUBLISH_MANIFEST.json'
+        self.acceptance = self.root / 'acceptance.json'
+        self.structural = self.root / 'validate_report.json'
+        for path in (self.manifest, self.acceptance, self.structural):
+            path.write_bytes(b'original evidence')
+        for module, key, value in (
+            (verify, 'MANIFEST_PATH', self.manifest),
+            (verify, 'ACCEPTANCE_REPORT', self.acceptance),
+            (validate, 'MANIFEST_PATH', self.manifest),
+            (validate, 'ACCEPTANCE_REPORT', self.structural),
+            (validate, 'RELEASE_ROOT', self.target),
+            (validate, 'TMP_ROOT', self.root / 'tmp'),
+        ):
+            guard = patch.object(module, key, value)
+            guard.start()
+            self.addCleanup(guard.stop)
+        self.args = ['--v10', str(self.source), '--v11', str(self.target),
+                     '--schema', str(self.schema)]
+
+    def view(self, root, sequence='s', camera='c'):
+        path = root / 'data' / sequence / camera / 'metainfo.json'
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_text('{}')
+
+    def rejected(self, module, args):
+        before = {p.relative_to(self.root): p.read_bytes()
+                  for p in self.root.rglob('*') if p.is_file()}
+        with contextlib.redirect_stdout(io.StringIO()), contextlib.redirect_stderr(io.StringIO()):
+            with self.assertRaises((SystemExit, AssertionError)):
+                module.main(args)
+        after = {p.relative_to(self.root): p.read_bytes()
+                 for p in self.root.rglob('*') if p.is_file()}
+        self.assertEqual(before, after, 'rejected runs must not modify evidence')
+
+    def test_missing_source(self):
+        self.view(self.target)
+        self.rejected(verify, self.args + ['--complete', '--report', str(self.acceptance)])
+
+    def test_empty_source(self):
+        (self.source / 'data').mkdir(parents=True)
+        self.view(self.target)
+        self.rejected(verify, self.args + ['--complete', '--report', str(self.acceptance)])
+
+    def test_missing_target(self):
+        self.view(self.source)
+        self.rejected(verify, self.args + ['--complete'])
+
+    def test_truncated_full_source(self):
+        self.view(self.source)
+        self.view(self.target)
+        self.rejected(verify, self.args + ['--complete', '--report', str(self.acceptance)])
+
+    def test_complete_requires_same_view_set(self):
+        self.view(self.source, 's1')
+        self.view(self.source, 's2')
+        self.view(self.target, 's1')
+        self.view(self.target, 'wrong')
+        with patch.object(verify, 'EXPECTED_VIEWS', 2):
+            self.rejected(verify, self.args + ['--complete'])
+
+    def test_partial_cannot_overwrite_acceptance(self):
+        self.rejected(verify, self.args + ['--complete', '--sample', '1',
+                                         '--report', str(self.acceptance)])
+
+    def test_manifest_is_never_a_report(self):
+        for module, args in ((verify, self.args), (validate, [])):
+            self.rejected(module, args + ['--report', str(self.manifest)])
+            self.rejected(module, args + ['--report', str(self.root / 'sub/../PUBLISH_MANIFEST.json')])
+        custom = self.root / 'custom-manifest.json'
+        self.rejected(verify, self.args + ['--manifest', str(custom), '--report', str(custom)])
+
+    def test_selection_must_be_nonempty_and_positive(self):
+        for module, option, args in ((verify, '--sample', self.args), (validate, '--limit', [])):
+            for value in ['0', '-1']:
+                self.rejected(module, args + [option, value])
+            self.rejected(module, args + ['--views'])
+
+    def test_unknown_source_sequence(self):
+        self.view(self.source)
+        (self.target / 'data').mkdir(parents=True)
+        self.rejected(verify, self.args + ['--views', 'unknown', '--no-manifest'])
+
+    def test_incremental_absent_view_stays_explicit_and_does_not_write_manifest(self):
+        self.view(self.source)
+        (self.target / 'data').mkdir(parents=True)
+        report = self.root / 'incremental.json'
+        with contextlib.redirect_stdout(io.StringIO()):
+            code = verify.main(self.args + ['--views', 's', '--report', str(report)])
+        self.assertEqual(code, 0)
+        result = json.loads(report.read_text())
+        self.assertEqual(result['views_absent_from_target'], ['s/c'])
+        self.assertEqual(result['views_migrated'], 0)
+        self.assertIn('skipped', result['manifest'])
+        self.assertEqual(self.manifest.read_bytes(), b'original evidence')
+        self.assertEqual(self.acceptance.read_bytes(), b'original evidence')
+
+
+if __name__ == '__main__':
+    unittest.main()

ARCTIC/shared-v1.1/script/arctic_v1_1/validate.py

@@ -339,6 +339,12 @@ def main(argv=None):
                              "without being asked. The accepting run must not pass this "
                              "(CONTRACT 6.4).")
     args = parser.parse_args(argv)
+    if args.limit is not None and args.limit < 1:
+        parser.error("--limit must be positive")
+    if args.views is not None and not args.views:
+        parser.error("--views requires at least one sequence")
+    if args.report is not None and args.report.resolve() == MANIFEST_PATH.resolve():
+        parser.error("--report must not overwrite the publication manifest")
 
     all_views = sorted((RELEASE_ROOT / "data").glob("*/*/metainfo.json"))
     wanted = set(args.views) if args.views else None

ARCTIC/shared-v1.1/script/arctic_v1_1/verify.py

@@ -49,7 +49,10 @@ DATASET_ROOT = RELEASE_ROOT.parent
 #: validate.py is what the documented chain rebuilds the shipping manifest with.
 MANIFEST_PATH = DATASET_ROOT / "tmp" / "arctic-tools" / "PUBLISH_MANIFEST.json"
 
+ACCEPTANCE_REPORT = MANIFEST_PATH.parent / "acceptance.json"
+
 V10_SCHEMA_VERSION = "hand-warehouse/1.0"
+EXPECTED_VIEWS = 3051
 #: metainfo keys the migration is allowed to rewrite at the top level.  Everything
 #: else is a claim the source made and must survive verbatim -- ARCTIC's release
 #: status flags, the omitted-component reasons, the split bookkeeping.  The version
@@ -384,10 +387,32 @@ def main(argv=None):
                              "the one to write it.")
     args = parser.parse_args(argv)
 
+    if args.sample is not None and args.sample < 1:
+        parser.error("--sample must be positive")
+    if args.views is not None and not args.views:
+        parser.error("--views requires at least one sequence")
+    acceptance_pass = args.complete and args.views is None and args.sample is None
+    if args.report is not None:
+        if args.report.resolve() in {MANIFEST_PATH.resolve(), args.manifest.resolve()}:
+            parser.error("--report must not overwrite a publication manifest")
+        if not acceptance_pass and args.report.resolve() == ACCEPTANCE_REPORT.resolve():
+            parser.error("the acceptance report requires --complete without --views/--sample")
+
     registry = Registry.load(args.schema)
     reader = BundleReader()
-    source_views = sorted(p.parent for p in (args.v10 / "data").glob("*/*/metainfo.json"))
-    target_data = args.v11 / "data"
+    source_data, target_data = args.v10 / "data", args.v11 / "data"
+    if not source_data.is_dir() or not target_data.is_dir():
+        parser.error("both source and target data directories must exist")
+    source_views = sorted(p.parent for p in source_data.glob("*/*/metainfo.json"))
+    if not source_views:
+        parser.error("source contains no views; no acceptance evidence was written")
+    if acceptance_pass:
+        source_set = {p.relative_to(source_data) for p in source_views}
+        target_set = {p.parent.relative_to(target_data)
+                      for p in target_data.glob("*/*/metainfo.json")}
+        if len(source_set) != EXPECTED_VIEWS or source_set != target_set:
+            parser.error(f"full acceptance requires {EXPECTED_VIEWS} source views and "
+                         "the exact same target view set")
     migrated, absent = [], []
     for view in source_views:
         relative = view.relative_to(args.v10 / "data")
@@ -402,6 +427,7 @@ def main(argv=None):
         known = {view.parent.name for view in source_views}
         assert wanted <= known, ("no such view", sorted(wanted - known))
         migrated = [v for v in migrated if v.parent.name in wanted]
+        absent = [v for v in absent if v.parent.name in wanted]
     if args.sample:
         migrated = migrated[:args.sample]
 
@@ -425,7 +451,7 @@ def main(argv=None):
             print(f"FAIL {relative}: {error}", file=sys.stderr)
 
     # 4: regenerated release manifest
-    if args.no_manifest:
+    if args.no_manifest or not acceptance_pass:
         #: reported as skipped, not as "0 files" -- an empty manifest in the report
         #: would read like a tree with nothing in it
         report["manifest"] = {"skipped": "incremental check; not an acceptance run"}