feat(HEL-478): 估值发布后晚间复核并原子追补上游修订

盘后成功发布后继续轻量比对 daily_basic 网站字段，发现修订才走质量门与整组原子切换，避免 17:10 快照落后于晚间上游改写。 Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: multica-agent <github@multica.ai>
2026-09-07 21:19:35 +08:00
parent 16ba83ec01
commit d175bb65d4
11 changed files with 982 additions and 4 deletions
@@ -0,0 +1,111 @@
+"""Post-publish revision review for datasets whose upstream may rewrite T-day fields.
+
+HEL-423 field evidence, not a whitelist of tolerated diffs:
+
+- 2026-09-07 valuation/daily_basic: hub published 003021.SZ turnover_rate=1.3565
+  at 17:10; website legacy and a direct Tushare read at 21:05 both showed 1.3572.
+  The other seven observed objects (daily, moneyflow, auction, stocks, status,
+  index_daily, calendar) matched. Hub had already stopped the day after the
+  first successful publish, so the revision never self-healed.
+- 2026-09-02: same dataset, opposite direction (hub already held the later
+  value). Confirms daily_basic is rewritten after the first complete dump.
+
+Daily bars, moneyflow, auction and index_daily have no same-evening field
+revision evidence. Stocks already refreshes at 20:00/23:10. Review therefore
+fetches only configured revision-risk datasets (default: valuation) and
+compares the website-requested field set. No numeric tolerance.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from datahub.db import DATASET_TABLES
+from datahub.normalize import VALUATION_FIELDS
+from datahub.numbers import finite_number, round4
+
+# Datasets with proven same-evening upstream rewrites. Config may replace this
+# list; it must not silently expand to a full EOD re-pull.
+DEFAULT_REVISION_DATASETS = ("valuation",)
+
+# Website daily_basic request (HEL-423): ts_code/trade_date plus the eight
+# value fields used by the old link and field_gates.
+WEBSITE_COMPARE_FIELDS: dict[str, tuple[str, ...]] = {
+    "valuation": VALUATION_FIELDS,
+}
+
+REVISION_STATES = ("waiting_review", "review_failed", "aligned", "cutoff")
+
+
+def revision_datasets(quality: dict[str, Any] | None) -> tuple[str, ...]:
+    raw = (quality or {}).get("revision_review_datasets")
+    if isinstance(raw, (list, tuple)) and raw:
+        names = tuple(str(item) for item in raw if str(item))
+        if names:
+            return names
+    return DEFAULT_REVISION_DATASETS
+
+
+def compare_fields(dataset: str) -> tuple[str, ...]:
+    fields = WEBSITE_COMPARE_FIELDS.get(dataset)
+    if fields:
+        return fields
+    gate = {}
+    return tuple(str(item) for item in (gate.get("fields") or []) if str(item))
+
+
+def _norm_value(field: str, value: Any) -> Any:
+    if field in {"ts_code", "trade_date"}:
+        return str(value or "")
+    number = round4(finite_number(value))
+    return number
+
+
+def row_signature(row: dict[str, Any], fields: tuple[str, ...]) -> tuple[Any, ...]:
+    return tuple(_norm_value(field, row.get(field)) for field in fields)
+
+
+def diff_published_vs_upstream(
+    dataset: str,
+    published: list[dict[str, Any]],
+    upstream: list[dict[str, Any]],
+    *,
+    max_diffs: int = 20,
+) -> dict[str, Any]:
+    """Exact compare on website-requested fields. No tolerance / exemption."""
+    fields = compare_fields(dataset)
+    if not fields:
+        fields = tuple(sorted({key for row in published + upstream for key in row if key != "batch_id"}))
+    pub_map = {str(row.get("ts_code") or "").upper(): row for row in published}
+    up_map = {str(row.get("ts_code") or "").upper(): row for row in upstream}
+    missing = sorted(code for code in pub_map if code not in up_map)
+    extra = sorted(code for code in up_map if code not in pub_map)
+    diffs: list[dict[str, Any]] = []
+    for code in sorted(set(pub_map) & set(up_map)):
+        left = row_signature(pub_map[code], fields)
+        right = row_signature(up_map[code], fields)
+        if left == right:
+            continue
+        for field, old, new in zip(fields, left, right):
+            if old == new:
+                continue
+            diffs.append({"ts_code": code, "field": field, "published": old, "upstream": new})
+            if len(diffs) >= max_diffs:
+                break
+        if len(diffs) >= max_diffs:
+            break
+    changed = bool(diffs or missing or extra)
+    return {
+        "changed": changed,
+        "dataset": dataset,
+        "fields": list(fields),
+        "published_rows": len(published),
+        "upstream_rows": len(upstream),
+        "missing_codes": missing[:max_diffs],
+        "extra_codes": extra[:max_diffs],
+        "diffs": diffs,
+    }
+
+
+def official_table(dataset: str) -> str:
+    return DATASET_TABLES[dataset][0]