fix(assistant): get_financials 按 period_end 排序后取最近 N 期, 不再依赖文件行序

排序候选是 report_date / end_date / ann_date / date, 本地财务表一个都没有
(报告期列是 period_end, 公告日是 announce_date), 于是不排序直接 tail(N)。
首次全量同步按数据源返回顺序原样落盘, 行序为新→旧时「最近 4 期」拿到的是最早的 4 期。

把 period_end 加为首选排序列, 与 financial_analyzer 按 period_end 取最新 N 期同口径。

Closes #355

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

backend/app/custom/assistant/tools.py

@@ -339,7 +339,8 @@ def _get_financials(args: dict[str, Any], ctx: ToolContext) -> dict[str, Any]:
     sub = df.filter(pl.col("symbol") == symbol)
     if sub.is_empty():
         return {"symbol": symbol, "table": table, "rows": [], "note": "本地财务数据中没有该标的。"}
-    sort_col = next((c for c in ("report_date", "end_date", "ann_date", "date") if c in sub.columns), None)
+    # 本地财务表的报告期列是 period_end; 物理行序取决于同步路径, 须显式排序后再取最近 N 期
+    sort_col = next((c for c in ("period_end", "report_date", "end_date", "ann_date", "date") if c in sub.columns), None)
     if sort_col:
         sub = sub.sort(sort_col)
     sub = sub.tail(periods)

backend/tests/test_assistant_financials_order.py

@@ -0,0 +1,56 @@
+"""AI 助手 get_financials 必须返回「最近 N 个报告期」, 不依赖 parquet 的物理行序。
+
+财务表的报告期列是 period_end (与 financial_sync 的合并键、financial_analyzer 的
+「按 period_end 降序取最新 N 期」一致)。_get_financials 的排序候选却是
+report_date / end_date / ann_date / date —— 本地财务表一个都没有, 于是不排序,
+直接 tail(N) 取物理行序的最后 N 行。首次全量同步 (_sync_table) 按数据源返回顺序
+原样落盘, 数据源按新→旧返回时, 助手答「最近 4 期」拿到的是最早的 4 期。
+"""
+from __future__ import annotations
+
+from datetime import date
+
+import polars as pl
+
+from app.custom.assistant import tools as assistant_tools
+
+_PERIODS = [
+    date(2024, 9, 30), date(2024, 12, 31), date(2025, 3, 31),
+    date(2025, 6, 30), date(2025, 9, 30), date(2025, 12, 31),
+]
+
+
+def _write_metrics(data_dir, periods: list[date]) -> None:
+    out = data_dir / "financials" / "metrics"
+    out.mkdir(parents=True)
+    pl.DataFrame({
+        "symbol": ["600519.SH"] * len(periods) + ["000001.SZ"],
+        "period_end": [*periods, date(2025, 12, 31)],
+        "announce_date": [date(p.year + (p.month == 12), (p.month % 12) + 1, 28) for p in periods] + [date(2026, 3, 20)],
+        "roe": [float(i) for i in range(len(periods))] + [9.9],
+    }).write_parquet(out / "part.parquet")
+
+
+async def _periods_returned(data_dir, periods: int) -> list[str]:
+    ctx = assistant_tools.ToolContext.build(data_dir=data_dir)
+    payload = await assistant_tools.execute_assistant_tool(
+        "get_financials", {"symbol": "600519.SH", "table": "metrics", "periods": periods}, ctx,
+    )
+    assert payload["ok"] is True
+    return [row["period_end"] for row in payload["result"]["rows"]]
+
+
+async def test_latest_periods_when_file_is_newest_first(tmp_path) -> None:
+    """数据源按新→旧返回、首次同步原样落盘。"""
+    _write_metrics(tmp_path, list(reversed(_PERIODS)))
+
+    assert await _periods_returned(tmp_path, 4) == [
+        "2025-03-31", "2025-06-30", "2025-09-30", "2025-12-31",
+    ]
+
+
+async def test_latest_periods_when_file_is_oldest_first(tmp_path) -> None:
+    """累积合并路径 (_merge_report_history) 已按 period_end 升序重写: 结果不变。"""
+    _write_metrics(tmp_path, _PERIODS)
+
+    assert await _periods_returned(tmp_path, 2) == ["2025-09-30", "2025-12-31"]