akg-factor-bridge/plan.py

"""每日选股计划（R4）：数据装配 / Markdown 渲染 / 产出，三段分离。

  collect()   -> dict   结构化计划——api.py 直接当 JSON 返回
  render_md() -> str    从 dict 渲染 Markdown
  generate()            CLI 与 cron 的入口：collect + render + 落盘 + 打印

数据全部来自已落库的表，不重算：
  平台因子表  t_factor_akg_score / _gate / _upside / _heat   —— 当日截面
  基座只读视图 v_factor_transmission                          —— 传导证据
  基座 industry_pools                                         —— 股票名称
升降档一节对比前一交易日的档位表——数据到达本身是信号（首次覆盖 /
新进传导链即升档）。

2026-09-03 起（《主观量化系统方案_2026-09-03》第 3.3 节）：候选卡每票多带数据基座的因果论断
证据线（只展示不进判决，sources.logic_claims）；generate 出计划时把市场四项（两市成交额、广度、
融资、恐贪，sources.market_context）写进快照的 market 段，与 08:45 追加的 regime 段并列，
接口 /plan 只从快照读这两段。
"""
from __future__ import annotations   # 注解不在定义时求值：开发机的 Python 3.9 也能导入本模块跑离线单测

import datetime as dt
import json
import os

import pandas as pd

import card
import common
import config
import db
import sources
import version


# 分数编码（与 factors.build_score 一致）：主榜 = 200 + 传导档位×20 + 组内分，
# 观察档 = 100 + 组内分，组内分 clip ±9.9。150 落在两带中间的空档上，用作分界。
_MAIN_MIN = 150.0


def _factor(table: str, ds: str) -> pd.Series:
    df = db.read_mysql(
        "factor",
        f"SELECT stock_code, factor_value FROM {table} WHERE trade_date = %s", (ds,))
    if df.empty:
        return pd.Series(dtype=float)
    return df.set_index("stock_code")["factor_value"].astype(float)


def _latest_date(table: str, upto: str | None = None):
    if upto:
        df = db.read_mysql(
            "factor", f"SELECT MAX(trade_date) d FROM {table} "
                      f"WHERE trade_date <= %s", (upto,))
    else:
        df = db.read_mysql("factor", f"SELECT MAX(trade_date) d FROM {table}")
    v = None if df.empty else df.iloc[0, 0]
    return None if v is None or pd.isna(v) else pd.Timestamp(v).date().isoformat()


def _prev_date(table: str, before: str):
    df = db.read_mysql(
        "factor", f"SELECT MAX(trade_date) d FROM {table} "
                  f"WHERE trade_date < %s", (before,))
    v = None if df.empty else df.iloc[0, 0]
    return None if v is None or pd.isna(v) else pd.Timestamp(v).date().isoformat()


def _names() -> dict:
    out = {}
    pools = db.read_pg("SELECT members FROM industry_pools")
    for _, r in pools.iterrows():
        ms = r["members"]
        if isinstance(ms, str):
            ms = json.loads(ms)
        for m in ms or []:
            ts, name = (m or {}).get("ts_code"), (m or {}).get("name")
            if ts and name:
                out.setdefault(common.to_prefix(ts), str(name))
    return out


def _evidence(ds: str):
    """每股最强一条传导证据：主题、源数、已动比例；另返回涉及的行情快照日。"""
    tr = db.read_pg(
        "SELECT ts_code, target, n_sources, moved_ratio, mkt_trade_date, "
        "members_total, moved FROM v_factor_transmission WHERE scan_date = %s", (ds,))
    if tr.empty:
        return {}, set()
    tr["k"] = tr["ts_code"].map(common.to_prefix)
    tr["n_sources"] = pd.to_numeric(tr["n_sources"], errors="coerce").fillna(0)
    tr["moved_ratio"] = pd.to_numeric(tr["moved_ratio"], errors="coerce").fillna(0)
    tr["strength"] = tr["n_sources"] * (1.0 - tr["moved_ratio"])
    tr = tr.sort_values("strength", ascending=False).drop_duplicates("k")
    # 元组前三项是既有口径（多处按下标取），第四、五项是环节的成员数与已动数（关注环节表用）
    ev = {r.k: (str(r.target), int(r.n_sources), float(r.moved_ratio),
                None if pd.isna(r.members_total) else int(r.members_total),
                None if pd.isna(r.moved) else int(r.moved))
          for r in tr.itertuples()}
    days = {str(x) for x in tr["mkt_trade_date"].dropna().unique()}
    return ev, days


def _tier_label(score: float) -> str:
    return {0: "无传导", 1: "弱传导", 2: "强传导"}.get(
        int((score - 190.0) // 20), "?")


def _val(series: pd.Series, k: str):
    v = series.get(k)
    return None if v is None or pd.isna(v) else float(v)


def _assemble_cards(ds: str, codes: list, ev: dict, upside: pd.Series,
                    mkt_days: set, risk: set | None = None) -> tuple[dict, list]:
    """候选卡装配（2026-09-02 方案第 2.2 节第三项）：对档位表里的全部票（主榜与观察档，
    裁剪之前）读三路证据、逐票判决。规则在 card.py，取数在 sources.py，这里只做对齐。

    返回 (cards, segments_pointed)：cards 按前缀码索引，含判决、理由、缺失、风险、卡内序
    与证据线原值；segments_pointed 是"关注环节"聚合——今日被传导指向的每个环节的源数、
    链符、成员数、已启动成员、领涨者、候选数。这是拍板记录第一项"强传导档降为关注环节"
    在计划文本层的落地。"""
    import factors

    moved = sources.moved_members(ds)
    daily = sources.stock_daily(ds)
    night = sources.night_conclusions(codes, ds)
    # 因果论断（2026-09-03）：数据基座抽取的论断挂在卡上作证据线，只展示不进判决；视图未建时为空。
    logic = sources.logic_claims(codes, ds)
    if risk is None:                              # collect 会传入读过一次的名单；单独调用时自己读
        try:
            risk = factors._risk_set() or set()  # noqa: SLF001 —— 同仓自用
        except Exception:  # noqa: BLE001 —— 风险名单拿不到时不当 ST 处理，与 gate 的宽容一致
            risk = set()
    stale = bool(mkt_days) and any(x != ds for x in mkt_days)

    cards: dict = {}
    for k in codes:
        mv, e, d, n = moved.get(k), ev.get(k), daily.get(k, {}), night.get(k, {})
        theme = (mv or {}).get("theme") or (e[0] if e else None)
        n_sources = (mv or {}).get("n_sources") or (e[1] if e else None)
        up = _val(upside, k)
        evd = {"pointed": bool(mv or e), "theme": theme, "n_sources": n_sources,
               "chain_fit": (mv or {}).get("chain_fit"),
               "pct0": d.get("pct0"), "covered": up is not None, "upside": up,
               "risk_name": k in risk,
               "accum_state": n.get("accum_state"), "accum_score": n.get("accum_score"),
               "accum_age": n.get("accum_age"), "y_signal": n.get("signal"),
               "stale_snapshot": stale, "logic": logic.get(k) or []}
        j = card.judge(evd, start_pct=config.CARD_START_PCT,
                       accum_max_age=config.CARD_ACCUM_MAX_AGE,
                       neg_tol=config.UPSIDE_NEG_TOLERANCE,
                       logic_stale_days=config.LOGIC_STALE_DAYS,
                       require_started=config.CARD_REQUIRE_STARTED)
        cards[k] = {
            **j,
            "theme": theme, "n_sources": n_sources, "chain_fit": evd["chain_fit"],
            "started_source": "moved_view" if mv else None,
            "logic_claims": evd["logic"],
            "pct0": d.get("pct0"), "net_z": d.get("net_z"), "heat_chg": d.get("heat_chg"),
            "accum": ({"state": n.get("accum_state"), "score": n.get("accum_score"),
                       "age": n.get("accum_age"), "pos_tag": n.get("accum_pos_tag"),
                       "date": n.get("conclusion_date")} if n else None),
            "night": ({"signal": n.get("signal"), "support": n.get("support"),
                       "pressure": n.get("pressure"), "date": n.get("conclusion_date")}
                      if n else None),
        }
    for i, (k, c) in enumerate(sorted(cards.items(), key=lambda kv: card.sort_key(kv[1])), 1):
        c["card_rank"] = i

    segs: dict = {}
    for k, mv in moved.items():
        s = segs.setdefault(mv["theme"], {
            "segment": mv["theme"], "n_sources": mv["n_sources"], "chain_fit": mv["chain_fit"],
            "members_total": mv.get("members_total"), "moved": mv.get("moved"),
            "started": [], "candidates": 0, "leader": None})
        s["started"].append(k)
        if cards.get(k, {}).get("verdict") == card.VERDICT_CANDIDATE:
            s["candidates"] += 1
    for e in ev.values():                       # 只有未动名单的环节也是"被指向"，列入但无已启动
        s = segs.setdefault(e[0], {"segment": e[0], "n_sources": e[1], "chain_fit": None,
                                   "members_total": None, "moved": None,
                                   "started": [], "candidates": 0, "leader": None})
        if s["members_total"] is None and len(e) > 4:
            s["members_total"], s["moved"] = e[3], e[4]
    for s in segs.values():
        best, best_pct = None, None
        for k in s["started"]:
            # 领涨者按行情视图找，不限于档位表里的票：被指向且已启动但不在主榜与观察档的
            # （无覆盖且不在赛道、或被风险闸挡）也要显示，这是"定位对不对"的读数
            p = (daily.get(k) or {}).get("pct0")
            if p is not None and (best_pct is None or p > best_pct):
                best, best_pct = k, p
        s["leader"] = {"code": best, "pct0": best_pct} if best else None
        s["started_count"] = len(s["started"])
        s["started_in_tiers"] = sum(1 for k in s["started"] if k in cards)
    ordered = sorted(segs.values(), key=lambda s: (-s["candidates"], -(s["n_sources"] or 0),
                                                   -(s["chain_fit"] or 0), s["segment"]))
    return cards, ordered


def collect(date: str | None = None, top: int = 20, obs_top: int = 10,
            theme_cap: int = 5) -> dict:
    """装配一天的计划为结构化字典。数据缺失抛 RuntimeError（api 侧转 404）。

    2026-09-02 起附带候选卡：main / observe 的装配、排序、裁剪一字不动（下游 PMS 只读
    这两段的既有字段），每行只是多联入判决类字段；顶层新增 generated_at、plan_version、
    card_counts、candidates（候选单全量，不受裁剪）、watch、segments_pointed。
    `_full` 是全量主榜与观察档行，只给 generate 落快照用，api 返回前会去掉。"""
    ds = date or _latest_date("t_factor_akg_score")
    if not ds:
        raise RuntimeError("t_factor_akg_score 还没有数据——先 build akg_score。")
    score = _factor("t_factor_akg_score", ds)
    gate = _factor("t_factor_akg_gate", ds)
    if score.empty or gate.empty:
        raise RuntimeError(f"{ds} 缺 akg_score / akg_gate——先 build 该日再出计划。")
    upside = _factor("t_factor_akg_upside", ds)
    if upside.empty:
        # 当日 upside 表为空时现算兜底（as-of 口径不变：consensus<=当日、当日收盘价）
        import factors
        df_up = factors.build_upside(ds, ds)
        if df_up is not None and not df_up.empty:
            x = df_up.copy()
            x["k"] = x["stock_code"].map(common.to_prefix)
            upside = x.groupby("k")["factor_value"].max().astype(float)
    hd = _latest_date("t_factor_akg_heat", ds)
    heat = _factor("t_factor_akg_heat", hd) if hd else pd.Series(dtype=float)
    names = _names()
    ev, mkt_days = _evidence(ds)

    main = score[score >= _MAIN_MIN].sort_values(ascending=False)
    obs = score[score < _MAIN_MIN].sort_values(ascending=False)
    generated_at = dt.datetime.now().isoformat(timespec="seconds")
    try:                                          # 风险名单只读一次：候选卡与升降档原因共用
        import factors
        risk = factors._risk_set() or set()      # noqa: SLF001 —— 同仓自用
    except Exception:  # noqa: BLE001
        risk = set()
    cards, segments_pointed = _assemble_cards(
        ds, list(main.index) + list(obs.index), ev, upside, mkt_days, risk=risk)

    def _pick(ranked: pd.Series, n: int):
        """分数从高到低取 n 条；每个传导主题最多 theme_cap 条（0=不设限）——
        传导目标是环节级、同环节成员共享同一条证据，不限额会被少数环节刷屏。"""
        out, cnt = [], {}
        for k, s in ranked.items():
            e = ev.get(k)
            theme = e[0] if e else "（无传导）"
            if theme_cap and cnt.get(theme, 0) >= theme_cap:
                continue
            cnt[theme] = cnt.get(theme, 0) + 1
            out.append((k, s))
            if len(out) >= n:
                break
        return out

    def _row(rank: int, k: str, s: float, with_tier: bool) -> dict:
        e = ev.get(k)
        c = cards.get(k) or {}
        evidence = ({"theme": e[0], "n_sources": e[1], "moved_ratio": round(e[2], 4)}
                    if e else None)
        # 已启动成员在传导视图里没有证据行（视图只摊平未动名单）：只在原本为空时用
        # 已动成员视图的目标环节与源数补上，不覆盖已有值——否则到 PMS 会全落进"无主题"桶。
        if evidence is None and c.get("started_source") == "moved_view" and c.get("theme"):
            evidence = {"theme": c["theme"], "n_sources": c.get("n_sources"),
                        "moved_ratio": None, "source": "moved_view"}
        r = {"rank": rank, "code": k, "name": names.get(k),
             "score": round(float(s), 2),
             "evidence": evidence,
             "heat": _val(heat, k), "upside": _val(upside, k)}
        if with_tier:
            r["tier"] = _tier_label(s)
        if c:
            # 2026-09-03 新增只联入：basis（判决依据一句话）、logic（因果论断带出处的文字行）、
            # card.chain_fit（链符，入池上下文用）、card.logic_claims（论断原值）、card.failed_gates。
            r.update(verdict=c["verdict"], reasons=c["reasons"], missing=c["missing"],
                     risk=c["risk"], card_rank=c["card_rank"],
                     basis=c.get("basis"), logic=c.get("logic") or [],
                     card={"pct0": c.get("pct0"), "net_z": c.get("net_z"),
                           "heat_chg": c.get("heat_chg"), "accum": c.get("accum"),
                           "night": c.get("night"), "gates": c.get("gates"),
                           "confirm": c.get("confirm"), "failed_gates": c.get("failed_gates"),
                           "chain_fit": c.get("chain_fit"),
                           "logic_claims": c.get("logic_claims") or []})
        return r

    def _full_rows(ranked: pd.Series, with_tier: bool) -> list:
        return [_row(i, k, s, with_tier) for i, (k, s) in enumerate(ranked.items(), 1)]

    def _by_verdict(v: str) -> list:
        rows = [_row(0, k, score[k], k in main.index) for k, c in cards.items()
                if c.get("verdict") == v]
        rows.sort(key=lambda r: r["card_rank"])
        for i, r in enumerate(rows, 1):
            r["rank"] = i
        return rows

    changes = None
    prev_ds = _prev_date("t_factor_akg_gate", ds)
    if prev_ds:
        prev = _factor("t_factor_akg_gate", prev_ds)
        both = pd.concat([prev.rename("prev"), gate.rename("cur")],
                         axis=1).fillna(-1.0)          # -1 = 当日不在面板
        lab = {-1.0: "池外", 0.0: "不采纳", 1.0: "观察档", 2.0: "主榜"}
        up_df = both[both["cur"] > both["prev"]].sort_values("cur", ascending=False)
        down_df = both[both["cur"] < both["prev"]].sort_values("prev", ascending=False)

        # ---- 升降原因（07-31 加）：区分首次覆盖 / 估值转正 / 新进传导链等。
        # 依据前一日的 upside / 传导因子表；表空时现算兜底。原因是启发式归类
        # （取最主要的一条），精确审计以档位表与因子表为准。
        prev_up = _factor("t_factor_akg_upside", prev_ds)
        if prev_up.empty:
            try:
                import factors
                dfu = factors.build_upside(prev_ds, prev_ds)
                if dfu is not None and not dfu.empty:
                    x = dfu.copy()
                    x["k"] = x["stock_code"].map(common.to_prefix)
                    prev_up = x.groupby("k")["factor_value"].max().astype(float)
            except Exception:  # noqa: BLE001 —— 兜底失败则原因退化为通用文案
                pass
        prev_tr = _factor("t_factor_akg_transmission", prev_ds)
        # 赛道闸与风险闸的成员集合（07-31 修：赛道闸开启后，"不在赛道"曾被
        # 误标成"风险闸/档位调整"、"赛道锚生效"曾被误标成"新进传导链"）。
        # 两个集合都尊重各自开关：闸没开时对应原因自然不会出现。
        tset = None
        try:
            import factors
            tset = factors._track_set()   # noqa: SLF001 —— 同仓自用
        except Exception:  # noqa: BLE001 —— 拿不到就退化为通用文案
            pass

        def _why(k: str, pg: float, cg: float) -> str:
            if cg == 2.0:                              # 升入主榜
                pu = _val(prev_up, k)
                if pu is None:
                    return "首次覆盖"
                return "估值转正" if pu < 0 else "重获资格"
            if cg == 1.0:                              # 升入观察档
                if k in ev:                            # 今天真在传导链上
                    pt = _val(prev_tr, k)
                    return "新进传导链" if pt is None or pt <= 0 else "档位调整"
                if tset is not None and k in tset:
                    return "赛道锚生效"                # 赛道成员身份给的锚（闸切换/成员变动）
                return "档位调整"
            if pg == 2.0:                              # 从主榜降出
                if k in risk:
                    return "风险闸"
                cu = _val(upside, k)
                if cu is None:
                    return "覆盖脱落"
                if cu < 0:
                    return "估值转负"
                if tset is not None and k not in tset:
                    return "赛道闸外"                  # 有覆盖也不贵，但不在十赛道内
                return "档位调整"
            if pg == 1.0:                              # 从观察档降出
                if k in risk:
                    return "风险闸"
                if k not in ev and (tset is None or k not in tset):
                    return "离开传导链"
                return "档位调整"
            return "出入面板"

        def _mv(d: pd.DataFrame):
            return [{"code": k, "name": names.get(k),
                     "from": lab.get(r["prev"], "?"), "to": lab.get(r["cur"], "?"),
                     "reason": _why(k, r["prev"], r["cur"])}
                    for k, r in d.iterrows()]

        changes = {"base_date": prev_ds,
                   "upgrades_total": int(len(up_df)),
                   "downgrades_total": int(len(down_df)),
                   "upgrades": _mv(up_df.head(15)),
                   "downgrades": _mv(down_df.head(15))}

    card_counts = {v: sum(1 for c in cards.values() if c.get("verdict") == v)
                   for v in (card.VERDICT_CANDIDATE, card.VERDICT_WATCH, card.VERDICT_SHOW)}
    return {
        "date": ds,
        "generated_at": generated_at,
        "plan_version": version.git_short_rev(),
        "counts": {"main": int(len(main)), "observe": int(len(obs)),
                   "gate_covered": int(len(gate))},
        "market_snapshot_days": sorted(mkt_days),
        "heat_date": hd,
        "theme_cap": theme_cap,
        "main": [_row(i, k, s, True)
                 for i, (k, s) in enumerate(_pick(main, top), 1)],
        "observe": [_row(i, k, s, False)
                    for i, (k, s) in enumerate(_pick(obs, obs_top), 1)],
        "changes": changes,
        "gate_on": bool(config.ENABLE_TRACK_GATE),
        "encoding": "主榜分=200+传导档位×20+组内分（还没热、还便宜）；"
                    "观察档分=100+0.6z(传导)+0.4z(−热度)",
        # ---- 候选卡（2026-09-02）：分数与档位之外的另一份产物，不受 top / theme_cap 裁剪 ----
        "card_params": {"start_pct": config.CARD_START_PCT,
                        "require_started": config.CARD_REQUIRE_STARTED,
                        "accum_max_age": config.CARD_ACCUM_MAX_AGE,
                        "neg_tol": config.UPSIDE_NEG_TOLERANCE,
                        "logic_per_stock": config.LOGIC_CLAIMS_PER_STOCK,
                        "rules": "候选=环节被指向∧当日涨幅达标∧券商覆盖且非ST∧明确吸筹∧无硬风险；"
                                 "关注=三门槛全过∧无硬风险∧确认线缺失或陈旧（系统无法判断，交人裁决）；"
                                 "其余仅展示（只差覆盖、潜在吸筹等非明确状态都不升格，台账 013）；"
                                 "因果论断只展示不进判决"},
        "card_counts": card_counts,
        "candidates": _by_verdict(card.VERDICT_CANDIDATE),
        "watch": _by_verdict(card.VERDICT_WATCH),
        "segments_pointed": segments_pointed,
        "_full": {"main": _full_rows(main, True), "observe": _full_rows(obs, False)},
    }


def _fmt_pct(v) -> str:
    return "—" if v is None else f"{v:+.0%}"


def _fmt_num(v) -> str:
    return "—" if v is None else f"{v:.2f}"


def _fmt_ev(e) -> str:
    if not e:
        return "—"
    if e.get("moved_ratio") is None:          # 已动成员视图补的证据行，没有已动比例
        return f"{e['theme']}（{e.get('n_sources') or '—'} 源，本票已启动）"
    return f"{e['theme']}（{e['n_sources']} 源，已动 {e['moved_ratio']:.0%}）"


def _fmt_pct0(v) -> str:
    return "—" if v is None else f"{v:+.1f}%"


def _fmt_accum(ac: dict) -> str:
    if not ac or not ac.get("state"):
        return "无评分"
    st = str(ac["state"]).split("·")[0]
    age = ac.get("age")
    return f"{st}（{age} 日前）" if isinstance(age, int) else st


def _fmt_logic(lines, width: int = 60) -> str:
    """候选单与关注单表格里的因果论断列：第一条截到 width 字，多于一条时带条数；表格里不能有竖线。"""
    lines = [str(x) for x in (lines or []) if x]
    if not lines:
        return "—"
    first = lines[0].replace("|", "／")
    if len(first) > width:
        first = first[:width] + "…"
    return f"{first}（共 {len(lines)} 条）" if len(lines) > 1 else first


def _fmt_yi(v) -> str:
    """已经是"亿元"的数直接显示。两市成交额由取数层换算好（原表单位是千元），不在这里换算。"""
    return "—" if v is None else f"{v:,.0f} 亿"


def _fmt_amount(v) -> str:
    """以元为单位的金额换算成亿显示。融资余额用它（那张表的单位是元）。"""
    if v is None:
        return "—"
    return f"{v / 1e8:,.0f} 亿" if abs(v) >= 1e8 else f"{v:,.0f}"


def _fmt_market(m: dict) -> str:
    """环境段市场四项的一行文字；缺的项写"—"，读数原因在 JSON 的 market.errors 里。"""
    t, b, mg, fg = m.get("turnover") or {}, m.get("breadth") or {}, m.get("margin") or {}, m.get("fear_greed") or {}
    parts = []
    if t:
        ratio = t.get("ratio_vs_prev5")
        parts.append(f"两市成交额 {_fmt_yi(t.get('amount_yi'))}"
                     + (f"（前五日均值的 {ratio:.2f} 倍）" if ratio else "")
                     + (f"，数据日 {t['data_date']}" if t.get("data_date") and t.get("data_date") != m.get("date") else ""))
    else:
        parts.append("两市成交额 —")
    if b:
        med = b.get("pct_median")
        parts.append(f"广度 上涨 {b.get('up')} / 下跌 {b.get('down')} 家，涨停近似 {b.get('limit_up_approx')} 家，"
                     f"涨幅中位数 {med:+.2f}%" if med is not None else
                     f"广度 上涨 {b.get('up')} / 下跌 {b.get('down')} 家")
    else:
        parts.append("广度 —")
    if mg:
        bal, chg = mg.get("financing_balance"), mg.get("change_percent_5d")
        parts.append(f"融资余额 {_fmt_amount(bal)}"
                     + (f"（五日变化 {chg:+.2f}%）" if chg is not None else "")
                     + (f"，{mg['date']}" if mg.get("date") else ""))
    else:
        parts.append("融资余额 —")
    if fg and fg.get("index_value") is not None:
        parts.append(f"恐贪指数 {fg['index_value']:.0f}" + (f"（{fg['date']}）" if fg.get("date") else ""))
    else:
        parts.append("恐贪指数 —")
    return "市场环境：" + "；".join(parts) + "（只展示与复盘分组，不作交易前置）。"


def render_md(d: dict) -> str:
    L = [f"# 每日选股计划 · {d['date']}", ""]
    c = d["counts"]
    L.append(f"主榜 {c['main']} 只 / 观察档 {c['observe']} 只 / "
             f"全池档位覆盖 {c['gate_covered']} 只。")
    stale = [x for x in d["market_snapshot_days"] if x != d["date"]]
    if stale:
        L.append(f"注：本日传导用的行情快照 = {'、'.join(stale)}"
                 f"（与计划日不同——历史降级日口径）。")
    L.append("")

    # ---- 候选单与关注环节（2026-09-02）：放在主榜之前，这是新的主产物 ----
    cc = d.get("card_counts") or {}
    cands = d.get("candidates") or []
    L.append(f"## 候选单（环节被指向、当日已启动、券商覆盖且非 ST、明确吸筹、无硬风险；"
             f"共 {cc.get('候选', len(cands))} 只，全量列出不受裁剪）")
    L.append("")
    if not cands:
        L.append("（今日无候选——候选为空不是故障：环节没被指向、成员没启动或没有明确吸筹，都会为空。）")
    else:
        L.append("| # | 代码 | 名称 | 环节 | 源数 | 当日涨幅 | 吸筹 | 预期空间 | 理由 | 因果论断（出处） |")
        L.append("|---|------|------|------|------|----------|------|----------|------|------------------|")
        for r in cands:
            c = r.get("card") or {}
            ac = c.get("accum") or {}
            ev_ = r.get("evidence") or {}
            L.append(f"| {r['rank']} | {r['code']} | {r['name'] or '—'} | {ev_.get('theme') or '—'} "
                     f"| {ev_.get('n_sources') or '—'} | {_fmt_pct0(c.get('pct0'))} "
                     f"| {_fmt_accum(ac)} | {_fmt_pct(r.get('upside'))} "
                     f"| {'；'.join(r.get('reasons') or [])} | {_fmt_logic(r.get('logic'))} |")
    L.append("")
    segs = d.get("segments_pointed") or []
    L.append(f"## 关注环节（今日被传导指向的 {len(segs)} 个环节：定位对不对看这里，挑票看候选单）")
    L.append("")
    if segs:
        L.append("| 环节 | 源数 | 链符 | 成员 | 已启动 | 领涨 | 候选 |")
        L.append("|------|------|------|------|--------|------|------|")
        for s in segs:
            ld = s.get("leader") or {}
            L.append(f"| {s['segment']} | {s.get('n_sources') or '—'} | {_fmt_num(s.get('chain_fit'))} "
                     f"| {s.get('members_total') if s.get('members_total') is not None else '—'} "
                     f"| {s.get('started_count', 0)} "
                     f"| {(ld.get('code') or '—') + (' ' + _fmt_pct0(ld.get('pct0')) if ld.get('code') else '')} "
                     f"| {s.get('candidates', 0)} |")
    L.append("")
    watch = d.get("watch") or []
    L.append(f"## 关注单（三门槛全过、无硬风险，但吸筹确认线缺失或陈旧——系统无法判断，交人裁决；"
             f"共 {cc.get('关注', len(watch))} 只，列前 20）")
    L.append("")
    if watch:
        L.append("| # | 代码 | 名称 | 环节 | 当日涨幅 | 吸筹 | 缺什么 | 因果论断（出处） |")
        L.append("|---|------|------|------|----------|------|--------|------------------|")
        for r in watch[:20]:
            c = r.get("card") or {}
            ev_ = r.get("evidence") or {}
            L.append(f"| {r['rank']} | {r['code']} | {r['name'] or '—'} | {ev_.get('theme') or '—'} "
                     f"| {_fmt_pct0(c.get('pct0'))} | {_fmt_accum(c.get('accum') or {})} "
                     f"| {'；'.join(r.get('missing') or [])} | {_fmt_logic(r.get('logic'))} |")
    L.append("")
    reg = d.get("regime")
    if reg:
        L.append(f"环境标签：{reg.get('status')}，弱势指数 {reg.get('weak_count')}/8"
                 f"{'，弱势日' if reg.get('weak_day') else ''}（只展示与复盘分组，不作交易前置）。")
        L.append("")
    mk = d.get("market")
    if mk:
        L.append(_fmt_market(mk))
        L.append("")

    cap_txt = f"，每主题限额 {d['theme_cap']}" if d["theme_cap"] else ""
    gate_txt = "、在十五五赛道内" if d.get("gate_on") else ""
    L.append(f"## 主榜 Top {len(d['main'])}"
             f"（有券商预期、目标价不低于现价{gate_txt}{cap_txt}）")
    L.append("")
    L.append("| # | 代码 | 名称 | 总分 | 档位 | 传导证据 | 热度 | 预期空间 |")
    L.append("|---|------|------|------|------|----------|------|----------|")
    for r in d["main"]:
        L.append(f"| {r['rank']} | {r['code']} | {r['name'] or '—'} | {r['score']:.1f} "
                 f"| {r['tier']} | {_fmt_ev(r['evidence'])} "
                 f"| {_fmt_num(r['heat'])} | {_fmt_pct(r['upside'])} |")
    L.append("")

    L.append(f"## 观察档 Top {len(d['observe'])}"
             f"（无券商预期、但{'在赛道或传导链上' if d.get('gate_on') else '在传导链上'}"
             f"——没有估值锚，置信度低{cap_txt}）")
    L.append("")
    L.append("| # | 代码 | 名称 | 分 | 传导证据 | 热度 |")
    L.append("|---|------|------|----|----------|------|")
    for r in d["observe"]:
        L.append(f"| {r['rank']} | {r['code']} | {r['name'] or '—'} | {r['score']:.1f} "
                 f"| {_fmt_ev(r['evidence'])} | {_fmt_num(r['heat'])} |")
    L.append("")

    L.append("## 今日升降档")
    L.append("")
    ch = d["changes"]
    if not ch:
        L.append("（没有更早的档位表可比，升降档从下一个交易日开始。）")
    else:
        L.append(f"对比 {ch['base_date']}：升档 {ch['upgrades_total']} 只，"
                 f"降档 {ch['downgrades_total']} 只。"
                 f"升档=拿到新锚（首次覆盖 / 新进传导链），本身就是值得看的信号。")
        if ch["upgrades"]:
            L.append("")
            L.append("**升档**：")
            L += [f"- {m['code']} {m['name'] or ''}：{m['from']} → {m['to']}"
                  f"（{m.get('reason', '—')}）"
                  for m in ch["upgrades"]]
            if ch["upgrades_total"] > len(ch["upgrades"]):
                L.append(f"- ……共 {ch['upgrades_total']} 只，其余见档位表")
        if ch["downgrades"]:
            L.append("")
            L.append("**降档**：")
            L += [f"- {m['code']} {m['name'] or ''}：{m['from']} → {m['to']}"
                  f"（{m.get('reason', '—')}）"
                  for m in ch["downgrades"]]
            if ch["downgrades_total"] > len(ch["downgrades"]):
                L.append(f"- ……共 {ch['downgrades_total']} 只，其余见档位表")
    L.append("")
    L.append("---")
    L.append(f"口径：{d['encoding']}。")
    return "\n".join(L)


def generate(date: str | None = None, top: int = 20, obs_top: int = 10,
             theme_cap: int = 5) -> str:
    try:
        data = collect(date, top, obs_top, theme_cap)
    except RuntimeError as e:
        raise SystemExit(str(e))
    # 环境段的市场四项（2026-09-03）：出计划时读一次落进快照，接口 /plan 只从快照读，盘中不再取数；
    # 每项读失败为空并把原因记在 market.errors，不阻断。区制段仍由 08:45 的追加步骤写入。
    data["market"] = sources.market_context(data["date"])
    text = render_md(data)
    os.makedirs(config.PLAN_SNAPSHOT_DIR, exist_ok=True)
    out = os.path.join(config.PLAN_SNAPSHOT_DIR, f"plan_{data['date']}.md")
    with open(out, "w", encoding="utf-8") as f:
        f.write(text + "\n")
    # ---- 当日 JSON 快照（2026-09-02 方案第 2.2 节第一项）：主榜与观察档全部行、全部证据线，
    # 不裁剪、不设主题限额；带生成时刻与代码版本。它是复盘与对账的唯一底本；
    # 08:45 的 regime-append 步骤会往里追加 regime 段，/plan 的 regime 段只读这份。----
    full = data.pop("_full", None) or {}
    snap = {**data, "main_shown": data["main"], "observe_shown": data["observe"],
            "main": full.get("main", []), "observe": full.get("observe", []),
            "shown_params": {"top": top, "obs_top": obs_top, "theme_cap": theme_cap}}
    jpath = os.path.join(config.PLAN_SNAPSHOT_DIR, f"plan_{data['date']}.json")
    tmp = jpath + ".tmp"
    with open(tmp, "w", encoding="utf-8") as f:
        json.dump(snap, f, ensure_ascii=False, indent=1, default=str)
    os.replace(tmp, jpath)
    print(text)
    print(f"\n已写入 {out} 与快照 {jpath}"
          f"（主榜 {len(snap['main'])} 行、观察档 {len(snap['observe'])} 行、"
          f"候选 {len(data.get('candidates') or [])} 只，版本 {data.get('plan_version')}）")
    return out