#!/usr/bin/env python3
# -*- coding: utf-8 -*-
r"""数值验证工具（判定端）——自动化评判 + 元进度爬升测量，替代「亲手试玩再反馈」循环。

三段采集（数据源 = smoke/verifier_batch.gd，本文件驱动它并做全部判定）：
  level 段     直入口单关基线（零加成）——验收带判定
  adventure 段 全冒险连探基线（零永久强化）——止步分布 + 败退节点归因（层/节点/敌人）
  campaign 段  营地循环战役（隔离档真实元进度账本：注入→冒险→收束入账→买锻造）——爬升曲线

评判标准（溯源）：
  ① 直入口基线通关率带（分带批 2026-09-20 用户拍板，难度曲线与Boss v0.16）：
     第一层 普通 [35,55]% · 第二层 普通 [25,45]% · 完美档 [35,65]%（两关同带）；第三层带未定——只测不判。
  ② 元循环目标（用户提案 2026-09-22，数字待确认）：累计通过 5 次第 k 层的冒险后，第 k+1 层的
     条件通关率（滑窗）应进入该层验收带——「5 次通过第一层的冒险后能过第二层；5 次通过第二层后能过第三层」。
     第三层带未定，只给爬升曲线与首次读数，不出判定。
  ③ 魔法道具面：道具工坊六件按通关次数解锁（阈值 1/1/2/2/3/3），局内效果未实装（归 Alpha 魔法道具批）——
     本工具只报解锁时点，不入判定。

表值硬编码溯源（单一事实源仍在 .tres / GDD，变更须同步）：
  · 锻造三档 flat（meta.tres）：生命 +3 / 攻击 +1 / 护甲 +1
  · 钥石入账（adventure.tres + meta.tres）：每层通关 10 + Boss 首杀/重复杀 5/1 · 8/2 · 12/3
  · 工坊阈值（meta.tres）：显影尘1 · 工兵令1 · 探宝罗盘2 · 探巢摆钟2 · 地穴全图卷3 · 爆破工兵包3

用法：
  python tools/balance_verifier.py                 # 全段默认规模（约 20–35 分钟）
  python tools/balance_verifier.py --quick         # 小规模快跑（约 5–10 分钟，读数精度低）
  python tools/balance_verifier.py --n 100 --na 60 --k 60 --seed0 9000 --spend atk_first
  python tools/balance_verifier.py --from-raw <raw.jsonl>   # 免重跑，对既有采集重新判定/出报告
输出：控制台判定结论 + production/qa/balance-verifier-<时间戳>.md（报告）+ 同名 .raw.jsonl（原始逐局数据）。
"""
import argparse
import json
import math
import subprocess
import sys
import time
from pathlib import Path

ROOT = Path(__file__).resolve().parent.parent
SLICE = ROOT / "prototypes" / "dungeon-sweeper-vertical-slice"
GODOT = r"D:\code\godot\Godot_v4.7.2-stable_win64_console.exe"
SCRIPT = "res://smoke/verifier_batch.gd"
QA_DIR = ROOT / "production" / "qa"
VJSON_PREFIX = "VJSON "
ERROR_MARKERS = ("SCRIPT ERROR", "Parse Error")

# ---- 判定标准区（溯源见文件头） ----
LEVEL_BANDS = {
    ("normal", 1): (35, 55),
    ("normal", 2): (25, 45),
    ("perfect", 1): (35, 65),
    ("perfect", 2): (35, 65),
}
CAMPAIGN_CLEARS_TARGET = 5    # 元循环检查点：累计通过第 k 层的冒险次数（用户提案 2026-09-22，待确认）
CAMPAIGN_WINDOW = 10          # 检查点滑窗局数（条件通关率的窗口）
CAMPAIGN_MIN_ATTEMPTS = 5     # 滑窗内少于此局数 = 数据不足（不出判定）
LEVEL_CN = {1: "第一层", 2: "第二层", 3: "第三层"}
NODE_TYPE_CN = {"battle": "战斗", "boss": "Boss", "rest": "恢复", "reward": "奖励", "shop": "商店房"}
FORGE_FLAT = {"max_hp": 3, "atk": 1, "armor_class": 1}   # meta.tres 锻造三档每级 flat 值
WORKSHOP_CN = {"probe_dust": "显影尘", "sapper_order": "工兵令", "loot_compass": "探宝罗盘",
               "nest_pendulum": "探巢摆钟", "catacomb_scroll": "地穴全图卷", "blast_kit": "爆破工兵包"}


def die(msg):
    print(f"[数值验证] FAIL：{msg}")
    sys.exit(1)


def disp_w(s):
    return sum(2 if ord(c) > 0x2E7F else 1 for c in str(s))


def pad(s, w, right=False):
    s = str(s)
    gap = w - disp_w(s)
    return (" " * gap + s) if right else (s + " " * gap)


# ---------------------------------------------------------------- 采集

def collect(args):
    segs = args.only.split(",")
    cmd = [args.godot, "--headless", "--path", str(SLICE), "--script", SCRIPT, "--",
           f"n={args.n}", f"na={args.na}", f"k={args.k}", f"seed0={args.seed0}",
           f"only={args.only}", f"spend={args.spend}"]
    print(f"[数值验证] 采集启动（{' · '.join(segs)}，约 {args.est_min} 分钟）…")
    t0 = time.time()
    try:
        proc = subprocess.run(cmd, capture_output=True, text=True, encoding="utf-8",
                              errors="replace", timeout=args.timeout)
    except subprocess.TimeoutExpired:
        die(f"外部超时（>{args.timeout}s）——可加大 --timeout 或缩小规模")
    out = (proc.stdout or "") + (proc.stderr or "")
    if proc.returncode != 0:
        print(out[-3000:])
        die(f"采集进程退出码 {proc.returncode}")
    hits = [ln.strip() for ln in out.splitlines() if any(k in ln for k in ERROR_MARKERS)]
    if hits:
        print("\n".join(hits[:5]))
        die(f"脚本错误标记命中 {len(hits)} 行（脚本错误静默中断断言/采集，必须人工排查）")
    rows, bad = [], 0
    for ln in out.splitlines():
        if ln.startswith(VJSON_PREFIX):
            try:
                rows.append(json.loads(ln[len(VJSON_PREFIX):]))
            except json.JSONDecodeError:
                bad += 1
    if bad:
        die(f"VJSON 行解析失败 {bad} 行（采集端输出契约被破坏）")
    if not rows:
        die("采集零数据行（verifier_batch.gd 未按契约输出）")
    print(f"[数值验证] 采集完成：{len(rows)} 局，耗时 {(time.time() - t0) / 60:.1f} 分钟")
    return rows, out


def load_raw(path):
    rows = []
    with open(path, encoding="utf-8") as f:
        for ln in f:
            ln = ln.strip()
            if ln:
                rows.append(json.loads(ln))
    return rows


# ---------------------------------------------------------------- 判定件

def binom_ci(wins, n):
    if n == 0:
        return 0.0
    p = wins / n
    return 1.96 * math.sqrt(p * (1 - p) / n)


def judge_band(rate_pct, band):
    """落带判定：返回 (结论, 是否达标)。band=None = 带未定，只报读数。"""
    if band is None:
        return "带未定（只测不判）", None
    lo, hi = band
    if rate_pct < lo:
        return f"越下界（低于 {lo}% 达 {lo - rate_pct:.1f} 点）", False
    if rate_pct > hi:
        return f"越上界（高于 {hi}% 达 {rate_pct - hi:.1f} 点）", False
    return f"落带 [{lo},{hi}] ✓", True


def node_label(defeat):
    """败退节点人读名：第X层·节点N·类型·敌人（敌人 = 死亡遭遇显示名，最多 3 个）。"""
    if not defeat:
        return "—"
    t = NODE_TYPE_CN.get(defeat.get("node_type", ""), defeat.get("node_type", "?"))
    s = f"{LEVEL_CN.get(defeat.get('level_index', -1), '?层')}·节点{defeat.get('node_id', '?')}·{t}"
    enemies = defeat.get("enemies") or []
    if enemies:
        seen, names = set(), []
        for e in enemies:
            if e not in seen:
                seen.add(e)
                names.append(e)
        s += "·" + "、".join(names[:3])
    return s


# ---------------------------------------------------------------- 分段分析

def analyze_level(rows):
    """直入口基线：按（画像, 层）聚合计数。"""
    agg = {}
    for r in rows:
        if r.get("seg") != "level":
            continue
        key = (r["policy"], int(r["level"]))
        a = agg.setdefault(key, {"n": 0, "wins": 0, "deaths": {}})
        a["n"] += 1
        if r.get("result") == "victory":
            a["wins"] += 1
        elif r.get("result") == "defeat":
            k = (int(r.get("defeat_node_id", -1)), r.get("defeat_node_type", ""))
            a["deaths"][k] = a["deaths"].get(k, 0) + 1
    return agg


def analyze_adventure(rows):
    out = {}
    for policy in ("normal", "perfect"):
        runs = [r for r in rows if r.get("seg") == "adventure" and r.get("policy") == policy]
        stopped, death_nodes, l1 = {}, {}, {"n": 0, "wins": 0}
        for r in runs:
            sa = int(r.get("stopped_at", -1))
            if sa == 0:
                stopped["通关"] = stopped.get("通关", 0) + 1
            elif sa > 0:
                stopped[f"止步第{sa}层"] = stopped.get(f"止步第{sa}层", 0) + 1
            d = r.get("defeat") or {}
            if d:
                lb = node_label(d)
                death_nodes[lb] = death_nodes.get(lb, 0) + 1
            if int(r.get("levels_cleared", -1)) >= 0:
                l1["n"] += 1
                if int(r.get("levels_cleared", 0)) >= 1:
                    l1["wins"] += 1
        out[policy] = {"n": len(runs), "stopped": stopped, "death_nodes": death_nodes, "l1": l1}
    return out


def campaign_series(runs):
    """逐局爬升序列：条件通关尝试流（第 k 层的尝试 = levels_cleared ≥ k-1 的局，按局序）。"""
    return {
        k: [(r["run_no"], int(r.get("levels_cleared", -1)) >= k, r)
            for r in sorted(runs, key=lambda x: x["run_no"])
            if int(r.get("levels_cleared", -1)) >= k - 1]
        for k in (1, 2, 3)
    }


def window_rate(series, end_idx, window):
    """尝试流截至 end_idx（含）往前 window 次尝试的通关率（%）。"""
    chunk = series[max(0, end_idx - window + 1): end_idx + 1]
    if not chunk:
        return None, 0
    return 100.0 * sum(1 for _, ok, _ in chunk if ok) / len(chunk), len(chunk)


def analyze_campaign(rows):
    out = {}
    for policy in ("normal", "perfect"):
        runs = [r for r in rows if r.get("seg") == "campaign" and r.get("policy") == policy
                and not r.get("anomaly")]
        anomalies = [r for r in rows if r.get("seg") == "campaign" and r.get("policy") == policy
                     and r.get("anomaly")]
        res = {"n": len(runs), "anomalies": len(anomalies), "runs": runs}
        if not runs:
            out[policy] = res
            continue
        series = campaign_series(runs)
        res["series"] = series
        # 检查点判定：累计通过第 k 层达目标后，第 k+1 层滑窗条件通关率 vs 验收带
        checkpoints = []
        for k in (1, 2):
            clears = 0
            trigger_run = None
            for r in sorted(runs, key=lambda x: x["run_no"]):
                if int(r.get("levels_cleared", 0)) >= k:
                    clears += 1
                    if clears >= CAMPAIGN_CLEARS_TARGET and trigger_run is None:
                        trigger_run = r["run_no"]
            cp = {"level": k, "trigger_run": trigger_run, "clears": clears}
            if trigger_run is None:
                cp["verdict"] = f"数据不足：全程通过{LEVEL_CN[k]}仅 {clears} 次（目标 {CAMPAIGN_CLEARS_TARGET}）——加大 --k 重跑"
            else:
                attempts = [(rn, ok, r) for rn, ok, r in series.get(k + 1, []) if rn > trigger_run]
                chunk = attempts[:CAMPAIGN_WINDOW]
                if len(chunk) < CAMPAIGN_MIN_ATTEMPTS:
                    cp["verdict"] = (f"触发于第 {trigger_run} 局，但其后通过{LEVEL_CN[k + 1]}的尝试仅 "
                                     f"{len(chunk)} 局（不足 {CAMPAIGN_MIN_ATTEMPTS}）——加大 --k 重跑")
                else:
                    rate = 100.0 * sum(1 for _, ok, _ in chunk if ok) / len(chunk)
                    cp["rate"] = rate
                    cp["attempts"] = len(chunk)
                    band = LEVEL_BANDS.get((policy, k + 1))
                    cp["verdict"], cp["pass"] = judge_band(rate, band)
                    cp["forge_at"] = chunk[0][2].get("override", {})
            checkpoints.append(cp)
        res["checkpoints"] = checkpoints
        # 工坊解锁时点（魔法道具面——只报不判）
        unlocks = {}
        for r in runs:
            for iid in r.get("new_unlocks", []):
                unlocks[WORKSHOP_CN.get(iid, iid)] = r["run_no"]
        res["unlocks"] = unlocks
        # 累计钥石与锻造终态
        last = max(runs, key=lambda x: x["run_no"])
        res["forge_final"] = last.get("forge", {})
        res["keystone_final"] = int(last.get("keystone_after", -1))
        res["income_total"] = sum(int(r.get("keystone_income", 0)) for r in runs)
        out[policy] = res
    return out


# ---------------------------------------------------------------- 报告

def fmt_table(header, rows):
    widths = [max(disp_w(h), max((disp_w(r[i]) for r in rows), default=0)) for i, h in enumerate(header)]
    def line(cells):
        return "| " + " | ".join(pad(c, w) for c, w in zip(cells, widths)) + " |"
    sep = "|" + "|".join("-" * (w + 2) for w in widths) + "|"
    out = [line(header), sep]
    out += [line(r) for r in rows]
    return "\n".join(out)


def report(args, rows, meta):
    lines = []
    A = lines.append
    A(f"# 数值验证报告（{meta['ts']}）")
    A("")
    A(f"- 采集规模：直入口每格 {args.n} 局 · 基线冒险 {args.na} 局/画像 · 战役 {args.k} 局/画像"
      f" · 种子 {args.seed0} 起逐局独立 · 锻造花费策略 = {args.spend}")
    A(f"- 切片版本：{meta['commit']} · 表状态：判定带与表值溯源见 tools/balance_verifier.py 文件头")
    A("")

    # ① 直入口基线
    A("## ① 直入口基线通关率（验收带判定）")
    A("")
    lvl = analyze_level(rows)
    tbl = []
    for (policy, li), a in sorted(lvl.items()):
        rate = 100.0 * a["wins"] / max(1, a["n"])
        ci = 1.96 * math.sqrt((rate / 100) * (1 - rate / 100) / max(1, a["n"])) * 100
        band = LEVEL_BANDS.get((policy, li))
        verdict, _ = judge_band(rate, band)
        tbl.append([f"{LEVEL_CN[li]}", "普通档" if policy == "normal" else "完美档",
                    f"{a['wins']}/{a['n']}", f"{rate:.1f}% ±{ci:.1f}",
                    f"[{band[0]},{band[1]}]" if band else "未定", verdict])
    A(fmt_table(["层", "画像", "通关/局数", "通关率±95%CI", "验收带", "判定"], tbl))
    A("")

    # ② 冒险基线：止步分布与败退节点
    A("## ② 零永久强化冒险：止步分布与败退节点（用户判据③）")
    A("")
    adv = analyze_adventure(rows)
    for policy in ("normal", "perfect"):
        d = adv[policy]
        cn = "普通档" if policy == "normal" else "完美档"
        A(f"### {cn}（{d['n']} 局）")
        A("")
        stop_line = "、".join(f"{k} {v}（{100.0 * v / max(1, d['n']):.0f}%）" for k, v in sorted(d["stopped"].items()))
        A(f"- 止步分布：{stop_line or '—'}")
        l1r = 100.0 * d["l1"]["wins"] / max(1, d["l1"]["n"])
        A(f"- 第一层通过率（冒险口径交叉核对）：{l1r:.1f}%")
        A("- 败退节点分布（失败局归因）：")
        for lb, c in sorted(d["death_nodes"].items(), key=lambda x: -x[1]):
            A(f"  - {lb} ×{c}")
        A("")

    # ③ 元循环爬升
    A("## ③ 营地循环爬升（用户判据④：5 次通关 → 下一层可过）")
    A("")
    camp = analyze_campaign(rows)
    for policy in ("normal", "perfect"):
        d = camp[policy]
        cn = "普通档" if policy == "normal" else "完美档"
        A(f"### {cn}（{d['n']} 局战役）")
        A("")
        if not d["runs"]:
            A("- 无数据")
            A("")
            continue
        A(f"- 钥石：全程入账 {d['income_total']} · 终态余额 {d['keystone_final']}"
          f" · 锻造终态 = 生命+{d['forge_final'].get('max_hp', 0) * FORGE_FLAT['max_hp']}"
          f"/攻击+{d['forge_final'].get('atk', 0) * FORGE_FLAT['atk']}"
          f"/护甲+{d['forge_final'].get('armor_class', 0) * FORGE_FLAT['armor_class']}")
        if d["unlocks"]:
            A("- 道具工坊解锁时点（魔法道具面，局内效果未实装只报不判）：" +
              "、".join(f"{k} 第 {v} 局" for k, v in sorted(d["unlocks"].items(), key=lambda x: x[1])))
        else:
            A("- 道具工坊解锁时点：无（未达阈值）")
        if d.get("anomalies"):
            A(f"- ⚠ 编排异常局 {d['anomalies']} 局（未结算零入账——timeout/advance_failed，须排查）")
        A("")
        for cp in d.get("checkpoints", []):
            A(f"- 判据：累计通过{LEVEL_CN[cp['level']]} {CAMPAIGN_CLEARS_TARGET} 次后，{LEVEL_CN[cp['level'] + 1]}"
              f"条件通关率（滑窗 {CAMPAIGN_WINDOW} 局）——{cp['verdict']}"
              + (f"（触发时注入 = 生命+{cp['forge_at'].get('max_hp', 0)}/攻击+{cp['forge_at'].get('atk', 0)}"
                 f"/护甲+{cp['forge_at'].get('armor_class', 0)}）" if "forge_at" in cp else ""))
        A("")
        tbl = []
        series = d.get("series", {})
        runs_by_no = {r["run_no"]: r for r in d["runs"]}
        for r in d["runs"][::max(1, len(d["runs"]) // 20)]:
            cells = [f"第{r['run_no']}局",
                     f"血{r['override'].get('max_hp', 0)}/攻{r['override'].get('atk', 0)}/甲{r['override'].get('armor_class', 0)}",
                     str(r.get("keystone_after", "-"))]
            for k in (1, 2, 3):
                s = series.get(k, [])
                idx = next((i for i, (rn, _, _) in enumerate(s) if rn == r["run_no"]), None)
                if idx is None:
                    cells.append("—")
                else:
                    rate, cnt = window_rate(s, idx, CAMPAIGN_WINDOW)
                    cells.append(f"{rate:.0f}%(n={cnt})" if rate is not None else "—")
            sa = int(r.get("stopped_at", -1))
            cells.append("通关" if sa == 0 else (f"止步第{sa}层" if sa > 0 else str(r.get("result", "?"))))
            tbl.append(cells)
        A(fmt_table(["局", "注入(flat)", "钥石", "第1层率(窗)", "第2层率(窗)", "第3层率(窗)", "终局"], tbl))
        A("")
    return "\n".join(lines)


def console_verdict(args, rows):
    """控制台判定结论（先读数后细节）。"""
    lvl = analyze_level(rows)
    print("\n判定结论（detail 见报告文件）")
    for (policy, li), a in sorted(lvl.items()):
        rate = 100.0 * a["wins"] / max(1, a["n"])
        band = LEVEL_BANDS.get((policy, li))
        verdict, ok = judge_band(rate, band)
        mark = {True: "✓", False: "✗", None: "·"}[ok]
        print(f"  [{mark}] 直入口{LEVEL_CN[li]}{'普通档' if policy == 'normal' else '完美档'}："
              f"{rate:.1f}%（{a['wins']}/{a['n']}）→ {verdict}")
    adv = analyze_adventure(rows)
    for policy in ("normal", "perfect"):
        d = adv[policy]
        cn = "普通档" if policy == "normal" else "完美档"
        top = sorted(d["death_nodes"].items(), key=lambda x: -x[1])[:2]
        print(f"  [·] 零强化冒险·{cn}：止步分布 {'、'.join(f'{k}{v}' for k, v in sorted(d['stopped'].items()))}"
              + (f"；主要败退点 = {'；'.join(f'{lb} ×{c}' for lb, c in top)}" if top else ""))
    camp = analyze_campaign(rows)
    for policy in ("normal", "perfect"):
        d = camp[policy]
        cn = "普通档" if policy == "normal" else "完美档"
        if not d["runs"]:
            print(f"  [·] 营地循环·{cn}：无数据")
            continue
        for cp in d.get("checkpoints", []):
            mark = {True: "✓", False: "✗", None: "·"}.get(cp.get("pass"), "·")
            print(f"  [{mark}] 营地循环·{cn}·{LEVEL_CN[cp['level']]}→{LEVEL_CN[cp['level'] + 1]}：{cp['verdict']}")
    print()


# ---------------------------------------------------------------- 入口

def main():
    if hasattr(sys.stdout, "reconfigure"):
        sys.stdout.reconfigure(encoding="utf-8", errors="replace")
    ap = argparse.ArgumentParser(description="数值验证工具（判定端）")
    ap.add_argument("--godot", default=GODOT)
    ap.add_argument("--n", type=int, default=100, help="直入口每格局数（±95%CI 约 ±9 点——带边判读须看 CI）")
    ap.add_argument("--na", type=int, default=60, help="基线冒险局数/画像")
    ap.add_argument("--k", type=int, default=60, help="战役局数/画像")
    ap.add_argument("--seed0", type=int, default=8000)
    ap.add_argument("--only", default="level,adventure,campaign")
    ap.add_argument("--spend", default="hp_first", choices=["hp_first", "atk_first"])
    ap.add_argument("--timeout", type=int, default=5400)
    ap.add_argument("--quick", action="store_true", help="小规模快跑（12/8/10）")
    ap.add_argument("--from-raw", default=None, help="跳过采集，对既有 raw.jsonl 重新判定出报告")
    args = ap.parse_args()
    if args.quick:
        args.n, args.na, args.k = 12, 8, 10
    args.est_min = (3 * 2 * args.n * 1.5 + 2 * args.na * 3.5 + 2 * args.k * 3.5) / 60

    ts = time.strftime("%Y%m%d-%H%M%S")
    if args.from_raw:
        rows = load_raw(args.from_raw)
        raw_path = Path(args.from_raw)
        print(f"[数值验证] 已载入既有采集 {len(rows)} 局：{raw_path}")
    else:
        rows, _ = collect(args)
        QA_DIR.mkdir(parents=True, exist_ok=True)
        raw_path = QA_DIR / f"balance-verifier-{ts}.raw.jsonl"
        with open(raw_path, "w", encoding="utf-8") as f:
            for r in rows:
                f.write(json.dumps(r, ensure_ascii=False) + "\n")

    commit = "?"
    try:
        commit = subprocess.run(["git", "-C", str(SLICE), "rev-parse", "--short", "HEAD"],
                                capture_output=True, text=True).stdout.strip()
    except OSError:
        pass
    meta = {"ts": ts, "commit": commit}

    console_verdict(args, rows)
    md = report(args, rows, meta)
    if args.from_raw:
        md_path = raw_path.with_suffix(".md")
    else:
        md_path = QA_DIR / f"balance-verifier-{ts}.md"
    md_path.write_text(md, encoding="utf-8")
    print(f"[数值验证] 报告：{md_path}")
    print(f"[数值验证] 原始逐局数据：{raw_path}（--from-raw 可免重跑重判）")
    return 0


if __name__ == "__main__":
    sys.exit(main())