#!/usr/bin/env python3
"""VERIFY: Step5「确认性吸收」回读强化（First Delivery 纪律#5, learn_queue #1330, 2026-10-12）

学到的知识（原文 L8）：「模型说「成功」不算成功——结果要看任务结束后的实际状态，
核对必须回读业务系统。」映射到 Hermes：`_verify_and_apply` 的确认性吸收 grounding 分支
旧逻辑拿「target skill 目录存在」直接判 `confirmed` → PASS。

病理 WHERE=step5_verify 确认性吸收 grounding 分支
     WHY=存在性是最弱奖励（Harness-R1）；「显示成功（目录在）」≠「实际状态（知识真在里面）」。
     实测 12 条近期 applied 中 11 条 created_skill/updated_skill 双空。

修复：存在性 + 内容回读双证据；仅存在性命中 → 标 `existence_only` 观测量 + 降 PASS_LAX。

TDD：
TC1 (RED/复现旧病理) 旧逻辑对「目录存在但不含知识」的 target 也判 confirmed
TC2 (GREEN/真实数据) 本轮吸收 #1330 的真实 outline 对真实目标 SKILL.md → 内容回读命中
TC3 (GREEN/判别力) 拿一个存在但无关的 skill 当 target → 判 existence_only（非无条件放行）
TC4 (源码守卫/防回退) app.py 含双证据分支 + 弱证据降级谓词 + reason code 注册
TC5 (冻结重跑 spot check) 同一输入跑 3 次，结果一致（无随机漂移）
"""
import pathlib
import sqlite3
import sys

SKILLS = pathlib.Path("/data/data/com.termux/files/home/.dsh/skills")
APP = pathlib.Path("/data/data/com.termux/files/home/wechat-viewer/app.py")
DB = pathlib.Path("/data/data/com.termux/files/home/wechat-viewer/circle.db")
TARGET = "ai-project-first-delivery"      # 本轮吸收的 target（去 new:/update: 前缀后）
UNRELATED = "feishu-todo-sync"            # 存在但与本知识无关的 skill（负对照）
FAIL = 0


def old_readback(tgt_files, fingerprint, outline):
    """复现旧逻辑：只判存在性。"""
    if tgt_files:
        return "confirmed (exists on disk)"      # ← 弱证据被当成交付证据
    return None


def new_readback(tgt_files, fingerprint, outline):
    """复现新逻辑：存在性 + 内容回读双证据。"""
    if not tgt_files:
        return None
    try:
        txt = tgt_files[0].read_text()
        if fingerprint and fingerprint in txt:
            return "confirmed (exists + content readback)"
        for ph in [outline[i:i + 20] for i in range(0, min(80, len(outline or "")), 20)]:
            if len(ph.strip()) >= 10 and ph.strip() in txt:
                return "confirmed (exists + content readback)"
    except Exception:
        pass
    return f"confirmed (exists only; readback=existence_only, READBACK_EXISTENCE_ONLY)"


def is_lax(verified):
    """复现修复后的 PASS/PASS_LAX 谓词。"""
    return any(m in v for m in ("unconfirmed", "existence_only") for v in verified)


def find(tgt):
    return list(SKILLS.rglob(f"*/{tgt}/SKILL.md")) or list(SKILLS.rglob(f"{tgt}/SKILL.md"))


# ── 取本轮真实吸收数据（learning_results 最新 #1330 reference 行）──
con = sqlite3.connect(DB)
row = con.execute(
    "SELECT result_summary, outline FROM learning_results "
    "WHERE learn_queue_id=1330 ORDER BY id DESC LIMIT 1").fetchone()
REAL_SUMMARY, REAL_OUTLINE = (row[0] or "", row[1] or "")
REAL_FP = REAL_SUMMARY.strip()[:30]

print("=== TC1(RED/复现旧病理): 只判存在性 ===")
tgt_files = find(TARGET)
old_res = old_readback(tgt_files, REAL_FP, REAL_OUTLINE)
print(f"  旧逻辑结果: {old_res!r}")
print(f"  {'✅' if old_res and 'readback' not in old_res else '❌'} "
      f"旧逻辑不区分「目录在」与「知识在」→ 弱证据直接当 confirmed（病理复现）")
if not (old_res and "readback" not in old_res):
    FAIL += 1
print(f"  ✅ 前提事实：目标 skill 在磁盘上 {[str(p) for p in tgt_files]}")
if not tgt_files:
    FAIL += 1

print("\n=== TC2(GREEN/真实数据): 存在性 + 内容回读 ===")
print(f"  fingerprint(result_summary[:30])={REAL_FP!r}")
print(f"  {'✅' if REAL_FP not in tgt_files[0].read_text() else '⚠️'} "
      f"主指纹不命中（与 SKILL.md Troubleshooting 记载的 FINGERPRINT 病理一致，预期）")
new_res = new_readback(tgt_files, REAL_FP, REAL_OUTLINE)
print(f"  新逻辑结果: {new_res!r}")
if "content readback" not in (new_res or ""):
    print("  ❌ 真实 outline 未能回读命中目标 SKILL.md")
    FAIL += 1
else:
    print("  ✅ 真实 outline[0:20] 回读命中 → 拿到「实际状态」证据（非仅存在性）")
print(f"  {'✅' if not is_lax([new_res]) else '❌'} 强证据 → PASS（非 PASS_LAX）")
if is_lax([new_res]):
    FAIL += 1

print("\n=== TC3(GREEN/判别力): 存在但无关的 target 必须判弱证据 ===")
un_files = find(UNRELATED)
neg = new_readback(un_files, REAL_FP, REAL_OUTLINE)
print(f"  负对照 {UNRELATED}: {neg!r}")
print(f"  {'✅' if un_files else '❌'} 前提：该 skill 在磁盘上")
if not un_files:
    FAIL += 1
print(f"  {'✅' if 'existence_only' in (neg or '') else '❌'} 仅存在性 → 标 existence_only（非无条件放行）")
if "existence_only" not in (neg or ""):
    FAIL += 1
print(f"  {'✅' if is_lax([neg]) else '❌'} 弱证据 → PASS_LAX（Step6 会带 double-check 警告）")
if not is_lax([neg]):
    FAIL += 1

print("\n=== TC4(源码守卫/防回退) ===")
src = APP.read_text(encoding="utf-8")
checks = [
    ("确认性分支含内容回读 _rb_hit", "_rb_hit" in src),
    ("强证据文案 exists + content readback", "exists + content readback" in src),
    ("弱证据标记 existence_only", "readback=existence_only" in src),
    ("reason code 定义", "READBACK_EXISTENCE_ONLY = " in src),
    ("reason code 注册", "VerifyReason.READBACK_EXISTENCE_ONLY:" in src),
    ("弱证据降级谓词含 existence_only", '_WEAK_EVIDENCE_MARKERS = ("unconfirmed", "existence_only")' in src),
]
for label, ok in checks:
    print(f"  {'✅' if ok else '❌'} {label}")
    if not ok:
        FAIL += 1

print("\n=== TC5(冻结重跑 spot check): 同一输入 3 次采样 ===")
runs = [new_readback(tgt_files, REAL_FP, REAL_OUTLINE) for _ in range(3)]
stable = len(set(runs)) == 1
print(f"  3 次结果: {runs}")
print(f"  {'✅' if stable else '❌'} 3/3 一致（无随机漂移）→ 结论置信度：中（确定性纯函数，非模型调用）")
if not stable:
    FAIL += 1

con.close()
print(f"\n{'✅ ALL PASS' if FAIL == 0 else f'❌ {FAIL} FAILED'}")
sys.exit(0 if FAIL == 0 else 1)