feat(drill): 缺陷样本库+选样引擎（W4-C4 .github#223，ADR-0069）

governance/drill/drill.py

@@ -0,0 +1,182 @@
+#!/usr/bin/env python3
+"""drill.py —— 周种子缺陷演习引擎·读面（宪法 §4B/§6 / ADR-0069 / .github#223 W4-C4）
+
+子命令（全部 fail-closed：任何失败非零退出，绝不静默降级为"通过"）:
+  select   随机选缺陷样本 + 随机目标仓（--seed 注入可复现——测试与事后复盘）
+  decode   owner 审阅用：解码样本缺陷内容（ADR-0069 决策 1 样本库 owner 直管）
+
+注入/独立验证/台账（inject/record/redrate）随后续 PR 落地——红绿判定与样本库
+分离（verify_gate.py，ADR-0069 决策 2"注入者与判定者分离"）。注入物只落在
+演习分支（隔离执行，不进 agent 工作区，ADR-0069 风险缓解）；分支验后即删。
+"""
+import argparse
+import base64  # decode 用（样本 defect_b64）
+import json
+import os
+import random
+import re
+import subprocess
+import sys
+import tempfile
+from datetime import datetime, timezone
+
+try:
+    import yaml
+except ImportError:  # pragma: no cover
+    print("FATAL 缺少 PyYAML（CI 预装；本地 pip install pyyaml）", file=sys.stderr)
+    raise SystemExit(2)
+
+ORG = os.environ.get("DRILL_ORG", "Cloudbird-Software")
+OWNER = "randypanding"  # 样本库唯一审批人（org owner，ADR-0069 决策 1）
+DIFFICULTIES = ("easy", "medium", "hard")
+SCOPES = ("org", "github")
+EXCLUDED_TARGETS = ("holdout",)  # owner 直管封存面（ADR-0056 隔离不变量）——演习分支不进
+ID_RE = re.compile(r"^[a-z0-9][a-z0-9-]*$")
+
+
+def die(msg, code=2):
+    print(f"::error::{msg}", file=sys.stderr)
+    raise SystemExit(code)
+
+
+def validate_samples(doc):
+    """样本库 schema 校验（AC-2：owner 审批记录 + 预期关卡 ID 必须逐条在）。
+
+    返回错误列表（空=合法）。tests/test-samples.sh 与 select/inject 共用本函数——
+    校验逻辑单一实现，防"测试过而引擎放行"的分叉。
+    """
+    errs = []
+    if not isinstance(doc, dict) or not isinstance(doc.get("samples"), list) or not doc["samples"]:
+        return ["顶层缺非空 samples 列表"]
+    seen = set()
+    for i, s in enumerate(doc["samples"]):
+        w = f"samples[{i}]"
+        if not isinstance(s, dict):
+            errs.append(f"{w}: 非对象"); continue
+        sid = s.get("id", "")
+        if not ID_RE.match(str(sid)):
+            errs.append(f"{w}: id 非法: {sid!r}")
+        if sid in seen:
+            errs.append(f"{w}: id 重复: {sid}")
+        seen.add(sid)
+        if s.get("difficulty") not in DIFFICULTIES:
+            errs.append(f"{sid}: difficulty 须为 {DIFFICULTIES} 之一")
+        if not str(s.get("gate", "")).strip():
+            errs.append(f"{sid}: 缺预期触发关卡 ID（gate）")
+        if s.get("scope") not in SCOPES:
+            errs.append(f"{sid}: scope 须为 {SCOPES} 之一")
+        kind = s.get("payload_kind")
+        if kind == "file":
+            b64 = str(s.get("defect_b64", "")).strip()
+            if not b64:
+                errs.append(f"{sid}: file 样本缺 defect_b64")
+            else:
+                try:
+                    base64.b64decode(b64, validate=True)
+                except Exception as e:
+                    errs.append(f"{sid}: defect_b64 非法 base64: {e}")
+        elif kind == "generated":
+            size = s.get("size_bytes")
+            if not isinstance(size, int) or not 1 <= size <= 20971520:
+                errs.append(f"{sid}: generated 样本 size_bytes 须为 1..20MB 整数")
+        else:
+            errs.append(f"{sid}: payload_kind 须为 file|generated")
+        path = str(s.get("payload_path", ""))
+        if not path or "{DATE}" not in path:
+            errs.append(f"{sid}: payload_path 须含 {{DATE}} 占位: {path!r}")
+        ap = s.get("approval")
+        if not isinstance(ap, dict) or ap.get("approved_by") != OWNER:
+            errs.append(f"{sid}: 缺 owner（{OWNER}）审批记录 approval.approved_by")
+        elif not re.match(r"^\d{4}-\d{2}-\d{2}$", str(ap.get("date", ""))):
+            errs.append(f"{sid}: approval.date 非 ISO 日期")
+        if not isinstance(s.get("pr_title_adr", False), bool):
+            errs.append(f"{sid}: pr_title_adr 须为布尔")
+    return errs
+
+
+def load_samples(path):
+    try:
+        doc = yaml.safe_load(open(path, encoding="utf-8"))
+    except Exception as e:
+        die(f"样本库 YAML 解析失败: {e}")
+    errs = validate_samples(doc)
+    if errs:
+        for e in errs:
+            print(f"::error::样本库校验失败: {e}", file=sys.stderr)
+        raise SystemExit(2)
+    return doc["samples"]
+
+
+def load_targets(repos_path, scope):
+    """目标池: REPOS.yaml active 仓 − holdout（隔离面）；scope=github 只打治理总仓。"""
+    if scope == "github":
+        return [".github"]
+    doc = yaml.safe_load(open(repos_path, encoding="utf-8"))
+    names = [r["name"] for r in doc["repos"]
+             if r.get("status") == "active" and r["name"] not in EXCLUDED_TARGETS]
+    if not names:
+        die("REPOS.yaml 无可用 active 目标仓")
+    return names
+
+
+def cmd_select(a):
+    samples = load_samples(a.samples)
+    pool = []
+    for s in samples:
+        if a.sample_id and s["id"] != a.sample_id:
+            continue  # 复盘/首演：指定样本仍走同一随机框架（仅固定样本维度）
+        targets = load_targets(a.repos, s["scope"])
+        if a.target_repo:
+            targets = [t for t in targets if t == a.target_repo]
+        if targets:
+            pool.append((s, targets))
+    if not pool:
+        die("无可执行样本（样本/目标池为空或 --sample-id/--target-repo 过滤后为空）")
+    s, targets = pool[a.rng.randrange(len(pool))]
+    target = targets[a.rng.randrange(len(targets))]
+    print(json.dumps({"seed": a.seed, "sample_id": s["id"], "difficulty": s["difficulty"],
+                      "gate": s["gate"], "scope": s["scope"], "target_repo": target},
+                     ensure_ascii=False))
+
+
+def cmd_decode(a):
+    samples = {s["id"]: s for s in load_samples(a.samples)}
+    if a.id not in samples:
+        die(f"样本不存在: {a.id}")
+    s = samples[a.id]
+    if s["payload_kind"] != "file":
+        print(f"#（生成物样本，无静态内容: size_bytes={s['size_bytes']}）")
+        return
+    sys.stdout.write(base64.b64decode(s["defect_b64"]).decode("utf-8"))
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    sub = ap.add_subparsers(dest="cmd", required=True)
+    here = os.path.dirname(os.path.abspath(__file__))
+
+    p = sub.add_parser("select", help="随机选样+目标（seed 可注入）")
+    p.add_argument("--samples", default=os.path.join(here, "samples", "registry.yaml"))
+    p.add_argument("--repos", default=os.path.join(here, "..", "REPOS.yaml"))
+    p.add_argument("--seed", type=int, default=None, help="缺省=run_id+日期能推导的种子")
+    p.add_argument("--sample-id", help="跳过随机，指定样本（复盘用）")
+    p.add_argument("--target-repo", help="跳过随机，指定目标仓（复盘/首演用）")
+    p.set_defaults(func=cmd_select)
+
+    p = sub.add_parser("decode", help="owner 审阅：解码样本内容")
+    p.add_argument("--samples", default=os.path.join(here, "samples", "registry.yaml"))
+    p.add_argument("--id", required=True)
+    p.set_defaults(func=cmd_decode)
+
+
+    a = ap.parse_args()
+    if a.cmd == "select":
+        if a.seed is None:
+            a.seed = int(os.environ.get("GITHUB_RUN_ID", "0") or 0) + int(
+                datetime.now(timezone.utc).strftime("%Y%m%d"))
+        a.rng = random.Random(a.seed)
+    a.func(a)
+
+
+if __name__ == "__main__":
+    main()

governance/drill/history.jsonl

@@ -0,0 +1,2 @@
+{"kind": "failclose-drill", "mode": "real", "note": "AUTO_MERGE_DISABLED 真置位→读回true→立即复位→读回false（窗口约3s）；#180 先例回归；复位时戳入档", "outcome": "pass", "reset_at": "2026-08-21T19:12:14Z", "run_id": "first-drill-20260821", "set_at": "2026-08-21T19:12:11Z", "ts": "2026-08-21T19:12:15Z"}
+{"branch": "drill/seed-20260822", "canary_link": "healthy(sweep_run=32516378688,registry=b07475ff,hits_drill=1,hits_real=0)", "checks": {"gate": "failure", "hygiene / hygiene": "failure", "org-adr-required": "success", "org-gate": "failure", "org-hygiene / hygiene": "failure"}, "difficulty": "easy", "gate": "org-hygiene", "head_sha": "78ee2201fc3f27e5841594e15634289a2185887d", "kind": "seed-drill", "limitation": "push 分支无 CI 触发面（org gate 仅 pull_request+push:main）——NO-SURFACE 实测后按 ADR-0069 原始形态降级 draft PR 面验证", "pr": 239, "run_id": "first-drill-20260821", "sample_id": "hygiene-gitleaks-aws-key", "seed": 20260822, "surface": "draft-pr", "target_repo": ".github", "ts": "2026-08-21T19:26:00Z", "verdict": "red"}

governance/drill/samples/registry.yaml

@@ -0,0 +1,98 @@
+version: 1
+# ============================================================================
+# 周种子缺陷演习——缺陷样本库（宪法 §4B / ADR-0069 / .github#223 W4-C4）
+#
+# 用途: 每周一 seed-drill workflow 从本库随机取样，注入到随机目标仓的演习
+#       分支/draft PR，断言"该缺陷应触发关卡 X 变红"——红=演习成功（关卡活着），
+#       绿=演习失败（关卡死了）→ P0。静态检查证明不了关卡活着，只有活体红能。
+#
+# 安全注意（ADR-0069 决策 1/风险缓解）:
+#   - 缺陷内容以 base64 存储（defect_b64）：本仓全公开（ADR-0020）不存在 agent
+#     不可读路径，编码仅为降低样本特征被 agent 上下文随手吸收的概率（等效缓解，
+#     与"owner 逐条审批 + 隔离分支执行 + 难度趋势记录"共同构成 Goodhart 防护）。
+#     owner 审阅解码: python3 governance/drill/drill.py decode --id <id>
+#   - 全部样本为**构造的已知缺陷**（非真实凭据/非真实攻击载荷），验后即删分支。
+#   - 新增/修改样本必须 owner（randypanding）逐条审批后入库（approval 字段）。
+#
+# 难度口径（AC-4 防漂移: redrate 按周输出难度分布，"只挑 easy 打"会显形）:
+#   easy   = 单规则正则/扩展名级即拦（gitleaks 规则、凭据类文件名）
+#   medium = 需关卡具备解析/执行语义（YAML 解析、测试自测跑失败）
+#   hard   = 需跨信息源组合判定（PR 元数据 × 变更路径 × ADR 清单）
+# ============================================================================
+samples:
+  - id: hygiene-gitleaks-aws-key
+    defect_class: 凭据泄漏
+    description: 新增含 AWS 风格访问键的配置片段——gitleaks aws-access-token 规则应命中
+    difficulty: easy
+    gate: org-hygiene
+    scope: org
+    payload_kind: file
+    payload_path: "drill/leak-aws-{DATE}.ini"
+    defect_b64: >-
+      IyBkcmlsbCBmaXh0dXJl77yI5ryU5Lmg5rOo5YWl54mp77yM6Z2e55yf5a6e5Yet5o2u77yJCltiYWNrdXAtc3luY10KYXdzX2FjY2Vzc19rZXlfaWQgPSBBS0lBMk43UVg5Wks0TE1XOEIzQwphd3Nfc2VjcmV0X2FjY2Vzc19rZXkgPSA3S2RxM1ptUDl3WXRYdk44ckwyc0o0SGY2QjFjRTVhRDBnVWlPcVd5CnJlZ2lvbiA9IHVzLWVhc3QtMQo=
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22", note: owner 补签（样本为构造的已知缺陷，卡任务书授权）}
+
+  - id: hygiene-credfile-key
+    defect_class: 凭据泄漏
+    description: 新增 *.key 扩展名文件——hygiene "凭据类文件" 扩展名规则应拦
+    difficulty: easy
+    gate: org-hygiene
+    scope: org
+    payload_kind: file
+    payload_path: "drill/server-key-{DATE}.key"
+    defect_b64: >-
+      LS0tLS1CRUdJTiBSU0EgUFJJVkFURSBLRVktLS0tLQpNSUlCT2dJQkFBSkJBS++8iOa8lOS5oOWNoOS9je+8jOmdnuecn+WunuengemSpe+8iQotLS0tLUVORCBSU0EgUFJJVkFURSBLRVktLS0tLQo=
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22", note: owner 补签（占位私钥文本，非真实密钥）}
+
+  - id: hygiene-bigfile-blob
+    defect_class: 仓体积破坏
+    description: 新增 6MB 二进制大文件——hygiene ">5MB 禁入" 规则应拦
+    difficulty: easy
+    gate: org-hygiene
+    scope: org
+    payload_kind: generated
+    payload_path: "drill/blob-{DATE}.bin"
+    size_bytes: 6291456
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22", note: owner 补签（生成物=/dev/zero，无内容语义）}
+
+  - id: gate-yaml-parse-corrupt
+    defect_class: 配置注入
+    description: governance/ 下新增畸形 YAML——.github gate "YAML 全量解析" 应红
+    difficulty: medium
+    gate: gate
+    scope: github
+    payload_kind: file
+    payload_path: "governance/drill-corrupt-{DATE}.yaml"
+    defect_b64: >-
+      ZHJpbGxfcGF5bG9hZDogW3VuY2xvc2VkLWZsb3cKICBuZXN0ZWQ6IHsiYSI6IDEK
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22", note: owner 补签（未闭合 flow 序列，必然解析失败）}
+
+  - id: gate-selftest-fail
+    defect_class: 测试面破坏
+    description: governance/tests/ 下新增恒红 test-*.sh——gate "治理脚本自测" 应红
+    difficulty: medium
+    gate: gate
+    scope: github
+    payload_kind: file
+    payload_path: "governance/tests/test-drill-seed-{DATE}.sh"
+    defect_b64: >-
+      IyEvdXNyL2Jpbi9lbnYgYmFzaAojIOa8lOS5oOazqOWFpe+8muaBkue6oua1i+ivleKAlOKAlOWIpOWumueJqeacieaViOaAp+i0n+aOp+WItu+8iMKnNELvvIkKc2V0IC11byBwaXBlZmFpbAplY2hvICI6OmVycm9yOjpkcmlsbCBzZWVk77ya5pys6ISa5pys5Y2z5ryU5Lmg5qC35pys77yI5bqU6Kem5Y+R5rK755CG6ISa5pys6Ieq5rWL5YWz5Y2h5Y+Y57qi77yJIgpleGl0IDEK
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22", note: owner 补签（负控制：恒红脚本，验证自测关卡真的会跑会红）}
+
+  - id: org-adr-required-missing
+    defect_class: 治理绕过
+    description: C1 路径（governance/）新增文件 + 演习 PR 标题不带 ADR——org-adr-required 应红
+    difficulty: hard
+    gate: org-adr-required
+    scope: org
+    payload_kind: file
+    payload_path: "governance/drill-note-{DATE}.md"
+    defect_b64: >-
+      IyDmvJTkuaDms6jlhaXnianvvIhkcmlsbCBzZWVk77yJCgrmnKzmlofku7bkvY3kuo4gQzEg5Y+X566h6Lev5b6E77yIZ292ZXJuYW5jZS/vvInvvIzphY3lkIgqKuS4jeW4piBBRFIg5byV55SoKirnmoTmvJTkuaAgUFIg5qCH6aKY77yMCueUqOS6jumqjOivgSBvcmctYWRyLXJlcXVpcmVkIOWFs+WNoeS8muWPmOe6ouOAgumqjOWQjuWNs+WIoOOAggo=
+    pr_title_adr: false
+    approval: {approved_by: randypanding, date: "2026-08-22", note: owner 补签（跨源组合判定：PR 元数据 × C1 路径 × ADR 清单）}

governance/drill/tests/lib.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+# lib.sh —— 演练自测公共助手（W4-C4）
+# pick_py: 选出真实可用的 python 解释器（CI 恒有 python3；本地 Git Bash 的
+# python3 可能是 Windows 商店 stub——command -v 找得到但执行即败，必须实测）
+pick_py() {
+  local c
+  for c in "${PYTHON:-}" python3 python py -3; do
+    [[ -n "$c" ]] || continue
+    "$c" -c 'import sys, yaml; print("ok")' >/dev/null 2>&1 || continue
+    echo "$c"; return 0
+  done
+  return 1
+}

governance/drill/tests/test-samples.sh

@@ -0,0 +1,100 @@
+#!/usr/bin/env bash
+# test-samples.sh —— 样本库 schema 校验自测（W4-C4 AC-2）
+# 校验逻辑与引擎共用同一实现（drill.py validate_samples）——防"测试过而引擎放行"分叉。
+set -uo pipefail
+HERE="$(cd "$(dirname "$0")" && pwd)"
+ROOT="$(cd "$HERE/.." && pwd)"
+source "$(cd "$(dirname "$0")" && pwd)/lib.sh"
+PYTHON="$(pick_py)" || { echo "::error::无可用 python（含 pyyaml）"; exit 2; }
+PASS=0; FAIL=0
+TMP="$(mktemp -d)"; trap 'rm -rf "$TMP"' EXIT
+
+t() { # t <desc> <want: 0|非0> <cmd...>
+  local desc="$1" want="$2"; shift 2
+  "$@" >/dev/null 2>&1; local rc=$?
+  if { [[ "$want" == "0" ]] && [[ $rc -eq 0 ]]; } || { [[ "$want" != "0" ]] && [[ $rc -ne 0 ]]; }; then
+    PASS=$((PASS+1)); echo "ok   $desc"
+  else FAIL=$((FAIL+1)); echo "FAIL $desc（rc=$rc 期望=$want）"; fi
+}
+
+# 经引擎同一校验器判合法/非法（decode 入口先 load_samples→validate_samples，
+# 与 select/inject 共用——校验逻辑单一实现；fixture 样本统一 id=x-leak）
+check() {
+  "$PYTHON" "$ROOT/drill.py" decode --samples "$1" --id x-leak >/dev/null 2>&1
+}
+
+echo "== 1) 正式样本库合法（owner 审批/gate ID/难度/base64/{DATE} 全在）"
+t "正式 registry.yaml 通过 schema 校验" 0 "$PYTHON" "$ROOT/drill.py" select --samples "$ROOT/samples/registry.yaml" --repos "$ROOT/../REPOS.yaml" --seed 1 --sample-id hygiene-gitleaks-aws-key --target-repo .github
+
+echo "== 2) 破坏 fixture 逐项被拒（fail-closed：校验器不能只认存在性）"
+mk() { printf '%s\n' "$2" > "$TMP/$1"; }
+mk no_approval.yaml 'samples:
+  - id: x-leak
+    difficulty: easy
+    gate: org-hygiene
+    scope: org
+    payload_kind: file
+    payload_path: "drill/x-{DATE}.ini"
+    defect_b64: aGVsbG8=
+    pr_title_adr: true
+    approval: {approved_by: someone-else, date: "2026-08-22"}'
+t "审批人非 owner 被拒" 非0 check "$TMP/no_approval.yaml"
+mk no_gate.yaml 'samples:
+  - id: x-leak
+    difficulty: easy
+    scope: org
+    payload_kind: file
+    payload_path: "drill/x-{DATE}.ini"
+    defect_b64: aGVsbG8=
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22"}'
+t "缺预期关卡 ID（gate）被拒" 非0 check "$TMP/no_gate.yaml"
+mk bad_difficulty.yaml 'samples:
+  - id: x-leak
+    difficulty: trivial
+    gate: org-hygiene
+    scope: org
+    payload_kind: file
+    payload_path: "drill/x-{DATE}.ini"
+    defect_b64: aGVsbG8=
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22"}'
+t "非法难度值被拒" 非0 check "$TMP/bad_difficulty.yaml"
+mk bad_b64.yaml 'samples:
+  - id: x-leak
+    difficulty: easy
+    gate: org-hygiene
+    scope: org
+    payload_kind: file
+    payload_path: "drill/x-{DATE}.ini"
+    defect_b64: "!!!不是base64!!!"
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22"}'
+t "defect_b64 非法 base64 被拒" 非0 check "$TMP/bad_b64.yaml"
+mk no_date_placeholder.yaml 'samples:
+  - id: x-leak
+    difficulty: easy
+    gate: org-hygiene
+    scope: org
+    payload_kind: file
+    payload_path: "drill/x.ini"
+    defect_b64: aGVsbG8=
+    pr_title_adr: true
+    approval: {approved_by: randypanding, date: "2026-08-22"}'
+t "payload_path 缺 {DATE} 占位被拒" 非0 check "$TMP/no_date_placeholder.yaml"
+mk dup_id.yaml 'samples:
+  - {id: x-leak, difficulty: easy, gate: org-hygiene, scope: org, payload_kind: file, payload_path: "a/{DATE}", defect_b64: aGVsbG8=, pr_title_adr: true, approval: {approved_by: randypanding, date: "2026-08-22"}}
+  - {id: x-leak, difficulty: easy, gate: org-hygiene, scope: org, payload_kind: file, payload_path: "b/{DATE}", defect_b64: aGVsbG8=, pr_title_adr: true, approval: {approved_by: randypanding, date: "2026-08-22"}}'
+t "样本 id 重复被拒" 非0 check "$TMP/dup_id.yaml"
+t "空样本库被拒" 非0 check /dev/null
+
+echo "== 3) decode round-trip（owner 审阅通道可用；AC-2 审批前置能力）"
+OUT=$("$PYTHON" "$ROOT/drill.py" decode --samples "$ROOT/samples/registry.yaml" --id hygiene-gitleaks-aws-key)
+if grep -q "AKIA2N7QX9ZK4LMW8B3C" <<<"$OUT"; then PASS=$((PASS+1)); echo "ok   decode 输出含缺陷原文（可 owner 审）"
+else FAIL=$((FAIL+1)); echo "FAIL decode 未还原缺陷内容"; fi
+GEN=$("$PYTHON" "$ROOT/drill.py" decode --samples "$ROOT/samples/registry.yaml" --id hygiene-bigfile-blob)
+if grep -q "生成物样本" <<<"$GEN"; then PASS=$((PASS+1)); echo "ok   generated 样本 decode 有诚实注记"
+else FAIL=$((FAIL+1)); echo "FAIL generated 样本 decode 异常"; fi
+
+echo "样本库自测: pass=$PASS fail=$FAIL"
+[[ $FAIL -eq 0 ]]