# -*- coding: utf-8 -*- """CCG 四要素接线 + 检索键分区（P33 · 设计态验收）。 对照交付计划： ① 契约层：VERIFICATION_BASIS 扩 textbook/public_kb，且显式分「可复现档 / 来源一致性档」 ② 回填来源：功能名补 state_attributes.name（旧白名单漏掉的来源），title 仅作回退 ③ 占位排除：骨架占位标记绝不渲染成正文事实，且单独计数（转待填充工单） ④ 接线：四要素写正文规范行 → 词面召回立即命中；幂等 + 可回滚 ⑤ 检索键分区：`# 不适用条件：` 是反例声明，LIKE 与 fuzzy 两路都不作召回键 ⑥ 反例的正确去向：仍由 judge_qualification 读 frontmatter 走 REJECT（不靠召回键） 独立临时根，重跑 ≡ 首跑；全程只读真实库之外的东西。 运行：python -m md_cg.test_p33_ccg_wiring """ from __future__ import annotations import json import os import shutil import tempfile from . import backfill, forgetting, nodefile, tokens, weights from .mdcg import (STATE_ACCEPT, STATE_BLINDSPOT, STATE_REJECT, TIER_GLOBAL_LIKE, TIER_GLOBAL_SCAN, MdCG) from .mdcos import MdCGOS PASS = FAIL = 0 FAILS = [] def ok(cond, label): global PASS, FAIL if cond: PASS += 1 else: FAIL += 1 FAILS.append(label) print(f" FAIL {label}") def ids(res): """从 search/fuzzy 结果里取节点 id（frontmatter.id 缺失时回退文件名）。""" out = [] for r in res: fm = r[0].get("frontmatter") or {} nid = fm.get("id") or os.path.splitext( os.path.basename(r[0].get("path") or ""))[0] out.append(nid) return out PH = "[历史·高中] 骨架锚点，内容待填充" # ④ 触发词只写在 comment.生效条件 里，正文没有 → apply 后才会进入检索键 GAP_DOC = "# 功能名：某专业知识点\n\n该节点正文不含触发词。\n" # ⑤⑥ 正条件 / 正负条件对照：负条件用 # 不适用条件： 行承载 POS_DOC = ("# 功能名：正条件节点\n# 生效条件：问ZXQ7\n# 子功能：子功能说明\n" "# 执行：执行说明\n# 验证方式：编译器/静态检查通过\n" "# 不适用条件：问完全无关主题\n") POSFX_DOC = ("# 功能名：正负条件节点\n# 生效条件：问通用流程\n# 子功能：子功能说明\n" "# 执行：执行说明\n# 验证方式：编译器/静态检查通过\n" "# 不适用条件：问ZXQ7\n") PH_DOC = "# 功能名：骨架节点\n\n迁移期空壳节点。\n" def _seed(root): cg = MdCGOS(root) cg.add("wire_gap", GAP_DOC, layer="knowledge", verification_basis="test", non_applicable_conditions=["问其它主题"], state_attributes={"comment": {"生效条件": ["问某某特殊主题"], "子功能": "子功能说明", "执行": "执行步骤说明"}}) cg.add("wire_pos", POS_DOC, layer="knowledge", verification_basis="compiler", non_applicable_conditions=["问完全无关主题"]) cg.add("wire_posfx", POSFX_DOC, layer="knowledge", verification_basis="compiler", non_applicable_conditions=["问ZXQ7"]) cg.add("wire_ph", PH_DOC, layer="knowledge", verification_basis="test", non_applicable_conditions=["问其它主题"], state_attributes={"comment": {"生效条件": ["问骨架"], "子功能": PH, "执行": PH}}) cg.add("wire_name", "无 CCG 行的正文。\n", layer="knowledge", verification_basis="data", non_applicable_conditions=["问其它主题"], state_attributes={"name": "规范功能名", "comment": {"生效条件": ["问规范名"], "子功能": "子功能说明", "执行": "执行步骤说明"}}) cg.flush() return cg def main(): global PASS, FAIL tmp = tempfile.mkdtemp(prefix="mdcg_p33_") try: root = os.path.join(tmp, "root") os.makedirs(root, exist_ok=True) cg = _seed(root) # ---------- ① 契约层：枚举扩展 + 两档划分 ---------- ok("textbook" in nodefile.VERIFICATION_BASIS and "public_kb" in nodefile.VERIFICATION_BASIS, "①VERIFICATION_BASIS 扩 textbook/public_kb") rep, cons = set(nodefile.REPRODUCIBLE_BASIS), set(nodefile.CONSISTENCY_BASIS) ok(not (rep & cons), "①可复现档与来源一致性档互斥") # #466（2026-10-15，Lead 裁定）：新增 untrusted 后**本契约升级** —— 从「两档 + other # 覆盖枚举」改为「两档 + other + **刻意在档外**集合 覆盖枚举」。理由：untrusted 是 # **外部内容**的基底，既非可复现（理科档）亦非来源一致（文科档），**必须落在两档之外** # （守卫 md_cg/test_issue466_basis_registry.py 的 A02 正是钉这一点）；把它塞进任一档 # 以满足旧等式，正是本件要防的行为。升级后本腿仍能抓「漏网值」，且能区分「漏网」与 # 「刻意在档外」。 _OFF_TIER = {"untrusted"} # 刻意在两档之外的枚举值（#466） ok(rep | cons | {"other"} | _OFF_TIER == set(nodefile.VERIFICATION_BASIS), "①两档 + other + 刻意在档外集合 恰好覆盖枚举（无漏网值）") ok(nodefile.verification_basis_valid({"verification_basis": "textbook"}) and nodefile.verification_basis_valid({"verification_basis": "public_kb"}), "①新档值通过 verification_basis_valid 校验") ok(set(backfill.BASIS_TEXT) <= set(nodefile.VERIFICATION_BASIS), "①BASIS_TEXT 只含合法枚举值") ok(all(backfill.BASIS_TEXT[v] != backfill.BASIS_TEXT["other"] for v in nodefile.CONSISTENCY_BASIS), "①新档有独立文本（不退化到 other）") ok(all(v in weights.BASIS_TRUST for v in nodefile.CONSISTENCY_BASIS), "①weights.BASIS_TRUST 认账新档（否则权重按缺失算）") ok(weights.BASIS_TRUST["textbook"] > weights.BASIS_TRUST["public_kb"] > weights.BASIS_TRUST["other"], "①可信度排序：教材 > 公开知识库 > other") ok(all(v in forgetting.VERIFIED_BASES for v in nodefile.CONSISTENCY_BASIS), "①长期分层认账新档（避免「白箱可 ACCEPT、分层算未验证」）") ok(not (set(nodefile.CONSISTENCY_BASIS) & set(forgetting.DETERMINISTIC_BASIS)), "①来源一致性 ≠ 内部确定性（两轴不混用）") # ---------- ③ 占位识别（纯函数） ---------- ok(nodefile.is_placeholder_text(PH), "③骨架占位标记被识别") ok(nodefile.is_placeholder_text("内容待填充"), "③裸占位短语被识别") ok(not nodefile.is_placeholder_text("论语（语录体，孔子）"), "③真实事实不被误判为占位") ok(nodefile.is_placeholder_text("") and nodefile.is_placeholder_text(None), "③空值不可渲染为事实") # ---------- ②④ name 来源优先级 + 占位过滤（单元级，精确定位） ---------- d = backfill.derive_fields( {"title": "标题功能名", "state_attributes": {"name": "规范功能名"}}, "正文\n", None) ok(d["功能名"] == ("规范功能名", "state_attributes.name"), "②功能名优先取 state_attributes.name") d2 = backfill.derive_fields({"title": "标题功能名"}, "正文\n", None) ok(d2["功能名"] == ("标题功能名", "frontmatter.title"), "②缺 state_attributes.name 时回退 frontmatter.title") ph_out = [] d3 = backfill.derive_fields( {"state_attributes": {"comment": { "执行": PH, "子功能": PH, "生效条件": ["问真实条件"]}}, "non_applicable_conditions": ["问其它"]}, "正文\n", None, placeholder_out=ph_out) ok("执行" not in d3 and "子功能" not in d3, "③占位字段不进入推导结果（绝不渲染成事实）") ok(set(ph_out) == {"执行", "子功能"}, "③占位字段逐名记录（可出待填充工单）") ok("生效条件" in d3 and "不适用条件" in d3, "③同节点内真实字段照常推导（不因占位而整体放弃）") ok(backfill.derive_fields({"state_attributes": {"comment": {"执行": PH}}}, "正文\n", None) == {}, "③placeholder_out 为可选出参（缺省不报错）") # ---------- ⑤ 检索键分区：positive_body / _like ---------- ok("ZXQ7" not in nodefile.positive_body(POSFX_DOC), "⑤positive_body 剥离 `# 不适用条件：` 行") ok("问通用流程" in nodefile.positive_body(POSFX_DOC), "⑤正条件行原样保留") plain = "正文提到不适用条件这个词，但不是 CCG 行。\n" ok(nodefile.positive_body(plain) == plain, "⑤只剥 CCG 行，不误伤普通句子") ok(MdCG._like(POSFX_DOC, {}, ["ZXQ7"]) is False, "⑤_like 不把反例当召回键") ok(MdCG._like(POSFX_DOC, {}, ["通用流程"]) is True, "⑤_like 正条件照常命中") ok(MdCG._like(POSFX_DOC, {"tags": ["ZXQ7"]}, ["ZXQ7"]) is True, "⑤tags 仍参与匹配（分区只作用于正文）") # ---------- ④ 接线前：触发词只在 comment，LIKE 层不命中 ---------- res0, meta0 = cg.search("某某特殊主题") ok(meta0["tier"] != TIER_GLOBAL_LIKE, "④接线前：comment 里的触发词尚未进入检索键") p0 = backfill.plan(cg) ok("wire_gap" in p0["planned_ids"], "④有来源节点进入可回填清单") ok(p0["nodes_with_placeholder"] >= 1, "③占位节点单独计数（不静默混入）") pp = backfill.plan(cg, include_partial=True) it_ph = next(i for i in pp["items"] if i["id"] == "wire_ph") ok(it_ph["class"] == "partial" and set(it_ph["placeholder_fields"]) == {"执行", "子功能"}, "③占位节点被标 partial 且注明占位字段") # ---------- ④ 接线落盘：四要素入正文 → 纳入检索键 ---------- a1 = backfill.apply(cg, ids=["wire_gap", "wire_name"], batch="p33a", actor="tester") ok(a1["written"] >= 2, "④apply 写入非零") _, c_gap = cg._read(cg.index["nodes"]["wire_gap"]) ok("# 生效条件：问某某特殊主题" in c_gap, "④生效条件渲染为正文规范行") ok(nodefile.ccg_completeness(c_gap)["complete"] is True, "④接线后 6 要素齐全（含生效条件；从仅可检索升级为可判定）") ok(nodefile.ccg_completeness( "# 功能名：X\n# 子功能：y\n# 执行：z\n" "# 验证方式：test\n# 不适用条件：n\n")["complete"] is False, "④生效条件列入必填：缺之即 not complete（不再可隐含）") res1, meta1 = cg.search("某某特殊主题") ok(meta1["tier"] == TIER_GLOBAL_LIKE, "④接线后：正文触发词进入 LIKE 检索层") ok("wire_gap" in ids(res1), "④接线后该节点被召回") _, c_name = cg._read(cg.index["nodes"]["wire_name"]) ok("# 功能名：规范功能名" in c_name, "②功能名由 state_attributes.name 写入正文") ok("# 执行：" in c_name, "④执行渲染为正文规范行") # 幂等 a2 = backfill.apply(cg, ids=["wire_gap"], batch="p33a", actor="tester") ok(a2["written"] == 0, "④幂等：已完成节点不重复写") # ---------- ④ 回滚：真逆操作 ---------- rb = backfill.rollback(cg, batch="p33a", actor="tester") ok(rb["reverted"] >= 2, "④回滚撤销非零") _, c_gap2 = cg._read(cg.index["nodes"]["wire_gap"]) ok("# 生效条件：" not in c_gap2, "④回滚移除正文规范行") res2, meta2 = cg.search("某某特殊主题") ok(meta2["tier"] != TIER_GLOBAL_LIKE, "④回滚后触发词退出检索键（接线可逆）") # ---------- ⑤ 集成：反例不作召回键，正条件照常 ---------- res3, meta3 = cg.search("ZXQ7") ok(meta3["tier"] == TIER_GLOBAL_LIKE, "⑤ZXQ7 在 LIKE 层有命中") ok("wire_pos" in ids(res3), "⑤正条件命中被召回") ok("wire_posfx" not in ids(res3), "⑤仅反例命中的节点不被召回（分区生效）") res4, _ = cg.search("通用流程") ok("wire_posfx" in ids(res4), "⑤正条件仍作召回键") fz, _ = cg._path_fuzzy("ZXQ7", cg._candidates()) ok("wire_pos" in ids(fz), "⑤fuzzy 路召回正条件命中") ok("wire_posfx" not in ids(fz), "⑤fuzzy 路同样不召回反例") # ---------- ⑥ 反例的正确去向：judge 走 REJECT ---------- fm_pfx, c_pfx = cg._read(cg.index["nodes"]["wire_posfx"]) j_rej = MdCG.judge_qualification({"frontmatter": fm_pfx, "content": c_pfx}, "问ZXQ7", {"query": "问ZXQ7"}) ok(j_rej["state"] == STATE_REJECT, "⑥反例命中走 REJECT（不依赖召回键，靠 frontmatter 负条件）") fm_pos, c_pos = cg._read(cg.index["nodes"]["wire_pos"]) j_acc = MdCG.judge_qualification({"frontmatter": fm_pos, "content": c_pos}, "问ZXQ7", {"query": "问ZXQ7"}) ok(j_acc["state"] == STATE_ACCEPT, "⑥正条件命中 + 基底已声明 → ACCEPT") fm_ph, c_ph = cg._read(cg.index["nodes"]["wire_ph"]) j_bl = MdCG.judge_qualification({"frontmatter": fm_ph, "content": c_ph}, "问骨架", {"query": "问骨架"}) ok(j_bl["state"] == STATE_BLINDSPOT, "⑥要素不全仍 BLINDSPOT（占位排除后不冒充齐全）") # ---------- ③ 占位节点端到端：include_partial 也不会写进占位 ---------- backfill.apply(cg, ids=["wire_ph"], batch="p33ph", include_partial=True, actor="tester") _, c_ph2 = cg._read(cg.index["nodes"]["wire_ph"]) ok("内容待填充" not in c_ph2, "③正文绝不出现占位标记") ok("# 执行：" not in c_ph2, "③占位字段不渲染成 `# 执行：` 行") ok("# 生效条件：问骨架" in c_ph2, "③同节点真实字段照常接线") # ---------- ⑦ 范围收窄：prefix 收窄 + 内部层出局 ---------- cg.add("note_x", "派生记忆正文。\n", layer="knowledge", verification_basis="test", state_attributes={"comment": {"生效条件": ["问派生主题"]}}) cg.add("anchor_x", "锚点脚手架正文。\n", layer="anchor", verification_basis="test", state_attributes={"comment": {"生效条件": ["问锚点主题"]}}) p_kp = backfill.plan(cg, prefix="wire_") ok("wire_gap" in p_kp["planned_ids"], "⑦prefix 内节点照常入列") ok("note_x" not in p_kp["planned_ids"], "⑦prefix 外节点出局（不混入工单）") p_all = backfill.plan(cg) ok("anchor_x" not in p_all["planned_ids"] and p_all["skipped_internal"] >= 1, "⑦内部 anchor/self 层出局并单独计数") ok(backfill.exempt_plan(cg)["skipped_internal"] >= 1, "⑦摘豁免同样排除内部层（不改写脚手架）") finally: shutil.rmtree(tmp, ignore_errors=True) print(f"\nPASS={PASS} FAIL={FAIL}") if FAILS: print("FAILED:") for f in FAILS: print(" -", f) raise SystemExit(1) print("全部通过：P33 四要素接线 + 检索键分区") if __name__ == "__main__": main()