foxclaw 0.1.0 → 0.1.2

package/python/tests/test_hkc_manifest_reuse.py

@@ -0,0 +1,317 @@
+"""
+HKC 持久化 manifest 校验测试(阶段 A 提交)。
+
+覆盖范围:
+  M1  save→load 8 步校验全过(roundtrip)
+  M2  旧 v1 manifest(无 format_version 字段)→ _reset + load 返 False
+  M3  format_version 升 3 → 校验失败 + _reset
+  M4  ku_text_version 不一致 → 校验失败 + _reset(_ku_text 拼接变了,旧 hash 无意义)
+  M5  backend 字段缺失 → 校验失败
+  M6  dim 字段缺失/为 0/非 int → 校验失败
+  M7  4 个 list 长度不一致 → 校验失败
+  M8  faiss 实际维度 ≠ meta["dim"] → 校验失败
+  M9  faiss 实际 ntotal ≠ len(ku_ids) → 校验失败
+  M10 meta 文件损坏(非 pickle) → _reset + load 返 False
+  M11 save invariant:4 list 长度不一致时抛 AssertionError
+  M12 text_hashes 与 _ku_text 内容一致(同 KU 不同时间 hash 相同)
+
+  B1  bm25 空 corpus → size=0,search 返 []
+  B2  bm25 全空 doc(5 条)→ 不崩,size 计数正确,search 返 []
+  B3  bm25 有一个非空 doc → 正常命中
+  B4  bm25 未 build 时 search 返 []
+
+设计原则(同 P0):
+  - 不 mock:真 faiss + StubBackend + 真 VectorIndex + 真 BM25Index。
+  - vendor 隔离硬阻断:模块级解析后立即自检 __file__,不是 vendor 就 RuntimeError。
+  - 不依赖网络 / 真实模型;全程 tmp_path。
+"""
+from __future__ import annotations
+
+import pickle
+import sys
+from pathlib import Path
+
+import faiss
+import numpy as np
+import pytest
+
+# ── vendor 隔离硬阻断(等价于 test_vendor_path_isolation.py 闸门 2-4) ────────
+from hkc_search.vector_search import (
+    VectorIndex,
+    _MANIFEST_FORMAT_VERSION,
+    _KU_TEXT_FORMAT_VERSION,
+    _hash_text,
+    _ManifestValidationError,
+)
+from hkc_search.bm25 import BM25Index
+from hkc_search.embedding_backends import StubBackend
+from hkc_core.models.ku import ConceptKU
+
+_VENDOR_ROOT = Path(__file__).resolve().parents[1] / "vendor" / "hkc"
+
+
+def _assert_vendor_path(obj, name: str) -> None:
+    mod = obj.__module__
+    f = getattr(sys.modules[mod], "__file__", "") or ""
+    assert str(Path(f).resolve()).startswith(str(_VENDOR_ROOT.resolve())), (
+        f"vendor isolation failed: {name}.__module__={mod!r} __file__={f!r} "
+        f"not under vendor/hkc. Test is running against wrong HKC copy."
+    )
+
+
+for _cls_name, _cls in [
+    ("VectorIndex",            VectorIndex),
+    ("BM25Index",              BM25Index),
+    ("StubBackend",            StubBackend),
+    ("ConceptKU",              ConceptKU),
+]:
+    _assert_vendor_path(_cls, _cls_name)
+
+
+# ── Fixtures ────────────────────────────────────────────────────────
+
+@pytest.fixture
+def stub_vector(tmp_path) -> VectorIndex:
+    """已 build 3 条 KU 的 vector,save 后留磁盘文件。"""
+    idx = VectorIndex(backend=StubBackend())
+    kus = [
+        ConceptKU(ku_id="KU_A", name="alpha",   summary="一种希腊字母",      domain="letter"),
+        ConceptKU(ku_id="KU_B", name="beta",    summary="另一种希腊字母",    domain="letter"),
+        ConceptKU(ku_id="KU_C", name="gamma",   summary="第三种希腊字母",    domain="letter"),
+    ]
+    idx.build(kus)
+    return idx
+
+
+def _ku_text(ku) -> str:
+    """与 VectorIndex._ku_text 等价的复制(测试侧重新计算 hash 用)。"""
+    parts = [ku.name]
+    if ku.summary:
+        parts.append(ku.summary)
+    if ku.tags:
+        parts.append(" ".join(ku.tags))
+    stmt = getattr(ku, "statement", "")
+    if stmt:
+        parts.append(stmt)
+    return " ".join(parts)
+
+
+# ── M1: roundtrip(全过) ───────────────────────────────────────────
+
+class TestManifestRoundtrip:
+    def test_save_load_roundtrip(self, tmp_path, stub_vector):
+        """save → load 完整,数据一致(ku_ids/names/summaries/text_hashes 全部保留)。"""
+        vec_path = str(tmp_path / "search_vector")
+        stub_vector.save(vec_path)
+        # 落盘了 v2 manifest
+        meta = pickle.loads((Path(vec_path + ".meta")).read_bytes())
+        assert meta["format_version"] == _MANIFEST_FORMAT_VERSION == 2
+        assert meta["ku_text_version"] == _KU_TEXT_FORMAT_VERSION == 1
+        # backend 字段以"Stub("开头(StubBackend.name 形如 "Stub(dim=384)")
+        assert meta["backend"].startswith("Stub("), meta["backend"]
+        assert meta["dim"] == stub_vector.dim
+        assert len(meta["ku_ids"]) == 3
+        assert len(meta["text_hashes"]) == 3
+        # 加载
+        fresh = VectorIndex(backend=StubBackend())
+        assert fresh.load(vec_path) is True
+        assert fresh.size() == 3
+        assert fresh._ku_ids == ["KU_A", "KU_B", "KU_C"]
+        assert fresh._text_hashes == stub_vector._text_hashes
+
+    def test_text_hash_matches_independent_compute(self, tmp_path, stub_vector):
+        """text_hashes[i] 必须等于独立计算 _hash_text(_ku_text(ku_i)) —— 这才是
+        阶段 B update_ku 能正确做差异检测的前提。"""
+        expected = []
+        for ku_id in stub_vector._ku_ids:
+            i = stub_vector._ku_ids.index(ku_id)
+            ku = ConceptKU(
+                ku_id=ku_id,
+                name=stub_vector._names[i],
+                summary=stub_vector._summaries[i],
+                domain="letter",
+            )
+            expected.append(_hash_text(_ku_text(ku)))
+        assert stub_vector._text_hashes == expected
+
+    def test_hash_stable_for_same_text(self):
+        """同文本 → 同 hash(阶段 B 跨进程比对、阶段 C 续跑检测都要靠这个稳定性)。"""
+        a = _hash_text("alpha 一种希腊字母")
+        b = _hash_text("alpha 一种希腊字母")
+        c = _hash_text("alpha 一种希腊字母 ")   # 末尾空格,文本不同
+        assert a == b
+        assert a != c
+
+
+# ── M2~M9: 8 步校验各失败路径 ───────────────────────────────────
+
+class TestManifestValidation:
+    def _fresh_v2_disk(self, tmp_path) -> str:
+        """构造一份合法的 v2 磁盘,返回 prefix path。"""
+        vec_path = str(tmp_path / "search_vector")
+        v = VectorIndex(backend=StubBackend())
+        kus = [ConceptKU(ku_id="KU_A", name="a", summary="a", domain="d"),
+               ConceptKU(ku_id="KU_B", name="b", summary="b", domain="d")]
+        v.build(kus)
+        v.save(vec_path)
+        return vec_path
+
+    def test_v1_manifest_no_format_version_resets(self, tmp_path):
+        """M2:旧 v1 manifest(无 format_version 字段)→ _reset。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        # 改写 meta 为 v1:仅 ku_ids/names/summaries/backend,无 format_version/dim/text_hashes
+        with open(vec_path + ".meta", "wb") as f:
+            pickle.dump({"ku_ids": ["X"], "names": ["x"],
+                         "summaries": ["x"], "backend": "stub"}, f)
+        v = VectorIndex(backend=StubBackend())
+        # 加载必失败 → 返 False
+        assert v.load(vec_path) is False
+        # _reset 已清空
+        assert v.size() == 0
+        assert v._ku_ids == []
+        assert v._text_hashes == []
+
+    def test_format_version_bump_resets(self, tmp_path):
+        """M3:format_version 升 3 → 校验失败 + _reset。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        with open(vec_path + ".meta", "rb") as f:
+            meta = pickle.load(f)
+        meta["format_version"] = 3
+        with open(vec_path + ".meta", "wb") as f:
+            pickle.dump(meta, f)
+        v = VectorIndex(backend=StubBackend())
+        assert v.load(vec_path) is False
+        assert v.size() == 0
+
+    def test_ku_text_version_drift_resets(self, tmp_path):
+        """M4:ku_text_version 改了 → 校验失败(旧 hash 在新拼接下不可信)。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        with open(vec_path + ".meta", "rb") as f:
+            meta = pickle.load(f)
+        meta["ku_text_version"] = 99
+        with open(vec_path + ".meta", "wb") as f:
+            pickle.dump(meta, f)
+        v = VectorIndex(backend=StubBackend())
+        assert v.load(vec_path) is False
+        assert v.size() == 0
+
+    def test_backend_field_missing_resets(self, tmp_path):
+        """M5:backend 字段缺失 → 校验失败。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        with open(vec_path + ".meta", "rb") as f:
+            meta = pickle.load(f)
+        del meta["backend"]
+        with open(vec_path + ".meta", "wb") as f:
+            pickle.dump(meta, f)
+        v = VectorIndex(backend=StubBackend())
+        assert v.load(vec_path) is False
+
+    def test_dim_field_invalid_resets(self, tmp_path):
+        """M6:dim 字段缺失/为 0/非 int → 校验失败。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        for bad in (None, 0, "not_an_int", -1):
+            with open(vec_path + ".meta", "rb") as f:
+                meta = pickle.load(f)
+            meta["dim"] = bad
+            with open(vec_path + ".meta", "wb") as f:
+                pickle.dump(meta, f)
+            v = VectorIndex(backend=StubBackend())
+            assert v.load(vec_path) is False, f"dim={bad!r} 应校验失败"
+            assert v.size() == 0
+
+    def test_list_length_mismatch_resets(self, tmp_path):
+        """M7:4 list 长度不一致 → 校验失败。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        with open(vec_path + ".meta", "rb") as f:
+            meta = pickle.load(f)
+        meta["text_hashes"] = meta["text_hashes"][:1]   # 只留 1 个,与 2 条 ku_ids 不一致
+        with open(vec_path + ".meta", "wb") as f:
+            pickle.dump(meta, f)
+        v = VectorIndex(backend=StubBackend())
+        assert v.load(vec_path) is False
+
+    def test_faiss_dim_mismatch_resets(self, tmp_path):
+        """M8:faiss 实际维度与 meta["dim"] 不一致(切后端后常见)→ 校验失败。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        # 重建一个维度错的 faiss 文件
+        wrong_idx = faiss.IndexFlatIP(999)  # 错维度
+        wrong_idx.add(np.random.rand(2, 999).astype("float32"))
+        faiss.write_index(wrong_idx, vec_path + ".faiss")
+        v = VectorIndex(backend=StubBackend())
+        assert v.load(vec_path) is False
+        assert v.size() == 0
+
+    def test_faiss_ntotal_mismatch_resets(self, tmp_path):
+        """M9:faiss 实际 ntotal 与 meta len(ku_ids) 不一致 → 校验失败。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        # 重建一个 ntotal=99 的 faiss(其他字段不动)
+        wrong_idx = faiss.IndexFlatIP(stub_vector_dim := VectorIndex(backend=StubBackend()).dim)
+        wrong_idx.add(np.random.rand(99, stub_vector_dim).astype("float32"))
+        faiss.write_index(wrong_idx, vec_path + ".faiss")
+        v = VectorIndex(backend=StubBackend())
+        assert v.load(vec_path) is False
+        assert v.size() == 0
+
+    def test_corrupt_meta_file_resets(self, tmp_path):
+        """M10:meta 文件非 pickle → _reset + load 返 False。"""
+        vec_path = self._fresh_v2_disk(tmp_path)
+        with open(vec_path + ".meta", "wb") as f:
+            f.write(b"\x00\x01not-a-pickle\xff")
+        v = VectorIndex(backend=StubBackend())
+        assert v.load(vec_path) is False
+        assert v.size() == 0
+
+    def test_save_invariant_assertion(self, tmp_path):
+        """M11:save() 时 4 list 长度不一致 → 抛 AssertionError(不让坏 manifest 落盘)。"""
+        v = VectorIndex(backend=StubBackend())
+        kus = [ConceptKU(ku_id="KU_A", name="a", summary="a", domain="d")]
+        v.build(kus)
+        # 故意破坏 invariant:删一个 text_hash
+        v._text_hashes = v._text_hashes[:0]   # 现在 _text_hashes=[] 但 _ku_ids=[KU_A]
+        with pytest.raises(AssertionError):
+            v.save(str(tmp_path / "search_vector"))
+
+    def test_validation_error_has_reason_code(self, tmp_path):
+        """_ManifestValidationError.reason 是机器可读码,供上层 logger/log metric 用。"""
+        try:
+            raise _ManifestValidationError("foo_reason", "foo msg")
+        except _ManifestValidationError as e:
+            assert e.reason == "foo_reason"
+            assert "foo msg" in str(e)
+
+
+# ── B1~B4: bm25 0 corpus 保护 ────────────────────────────────────
+# B 阶段把 BM25Index 重写为自实现倒排索引(不再包装 rank_bm25),旧私有属性
+# `_bm25` / `_ku_ids` / `_names` / `_summaries` / `_corpus` / `_rebuild_bm25`
+# 均已不存在。故断言改为验证**对外行为**(size / search),不再耦合私有结构 ——
+# 旧测试直接拼私有 list 再调私有 rebuild,是它在此次重构后失效的根因。
+
+class TestBM25EmptyCorpus:
+    def test_build_empty_list(self):
+        """B1:build([]) → 索引为空,search 返 []。"""
+        b = BM25Index()
+        b.build([])
+        assert b.size() == 0
+        assert b.search("anything") == []
+
+    def test_build_all_empty_docs(self):
+        """B2:全空 doc(name/summary/tags 皆空)→ 不崩,size 计数正确,search 返 []。
+        触发场景:append_ku 首个 KU 时拼出的 token 列表为空。"""
+        b = BM25Index()
+        for i in range(5):
+            b.append_ku(ConceptKU(ku_id=f"KU_{i}", name="", summary=""))
+        assert b.size() == 5
+        assert b.search("x") == []
+
+    def test_build_one_valid_doc(self):
+        """B3:哪怕只有一个非空 doc 也能正常命中。"""
+        b = BM25Index()
+        b.append_ku(ConceptKU(ku_id="KU_X", name="x", summary="x is a letter"))
+        r = b.search("x letter", top_k=1)
+        assert len(r) == 1 and r[0].ku_id == "KU_X"
+
+    def test_search_with_no_bm25(self):
+        """B4:未 build 时 search 返 []。"""
+        b = BM25Index()
+        # 不 build,直接 search
+        assert b.search("anything") == []