ltcai 11.4.0 → 11.5.1

package/scripts/generate_rust_parity_fixtures.py

@@ -6,27 +6,23 @@ something keeps proving it is still one. This script is the Python half of that
 proof: it builds a small, fully deterministic Brain with the **real** write path
 (``KnowledgeGraphStore._upsert_node`` / ``_upsert_chunk`` / ``_upsert_edge``,
 the real hash embedder, the real v2 projection and trigram FTS index), then runs
-the real ``hybrid_search`` / ``search`` / ``vector_search`` over it and writes
-their answers to ``rust/fixtures/golden/``.
-
-Two consumers read what it writes:
-
-* ``tests/unit/test_rust_parity_contract.py`` re-runs the Python engines against
-  the committed database and asserts the goldens still hold — so a change to
-  Python retrieval semantics fails loudly instead of silently invalidating the
-  contract the Rust side is pinned to;
-* ``rust/lattice-retrieval/tests/parity.rs`` runs the Rust port against the same
-  database and the same goldens.
-
-Determinism is the whole design constraint:
-
-* every timestamp is written by the real code and then **backdated** to a fixed
-  value, so nothing in the fixture depends on when it was generated;
-* ``hybrid_search``'s recency decay calls ``datetime.now()``, so the clock is
-  frozen at :data:`FROZEN_NOW` (recorded in the manifest for the Rust side);
-* LLM concept extraction is forced off, so ``_topic_candidates`` always takes
-  the rule-based path a port can reproduce;
-* every environment knob the retrieval stack reads is pinned to its default.
+the real engines over it and writes their answers to ``rust/fixtures/golden/``.
+The suite set widens with the port (:data:`SUITES`): the search engines
+(v11.4.0); the KG relationship/traverse reads, the service layer, the durable
+history reads and the context assembler (v11.5.0); the document-generation
+search, traversal and context builder (v11.5.1, whose corpus and runners live
+in ``scripts/parity_fixture_corpus_docgen.py``).
+
+Two consumers read what it writes: ``tests/unit/test_rust_parity_contract.py``
+re-runs the Python entry points against the committed database and asserts the
+goldens still hold — so a change to Python semantics fails loudly instead of
+silently invalidating the contract the Rust side is pinned to — and
+``rust/lattice-retrieval/tests/{parity,suites}.rs`` run the Rust ports against
+the same database and the same goldens. Determinism is the whole design
+constraint: every timestamp is written by the real code and then **backdated**;
+every ``datetime.now()`` the ports reach is frozen at :data:`FROZEN_NOW`
+(recorded in the manifest); LLM concept extraction is forced off; every
+environment knob is pinned to its default.
 
 Usage::
 
@@ -36,6 +32,7 @@ Usage::
 from __future__ import annotations
 
 import json
+import logging
 import os
 import shutil
 import sqlite3
@@ -46,20 +43,24 @@ from pathlib import Path
 from typing import Any, Callable, Dict, Iterator, List, Optional, Tuple
 
 REPO_ROOT = Path(__file__).resolve().parents[1]
-if str(REPO_ROOT) not in sys.path:
-    sys.path.insert(0, str(REPO_ROOT))
+# The repo (for the product packages) and this directory: `scripts` is not a
+# package, so the docgen corpus is imported by name off the script directory —
+# the same by-path convention the contract test uses to import this file.
+for _import_root in (REPO_ROOT, Path(__file__).resolve().parent):
+    if str(_import_root) not in sys.path:
+        sys.path.insert(0, str(_import_root))
+
+import parity_fixture_corpus_docgen as docgen  # noqa: E402
 
 FIXTURE_DIR = REPO_ROOT / "rust" / "fixtures"
 GOLDEN_DIR = FIXTURE_DIR / "golden"
 STORE_PATH = FIXTURE_DIR / "parity_store.sqlite"
 
-#: The wall clock ``hybrid_search`` sees. Recency decay is a function of "now",
-#: so a golden generated against a moving clock is not a golden.
+#: The wall clock the ports see: a golden built against a moving clock is none.
 FROZEN_NOW = "2026-08-01T12:00:00"
 
-#: Every environment variable the ported path reads, pinned to the default
-#: configuration the port targets (brute backend, RRF off, graph expansion off,
-#: cross-encoder rerank off, rewrite on).
+#: Every environment variable the ported path reads, pinned to the configuration
+#: it targets (brute backend, RRF/expansion/rerank off, rewrite on).
 PINNED_ENV: Dict[str, str] = {
     "LATTICEAI_VECTOR_DIM": "384",
     "LATTICEAI_VECTOR_INDEX": "brute",
@@ -80,16 +81,12 @@ WS_BETA = "ws-beta"
 # ── the corpus ───────────────────────────────────────────────────────────────
 # (node_id, type, title, summary, metadata, workspace_id, updated_at)
 #
-# Shaped on purpose:
-#   * every type in ``search()``'s fixed ``type_boost`` set appears, and so do
-#     types outside it, so the boost is observable;
-#   * titles/summaries are half Korean and half English, because the tokenizer,
-#     the query classifier and the concept extractor all branch on script;
-#   * the ``tie:`` block is five rows sharing one timestamp and one type with
-#     nothing to match, which pins the (hits, type_boost, updated_at) → id ASC
-#     tie-break that both engines have to reproduce;
-#   * two workspaces plus NULL-workspace legacy rows cover all three scoping
-#     answers (no scoping / empty set / a specific workspace).
+# Shaped on purpose: every ``type_boost`` type appears and so do types outside it;
+# titles/summaries are half Korean (tokenizer, classifier and extractor all branch
+# on script); the ``tie:`` block is five rows sharing one timestamp, one type and
+# nothing to match, pinning the (hits, boost, updated_at) → id ASC tie-break; two
+# workspaces plus NULL-workspace rows cover all three scoping answers (no scoping
+# / empty set / a specific workspace). The v11.5.1 tail is in the docgen corpus.
 NODES: List[Tuple[str, str, str, str, Dict[str, Any], Optional[str], str]] = [
     ("dec:fusion-alpha", "Decision", "Hybrid retrieval fusion stays alpha weighted",
      "We decided the ranking keeps alpha fusion: lexical rank plus max normalized vector score.",
@@ -191,15 +188,14 @@ NODES: List[Tuple[str, str, str, str, Dict[str, Any], Optional[str], str]] = [
     ("legacy:global-note", "Document", "Legacy global note",
      "A legacy row with no workspace, visible only with include_legacy_global.",
      {}, None, "2026-06-20T10:00:00"),
+    *docgen.DOCGEN_NODES,
 ]
 
 # (chunk_id, parent_node_id, index, node_title, text, chunk_fields, workspace, updated_at)
 #
-# The shape mirrors ``KnowledgeGraphIngestMixin``: every chunk is BOTH a
-# ``Chunk`` node (so the lexical lane can match it and workspace scoping applies)
-# and a ``chunks`` row with its own embedding (so the vector lane returns it and
-# has to roll it up to its parent). Getting that duality wrong is the whole
-# reason chunk-heavy queries are in the query set.
+# The shape mirrors ``KnowledgeGraphIngestMixin``: every chunk is BOTH a ``Chunk``
+# node (lexical lane + workspace scoping) and a ``chunks`` row with its own embedding
+# (vector lane, rolled up to its parent) — the duality chunk-heavy queries pin.
 CHUNKS: List[Tuple[str, str, int, str, str, Dict[str, Any], Optional[str], str]] = [
     ("chunk:handbook:1", "doc:handbook", 0, "handbook.pdf chunk 1",
      "온보딩 체크리스트: 첫째, 폴더를 연결합니다. 둘째, 질문을 합니다. 셋째, 근거를 확인합니다.",
@@ -218,14 +214,75 @@ CHUNKS: List[Tuple[str, str, int, str, str, Dict[str, Any], Optional[str], str]]
      {"start_char": 900}, WS_BETA, "2026-05-28T13:15:02"),
 ]
 
-# (from, to, type)
-EDGES: List[Tuple[str, str, str]] = [
-    ("dec:fusion-alpha", "concept:retrieval", "mentions"),
-    ("dec:fusion-alpha", "concept:ranking", "mentions"),
-    ("task:parity-harness", "dec:rust-foundation", "relates_to"),
-    ("doc:handbook", "page:onboarding", "contains"),
-    ("meeting:weekly", "dec:fusion-alpha", "discusses"),
-    ("person:jiwon", "task:parity-harness", "owns"),
+# (from, to, type, weight, created_at)
+#
+# Weights and timestamps are assigned rather than inherited: ``relationship_search``
+# orders by ``weight DESC, created_at DESC, id ASC`` and ``traverse`` caps every BFS
+# round with ``ORDER BY weight DESC, id ASC``, so an edge set sharing one weight and
+# one clock proves nothing. Three shapes are deliberate — a weight tie broken by
+# ``created_at`` (``org:lattice``), a weight *and* clock tie broken by edge id
+# (``topic:quality``/``deck:review``), and legacy-global endpoints for scoping. The
+# write door canonicalizes types (``relates_to``/``owns`` → ``MENTIONS``), which is
+# why the goldens record uppercase names the fixture never spells.
+EDGES: List[Tuple[str, str, str, float, str]] = [
+    ("dec:fusion-alpha", "concept:retrieval", "mentions", 0.9, "2026-07-01T00:00:00"),
+    ("dec:fusion-alpha", "concept:ranking", "mentions", 0.8, "2026-07-02T00:00:00"),
+    ("task:parity-harness", "dec:rust-foundation", "relates_to", 1.0, "2026-07-03T00:00:00"),
+    ("doc:handbook", "page:onboarding", "contains", 0.7, "2026-07-04T00:00:00"),
+    ("meeting:weekly", "dec:fusion-alpha", "discusses", 0.95, "2026-07-05T00:00:00"),
+    ("person:jiwon", "task:parity-harness", "owns", 0.6, "2026-07-06T00:00:00"),
+    ("person:minseo", "code:build-failure", "owns", 0.6, "2026-07-07T00:00:00"),
+    ("meeting:weekly", "task:onboarding-checklist", "discusses", 0.55, "2026-07-08T00:00:00"),
+    ("meeting:kickoff", "dec:rust-foundation", "discusses", 0.5, "2026-06-01T00:00:00"),
+    ("doc:retrieval-spec", "concept:ranking", "mentions", 0.45, "2026-06-02T00:00:00"),
+    ("concept:retrieval", "concept:ranking", "relates_to", 0.4, "2026-06-03T00:00:00"),
+    ("file:ranking-notes", "dec:fusion-alpha", "relates_to", 0.35, "2026-06-04T00:00:00"),
+    ("code:hybrid-search", "file:ranking-notes", "relates_to", 0.3, "2026-06-05T00:00:00"),
+    # Same weight, different clock → created_at DESC decides.
+    ("org:lattice", "person:jiwon", "contains", 0.25, "2026-06-06T00:00:00"),
+    ("org:lattice", "person:minseo", "contains", 0.25, "2026-06-07T00:00:00"),
+    ("slide:fusion", "concept:retrieval", "mentions", 0.25, "2026-06-08T00:00:00"),
+    # Same weight AND same clock → the id ASC tie-break is the only thing left.
+    ("topic:quality", "deck:review", "relates_to", 0.2, "2026-06-09T00:00:00"),
+    ("deck:review", "meeting:weekly", "relates_to", 0.2, "2026-06-09T00:00:00"),
+    ("task:onboarding-checklist", "page:onboarding", "relates_to", 0.15, "2026-05-01T00:00:00"),
+    ("doc:handbook", "doc:retrieval-spec", "relates_to", 0.1, "2026-05-02T00:00:00"),
+    ("tie:a", "tie:b", "relates_to", 1.0, "2026-07-03T00:00:00"),
+    *docgen.DOCGEN_EDGES,
+]
+
+# ── the conversation corpus (episodic memory, same database file) ────────────
+# (conversation_id, role, content, user_email, nickname, source, timestamp,
+#  workspace_id, organization_id, extra)
+#
+# Every branch the history reads take: two users × three workspaces, NULL and
+# empty-string workspaces (the legacy rows ``_scope_sql`` admits), rows with no
+# ``conversation_id`` (the ``legacy-previous-history`` bucket), a whitespace-only
+# first message (the ``새 대화`` placeholder and its later upgrade), an
+# assistant-first conversation (no upgrade), an empty timestamp (the ``or ""``
+# fallbacks), extra keys ``metadata_json`` merges flat, ko/en content.
+MESSAGES: List[Tuple[Any, ...]] = [
+    ("conv-a", "user", "온보딩 체크리스트 어떻게 시작해?", "jiwon@lattice.ai", "지원", "web", "2026-07-20T09:00:00", WS_ALPHA, "org-1", {}),
+    ("conv-a", "assistant", "먼저 폴더를 연결하세요. 그다음 질문하면 됩니다.", "jiwon@lattice.ai", None, "web", "2026-07-20T09:00:05", WS_ALPHA, "org-1", {}),
+    ("conv-a", "user", "고마워", "jiwon@lattice.ai", "지원", "web", "2026-07-20T09:01:00", WS_ALPHA, "org-1", {"trace_id": "t-1"}),
+    ("conv-b", "user", "How does hybrid retrieval ranking work?", "minseo@lattice.ai", "Minseo", "telegram", "2026-07-21T10:00:00", WS_BETA, "org-1", {}),
+    ("conv-b", "assistant", "The lexical channel scores one over rank.", "minseo@lattice.ai", None, "telegram", "2026-07-21T10:00:07", WS_BETA, "org-1", {"tokens": 42, "cited": ["doc:retrieval-spec"]}),
+    ("conv-c", "user", "   \n  ", None, None, None, "2026-07-22T08:00:00", None, None, {}),
+    ("conv-c", "assistant", "무엇을 도와드릴까요?", None, None, None, "2026-07-22T08:00:05", None, None, {}),
+    ("conv-c", "user", "지난주 회의 기록 보여줘", None, None, None, "2026-07-22T08:01:00", None, None, {}),
+    (None, "user", "이전 대화 기록입니다", "jiwon@lattice.ai", "지원", "web", "2026-06-01T09:00:00", "", None, {}),
+    (None, "assistant", "네, 확인했습니다.", "jiwon@lattice.ai", None, "web", "2026-06-01T09:00:03", "", None, {}),
+    (None, "user", "legacy english message about ranking", None, None, None, "", None, None, {}),
+    ("conv-d", "user", "빌드 실패 원인 알려줘", "minseo@lattice.ai", "Minseo", "vscode", "2026-07-23T11:00:00", WS_ALPHA, "org-1", {}),
+    ("conv-d", "assistant", "컴파일 오류 로그를 확인하세요.", "minseo@lattice.ai", None, "vscode", "2026-07-23T11:00:04", WS_ALPHA, "org-1", {}),
+    ("conv-e", "user", "Ranking ties are broken by node id", "jiwon@lattice.ai", "지원", "web", "2026-07-24T12:00:00", WS_BETA, "org-2", {}),
+    ("conv-e", "assistant", "Yes — id ascending keeps it stable.", "jiwon@lattice.ai", None, "web", "2026-07-24T12:00:05", WS_BETA, "org-2", {}),
+    ("conv-f", "user", "검색 품질을 어떻게 측정하나요? 재현율과 정밀도를 모두 보고 싶고 주간 회의에서 공유할 예정입니다.", None, None, "web", "2026-07-25T13:00:00", None, None, {}),
+    ("conv-f", "assistant", "재현율/정밀도 지표는 분기 리뷰 발표자료에 있습니다.", None, None, "web", "2026-07-25T13:00:05", None, None, {}),
+    ("conv-g", "assistant", "assistant-first conversation", None, None, "web", "2026-07-26T14:00:00", None, None, {}),
+    ("conv-g", "user", "follow up question about ranking", None, None, "web", "2026-07-26T14:00:10", None, None, {}),
+    ("conv-h", "user", "Empty user and empty workspace row", "", "", "", "2026-07-27T15:00:00", "", "", {}),
+    ("conv-h", "assistant", "회의 결정 사항을 정리했습니다.", "", "", "", "2026-07-27T15:00:06", "", "", {}),
 ]
 
 # ── the query set ────────────────────────────────────────────────────────────
@@ -253,25 +310,178 @@ QUERIES: List[Dict[str, Any]] = [
     {"key": "ws_alpha", "query": "hybrid retrieval ranking", "allowed": [WS_ALPHA]},
     {"key": "ws_alpha_legacy", "query": "회의 결정 사항", "allowed": [WS_ALPHA], "legacy": True},
     {"key": "ws_beta", "query": "온보딩 체크리스트", "allowed": [WS_BETA]},
-    # A vector floor nothing clears: the lane returns zero rows, which is the
-    # only way to reach the stale-embedder probe and the lexical-only fusion
-    # label without breaking the store.
+    # A vector floor nothing clears: the stale-embedder probe and the
+    # lexical-only fusion label, without breaking the store.
     {"key": "min_vector_floor", "query": "hybrid retrieval ranking", "min_vector": 0.95},
-    # A small top_k so the rerank window (top_k * 2) is narrower than the
-    # candidate list and the cut is observable.
+    # top_k small enough that the rerank window (top_k * 2) cuts the candidates.
     {"key": "top_k_small", "query": "hybrid retrieval ranking", "top_k": 3},
-    # An explicitly pinned alpha: no policy, so no class, no rewrite and — on a
-    # query that would otherwise be recency-classed — no age decay either.
+    # Pinned alpha: no policy, so no class/rewrite/age decay on a recency query.
     {"key": "alpha_pinned", "query": "지난주 회의 기록", "alpha": 0.2},
-    # A limit below the FTS hit count, so `ORDER BY rank LIMIT ?` decides which
-    # rows exist at all. This is the one place bm25 ordering is observable
-    # (`search()` re-sorts by id afterwards), and therefore the one place a
-    # SQLite version difference between the two runtimes could show up.
+    # A limit below the FTS hit count, so `ORDER BY rank LIMIT ?` decides which rows
+    # exist at all — the one place bm25 ordering (a SQLite version difference) shows.
     {"key": "fts_rank_cut", "query": "Tie candidate", "limit": 2, "top_k": 2},
 ]
 
-#: Texts whose tokenizer output, hash pairs and full vectors pin the Rust
-#: embedding port bit-for-bit (Korean, English, mixed, symbols, digits).
+#: Every branch of the memories section: a workspace hit, one with no kind and an
+#: empty snippet, a non-workspace row to drop, and one past the limit.
+CONTEXT_MEMORIES: Dict[str, Any] = {
+    "results": [
+        {"id": "mem-1", "kind": "preference", "snippet": "답변은 한국어로", "score": 0.91, "source": "workspace"},
+        {"id": "mem-2", "kind": None, "snippet": "", "score": 0.0, "source": "workspace"},
+        {"id": "mem-3", "kind": "decision", "snippet": "ranking keeps alpha fusion", "score": 0.4, "source": "personal"},
+        {"id": "mem-4", "kind": "fact", "snippet": "온보딩은 다섯 걸음", "score": 0.3, "source": "workspace"},
+    ]
+}
+
+#: A ledger with a pathless row, a non-dict row, and more than the ten-row cut.
+CONTEXT_ARTIFACTS: List[Any] = [
+    {"path": "notes/ranking.md", "at": "2026-07-20T09:00:00", "run_id": "run-1"},
+    {"path": "notes/onboarding.md", "run_id": "run-2"},
+    {"path": "", "at": "2026-07-20T09:00:01"},
+    "not-a-dict",
+] + [{"path": f"out/file-{index}.md", "at": None, "run_id": f"r{index}"} for index in range(10)]
+
+#: The Phase-2/3 suites: one spec list per ported entry point.
+SUITES: Dict[str, List[Dict[str, Any]]] = {
+    "relationship": [
+        {"key": "all"},
+        {"key": "by_type_mention", "relationship_type": "mention"},
+        {"key": "by_type_contains", "relationship_type": "CONTAINS"},
+        {"key": "by_type_unknown", "relationship_type": "relates_to"},
+        {"key": "by_node", "node_id": "dec:fusion-alpha"},
+        {"key": "by_query_ko", "query": "회의"},
+        {"key": "by_query_en", "query": "ranking"},
+        {"key": "by_query_meta", "query": "lattice"},
+        {"key": "combined", "node_id": "dec:fusion-alpha", "relationship_type": "mention", "query": "retrieval"},
+        {"key": "limit_one", "limit": 1},
+        {"key": "limit_zero", "limit": 0},
+        {"key": "limit_over", "limit": 500},
+        {"key": "scoped_alpha", "allowed": [WS_ALPHA]},
+        {"key": "scoped_alpha_legacy", "allowed": [WS_ALPHA], "legacy": True},
+        {"key": "scoped_beta", "allowed": [WS_BETA]},
+        {"key": "scoped_empty", "allowed": []},
+        {"key": "no_hit", "query": "zzqq wumpus"},
+    ],
+    "traverse": [
+        # 9 clamps to 4 and -1 clamps to 0; both are on purpose.
+        *[{"key": f"hub_d{depth}", "node_id": "dec:fusion-alpha", "depth": depth}
+          for depth in (0, 1, 2, 3, 9)],
+        {"key": "hub_dneg", "node_id": "dec:fusion-alpha", "depth": -1},
+        {"key": "leaf_d2", "node_id": "code:hybrid-search", "depth": 2},
+        {"key": "isolated", "node_id": "tie:c", "depth": 2},
+        {"key": "limit_two", "node_id": "dec:fusion-alpha", "depth": 3, "limit": 2},
+        {"key": "limit_five", "node_id": "dec:fusion-alpha", "depth": 3, "limit": 5},
+        {"key": "limit_zero", "node_id": "dec:fusion-alpha", "depth": 2, "limit": 0},
+        {"key": "limit_over", "node_id": "dec:fusion-alpha", "depth": 2, "limit": 900},
+        {"key": "org_hub", "node_id": "org:lattice", "depth": 2},
+        {"key": "tie_pair", "node_id": "tie:a", "depth": 2},
+        {"key": "scoped_alpha", "node_id": "dec:fusion-alpha", "depth": 2, "allowed": [WS_ALPHA]},
+        {"key": "scoped_alpha_legacy", "node_id": "dec:fusion-alpha", "depth": 2, "allowed": [WS_ALPHA], "legacy": True},
+        {"key": "scoped_seed_hidden", "node_id": "dec:fusion-alpha", "allowed": [WS_BETA]},
+        {"key": "scoped_empty", "node_id": "dec:fusion-alpha", "allowed": []},
+        {"key": "empty_id", "node_id": ""},
+        {"key": "missing_seed", "node_id": "nope:missing"},
+    ],
+    "graph_search": [
+        {"key": "en_fact", "query": "hybrid retrieval ranking"},
+        {"key": "ko_fact", "query": "회의 결정 사항"},
+        {"key": "person", "query": "who owns the onboarding checklist"},
+        {"key": "code", "query": "빌드 실패 원인"},
+        {"key": "expand0", "query": "hybrid retrieval ranking", "expand_depth": 0},
+        {"key": "expand3", "query": "hybrid retrieval ranking", "expand_depth": 3},
+        {"key": "expand_clamp", "query": "회의 결정 사항", "expand_depth": 9},
+        {"key": "limit_small", "query": "hybrid retrieval ranking", "limit": 3},
+        {"key": "limit_over", "query": "회의 결정 사항", "limit": 500},
+        {"key": "scoped_alpha", "query": "hybrid retrieval ranking", "allowed": [WS_ALPHA]},
+        {"key": "scoped_alpha_legacy", "query": "회의 결정 사항", "allowed": [WS_ALPHA], "legacy": True},
+        {"key": "scoped_empty", "query": "hybrid retrieval ranking", "allowed": []},
+        {"key": "no_hit", "query": "zzqq wumpus nonsense"},
+        {"key": "empty_query", "query": ""},
+    ],
+    "service_hybrid": [
+        {"key": "en_fact", "query": "hybrid retrieval ranking"},
+        {"key": "ko_recency", "query": "지난주 회의 기록"},
+        {"key": "en_code", "query": "vector_search() returns"},
+        {"key": "ko_person", "query": "담당자 누구"},
+        {"key": "en_filler", "query": "  what is   the retrieval specification please  "},
+        # Explicit weights disable BOTH rewrite and age decay on a query that
+        # would otherwise get both — that asymmetry is the contract.
+        {"key": "pinned_recency", "query": "지난주 회의 기록", "weights": {"keyword": 0.5, "vector": 0.3, "graph": 0.2}},
+        {"key": "pinned_partial", "query": "hybrid retrieval ranking", "weights": {"graph": 1.0}},
+        {"key": "pinned_zero", "query": "회의 결정 사항", "weights": {"keyword": 0.0, "vector": 0.0, "graph": 0.0}},
+        {"key": "limit_small", "query": "hybrid retrieval ranking", "limit": 3},
+        {"key": "limit_over", "query": "회의 결정 사항", "limit": 500},
+        {"key": "channel_limits", "query": "hybrid retrieval ranking", "keyword_limit": 5, "vector_limit": 5, "graph_limit": 5},
+        {"key": "scoped_alpha", "query": "hybrid retrieval ranking", "allowed": [WS_ALPHA]},
+        {"key": "scoped_alpha_legacy", "query": "회의 결정 사항", "allowed": [WS_ALPHA], "legacy": True},
+        {"key": "scoped_empty", "query": "hybrid retrieval ranking", "allowed": []},
+        {"key": "no_hit", "query": "zzqq wumpus nonsense"},
+        {"key": "empty_query", "query": ""},
+    ],
+    "history": [
+        {"key": "all"},
+        {"key": "limit_two", "limit": 2},
+        {"key": "limit_zero", "limit": 0},
+        {"key": "conv_a", "conversation_id": "conv-a"},
+        {"key": "conv_missing", "conversation_id": "nope"},
+        {"key": "conv_null", "conversation_id": ""},
+        {"key": "user_jiwon", "user_email": "jiwon@lattice.ai"},
+        {"key": "user_jiwon_strict", "user_email": "jiwon@lattice.ai", "legacy": False},
+        {"key": "user_unknown", "user_email": "ghost@lattice.ai"},
+        {"key": "ws_alpha", "allowed": [WS_ALPHA]},
+        {"key": "ws_alpha_strict", "allowed": [WS_ALPHA], "legacy": False},
+        {"key": "ws_both_strict", "allowed": [WS_ALPHA, WS_BETA], "legacy": False},
+        {"key": "ws_empty_legacy", "allowed": []},
+        {"key": "ws_empty_strict", "allowed": [], "legacy": False},
+        {"key": "user_and_ws", "user_email": "jiwon@lattice.ai", "allowed": [WS_ALPHA], "legacy": False},
+        {"key": "conv_and_user", "conversation_id": "conv-a", "user_email": "jiwon@lattice.ai", "legacy": False},
+        {"key": "ws_blank_only", "allowed": [""], "legacy": False},
+    ],
+    "conversations": [
+        {"key": "all"},
+        {"key": "user_jiwon", "user_email": "jiwon@lattice.ai"},
+        {"key": "user_jiwon_strict", "user_email": "jiwon@lattice.ai", "legacy": False},
+        {"key": "ws_alpha_strict", "allowed": [WS_ALPHA], "legacy": False},
+        {"key": "ws_beta_strict", "allowed": [WS_BETA], "legacy": False},
+        {"key": "ws_empty_strict", "allowed": [], "legacy": False},
+    ],
+    "conversation_messages": [
+        {"key": "conv_a", "conversation_id": "conv-a"},
+        {"key": "legacy_bucket", "conversation_id": "legacy-previous-history"},
+        {"key": "missing", "conversation_id": "nope"},
+        {"key": "scoped_alpha_strict", "conversation_id": "conv-a", "allowed": [WS_ALPHA], "legacy": False},
+        {"key": "legacy_bucket_scoped", "conversation_id": "legacy-previous-history", "allowed": [WS_ALPHA], "legacy": False},
+    ],
+    "history_search": [
+        {"key": "ko_hit", "query": "회의"},
+        {"key": "ko_partial", "query": "체크리스트"},
+        {"key": "en_hit", "query": "ranking"},
+        {"key": "case_insensitive", "query": "RANKING"},
+        {"key": "blank", "query": "   "},
+        {"key": "no_hit", "query": "zzqq"},
+        {"key": "limit_one", "query": "ranking", "limit": 1},
+        {"key": "scoped_strict", "query": "ranking", "allowed": [WS_BETA], "legacy": False},
+    ],
+    "context_assemble": [
+        {"key": "all_seams", "query": "회의 결정 사항", "memories": CONTEXT_MEMORIES, "artifacts": CONTEXT_ARTIFACTS, "notes": "정원 노트: 랭킹은 alpha 융합을 유지한다.", "recent": {"limit": 4}},
+        {"key": "knowledge_only", "query": "hybrid retrieval ranking"},
+        {"key": "no_seams", "query": "hybrid retrieval ranking", "knowledge": False},
+        {"key": "memories_only", "query": "온보딩", "knowledge": False, "memories": CONTEXT_MEMORIES, "memory_limit": 2},
+        {"key": "artifacts_only", "query": "온보딩", "knowledge": False, "artifacts": CONTEXT_ARTIFACTS},
+        {"key": "notes_blank", "query": "온보딩", "knowledge": False, "notes": "   "},
+        {"key": "recent_conversation", "query": "온보딩", "knowledge": False, "recent": {"conversation_id": "conv-a", "limit": 10}},
+        {"key": "recent_personal_workspace", "query": "온보딩", "knowledge": False, "recent": {"workspace_id": "personal", "limit": 6}},
+        {"key": "recent_user_scoped", "query": "온보딩", "knowledge": False, "recent": {"user_email": "jiwon@lattice.ai", "limit": 5}},
+        {"key": "budget_tiny", "query": "회의 결정 사항", "memories": CONTEXT_MEMORIES, "artifacts": CONTEXT_ARTIFACTS, "notes": "정원 노트: 랭킹은 alpha 융합을 유지한다.", "recent": {"limit": 4}, "budget": 20},
+        {"key": "budget_one", "query": "회의 결정 사항", "memories": CONTEXT_MEMORIES, "artifacts": CONTEXT_ARTIFACTS, "notes": "노트", "recent": {"limit": 4}, "budget": 1},
+        {"key": "budget_zero", "query": "회의 결정 사항", "memories": CONTEXT_MEMORIES, "notes": "노트", "budget": 0},
+        {"key": "knowledge_limit_one", "query": "hybrid retrieval ranking", "knowledge_limit": 1},
+    ],
+    # v11.5.1: the document-generation ports, specs and runners alike.
+    **docgen.DOCGEN_SUITES,
+}
+
+#: Texts whose tokenizer output, hashes and vectors pin the embedding port.
 EMBEDDING_TEXTS: List[str] = [
     "hybrid retrieval ranking",
     "회의 결정 사항",
@@ -283,8 +493,7 @@ EMBEDDING_TEXTS: List[str] = [
     "a",
 ]
 
-#: Values that separate CPython's round-half-even from naive scaling, plus the
-#: shapes real fusion arithmetic produces.
+#: Values separating CPython round-half-even from naive scaling, plus real shapes.
 ROUNDING_VALUES: List[float] = [
     0.0, 1.0, 5e-07, 1.5e-06, 2.5e-06, 2.6535895, 1 / 3, 1 / 7,
     0.1234565, 0.1234575, 0.9999999999, 123456.7890625, -0.0000005,
@@ -309,13 +518,16 @@ def pinned_environment() -> Iterator[None]:
 
 @contextmanager
 def frozen_clock() -> Iterator[None]:
-    """Freeze ``hybrid_search``'s ``datetime.now()`` at :data:`FROZEN_NOW`.
+    """Freeze every ported ``datetime.now()`` at :data:`FROZEN_NOW`.
 
-    Only the recency-class age decay reads the clock, and it reads it through
-    the ``datetime`` name that ``hybrid.py`` imported — so rebinding that name
-    is the whole patch.
+    Only the recency decay reads the clock, through the ``datetime`` name its
+    module imported, so rebinding that name is the whole patch. Three modules
+    reach it: the graph-layer and service-layer ``hybrid_search``, and the
+    document-generation search's own 14-day half-life.
     """
+    from lattice_brain.graph import retrieval_docgen as docgen_module
     from lattice_brain.graph.retrieval import hybrid as hybrid_module
+    from latticeai.services import search_service as service_module
 
     frozen = datetime.fromisoformat(FROZEN_NOW)
 
@@ -324,20 +536,23 @@ def frozen_clock() -> Iterator[None]:
         def now(cls, tz=None):  # noqa: ARG003 — mirrors datetime.now's signature
             return frozen
 
-    original = hybrid_module.datetime
-    hybrid_module.datetime = _FrozenDatetime
+    modules = (hybrid_module, service_module, docgen_module)
+    originals = [module.datetime for module in modules]
+    for module in modules:
+        module.datetime = _FrozenDatetime
     try:
         yield
     finally:
-        hybrid_module.datetime = original
+        for module, original in zip(modules, originals, strict=True):
+            module.datetime = original
 
 
 @contextmanager
 def rules_only_extraction() -> Iterator[None]:
     """Force ``_topic_candidates`` down its rule-based path.
 
-    The LLM path needs a bound router, which no fixture run has — but "no router
-    happens to be bound" is an accident, and this contract cannot rest on one.
+    The LLM path needs a bound router no fixture run has, but "no router
+    happens to be bound" is an accident this contract cannot rest on.
     """
     from lattice_brain.graph._kg_common import extraction
 
@@ -352,10 +567,27 @@ def rules_only_extraction() -> Iterator[None]:
 def open_store(db_path: Path):
     """A ``KnowledgeGraphStore`` over ``db_path`` (blobs beside it)."""
     from lattice_brain.graph import KnowledgeGraphStore
-
     return KnowledgeGraphStore(Path(db_path), Path(db_path).parent / "blobs")
 
 
+def open_conversations(db_path: Path):
+    """A ``ConversationStore`` over ``db_path`` (the same file as the graph)."""
+    from lattice_brain.conversations import ConversationStore
+    return ConversationStore(Path(db_path))
+
+
+def write_conversations(db_path: Path) -> None:
+    """Append :data:`MESSAGES` through the real durable-history write path."""
+    conversations = open_conversations(db_path)
+    for conv_id, role, content, email, nick, source, stamp, workspace, org, extra in MESSAGES:
+        conversations.append({
+            "conversation_id": conv_id, "role": role, "content": content,
+            "user_email": email, "user_nickname": nick, "source": source,
+            "timestamp": stamp, "workspace_id": workspace, "organization_id": org,
+            **extra,
+        })
+
+
 def _backdate(conn: sqlite3.Connection, node_id: str, stamp: str) -> None:
     conn.execute("UPDATE nodes SET created_at=?, updated_at=? WHERE id=?", (stamp, stamp, node_id))
     conn.execute(
@@ -388,15 +620,15 @@ def build_store(db_path: Path) -> None:
                 conn, chunk_id=chunk_id, source_node=parent, text=text, metadata=metadata
             )
             _backdate(conn, chunk_id, stamp)
-        for from_node, to_node, edge_type in EDGES:
-            store._upsert_edge(conn, from_node, to_node, edge_type, 1.0, {})
-            conn.execute(
-                "UPDATE edges SET created_at=? WHERE from_node=? AND to_node=?",
-                ("2026-07-01T00:00:00", from_node, to_node),
-            )
+        for from_node, to_node, edge_type, weight, stamp in EDGES:
+            # ``_upsert_edge`` stamps ``created_at`` from the wall clock in BOTH
+            # tables, and the read path is the ``kgv2_edges`` view over ``edges_v2``:
+            # backdating only the legacy table left the ordering moving with the clock.
+            edge_id = store._upsert_edge(conn, from_node, to_node, edge_type, weight, {})
+            conn.execute("UPDATE edges SET created_at=? WHERE id=?", (stamp, edge_id))
+            conn.execute("UPDATE edges_v2 SET created_at=? WHERE id=?", (stamp, edge_id))
         # ``indexed_at`` decides the candidate scan order (and, when the cap
-        # bites, which candidates exist at all), so it is assigned explicitly
-        # rather than inherited from the clock.
+        # bites, which candidates exist), so it is assigned rather than inherited.
         item_ids = [row["item_id"] for row in conn.execute(
             "SELECT item_id FROM vector_embeddings ORDER BY item_id ASC"
         ).fetchall()]
@@ -406,9 +638,10 @@ def build_store(db_path: Path) -> None:
                 "UPDATE vector_embeddings SET indexed_at=? WHERE item_id=?", (stamp, item_id)
             )
     store.record_embedder_fingerprint()
+    write_conversations(db_path)
 
-    # Leave one self-contained file behind: checkpoint the WAL, drop back to a
-    # rollback journal so no -wal/-shm sidecar has to be committed, and compact.
+    # One self-contained file: checkpoint the WAL, drop to a rollback journal
+    # so no sidecar is committed, and compact.
     with sqlite3.connect(str(db_path)) as conn:
         conn.execute("PRAGMA wal_checkpoint(TRUNCATE)")
         conn.execute("PRAGMA journal_mode=DELETE")
@@ -429,17 +662,12 @@ def _allowed(spec: Dict[str, Any]):
 
 ENGINES: Dict[str, Callable[[Any, Dict[str, Any]], Dict[str, Any]]] = {
     "hybrid": lambda store, spec: store.hybrid_search(
-        spec["query"],
-        top_k=spec.get("top_k", 20),
-        alpha=spec.get("alpha"),
-        allowed_workspaces=_allowed(spec),
+        spec["query"], top_k=spec.get("top_k", 20), alpha=spec.get("alpha"),
+        allowed_workspaces=_allowed(spec), min_vector_score=spec.get("min_vector", 0.0),
         include_legacy_global=spec.get("legacy", False),
-        min_vector_score=spec.get("min_vector", 0.0),
     ),
     "keyword": lambda store, spec: store.search(
-        spec["query"],
-        spec.get("limit", 30),
-        allowed_workspaces=_allowed(spec),
+        spec["query"], spec.get("limit", 30), allowed_workspaces=_allowed(spec),
         include_legacy_global=spec.get("legacy", False),
     ),
     "vector": lambda store, spec: store.vector_search(
@@ -454,6 +682,205 @@ def run_engine(store, engine: str, spec: Dict[str, Any]) -> Dict[str, Any]:
         return ENGINES[engine](store, spec)
 
 
+# ── the suites beyond search (v11.5.0/v11.5.1) ───────────────────────────────
+#
+# The Phase-1 engines share one query shape (a query set × an engine set); the
+# later ports do not, so each gets its own spec list and runner, both carried in
+# the manifest so Rust and Python enumerate exactly the same work.
+
+
+class Harness:
+    """Every Python entry point the v11.5.0 goldens are produced from.
+
+    ``require_auth=False`` is the loopback-owner configuration the native routes
+    reproduce: the history scope is whatever the caller passes.
+    """
+
+    def __init__(self, db_path: Path):
+        from latticeai.runtime.history_runtime import build_history_query_runtime
+        from latticeai.services.chat_service import ChatService
+        from latticeai.services.search_service import SearchService
+        self.store = open_store(db_path)
+        self.service = SearchService(graph_store=self.store)
+        self.conversations = open_conversations(db_path)
+        self.history_runtime = build_history_query_runtime(
+            conversations=self.conversations,
+            workspace_service=None,
+            require_auth=False,
+            logging=logging,
+        )
+        self.chat = ChatService(store=None, get_history=self.history_runtime["get_history"])
+
+
+def _allowed_list(spec: Dict[str, Any]):
+    """``allowed`` as the graph layer wants it: ``None`` or a set."""
+    allowed = spec.get("allowed")
+    return None if allowed is None else set(allowed)
+
+
+def _history_scope(spec: Dict[str, Any]) -> Dict[str, Any]:
+    """The identity/workspace scope every history read takes.
+
+    ``include_legacy_global`` defaults to ``True`` — ``ConversationStore``'s own
+    default, the opposite of the graph layer's and an asymmetry a port gets wrong.
+    """
+    allowed = spec.get("allowed")
+    return {
+        "user_email": spec.get("user_email"),
+        "allowed_workspaces": None if allowed is None else list(allowed),
+        "include_legacy_global": spec.get("legacy", True),
+    }
+
+
+def _run_relationship(h: Harness, spec: Dict[str, Any]) -> Dict[str, Any]:
+    return h.store.relationship_search(
+        query=spec.get("query", ""), node_id=spec.get("node_id", ""),
+        relationship_type=spec.get("relationship_type", ""), limit=spec.get("limit", 30),
+        allowed_workspaces=_allowed_list(spec),
+        include_legacy_global=spec.get("legacy", False),
+    )
+
+
+def _run_traverse(h: Harness, spec: Dict[str, Any]) -> Dict[str, Any]:
+    try:
+        return h.store.traverse(
+            spec.get("node_id", ""), depth=spec.get("depth", 1),
+            limit=spec.get("limit", 100), allowed_workspaces=_allowed_list(spec),
+            include_legacy_global=spec.get("legacy", False),
+        )
+    except ValueError as exc:
+        # The two documented refusals (blank id, seed invisible to the caller's
+        # scope) are contract, so they are recorded rather than skipped; a payload
+        # never carries an "error" key, so the golden stays unambiguous.
+        return {"error": str(exc)}
+
+
+def _run_graph_search(h: Harness, spec: Dict[str, Any]) -> Dict[str, Any]:
+    return h.service.graph_search(
+        spec["query"], limit=spec.get("limit", 30),
+        expand_depth=spec.get("expand_depth", 1), allowed_workspaces=_allowed_list(spec),
+        include_legacy_global=spec.get("legacy", False),
+    )
+
+
+def _run_service_hybrid(h: Harness, spec: Dict[str, Any]) -> Dict[str, Any]:
+    return h.service.hybrid_search(
+        spec["query"], limit=spec.get("limit", 30),
+        keyword_limit=spec.get("keyword_limit", 30),
+        vector_limit=spec.get("vector_limit", 30),
+        graph_limit=spec.get("graph_limit", 30), weights=spec.get("weights"),
+        allowed_workspaces=_allowed_list(spec),
+        include_legacy_global=spec.get("legacy", False),
+    )
+
+
+def _run_history(h: Harness, spec: Dict[str, Any]) -> List[Dict[str, Any]]:
+    return h.conversations.history(
+        conversation_id=spec.get("conversation_id"), limit=spec.get("limit"),
+        **_history_scope(spec),
+    )
+
+
+def _run_conversations(h: Harness, spec: Dict[str, Any]) -> List[Dict[str, Any]]:
+    history = h.history_runtime["get_history"](**_history_scope(spec))
+    return h.history_runtime["group_history_conversations"](history)
+
+
+def _run_conversation_messages(h: Harness, spec: Dict[str, Any]) -> List[Dict[str, Any]]:
+    return h.history_runtime["get_conversation_messages"](
+        spec["conversation_id"], **_history_scope(spec)
+    )
+
+
+def _run_history_search(h: Harness, spec: Dict[str, Any]) -> List[Dict[str, Any]]:
+    return h.chat.search_history(
+        spec["query"], scope=_history_scope(spec), limit=spec.get("limit", 30),
+        conversation_title=h.history_runtime["conversation_title"],
+    )
+
+
+def _context_seams(h: Harness, spec: Dict[str, Any]) -> Dict[str, Any]:
+    """The seam set for one context spec — data seams plus the real engines.
+
+    ``memories`` / ``artifacts`` / ``notes`` are *data* seams: the payload is the
+    spec, so both runtimes feed the assembler the same bytes and what is under
+    test is the assembler. ``knowledge`` and ``recent`` are real engines — the
+    service-layer hybrid search and the durable history reader.
+    """
+    from latticeai.api.chat_helpers import build_recent_chat_context
+
+    # Signatures matter: the assembler inspects them to decide which context
+    # fields a seam may be handed, so each one declares exactly what it accepts.
+    seams: Dict[str, Any] = {}
+    if spec.get("memories") is not None:
+        memories = spec["memories"]
+        seams["memory_recall"] = (
+            lambda query, *, user_email=None, workspace_id=None, limit=5: memories
+        )
+    if spec.get("artifacts") is not None:
+        artifacts = spec["artifacts"]
+        seams["recent_artifacts"] = (
+            lambda *, user_email=None, conversation_id=None, workspace_id=None: artifacts
+        )
+    if spec.get("knowledge", True):
+        # Loopback trust: no workspace scoping, exactly as on the native route.
+        seams["hybrid_search"] = (
+            lambda query, *, limit=5, user_email=None, workspace_id=None:
+            h.service.hybrid_search(query, limit=limit)
+        )
+    if spec.get("notes") is not None:
+        notes = spec["notes"]
+        seams["notes_context"] = lambda query, *, user_email=None, workspace_id=None: notes
+    if spec.get("recent") is not None:
+        recent = spec["recent"]
+        seams["recent_chat"] = (
+            lambda *, user_email=None, conversation_id=None, workspace_id=None:
+            build_recent_chat_context(
+                get_history=h.history_runtime["get_history"],
+                limit=recent.get("limit", 10),
+                include_image_missing_replies=recent.get("images", True),
+                user_email=recent.get("user_email"),
+                conversation_id=recent.get("conversation_id"),
+                workspace_id=recent.get("workspace_id"),
+            )
+        )
+    return seams
+
+
+def _run_context_assemble(h: Harness, spec: Dict[str, Any]) -> Dict[str, Any]:
+    from lattice_brain.context import ContextAssembler
+
+    assembled = ContextAssembler(**_context_seams(h, spec)).assemble(
+        spec["query"], user_email=spec.get("user_email"),
+        workspace_id=spec.get("workspace_id"),
+        conversation_id=spec.get("conversation_id"), budget=spec.get("budget", 2000),
+        memory_limit=spec.get("memory_limit", 5),
+        knowledge_limit=spec.get("knowledge_limit", 5),
+    )
+    return {"text": assembled.text, "approx_tokens": assembled.approx_tokens,
+            "trace": assembled.trace()}
+
+
+SUITE_RUNNERS: Dict[str, Callable[[Harness, Dict[str, Any]], Any]] = {
+    "relationship": _run_relationship,
+    "traverse": _run_traverse,
+    "graph_search": _run_graph_search,
+    "service_hybrid": _run_service_hybrid,
+    "history": _run_history,
+    "conversations": _run_conversations,
+    "conversation_messages": _run_conversation_messages,
+    "history_search": _run_history_search,
+    "context_assemble": _run_context_assemble,
+    **docgen.DOCGEN_RUNNERS,
+}
+
+
+def run_suite(harness: Harness, suite: str, spec: Dict[str, Any]) -> Any:
+    """Run one suite spec under the frozen clock and the rule-based extractor."""
+    with frozen_clock(), rules_only_extraction():
+        return SUITE_RUNNERS[suite](harness, spec)
+
+
 def golden_path(engine: str, key: str) -> Path:
     return GOLDEN_DIR / f"{engine}__{key}.json"
 
@@ -464,10 +891,8 @@ def golden_payload(engine: str, spec: Dict[str, Any], result: Dict[str, Any]) ->
         "key": spec["key"],
         "query": spec["query"],
         "params": {
-            "top_k": spec.get("top_k", 20),
-            "alpha": spec.get("alpha"),
-            "limit": spec.get("limit", 30),
-            "min_score": spec.get("min_score", 0.0),
+            "top_k": spec.get("top_k", 20), "alpha": spec.get("alpha"),
+            "limit": spec.get("limit", 30), "min_score": spec.get("min_score", 0.0),
             "min_vector_score": spec.get("min_vector", 0.0),
             "allowed_workspaces": spec.get("allowed"),
             "include_legacy_global": spec.get("legacy", False),
@@ -476,6 +901,15 @@ def golden_payload(engine: str, spec: Dict[str, Any], result: Dict[str, Any]) ->
     }
 
 
+def suite_payload(suite: str, spec: Dict[str, Any], result: Any) -> Dict[str, Any]:
+    """One suite golden: the spec that produced it, verbatim, and the answer.
+
+    The spec rides along rather than being flattened into a fixed ``params``
+    block — these entry points share no parameter shape.
+    """
+    return {"suite": suite, "key": spec["key"], "spec": spec, "result": result}
+
+
 def _dump(path: Path, payload: Any) -> None:
     path.write_text(
         json.dumps(payload, ensure_ascii=False, sort_keys=True, indent=2) + "\n",
@@ -511,12 +945,11 @@ def embeddings_golden() -> Dict[str, Any]:
 
 def rounding_golden() -> List[Dict[str, float]]:
     """``round(x, 6)`` for values where the tie rule is observable."""
-    return [{"input": value, "expected": round(value, 6)} for value in ROUNDING_VALUES]
+    return [{"input": v, "expected": round(v, 6)} for v in ROUNDING_VALUES]
 
 
 def manifest() -> Dict[str, Any]:
     from lattice_brain.embeddings import LocalEmbeddingModel
-
     model = LocalEmbeddingModel()
     return {
         "frozen_now": FROZEN_NOW,
@@ -525,6 +958,7 @@ def manifest() -> Dict[str, Any]:
         "embedding_dim": model.dim,
         "engines": sorted(ENGINES),
         "queries": QUERIES,
+        "suites": {suite: SUITES[suite] for suite in sorted(SUITES)},
         "pinned_env": PINNED_ENV,
     }
 
@@ -543,12 +977,16 @@ def main() -> int:
                 result = run_engine(store, engine, spec)
                 _dump(golden_path(engine, spec["key"]), golden_payload(engine, spec, result))
                 written += 1
+        harness = Harness(STORE_PATH)
+        for suite in sorted(SUITES):
+            for spec in SUITES[suite]:
+                result = run_suite(harness, suite, spec)
+                _dump(golden_path(suite, spec["key"]), suite_payload(suite, spec, result))
+                written += 1
         _dump(GOLDEN_DIR / "embeddings_golden.json", embeddings_golden())
         _dump(GOLDEN_DIR / "rounding_golden.json", rounding_golden())
         _dump(GOLDEN_DIR / "manifest.json", manifest())
-    # ``KnowledgeGraphStore.__init__`` creates its blob directory eagerly; the
-    # fixture has no blobs, and an empty directory beside a committed artefact is
-    # just a thing for the next reader to wonder about.
+    # The store creates its blob directory eagerly and the fixture has no blobs.
     blob_dir = STORE_PATH.parent / "blobs"
     if blob_dir.is_dir() and not any(blob_dir.iterdir()):
         blob_dir.rmdir()