"""K05 的字段基线 B:首次接受导入时把**被接受的导入值**按 metric_id + field_path 存成不可变工件。 这份文件盯住任务书 4.7 的三件硬要求: 1. **首次接受导入就是 B 的那一次。** 导入成功规范化、落进初始草稿图的那一刻,被接受的字段值(含引用原文这类不可变 引用)按 metric_id + field_path 存下来,逐格带值哈希,并连着来源快照与 accepted_revision。读得回来——不是只 存一个哈希(哈希还原不了 B)。 2. **人工改 L 不动 B。** 字段操作(K03)改的是当前内容;B 还是当初被接受的那一份,一个字不动(也不因为改过了就 把当前值写回基线)。 3. **旧资产 baseline_unknown。** 保存基线之前建的草稿没有引用:如实标 unknown,不拿它现在的内容倒填历史基线。 模型一律注入假模型、试跑换成假的;这里不连真实业务库。 """ from __future__ import annotations import copy import uuid import pytest from sqlalchemy import delete, select from app.db.base import SessionLocal from app.db.models import CaliberDraft from app.ledger import fieldbaseline from tests.test_ledger_docdraft_api import ( # noqa: F401 - fixture 和起草用例共用 client, datasource, documents, extract, fake_model, fake_trial, plan_reply, ) async def accepted_draft(client, documents, datasource, monkeypatch) -> dict: """走一次真实起草(假模型 + 假试跑),返回草稿详情。""" doc = await documents("baseline-" + uuid.uuid4().hex[:8]) source_id = await datasource() fake_model(monkeypatch, plan_reply(f"{doc}-2")) body = await extract(client, doc, source_id) return (await client.get(f"/api/ledger/drafts/{body['draft_id']}")).json() @pytest.fixture(autouse=True) def _fake_trial(fake_trial): """本模块的草稿都走接口起草:试跑必须换成假的,不然会去打夹具里不存在的库。 (fake_trial 是从别的模块 import 进来的普通夹具,它自带的 autouse 在本模块不生效;这里替本模块的每个用例把它拉起来。)""" return fake_trial async def test_the_first_accepted_import_is_stored_as_a_readable_value_level_baseline( client, documents, datasource, monkeypatch): """接受导入那一刻存下的是**逐字段的值**:读得回来、哈希对得上、连着来源快照和接受版本。""" made = await accepted_draft(client, documents, datasource, monkeypatch) baseline = made["field_baseline"] assert baseline["status"] == "known" and baseline["accepted_revision"] == 1 assert baseline["metric_ids"] == ["m1"] and baseline["field_count"] >= 5 # payload 里另有一份追加式的历史(这次接受是第一条) assert [item["artifact_id"] for item in made["field_baseline_history"]] == [baseline["artifact_id"]] assert baseline["source_snapshot_artifact_id"] == made["source_snapshot"]["artifact_id"] # 按引用读回工件:内容就是当时被接受的那一份,逐格有值、有哈希(不是一串哈希) stored = fieldbaseline.read({"artifact_id": baseline["artifact_id"]}) assert stored["card_id"] == made["card_id"] and stored["accepted_revision"] == 1 fields = stored["metrics"]["m1"]["fields"] assert fields["/name"]["value"] == "已支付退款率" assert fields["/unit"]["value"] == "%" and fields["/decimals"]["value"] == 2 assert fields["/definition/text"]["value"] == "已支付订单里发生退款的比例" assert fields["/synonyms"]["value"] == ["退款率"] # 引用原文是「不可变值引用」:整条留着(哪一段、原话、span),重导入要拿它比 citation = fields["/definition/citations"]["value"][0] assert citation["quote"] == "只算已支付的订单" and citation["span"] == [8, 16] assert all(cell["hash"] for cell in fields.values()) # 同一份材料只存一份:引用里的 content_hash 就是工件 id(内容寻址) assert baseline["content_hash"] == baseline["artifact_id"] assert fieldbaseline.values({"artifact_id": baseline["artifact_id"]}, "m1")["/name"] == "已支付退款率" async def test_editing_fields_moves_l_and_never_touches_b(client, documents, datasource, monkeypatch): """人工按字段改(K03)只改 L:B 的引用、值、哈希一个字不动,覆盖判断跟着定义指纹走。""" made = await accepted_draft(client, documents, datasource, monkeypatch) before_ref = copy.deepcopy(made["field_baseline"]) before_fields = fieldbaseline.values({"artifact_id": before_ref["artifact_id"]}) assert made["validation"]["covers_current"] is True changed = await client.patch(f"/api/ledger/drafts/{made['id']}", json={ "operations": [{"op": "set", "metric_id": "m1", "path": "/name", "value": "退款率(人工改过)"}, {"op": "set", "metric_id": "m1", "path": "/definition/text", "value": "人工写的新说明"}], "expected_revision": made["revision"], "request_id": "baseline-edit-1"}) assert changed.status_code == 200, changed.text assert changed.json()["change"]["applied"] is True after = (await client.get(f"/api/ledger/drafts/{made['id']}")).json() assert after["metrics"][0]["name"] == "退款率(人工改过)" # L 变了 assert after["metrics"][0]["definition"]["text"] == "人工写的新说明" assert after["field_baseline"] == before_ref # B 的引用原样 assert fieldbaseline.values({"artifact_id": after["field_baseline"]["artifact_id"]}) == before_fields # B 里还是当初被接受的那一份,不是「现在长什么样」 assert fieldbaseline.values({"artifact_id": after["field_baseline"]["artifact_id"]}, "m1")["/name"] \ == "已支付退款率" # 定义改了,旧验证记录不再覆盖当前版本(覆盖是读的时候按指纹现算的) assert after["validation"]["covers_current"] is False assert after["validation"]["verdict"] == "stale" async def test_an_old_draft_reads_as_baseline_unknown_and_is_never_backfilled(client): """没有可还原基线的老草稿:如实说 unknown,不拿它现在的内容当 B,也不因为读一次就补一份出来。""" draft_id = "old-" + uuid.uuid4().hex[:10] async with SessionLocal() as session: session.add(CaliberDraft(id=draft_id, title="老草稿", status="path_pending", payload={"metrics": [{"id": "m1", "name": "旧指标", "definition": {"text": "老草稿里现在写的说明"}}]})) await session.commit() try: detail = (await client.get(f"/api/ledger/drafts/{draft_id}")).json() assert detail["field_baseline"]["status"] == "baseline_unknown" assert "artifact_id" not in detail["field_baseline"] assert "无法还原" in detail["field_baseline"]["reason"] listed = (await client.get(f"/api/ledger/drafts/{draft_id}/validations")).json() assert listed["field_baseline"]["status"] == "baseline_unknown" assert listed["records"] == [] and listed["summary"]["verdict"] == "none" # 再读一次也不会「补」出一份基线:草稿里现在有指标,B 仍旧是 unknown(当前值不是历史基线) again = (await client.get(f"/api/ledger/drafts/{draft_id}")).json() assert again["field_baseline"]["status"] == "baseline_unknown" assert fieldbaseline.ref_of(again) is None finally: async with SessionLocal() as session: row = await session.get(CaliberDraft, draft_id) if row is not None: await session.delete(row) await session.commit() async def test_the_baseline_is_content_addressed_and_keeps_every_acceptance(client, documents, datasource, monkeypatch): """同一份被接受的材料只存一份工件;每次接受追加一条引用,当前 B 是最后一条。""" made = await accepted_draft(client, documents, datasource, monkeypatch) ref = dict(made["field_baseline"]) metrics = [{"id": "m1", "name": "已支付退款率", "unit": "%", "decimals": 2, "definition": {"text": "已支付订单里发生退款的比例"}}] again = await fieldbaseline.put_metrics(metrics=metrics, card_id=ref["card_id"], workflow_id=ref["workflow_id"], accepted_revision=1, draft_id=made["id"]) assert again is not None and again["artifact_id"] != ref["artifact_id"] # 字段少了,内容不同 → 另一份工件 assert ref["artifact_id"] not in (again["artifact_id"],) history = fieldbaseline.history({**made, fieldbaseline.HISTORY_KEY: [ref, again]}) assert [item["artifact_id"] for item in history] == [ref["artifact_id"], again["artifact_id"]] async def test_a_baseline_that_cannot_be_stored_leaves_the_draft_readable(monkeypatch, client, documents, datasource): """基线落库失败不把草稿判死:照实说没有基线(baseline_unknown),不假装有。""" from app.core import artifact_store async def boom(*args, **kwargs): raise OSError("工件库写不进去") monkeypatch.setattr(artifact_store, "put_json", boom) broken = await accepted_draft(client, documents, datasource, monkeypatch) assert broken["status"] == "ready" # 草稿照出 assert broken["field_baseline"]["status"] == "baseline_unknown" assert "无法还原" in broken["field_baseline"]["reason"] @pytest.fixture(autouse=True) async def _cleanup_records(): """本模块造的草稿由 documents 夹具统一清掉;这里兜一下记录表,避免留下跨用例的残行。""" yield from app.db.models import CaliberReviewDecision, CaliberValidationRecord async with SessionLocal() as session: drafts = list((await session.execute(select(CaliberDraft.id))).scalars()) if drafts: await session.execute(delete(CaliberValidationRecord).where( CaliberValidationRecord.draft_id.in_(drafts))) await session.execute(delete(CaliberReviewDecision).where( CaliberReviewDecision.draft_id.in_(drafts))) await session.commit()