# ============================================================ # 参数优化闭环黄金测试（矩阵 87 行 · 方向 E） # 覆盖：集隔离、回放（沙盒重排）、灰度受限生效、退化自动回滚、 # 验证通过全量生效、敏感性分析衔接（候选方向） # ============================================================ from __future__ import annotations import json import pytest from server.agent_core.param_opt import ParameterOptimizer, optimize from server.aps_domain.sensitivity import run_sensitivity from server.state.seed import seed_world from server.timeutil import add_minutes, fmt_date, today0 def _tight_world(): """6 单紧张交期世界：基线（冻结 24h）总延期 > 0，参数方向稳定可复现。 构造口径与 test_params.py 的 _two_orders_world 一致（在演示主数据上 覆盖订单集），qty=300 使单产线满负荷，2/3/4/5 天交期产生延期压力。 """ w = seed_world() base = today0() order_date = fmt_date(base) product = next(m for m in w["materials"] if m["type"] == "FINISHED_PRODUCT") sos = [] for i in range(6): oid = i + 1 level = "VIP" if i % 2 == 0 else "C" due = fmt_date(add_minutes(base, (2, 2, 3, 3, 4, 5)[i] * 24 * 60)) sos.append({ "id": oid, "orderNo": f"SO-PT-{oid}", "customerId": f"C{oid}", "customerName": f"客户{level}", "customerLevel": level, "orderDate": order_date, "deliveryDate": due, "priority": 5, "manualPriority": None, "status": "CONFIRMED", "source": "MANUAL", "specialRequirements": "", "totalAmount": 1000, "isRush": False, "rushStrategy": None, "changes": [], "createdBy": "test", "createdAt": order_date + " 09:00", "updatedAt": order_date + " 09:00", "items": [{ "id": oid * 10, "orderId": oid, "lineNo": 1, "productId": product["id"], "productName": product["name"], "productCode": product["code"], "quantity": 300, "unit": "件", "bomVersion": "V1.0", "routingVersion": "V1.0", "status": "PENDING", "note": "", }], }) w["salesOrders"] = sos w["scheduleVersions"] = [] w["productionOrders"] = [] w["workOrders"] = [] w["conflicts"] = [] return w def test_split_orders_isolation(): """训练/验证集隔离：互斥、并集全量、同种子可复现、不同种子不同切分。""" w = _tight_world() opt = ParameterOptimizer(seed=20260802) s1 = opt.split_orders(w) s2 = opt.split_orders(w) # 同种子 → 相同切分 assert s1 == s2 assert not (set(s1["train"]) & set(s1["validation"])) all_ids = {so["id"] for so in w["salesOrders"]} assert set(s1["train"]) | set(s1["validation"]) == all_ids assert len(s1["train"]) >= 1 and len(s1["validation"]) >= 1 s3 = opt.split_orders(w, seed=42) # 不同种子 → 不同切分 assert s1["train"] != s3["train"] def test_replay_sandbox_no_mutation_and_directions(): """回放：沙盒重排零副作用；改善/退化方向与引擎信号一致；越界参数拒绝。""" w = _tight_world() before = json.dumps(w, ensure_ascii=False, sort_keys=True) opt = ParameterOptimizer() improved = opt.replay(w, {"freezeWindowHours": 0.0}) assert improved["degraded"] is False assert improved["deltas"]["tardiness"] < -1 # 冻结 0h 相对 24h 明显改善延期 degraded = opt.replay(w, {"freezeWindowHours": 48.0}) assert degraded["degraded"] is True assert degraded["deltas"]["tardiness"] > 1 # 冻结 48h 明显恶化延期 after = json.dumps(w, ensure_ascii=False, sort_keys=True) assert before == after # 回放不写主干 with pytest.raises(ValueError): opt.replay(w, {"freezeWindowHours": 999.0}) # 越界参数拒绝 with pytest.raises(ValueError): opt.replay(w, {"unknownKey": 1.0}) # 未知键拒绝 def test_gray_restricted_effect(): """灰度受限生效：命中作用域的查询才拿到候选值，其余保持基线。""" w = _tight_world() opt = ParameterOptimizer() g = opt.promote_gray(w, {"freezeWindowHours": 0.0}, scope={"lineIds": [1]}) assert g["accepted"] is True assert g["experiment"]["status"] == "GRAY" assert opt.active_params(w, {"lineIds": [1]})["freezeWindowHours"] == 0.0 assert opt.active_params(w, {"lineIds": [2]})["freezeWindowHours"] == 24.0 assert opt.active_params(w)["freezeWindowHours"] == 24.0 # 无作用域不扩散 # 灰度阶段不动 scheduleParams（全量写入在 finalize） assert w["scheduleParams"]["freezeWindowHours"] == 24.0 def test_replay_gate_rejects_degrading_candidate(): """回放门禁：训练集回放退化的候选在灰度前被拦截，世界零写入。""" w = _tight_world() opt = ParameterOptimizer() g = opt.promote_gray(w, {"freezeWindowHours": 48.0}) assert g["accepted"] is False assert g["reason"] == "replay-degraded" assert g["replay"]["degraded"] is True assert len(w.get("paramExperiments") or []) == 0 def test_rollback_on_degradation(): """效果退化自动回滚：金丝雀灰度后验证集回放退化 → 自动回滚到上一版本。""" w = _tight_world() opt = ParameterOptimizer() g = opt.promote_gray(w, {"freezeWindowHours": 48.0}, scope={"lineIds": [1]}, require_replay_pass=False) assert g["accepted"] is True assert g["experiment"]["trainReplay"]["degraded"] is True rec = opt.finalize(w, g["experiment"]["id"]) assert rec["status"] == "ROLLED_BACK" assert rec["rollbackReason"] == "validation-degraded" assert "validationReplay" in rec and rec["validationReplay"]["degraded"] is True # 回滚后：受限参数不再生效，基线参数保持 assert opt.active_params(w, {"lineIds": [1]})["freezeWindowHours"] == 24.0 assert w["scheduleParams"]["freezeWindowHours"] == 24.0 assert len(opt.history(w)) == 1 def test_finalize_full_apply_and_manual_rollback(): """验证通过 → 全量生效（留 before/after 证据）；显式回滚恢复上一版本。""" w = _tight_world() opt = ParameterOptimizer() g = opt.promote_gray(w, {"freezeWindowHours": 0.0}, scope={"lineIds": [1]}) rec = opt.finalize(w, g["experiment"]["id"]) assert rec["status"] == "FULL" assert w["scheduleParams"]["freezeWindowHours"] == 0.0 assert rec["fullApply"]["before"]["freezeWindowHours"] == 24.0 assert rec["fullApply"]["after"]["freezeWindowHours"] == 0.0 rb = opt.rollback(w, g["experiment"]["id"]) assert rb["status"] == "ROLLED_BACK" assert w["scheduleParams"]["freezeWindowHours"] == 24.0 # 恢复上一参数版本 def test_candidates_from_sensitivity(): """与敏感性分析衔接：Tornado 产出 → 候选方向，且候选回放不退化。""" w = _tight_world() opt = ParameterOptimizer() report = run_sensitivity(w, strategy="COMPREHENSIVE") cands = opt.candidates_from_sensitivity(report) assert cands, "敏感性报告应产出候选" # 摆幅最大因子优先：冻结窗口（低档 0h 改善延期） assert cands[0]["factorId"] == "freezeWindowHours" assert cands[0]["candidate"]["freezeWindowHours"] == 0.0 # 效率因子高档（×1.2）方向 eff = next(c for c in cands if c["factorId"] == "lineEfficiency") assert eff["candidate"]["efficiencyScale"] == 1.2 # 每个候选按改善方向回放都不退化 for c in cands[:2]: r = opt.replay(w, c["candidate"]) assert r["degraded"] is False def test_pipeline_end_to_end(): """一键闭环：敏感性取方向 → 逐候选灰度→收口，产出实验记录。""" w = _tight_world() out = optimize(w, scope={"lineIds": [1]}) assert out["summary"]["total"] >= 3 assert out["summary"]["rolledBack"] == 0 # 敏感性方向候选不应退化 assert out["summary"]["active"] >= 1 assert any(r["stage"] == "full" for r in out["results"]) assert len(w.get("paramExperiments") or []) == out["summary"]["total"] # ============================================================ # 矩阵 87 补充：线上观测回调单元（观测落盘 / 基线 / 连续劣化自动回滚 / 审计） # ============================================================ def _activate_full(w, candidate=None): """helper: 灰度→finalize 升级为 FULL，返回实验 id。""" opt = ParameterOptimizer() cand = candidate or {"freezeWindowHours": 0.0} g = opt.promote_gray(w, cand) assert g["accepted"] is True rec = opt.finalize(w, g["experiment"]["id"]) assert rec["status"] == "FULL" return g["experiment"]["id"] def test_observation_persisted_and_self_calibrating_baseline(): """观测回调：观测落盘 paramObservations；首个观测自校准为生效前基线。""" w = _tight_world() opt = ParameterOptimizer() exp_id = _activate_full(w) rec0 = w["paramExperiments"][0] assert rec0["productionBaseline"] is None # 激活时无观测 out = opt.record_observation(w, {"tardiness": 5.0}) assert out["autoRolledBack"] == [] assert len(w["paramObservations"]) == 1 obs = w["paramObservations"][0] assert obs["kpi"] == {"tardiness": 5.0} assert obs["evaluations"][0]["experimentId"] == exp_id assert obs["evaluations"][0]["degraded"] is False # 首个观测即基线，不算劣化 assert rec0["productionBaseline"] == {"tardiness": 5.0} def test_observation_uses_pre_activation_baseline(): """激活前已有生产观测：finalize 捕获最近一次观测为生效前基线。""" w = _tight_world() opt = ParameterOptimizer() opt.record_observation(w, {"tardiness": 10.0}) # 生产基线观测（激活前） g = opt.promote_gray(w, {"freezeWindowHours": 0.0}) rec = opt.finalize(w, g["experiment"]["id"]) assert rec["productionBaseline"] == {"tardiness": 10.0} assert rec["degradedStreak"] == 0 def test_consecutive_degradation_auto_rollback_with_audit(): """连续劣化自动回滚：streak 达阈值 → ROLLED_BACK + 恢复基线参数 + 审计。""" w = _tight_world() opt = ParameterOptimizer(max_consecutive_degraded=2) exp_id = _activate_full(w) opt.record_observation(w, {"tardiness": 10.0}) # 基线自校准 opt.record_observation(w, {"tardiness": 50.0}) # 劣化 1 rec = w["paramExperiments"][0] assert rec["degradedStreak"] == 1 assert rec["status"] == "FULL" out = opt.record_observation(w, {"tardiness": 80.0}) # 劣化 2 → 回滚 assert out["autoRolledBack"] == [exp_id] rec = w["paramExperiments"][0] assert rec["status"] == "ROLLED_BACK" assert rec["rollbackReason"] == "production-degraded" assert rec["autoRollback"]["degradedStreak"] == 2 assert w["scheduleParams"]["freezeWindowHours"] == 24.0 # 恢复上一版本 events = [e for e in w["auditEvents"] if e["action"] == "param.experiment.auto_rolled_back"] assert len(events) == 1 assert events[0]["target"] == {"type": "PARAM_EXPERIMENT", "id": exp_id} assert events[0]["power"] == "P1" assert events[0]["rationale"]["degradedStreak"] == 2 assert events[0]["rationale"]["reason"] == "production-degraded" assert len(w["paramObservations"]) == 3 def test_observation_recovery_resets_streak(): """观测恢复：非劣化观测清零连败计数，不触发回滚。""" w = _tight_world() opt = ParameterOptimizer(max_consecutive_degraded=2) _activate_full(w) opt.record_observation(w, {"tardiness": 10.0}) opt.record_observation(w, {"tardiness": 60.0}) # 劣化 1 opt.record_observation(w, {"tardiness": 8.0}) # 恢复 → 清零 rec = w["paramExperiments"][0] assert rec["degradedStreak"] == 0 assert rec["status"] == "FULL" assert w["scheduleParams"]["freezeWindowHours"] == 0.0 assert [e["action"] for e in w["auditEvents"]] == [] # 无自动回滚审计 assert len(w["paramObservations"]) == 3 def test_observation_monitors_only_active_full(): """只监控 ACTIVE/FULL：GRAY 与已回滚实验不参与观测评估。""" w = _tight_world() opt = ParameterOptimizer() g = opt.promote_gray(w, {"freezeWindowHours": 0.0}) # GRAY out = opt.record_observation(w, {"tardiness": 10.0}) assert out["observation"]["evaluations"] == [] assert w["paramExperiments"][0]["status"] == "GRAY" opt.finalize(w, g["experiment"]["id"]) # → FULL opt.rollback(w, g["experiment"]["id"]) # → ROLLED_BACK out2 = opt.record_observation(w, {"tardiness": 100.0}) assert out2["observation"]["evaluations"] == [] assert out2["autoRolledBack"] == []