feat: preserve Quant OS target-package vertical slice

gate_scorecard.json

@@ -33,22 +33,31 @@
   "current_delivery": {
     "score": null,
     "status": "not_scored",
-    "reason": "Quant OS 已有可执行本地流水线、真实 Qlib fixture smoke、带校验但不完整的 Tushare→Qlib 技术 smoke，以及一次完成的真实聚宽 hosted smoke；但仍缺完整复权 PIT 数据、同输入聚宽比较、真实 QMT、券商影子盘、连续对账与合规证据，因此不能计入 Baseline 60。",
+    "reason": "Quant OS 已接通本地 synthetic 五层候选链，能冻结 Ridge ModelBundle、逐层 trace 和无股数 TargetPackageV1；一个 synthetic-research TargetPackage 已在真实 JoinQuant hosted 环境完成消费、账户/开盘价绑定、3 笔委托和 3 笔成交，因此 execution 层已有真实平台观察。但真实 provider-data 长样本/OOS 尚未接入，JoinQuant 没有重跑上游 Universe/Alpha/Portfolio/Risk，QMT peer 与逐层真实数据 parity 仍缺失，factor risk、CVXPY 和 guarded TWAP/POV 也尚未进入当前主链。因此它仍是 NOT_BASELINE_60，G9 未通过。",
     "verified_supporting_facts": [
-      "The latest standard-library suite completed 224 tests with OK and 6 optional-runtime skips; the CPython 3.12 pyqlib environment completed the same 224 tests with the native Qlib smoke enabled and no skips.",
+      "The current standard-library suite completed 285 tests with OK and 6 optional-runtime skips; successful local tests do not prove real-platform, broker or compliance gates.",
+      "The local synthetic research entrypoint now executes one causally linked Universe → Ridge Alpha → cost/risk-aware Portfolio → post-risk gate → reference Execution path, freezes a replayable ModelBundle, and publishes a hash-linked TargetPackageV1 and five-layer trace. This is structural synthetic evidence only.",
+      "The connected candidate currently uses 60-session diagonal realized variance, explicit spread/fee/square-root-impact cost forecasts, the deterministic constrained optimizer, a post-target risk gate, and reference lot/T+1 order deltas. factor_risk.py, the optional CVXPY solver mode, and guarded TWAP/POV are not selected by the current main slice.",
+      "JoinQuant and QMT target-package modes verify an exact embedded package tape, match the declared dual clock, bind platform equity/positions/sellable quantities/prices, and reject missing, duplicate or tampered packages without falling back to momentum. Both modes have local harness/contract evidence; JoinQuant execution alone now also has one narrow real hosted smoke.",
+      "JoinQuant backtest 2457a7c39a276e09e0fabf99e28978e1 ran the exact FINAL4 bundle for 2024-03-11 through 2024-03-13 with CNY 1,000,000, daily Python3 and CSI 500 benchmark. It observed 3/11 and 3/13 no-op, an exact package hit on 3/12 using CURRENT_DATA_DAY_OPEN, exact package weights, three accepted orders and three fills. The input evidence class remains synthetic-research; only execution.real_platform_observed is true, performance_claim is false, and this run earns no G9 or Baseline score.",
       "The JQData adapter maps raw bars, adjustment factor, previous close, daily limits, pause state, PIT is_st and daily PIT index membership into a semantically revalidated snapshot; it rejects an end date that has not reached the provider's complete daily-bar boundary, and only fake-provider tests have been run so far.",
       "The local snapshot can drive snapshot-backtest and snapshot-decision with manifest verification; it freezes a cross-checked provider trading calendar, and a T-close signal can bind only to a broker observation in the unique next session's Asia/Shanghai [09:00,09:30) window without collapsing the two clocks.",
       "CPython 3.12 with pyqlib 0.9.7 completed a native momentum fixture smoke with 24 signals and saved audited portfolio-table hashes, bounds and recomputed metrics.",
       "Qlib Alpha158, LightGBM and Recorder completed a real local fixture smoke with audited portfolio output; the deterministic two-stock synthetic fixture has no investment-value meaning.",
       "The local Tushare inventory verified all 93 completed Parquet files against recorded size, SHA-256 and row count. An immutable managed Qlib provider and the same early-history momentum command produced byte-identical evidence JSON twice with PYTHONHASHSEED=0. The mirror is only about 8.2% complete for daily partitions and lacks adjustment, real benchmark, point-in-time membership, ST, suspension and limit data, so production_ready and investment_value_claim remain false and no gate credit is earned.",
-      "A real JoinQuant daily hosted backtest completed for 2024-01-02 through 2024-12-31 with CNY 1,000,000; the visible completed-run log had zero ERROR/Traceback/order-rejection entries. This proves hosted runtime execution only: exact input/plan/order/fill export and a local same-input peer are still missing.",
+      "A real JoinQuant daily hosted backtest completed for 2024-01-02 through 2024-12-31 with CNY 1,000,000; the visible completed-run log had zero ERROR/Traceback/order-rejection entries. That run used portable_momentum_smoke, not the Ridge/TargetPackage Baseline candidate, and earns no candidate or cross-engine gate credit.",
       "The QMT shadow CLI is strictly read-only and produces HMAC account-bound broker observation, broker snapshot and target-diff artifacts. QMT_SHADOW_EVIDENCE_V1 authenticates semantic content, deterministic published bytes, the original decision, planner/engine/operator source and fixed policy; semantic replay verifies exact derived state. Independent local QA completed 58 deterministic scenarios, 500 malformed fuzz cases and 300 legal-observation fuzz cases with no P0/P1. This is Quant OS publisher integrity, not broker attestation, and it still has only fake-broker evidence.",
       "JoinQuant/QMT mock parity is mock_contract_only, real_platform_pass is false and it earns no G9 gate credit."
     ],
+    "internal_delivery_gaps": [
+      "The authorized provider-data path has not yet been connected end to end to the frozen Ridge ModelBundle, calibrated risk/cost inputs, five-layer Baseline candidate and TargetPackage tape; the current connected candidate remains synthetic.",
+      "factor_risk.py, the optional CVXPY optimizer mode and guarded TWAP/POV exist as components or optional paths but are not selected by the current Baseline main slice; the connected risk model is 60-session diagonal realized variance and execution is the reference one-shot planner.",
+      "One synthetic-research TargetPackageV1 has been observed at the real JoinQuant execution boundary, but JoinQuant did not recompute the upstream four layers; there is no authorized real-data multi-period tape, same-input layer-by-layer parity, or real QMT peer."
+    ],
     "external_evidence_missing": [
-      "authorized JQData market snapshot and licence/lineage record",
-      "complete JoinQuant hosted input/plan/order/fill export and same-input local comparison",
-      "real QMT built-in and qmttools backtest exports",
+      "authorized JQData market snapshot, licence/lineage record and a frozen provider-data candidate tape",
+      "authorized real-data multi-period JoinQuant TargetPackage run plus same-input local/JoinQuant layer-by-layer reconciliation",
+      "real QMT TargetPackage built-in and qmttools backtest exports",
       "real QMT historical-ST positive control, terminal/xtquant build record, broker query contract and trade-amount semantic probes, callback/restart and at least 20 trading days of shadow evidence",
       "continuous reconciliation, recovery drills and broker compliance confirmation"
     ]