fix: derive crypto drift baseline from walk-forward returns

Co-Authored-By: Codex <noreply@openai.com>

Author: Pigbibi

.github/workflows/drift-check.yml

@@ -16,6 +16,8 @@ jobs:
     if: github.ref == format('refs/heads/{0}', github.event.repository.default_branch)
     runs-on: ubuntu-latest
     timeout-minutes: 30
+    outputs:
+      snapshot_repository_ref: ${{ steps.snapshot-input.outputs.snapshot_repository_ref }}
     steps:
       - name: Checkout
         uses: actions/checkout@v6
@@ -24,7 +26,7 @@ jobs:
         uses: actions/checkout@v6
         with:
           repository: QuantStrategyLab/QuantPlatformKit
-          ref: 9bb8f31e898ea238a6446472f9f5e58133128d0c
+          ref: bda6afdab0a2dd693c35d14493176829f4da1231
           path: external/QuantPlatformKit
 
       - name: Set up Python
@@ -40,6 +42,7 @@ jobs:
           python -m pip install --no-deps -e external/QuantPlatformKit
 
       - name: Download latest trusted lifecycle inputs
+        id: snapshot-input
         env:
           GH_TOKEN: ${{ secrets.SNAPSHOT_REPOSITORY_TOKEN }}
           INPUT_ROOT: ${{ runner.temp }}/crypto-lifecycle-inputs
@@ -62,17 +65,24 @@ jobs:
           pages = json.loads((Path(os.environ["RUNNER_TEMP"]) / "snapshot-artifacts.json").read_text())
           run_pages = json.loads((Path(os.environ["RUNNER_TEMP"]) / "trusted-snapshot-runs.json").read_text())
           artifacts = [item for page in pages for item in page.get("artifacts", [])]
-          trusted_run_ids = {run["id"] for page in run_pages for run in page.get("workflow_runs", [])}
-          candidates = [
-              item for item in artifacts
-              if not item.get("expired")
-              and str(item.get("name", "")).startswith("crypto-lifecycle-inputs-")
-              and item.get("workflow_run", {}).get("head_branch") == "main"
-              and item.get("workflow_run", {}).get("id") in trusted_run_ids
-          ]
-          if not candidates:
+          trusted_runs = [run for page in run_pages for run in page.get("workflow_runs", [])]
+          selected = None
+          for run in sorted(
+              trusted_runs,
+              key=lambda item: (item.get("run_number", 0), item.get("run_attempt", 0)),
+              reverse=True,
+          ):
+              matches = [
+                  item for item in artifacts
+                  if not item.get("expired")
+                  and str(item.get("name", "")).startswith("crypto-lifecycle-inputs-")
+                  and item.get("workflow_run", {}).get("id") == run["id"]
+              ]
+              if matches:
+                  selected = max(matches, key=lambda item: item["created_at"])
+                  break
+          if selected is None:
               raise SystemExit("no trusted crypto lifecycle input artifact is available")
-          selected = max(candidates, key=lambda item: item["created_at"])
           print(selected["id"], selected["workflow_run"]["id"])
           PY
           read -r artifact_id workflow_run_id < "${RUNNER_TEMP}/snapshot-artifact-selection.txt"
@@ -94,6 +104,19 @@ jobs:
           if mismatches:
               raise SystemExit(f"lifecycle input provenance check failed: {mismatches}")
           PY
+          snapshot_repository_ref="$(python - <<'PY'
+          import json
+          import os
+          from pathlib import Path
+          run = json.loads((Path(os.environ["RUNNER_TEMP"]) / "snapshot-workflow-run.json").read_text())
+          print(run["head_sha"])
+          PY
+          )"
+          if [[ ! "${snapshot_repository_ref}" =~ ^[0-9a-f]{40}$ ]]; then
+            echo "::error::Invalid snapshot producer head SHA"
+            exit 1
+          fi
+          echo "snapshot_repository_ref=${snapshot_repository_ref}" >> "${GITHUB_OUTPUT}"
           gh api "/repos/QuantStrategyLab/CryptoLivePoolPipelines/actions/artifacts/${artifact_id}/zip" \
             > "${RUNNER_TEMP}/snapshot-artifact.zip"
           python - <<'PY'
@@ -176,15 +199,16 @@ jobs:
       contents: read
       issues: write
       id-token: write
-    uses: QuantStrategyLab/QuantPlatformKit/.github/workflows/reusable-drift-check.yml@9bb8f31e898ea238a6446472f9f5e58133128d0c
+    uses: QuantStrategyLab/QuantPlatformKit/.github/workflows/reusable-drift-check.yml@bda6afdab0a2dd693c35d14493176829f4da1231
     with:
       strategy_domain: crypto
       caller_event_name: ${{ github.event_name }}
       caller_pr_head_repository: ${{ github.event.pull_request.head.repo.full_name || '' }}
       snapshot_repository: QuantStrategyLab/CryptoLivePoolPipelines
       snapshot_checkout_path: external/CryptoLivePoolPipelines
+      snapshot_repository_ref: ${{ needs.preflight_backtests.outputs.snapshot_repository_ref }}
       ai_gateway_service_url: ${{ vars.AI_GATEWAY_SERVICE_URL }}
-      quant_platform_kit_ref: 9bb8f31e898ea238a6446472f9f5e58133128d0c
+      quant_platform_kit_ref: bda6afdab0a2dd693c35d14493176829f4da1231
       lifecycle_preflight_artifact: lifecycle-preflight-${{ github.run_id }}-${{ github.run_attempt }}
     secrets:
       codex_audit_service_url: ${{ secrets.CODEX_AUDIT_SERVICE_URL }}

scripts/run_walk_forward_backtest.py

@@ -233,34 +233,17 @@ def run_walk_forward(
             panel=panel,
             market_history=market_history,
         )
-    return_matrix_runner = _build_runner(
-        profile=profile,
-        panel=shared_panel,
-        market_history=shared_market_history,
-        synthetic_days=synthetic_days,
-    )
-    full_start = min(start for start, _ in windows)
-    baseline_end = max(end for _, end in windows)
-    full_window_raw = return_matrix_runner.run(
-        profile,
-        copy.deepcopy(baseline_params),
-        start_date=full_start,
-        end_date=baseline_end,
-    )
-    full_window_returns = return_matrix_runner.last_daily_returns
-    if len(full_window_returns) < DRIFT_BASELINE_HORIZON_DAYS:
-        raise ValueError("full-window returns do not cover the 126-day drift baseline")
-    baseline_raw = _baseline_from_return_tail(full_window_raw, full_window_returns)
     with tempfile.TemporaryDirectory(prefix=f"{profile}_wf_", dir=target_root) as scratch_dir:
         scratch_orchestrator = BacktestOrchestrator(store=PerformanceStore(local_root=Path(scratch_dir)))
+        walk_forward_runner = _build_runner(
+            profile=profile,
+            panel=shared_panel,
+            market_history=shared_market_history,
+            synthetic_days=synthetic_days,
+        )
         scratch_orchestrator.register_runner(
             "crypto",
-            _build_runner(
-                profile=profile,
-                panel=shared_panel,
-                market_history=shared_market_history,
-                synthetic_days=synthetic_days,
-            ),
+            walk_forward_runner,
         )
         wf_results = scratch_orchestrator.walk_forward(
             profile,
@@ -269,6 +252,10 @@ def run_walk_forward(
             windows=windows,
             param_set_id=f"{profile}_wf",
         )
+        full_window_returns = pd.concat(walk_forward_runner.run_return_history).sort_index()
+    if len(full_window_returns) < DRIFT_BASELINE_HORIZON_DAYS:
+        raise ValueError("walk-forward returns do not cover the 126-day drift baseline")
+    baseline_raw = _baseline_from_return_tail(wf_results[-1], full_window_returns)
     orchestrator = BacktestOrchestrator(store=PerformanceStore(local_root=target_root))
     baseline = orchestrator.persist_result(
         baseline_raw,

src/crypto_strategies/backtest/orchestrator_runner.py

@@ -145,11 +145,16 @@ def __init__(self, *, panel: pd.DataFrame | None = None, synthetic_days: int = 1
         self._panel = panel
         self._synthetic_days = int(synthetic_days)
         self._last_daily_returns = pd.Series(dtype=float)
+        self._run_return_history: list[pd.Series] = []
 
     @property
     def last_daily_returns(self) -> pd.Series:
         return self._last_daily_returns.copy()
 
+    @property
+    def run_return_history(self) -> tuple[pd.Series, ...]:
+        return tuple(item.copy() for item in self._run_return_history)
+
     def run(
         self,
         strategy_profile: str,
@@ -186,6 +191,7 @@ def run(
             start_date=start_date,
             end_date=end_date,
         )
+        self._run_return_history.append(self._last_daily_returns.copy())
         elapsed = (datetime.now(timezone.utc) - started).total_seconds()
         eval_dates = sliced.index.get_level_values("date")
         return _metrics_to_result(
@@ -210,11 +216,16 @@ def __init__(
         self._market_history = market_history
         self._synthetic_days = int(synthetic_days)
         self._last_daily_returns = pd.Series(dtype=float)
+        self._run_return_history: list[pd.Series] = []
 
     @property
     def last_daily_returns(self) -> pd.Series:
         return self._last_daily_returns.copy()
 
+    @property
+    def run_return_history(self) -> tuple[pd.Series, ...]:
+        return tuple(item.copy() for item in self._run_return_history)
+
     def run(
         self,
         strategy_profile: str,
@@ -260,6 +271,7 @@ def run(
             start_date=start_date,
             end_date=end_date,
         )
+        self._run_return_history.append(self._last_daily_returns.copy())
         elapsed = (datetime.now(timezone.utc) - started).total_seconds()
         eval_frame = sliced
         if start_date is not None:

tests/test_drift_workflow_config.py

@@ -6,6 +6,10 @@ def test_drift_workflow_wires_real_pipeline_inputs_and_preflight_bundle() -> Non
 
     assert "preflight_backtests:" in workflow
     assert "needs: preflight_backtests" in workflow
+    assert "snapshot_repository_ref: ${{ steps.snapshot-input.outputs.snapshot_repository_ref }}" in workflow
+    assert "id: snapshot-input" in workflow
+    assert 'print(run["head_sha"])' in workflow
+    assert "snapshot_repository_ref: ${{ needs.preflight_backtests.outputs.snapshot_repository_ref }}" in workflow
     assert "Download latest trusted lifecycle inputs" in workflow
     assert "gh api --paginate --slurp" in workflow
     assert "trusted-snapshot-runs.json" in workflow
@@ -15,7 +19,7 @@ def test_drift_workflow_wires_real_pipeline_inputs_and_preflight_bundle() -> Non
     assert "research_panel.csv.gz" in workflow
     assert "market_history.csv.gz" in workflow
     assert "repository: QuantStrategyLab/QuantPlatformKit" in workflow
-    assert "ref: 9bb8f31e898ea238a6446472f9f5e58133128d0c" in workflow
+    assert "ref: bda6afdab0a2dd693c35d14493176829f4da1231" in workflow
     assert "python -m pip install --no-deps -e external/QuantPlatformKit" in workflow
     assert "scripts/run_walk_forward_backtest.py" in workflow
     assert '"--list-profiles"' in workflow
@@ -25,7 +29,7 @@ def test_drift_workflow_wires_real_pipeline_inputs_and_preflight_bundle() -> Non
     assert "Upload lifecycle preflight artifact" in workflow
     assert "lifecycle-preflight-${{ github.run_id }}-${{ github.run_attempt }}" in workflow
     assert workflow.count("github.ref == format('refs/heads/{0}', github.event.repository.default_branch)") == 2
-    assert "uses: QuantStrategyLab/QuantPlatformKit/.github/workflows/reusable-drift-check.yml@9bb8f31e898ea238a6446472f9f5e58133128d0c" in workflow
+    assert "uses: QuantStrategyLab/QuantPlatformKit/.github/workflows/reusable-drift-check.yml@bda6afdab0a2dd693c35d14493176829f4da1231" in workflow
     assert "strategy_domain: crypto" in workflow
     assert "snapshot_repository: QuantStrategyLab/CryptoLivePoolPipelines" in workflow
     assert "snapshot_checkout_path: external/CryptoLivePoolPipelines" in workflow

tests/test_orchestrator_runner.py

@@ -43,6 +43,7 @@ def test_run_returns_backtest_result(self) -> None:
         self.assertGreaterEqual(runner.last_daily_returns.index.min().date(), date(2023, 6, 1))
         self.assertLessEqual(runner.last_daily_returns.index.max().date(), date(2024, 6, 1))
         self.assertEqual(result.observation_count, len(runner.last_daily_returns))
+        self.assertEqual(len(runner.run_return_history), 1)
 
     def test_walk_forward_produces_one_result_per_window(self) -> None:
         from pathlib import Path
@@ -82,6 +83,7 @@ def test_run_returns_backtest_result(self) -> None:
         self.assertGreaterEqual(runner.last_daily_returns.index.min().date(), date(2023, 6, 1))
         self.assertLessEqual(runner.last_daily_returns.index.max().date(), date(2024, 6, 1))
         self.assertEqual(result.observation_count, len(runner.last_daily_returns))
+        self.assertEqual(len(runner.run_return_history), 1)
 
     def test_invalid_combo_mode_raises(self) -> None:
         runner = CryptoEquityComboBacktestRunner(synthetic_days=1600)

tests/test_run_walk_forward_backtest.py

@@ -131,6 +131,7 @@ def test_run_walk_forward_uses_real_panel_and_writes_return_matrix(
     assert payload["baseline"]["observation_count"] == 126
     assert {"as_of", "crypto_live_pool_rotation", "buy_hold_BTC"} <= set(return_matrix.columns)
     assert len(return_matrix) > payload["baseline"]["observation_count"]
+    assert len(return_matrix) == sum(item["observation_count"] for item in payload["walk_forward_folds"])
 
 
 def test_baseline_uses_exact_tail_of_full_return_stream() -> None: