fix(metrics): count orphaned-workflow governance incidents in the window rollup (#17)

The window governance counters (blocked_budget, blocked_policy, failed)
were incremented inside the per-workflow loop, so a run whose workflow_id
is no longer in the config (a pilot deleted/renamed while its historical
runs remain) was never counted — while the all-time counters, tallied in a
per-run pass with no workflow lookup, DO count it. So for an in-window
incident on a since-deleted workflow the window counter read 0 while the
all-time counter read 1, breaking the documented "_all is a superset of the
window" relationship and understating the window's governance posture.

Add a small pass over the window runs for workflows absent from the config,
classifying their blocks/failures the same way. Runs of known workflows are
already tallied above, so the pass is disjoint — no double counting, and no
change for any org whose runs all belong to current workflows.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: sturlese

src/flightdeck/metrics.py

@@ -321,6 +321,22 @@ def build_report(org: Org, store: Store, ledger: Ledger, days: int = 30, now: da
         entry.health = _health(workflow, entry)
         report.workflows.append(entry)
 
+    # Governance incidents for a workflow since deleted from config still belong in
+    # the window rollup: the all-time counters already include them, so the window
+    # counters must too (they have no per-workflow row to be counted under). Runs of
+    # known workflows were already tallied in the loop above, so this pass is disjoint.
+    known_workflows = set(org.workflows)
+    for run in window_runs:
+        if run.workflow_id in known_workflows:
+            continue
+        if run.status == "blocked":
+            if "budget" in (run.reason or ""):
+                report.governance.blocked_budget += 1
+            else:
+                report.governance.blocked_policy += 1
+        elif run.status == "failed":
+            report.governance.failed += 1
+
     completed = [run for run in window_runs if run.status == "completed"]
     if completed:
         no_training = sum(

tests/test_metrics.py

@@ -126,6 +126,23 @@ def test_governance_rollup(self, seeded):
         assert gov.no_training_share == 1.0
         assert gov.ledger_ok and gov.ledger_entries == 1
 
+    def test_window_governance_counts_orphaned_workflow_incidents(self, org, store, ledger):
+        # A workflow deleted from config leaves its historical runs behind. In-window
+        # governance incidents for it still belong in the window rollup — the all-time
+        # counters already include them, so the window counters must too.
+        when = NOW - timedelta(days=1)
+        store.add_run(_run("gb", when, workflow_id="deleted-wf", status="blocked",
+                           reason="monthly budget exhausted", model_id="", cost=0))
+        store.add_run(_run("gp", when, workflow_id="deleted-wf", status="blocked",
+                           reason="no policy-compliant model", model_id="", cost=0))
+        store.add_run(_run("gf", when, workflow_id="deleted-wf", status="failed",
+                           reason="provider: timeout", model_id="", cost=0))
+        gov = build_report(org, store, ledger, days=30, now=NOW).governance
+        # Window counters now match the all-time counters for these in-window incidents.
+        assert gov.blocked_budget == gov.blocked_budget_all == 1
+        assert gov.blocked_policy == gov.blocked_policy_all == 1
+        assert gov.failed == 1
+
     def test_health_flags_the_underperformer(self, seeded):
         # acceptance 0.67 vs target 0.80 and ~1.5 weekly actives vs target 6 → worst
         # ratio far below 0.75.