fix(admin): derive EM experiment KPI total from bucketed counts, not metadata

The EM EXPERIMENT_COUNT total (and its workspace roll-up) was sourced from
`numberOfExperiments` metadata while the chart series was built from
experiments bucketed by start_server_timestamp. When those disagree, the
adoption KPI total didn't equal the cumulative chart sum. Derive `total`
from the same counts map that feeds the series so they always match.

Add a regression test covering the metadata-disagrees case.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: Douglas Blank

cometx/cli/admin_growth_report.py

@@ -851,17 +851,21 @@ def _collect_em(self, workspaces):
             ws_experiment_total = 0
             ws_counts: dict = defaultdict(int)
 
-            for project in tqdm(
-                projects, desc=f"EM {ws}", unit="proj", leave=False
-            ):
+            for project in tqdm(projects, desc=f"EM {ws}", unit="proj", leave=False):
                 proj_name = project.get("projectName")
                 try:
                     # Fetch the project's experiments once and reuse the list
                     # for both the creation proxy and the over-time series.
                     experiments = self.api.get_experiments(ws, proj_name) or []
                     created = self._em_project_created(project, experiments)
                     counts = self._em_experiment_counts(experiments)
-                    total = project.get("numberOfExperiments", sum(counts.values()))
+                    # Derive the KPI total from the SAME bucketed counts that
+                    # feed the chart series -- NOT the `numberOfExperiments`
+                    # metadata, which can disagree with the experiments that
+                    # actually carry a `start_server_timestamp`. This keeps the
+                    # EM adoption KPI total (and its workspace roll-up) exactly
+                    # equal to the cumulative chart sum.
+                    total = sum(counts.values())
 
                     events.append(
                         CreationEvent(
@@ -1017,9 +1021,7 @@ def _collect_opik(self, workspaces):
 
             ws_counts: dict = defaultdict(float)
 
-            for project in tqdm(
-                projects, desc=f"Opik {ws}", unit="proj", leave=False
-            ):
+            for project in tqdm(projects, desc=f"Opik {ws}", unit="proj", leave=False):
                 try:
                     if project.created_at is not None:
                         events.append(
@@ -1045,9 +1047,9 @@ def _collect_opik(self, workspaces):
                     counts: dict = defaultdict(float)
                     for result in resp.results or []:
                         for dp in result.data or []:
-                            counts[
-                                format_time_key(dp.time, self.units)
-                            ] += _as_float(dp.value)
+                            counts[format_time_key(dp.time, self.units)] += _as_float(
+                                dp.value
+                            )
 
                     usage.append(
                         UsageMetric(
@@ -1128,8 +1130,7 @@ def _collect_mpm(self, workspaces):
             for ws_entry in all_workspaces
             # The MPM inventory shape is not verifiable live; guard against
             # malformed elements so one bad entry can't crash the report.
-            if isinstance(ws_entry, dict)
-            and ws_entry.get("workspaceName") in requested
+            if isinstance(ws_entry, dict) and ws_entry.get("workspaceName") in requested
         ]
         for ws_entry in tqdm(selected_entries, desc="MPM workspaces", unit="ws"):
             ws = ws_entry.get("workspaceName")
@@ -1138,9 +1139,7 @@ def _collect_mpm(self, workspaces):
             ws_counts: dict = defaultdict(float)
             ws_total = 0.0
 
-            for model in tqdm(
-                models, desc=f"MPM {ws}", unit="model", leave=False
-            ):
+            for model in tqdm(models, desc=f"MPM {ws}", unit="model", leave=False):
                 if not isinstance(model, dict):
                     continue
                 model_name = model.get("modelName")

tests/unit/test_admin_growth_report.py

@@ -212,6 +212,48 @@ def test_collect_em_usage_metrics_experiment_count_and_registry_snapshot():
     assert not any(m.metric.startswith("REGISTRY") and m.series for m in usage)
 
 
+def test_collect_em_kpi_total_matches_chart_sum_when_metadata_disagrees():
+    # Regression: the EXPERIMENT_COUNT total must come from the SAME bucketed
+    # timestamps that feed the chart series, NOT from `numberOfExperiments`.
+    # Here the metadata (10) disagrees with the experiments that actually carry
+    # a start_server_timestamp (3), so the old fallback would have produced a
+    # KPI total that didn't equal the cumulative chart sum.
+    from cometx.cli.admin_growth_report import GrowthReporter
+
+    api = MagicMock()
+    api._client.get_from_endpoint.return_value = {
+        "projects": [
+            {
+                "projectName": "proj1",
+                "projectId": "p1",
+                "workspaceName": "ws1",
+                "numberOfExperiments": 10,  # metadata, intentionally != 3
+                "lastUpdated": 1700000000000,
+            }
+        ]
+    }
+    api.get_experiments.return_value = [
+        MagicMock(start_server_timestamp=1695000000000),
+        MagicMock(start_server_timestamp=1695100000000),
+        MagicMock(start_server_timestamp=1695200000000),
+    ]
+    api.get_registry_model_names.return_value = []
+    reporter = GrowthReporter(api, window="7d", units="month", platforms="em")
+    _events, usage = reporter._collect_em(["ws1"])
+
+    proj_metric = next(
+        m for m in usage if m.metric == "EXPERIMENT_COUNT" and m.project == "proj1"
+    )
+    ws_total = next(
+        m for m in usage if m.metric == "EXPERIMENT_COUNT" and m.project is None
+    )
+    # total is derived from counts (3), NOT numberOfExperiments (10)
+    assert proj_metric.value == 3
+    assert ws_total.value == 3
+    # the KPI total equals the cumulative chart sum for the workspace
+    assert ws_total.value == sum(v for _k, v in ws_total.series)
+
+
 def test_collect_em_respects_limit_on_workspaces():
     from cometx.cli.admin_growth_report import GrowthReporter