diff --git a/hermes_cli/observability/shared_metrics_update.py b/hermes_cli/observability/shared_metrics_update.py
index 25841c7de7..ebb12a44b0 100644
--- a/hermes_cli/observability/shared_metrics_update.py
+++ b/hermes_cli/observability/shared_metrics_update.py
@@ -11,6 +11,11 @@ Desktop's packaged updaters (electron-updater, App Installer, Store) never run `
 Desktop reports their outcome through the ``shared_metrics.update_run`` RPC instead. Desktop's
 source-checkout hand-off DOES run ``hermes update``; that receipt is tagged ``initiator=desktop``
 and counted here, never by the RPC, so no run is counted twice.
+
+Run and stage rows agree on where a FAILED run stopped: a run whose ``failed_stage`` names a stage it
+never marked (it exited inside it, before the stage's END mark) also gets one failed stage row for it.
+A run that passed the commit point and only owes follow-ups (contract C3) is a ``success`` run, while
+the stage that failed (``deps``, ``build``, ``restart``, ``verify``) still reads ``failed`` at stage level.
 """
 
 from __future__ import annotations
@@ -293,9 +298,31 @@ def update_receipt_fields(receipt: dict[str, Any]) -> tuple[dict[str, str], list
         }
         for stage in stages
     ]
+    if died_in := _unmarked_failed_stage(run, stages):
+        last_mark = next((s["at"] for s in reversed(stages) if s.get("at")), receipt.get("started_at"))
+        stage_rows.append({
+            "duration_bucket": update_duration_bucket(_elapsed_ms(last_mark, receipt.get("finished_at"))),
+            "outcome": "failed", "stage": died_in,
+        })
     return run, stage_rows
 
 
+def _unmarked_failed_stage(run: dict[str, str], stages: list[dict[str, Any]]) -> str | None:
+    """The stage a failed run stopped in when the receipt holds no mark for it, else None.
+
+    Stage marks are END marks, so a run that exits inside a stage (every pre-apply ``sys.exit``:
+    fetch, channel, branch, merge, HEAD checks) leaves none for it and the stage rows would show no
+    failure where the run row says it died. Only ``failed`` runs: a refusal is not a stage failure,
+    and a committed run that owes follow-ups is a ``success`` whose failed stage already has its mark.
+    """
+    from .shared_metrics_contract import UPDATE_STAGES
+
+    died_in = run["failed_stage"]
+    if run["outcome"] != "failed" or died_in not in UPDATE_STAGES or any(s["name"] == died_in for s in stages):
+        return None
+    return died_in
+
+
 def _collection_on() -> bool:
     """Cheap pre-gate so a disabled install never loads the Relay runtime; `_emit` re-checks."""
     from hermes_cli.config import read_raw_config_readonly
diff --git a/tests/hermes_cli/test_shared_metrics_install_failures.py b/tests/hermes_cli/test_shared_metrics_install_failures.py
index 6dbb64d7fa..7bea2a0cf7 100644
--- a/tests/hermes_cli/test_shared_metrics_install_failures.py
+++ b/tests/hermes_cli/test_shared_metrics_install_failures.py
@@ -116,6 +116,23 @@ def test_a_partial_run_with_a_failed_stage_keeps_that_stage_and_its_verify_row_p
     assert [(s["stage"], s["outcome"]) for s in stage_rows if s["outcome"] == "failed"] == [(failed, "failed")]
 
 
+def test_a_run_that_died_before_a_stage_mark_has_a_failed_row_for_that_stage():
+    """Invariant: stage rows agree with the run row on where a FAILED run stopped, and a committed run
+    that owes follow-ups (C3) is a run-level success whose failed stage still shows at stage level."""
+    fetch_failed = _receipt("failed", [("plan", "success"), ("snapshot", "skipped")], exit_code=1,
+                            stop_reason="sys.exit(1)")
+    run, stage_rows = update_metrics.update_receipt_fields(fetch_failed)
+    failed_rows = [s for s in stage_rows if s["outcome"] == "failed"]
+    assert run["failed_stage"] == "apply" and [s["stage"] for s in failed_rows] == ["apply"]
+    assert all(contract.counter_dimensions_are_valid(contract.UPDATE_STAGE_METRIC, s) for s in stage_rows)
+
+    deps_owed = _receipt("success", [*_ALL_PASSED[:3], ("deps", "failed")],
+                         followups=[{"step": "dependencies", "reason": "uv sync exited 2"}])
+    run, stage_rows = update_metrics.update_receipt_fields(deps_owed)
+    assert (run["outcome"], run["failed_stage"]) == ("success", "none")
+    assert [s["stage"] for s in stage_rows if s["outcome"] == "failed"] == ["deps"]
+
+
 @pytest.mark.parametrize(("receipt", "expected"), [
     (_receipt("success", _ALL_PASSED), "none"),
     (_receipt("refused", [], steps=[{"name": "admission", "ok": False}], stop_reason="docker"), "managed_install"),
diff --git a/tests/hermes_cli/test_shared_metrics_reliability.py b/tests/hermes_cli/test_shared_metrics_reliability.py
index 0946d4fd51..bf2237e238 100644
--- a/tests/hermes_cli/test_shared_metrics_reliability.py
+++ b/tests/hermes_cli/test_shared_metrics_reliability.py
@@ -75,8 +75,9 @@ def test_failed_update_reports_the_stage_it_never_reached(marks, monkeypatch):
         "apply_mode": "git", "duration_bucket": "lt_30s", "failed_stage": "deps",
         "from_version_age_bucket": "7d_to_30d", "kind": "cli", "outcome": "failed", "failure_class": "deps_failed",
     }]
+    # The stage the run died in carries a failed row, so run and stage counts agree.
     assert [(s["stage"], s["outcome"]) for s in stages] == [
-        ("plan", "success"), ("snapshot", "skipped"), ("apply", "success")]
+        ("plan", "success"), ("snapshot", "skipped"), ("apply", "success"), ("deps", "failed")]
     _assert_valid(marks.rows)