diff --git a/docs/ci/health-report.md b/docs/ci/health-report.md
index 6c40bd855d5b..093741aba79c 100644
--- a/docs/ci/health-report.md
+++ b/docs/ci/health-report.md
@@ -33,6 +33,9 @@ Related: #13095 (CI cost and capacity) and #13325 (CI waste).
 Minutes are `completed_at - started_at` per job, summed, and split by
 conclusion into success / failure / cancelled / skipped. `timed_out` and
 `startup_failure` count as failure, because they cost what a failure costs.
+A job cancelled before any runner picked it up (`runner_id: 0`, no steps) has
+zero minutes; the API stamps its `started_at` at creation, so its whole wait
+until the cancel counts as queue wait instead.
 
 These tables are **sampled**: job listings are the expensive API call, so the
 report reads jobs for a bounded number of runs spread across workflows (macOS-
@@ -41,11 +44,18 @@ which workflow, which job, which pool, which conclusion is eating the minutes 
 not as a billing total. The header states how many runs were sampled.
 
 Failure and cancelled minutes are the interesting columns. Success minutes are
-the price of CI; the rest is the price of CI not working.
+the price of CI; the rest is the price of CI not working. One exception:
+`cmux-tui-testbox-warmup.yml` holds a Testbox VM for a maintainer session, and
+the session's cleanup (`scripts/blacksmith-testbox-demo.sh`) cancels the run on
+purpose, so its cancelled minutes are the session itself.
 
 ### Queue wait (created → started) per runner label
 
 `started_at - created_at` per job, as p50 / p90 / p99 / worst, per runner label.
+A job cancelled while still queued counts with its wait up to the cancel, a
+lower bound on what it would have waited. These used to count as zero, so the
+first reports after that change show a higher macOS p90 without any real
+regression.
 Labels are joined when a job asks for several, so `self-hosted+macos` is not
 silently pooled with `macos`.
 
diff --git a/scripts/ci/ci_health_report.py b/scripts/ci/ci_health_report.py
index 8815012d2444..dd964e936076 100644
--- a/scripts/ci/ci_health_report.py
+++ b/scripts/ci/ci_health_report.py
@@ -246,6 +246,16 @@ class JobRow:
     fork: bool
 
 
+def never_got_a_runner(job: Mapping[str, Any]) -> bool:
+    """A job that finished without ever being assigned a runner.
+
+    The Actions API reports such a job (cancelled while queued) with
+    `runner_id: 0`, an empty runner name and no steps. A skipped job has
+    `runner_id: null` instead and no start time, so it is not matched here.
+    """
+    return job.get("runner_id") == 0 and not job.get("runner_name") and not job.get("steps")
+
+
 def job_rows(run: Mapping[str, Any], jobs: Iterable[Mapping[str, Any]], repo: str) -> list[JobRow]:
     """Flatten one run's jobs into rows the aggregations read.
 
@@ -258,6 +268,10 @@ def job_rows(run: Mapping[str, Any], jobs: Iterable[Mapping[str, Any]], repo: st
         created = parse_time(job.get("created_at"))
         started = parse_time(job.get("started_at"))
         completed = parse_time(job.get("completed_at"))
+        if never_got_a_runner(job):
+            # GitHub stamps started_at = created_at on a job cancelled while
+            # still queued, so started -> completed would be the whole wait.
+            started = completed
         minutes = 0.0
         if started and completed and completed > started:
             minutes = (completed - started).total_seconds() / 60.0
diff --git a/tests/test_ci_health_report.py b/tests/test_ci_health_report.py
index a6d6b31c0833..ec999f92891c 100644
--- a/tests/test_ci_health_report.py
+++ b/tests/test_ci_health_report.py
@@ -85,6 +85,30 @@ def test_a_job_that_never_started_has_no_wait_and_no_minutes(self):
         self.assertEqual(skipped.minutes, 0.0)
         self.assertEqual(skipped.bucket, "skipped")
 
+    def test_a_job_cancelled_while_queued_is_wait_not_minutes(self):
+        # Shape of iroh-v2 run 35941162262's client job: superseded after 53
+        # minutes in the macOS queue, never assigned a runner. The API stamps
+        # started_at = created_at on it.
+        queued = {
+            "name": "client",
+            "conclusion": "cancelled",
+            "created_at": "2026-09-24T01:02:41Z",
+            "started_at": "2026-09-24T01:02:41Z",
+            "completed_at": "2026-09-24T01:55:37Z",
+            "runner_id": 0,
+            "runner_name": "",
+            "steps": [],
+            "labels": ["blacksmith-6vcpu-macos-26"],
+        }
+        ran = dict(queued, runner_id=1195080, runner_name="mac-runner-1", steps=[{"name": "Set up job"}])
+        run = self.fixture["runs"][0]
+        [queued_row, ran_row] = report.job_rows(run, [queued, ran], REPO)
+        self.assertEqual(queued_row.minutes, 0.0)
+        self.assertAlmostEqual(queued_row.queue_seconds, (52 * 60) + 56)
+        self.assertEqual(queued_row.bucket, "cancelled")
+        self.assertEqual(report.cancelled_macos_waste([queued_row]), [])
+        self.assertAlmostEqual(ran_row.minutes, 52 + 56 / 60)
+
     def test_fork_head_repository_marks_the_row(self):
         fork = next(row for row in self.rows if row.run_id == 1013)
         self.assertTrue(fork.fork)