diff --git a/.github/copilot-instructions.md b/.github/copilot-instructions.md
index e734e6a13..f6f9caaa1 100644
--- a/.github/copilot-instructions.md
+++ b/.github/copilot-instructions.md
@@ -109,7 +109,8 @@
   matrix. Converting an active pull request to draft does not cancel its existing
   run. Prefer Draft for workload-affecting changes until local deterministic gates
   pass.
-- Pull-request concurrency cancels superseded runs. Before debugging a cancelled
+- Pull-request concurrency cancels superseded runs, while `dev` pushes queue in
+  order and never cancel one another. Before debugging a cancelled pull-request
   run, compare its head SHA with the current pull-request head and classify an
   expected supersession separately from a failure. Use `nvx-ci-investigation` for
   failed or unexpectedly cancelled runs.
diff --git a/.github/skills/nvx-ci-investigation/SKILL.md b/.github/skills/nvx-ci-investigation/SKILL.md
index 51cdcbb01..a515a0443 100644
--- a/.github/skills/nvx-ci-investigation/SKILL.md
+++ b/.github/skills/nvx-ci-investigation/SKILL.md
@@ -36,6 +36,10 @@ cancels an older run when a newer head is synchronized.
 - If a cancelled run's head differs from the current pull-request head and a newer run
   exists, classify it as expected supersession. Report the replacement run and stop
   unless cancellation cleanup itself failed.
+- When the workflow at its head sets `queue: max` for pushes, a `dev` push run waits
+  behind earlier pushes instead of being replaced, so its cancellation is never a
+  supersession. Under earlier workflows, a newer `dev` push replaced the one waiting
+  run.
 - A run started while a pull request was Ready remains active after conversion to
   Draft because `converted_to_draft` is not a workflow trigger. Do not diagnose this
   as the draft guard being ignored.
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 7db9b76c9..2686a5911 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -19,6 +19,10 @@ on:
 concurrency:
   group: ci-${{ github.workflow }}-${{ (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && format('full-{0}', github.run_id) || github.event.pull_request.number || github.ref }}
   cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+  # dev pushes wait in order instead of cancelling the run that waits. GitHub
+  # rejects queueing together with cancel-in-progress, so pull requests keep
+  # the single waiting run that a newer push replaces.
+  queue: ${{ github.event_name == 'pull_request' && 'single' || 'max' }}
 
 permissions:
   contents: read
diff --git a/doc/ci.md b/doc/ci.md
index 45f0102af..8e9bc513f 100644
--- a/doc/ci.md
+++ b/doc/ci.md
@@ -532,6 +532,29 @@ still 13% with a warm sccache, which would take over 2 GiB of the repository's
 repositories pay for, or a self-hosted Windows runner without a hypervisor
 would cost about as much as another general WHP runner.
 
+## Run concurrency
+
+Each pull request runs in a concurrency group of its own, and a push to it
+cancels the run of the previous head, whether that run waits or runs. Pushes
+to `dev` share one group and run one at a time in the order in which they
+arrive. The group keeps up to 100 waiting runs (`queue: max`), so a merge does
+not cancel the waiting run of an earlier merge. Until October 2026 the group
+kept a single waiting run, which each newer push replaced: 12 of the 199 `dev`
+pushes from October 4 to 21 lost their run that way, five of them merged
+within four minutes on October 20. Those commits got no development release
+and no performance results, and because the detector compares a push with its
+previous commit, the run that followed classified only its own changes.
+GitHub rejects `queue: max` together with `cancel-in-progress: true`, so the
+workflow derives both settings from the event. Each scheduled or manually
+dispatched full run has a group of its own (see
+[Change classification](#change-classification)).
+
+A queued run waits for the runs ahead of it. Replaying the `dev` pushes of
+October 4 to 21 one at a time, at their measured durations and the median one
+for those that were cancelled, half would have started at once, one in ten
+would have waited at least 50 minutes, and the longest wait, on October 20,
+would have been about three hours.
+
 ## Change classification
 
 `openvmm-changes` lists the files that changed between the pull request's base
diff --git a/scripts/test_nvx_tools.py b/scripts/test_nvx_tools.py
index 7a553f8c4..394e91fca 100644
--- a/scripts/test_nvx_tools.py
+++ b/scripts/test_nvx_tools.py
@@ -4030,6 +4030,26 @@ def test_ci_full_run_triggers_and_concurrency_are_separate(self):
             header,
         )
 
+    def test_ci_dev_pushes_queue_while_pull_requests_cancel(self):
+        workflow = (
+            BuildConstants.REPO_ROOT / ".github" / "workflows" / "ci.yml"
+        ).read_text(encoding="utf-8")
+        header = workflow.split("\njobs:\n", 1)[0]
+        concurrency = header.split("\nconcurrency:\n", 1)[1].split("\n\n", 1)[0]
+
+        # GitHub rejects `queue: max` with `cancel-in-progress: true`, so a
+        # static queue mode would stop every pull-request run from starting.
+        self.assertIn(
+            "\n  cancel-in-progress: ${{ github.event_name == 'pull_request' }}\n",
+            concurrency,
+        )
+        self.assertIn(
+            "\n  queue: ${{ github.event_name == 'pull_request' && "
+            "'single' || 'max' }}",
+            concurrency,
+        )
+        self.assertNotRegex(concurrency, r"(?m)^\s*queue:\s*['\"]?max")
+
     def test_ci_full_runs_short_circuit_change_detection(self):
         workflow = (
             BuildConstants.REPO_ROOT / ".github" / "workflows" / "ci.yml"