Preserve pending post-merge benchmark runs

.github/workflows/benchmarks.yml

@@ -12,6 +12,8 @@ permissions:
 
 concurrency:
   group: benchmarks-${{ github.ref }}
+  # Keep each merged revision; the default single pending run replaces its predecessor.
+  queue: max
   cancel-in-progress: false
 
 jobs:

TODO.md

@@ -797,6 +797,10 @@ EP/JC/P work; candidates are not speedup claims until measured on both platforms
 
 ## Benchmark publication follow-up
 
+- [x] Preserve up to 100 pending main benchmark runs instead of replacing the previous pending
+      merge's measurement; keep active runs and exact revision identities.
+- [ ] Reconcile missing main measurements after queue cancellation or overflow against retained
+      reports and source artifacts; backfill only missing revisions, without replacing existing data.
 - [ ] Retain separate macOS and Linux HTML reports plus raw JSON, with commit, build,
       hardware/runner and measurement-contract identities. Local temporary previews are not release assets.
 - [ ] Publish post-merge reports as official results and pre-merge reports as labeled PR previews.

docs/benchmark-comparisons.md

@@ -42,6 +42,20 @@ local history is needed.
 Use the [historical backfill workflow](benchmark-backfill.md) to build and measure a frozen revision
 list in one host session, with a common driver and reference-tool set.
 
+## Retaining measurements for successive merges
+
+Each merge into `main` requests its own post-merge measurement. The `Benchmarks` workflow
+serializes runs for the same ref with `queue: max` and `cancel-in-progress: false`. This keeps
+up to 100 pending runs alongside the active run. GitHub's default `queue: single` replaces
+an older pending run when a newer merge arrives, even when cancellation of the active run
+is disabled. See [GitHub's concurrency queue rules](https://docs.github.com/en/actions/how-tos/write-workflows/choose-when-workflows-run/control-workflow-concurrency).
+
+The queue is bounded: additional runs are canceled once 100 are pending. This change does not
+recover earlier canceled runs or guarantee measurement completion after runner or workflow
+failures. Use the [backfill workflow](benchmark-backfill.md) for missing revisions. Queue order
+also does not establish revision order; retained reports keep their exact commit and run identity.
+PR measurements and publication use their own concurrency policies.
+
 ## Relative performance and advisory alarm
 
 Each shape also has a CPU-grouped ratio table for participants doing the same task. It shows

docs/performance-analysis.md

@@ -4970,3 +4970,18 @@ manual metadata reconciliation continues normally. API failures still fail rathe
 than masquerading as obsolete data. Fresh previews pay one bounded source/PR API check;
 no fixed elapsed-time improvement is claimed. Old queued workflow definitions retain
 their original steps, and rejected producer events still need independent recovery.
+
+### Post-merge measurement queue admission
+
+Main benchmark run `42354657340` for merge `1278c8943a6abd236a2f20e3937cc8e23a0f95a7`
+was automatically canceled while pending when the next main merge requested a run. The workflow
+already disabled active-run cancellation, but omitted `queue`, whose default permits only one
+pending run. This loses a requested revision before measurement and is separate from CPU-count
+agreement, runner performance, and publication backlogs. No agent cancellation caused this event.
+
+Use `queue: max` for the existing per-ref main measurement concurrency group, keeping
+`cancel-in-progress: false`. A committed workflow regression checks all three properties.
+GitHub permits up to 100 pending runs; overflow still cancels additional requests. Existing
+canceled revisions need explicit reconciliation/backfill, not a claim that changing the workflow
+recovered them. Measurement commands, sampling, CPU selection, artifact identity, PR checks and
+publication policies are unchanged. There is no runtime or elapsed-time improvement claim.

tools/benchmark_history_test.py

@@ -185,6 +185,13 @@ def test_comparisons_are_rendered_and_bound_to_head_binary(self):
             with self.assertRaisesRegex(ValueError, "head binary"):
                 history.retain(Path(directory), value, source(2))
 
+    def test_main_measurement_queue_preserves_pending_merged_revisions(self):
+        workflow = repository_file('.github/workflows/benchmarks.yml').read_text()
+        concurrency = workflow.split('\nconcurrency:\n', 1)[1].split('\njobs:', 1)[0]
+        self.assertIn('group: benchmarks-${{ github.ref }}', concurrency)
+        self.assertRegex(concurrency, r'(?m)^  queue: max$')
+        self.assertRegex(concurrency, r'(?m)^  cancel-in-progress: false$')
+
     def test_workflow_keeps_measurement_unprivileged_and_publication_serialized(self):
         measure = repository_file(".github/workflows/benchmarks.yml").read_text()
         shard_action = repository_file(".github/actions/benchmark-measure/action.yml").read_text()