Preserve active benchmark cohorts across automatic pushes

Automatic main pushes must queue behind an active benchmark cohort instead of discarding its unfinished measurements. Manual dispatches still supersede old runs, and intentional replacement after a repair cancels the obsolete run first.

Upload partial original worker reports and logs on cancellation without finalizing or qualifying them. Publication still requires complete valid evidence. Reference the canonical runner policy from session documentation.

Verified: actionlint, repository governance and diff whitespace checks. This checkpoint changes workflow and documentation only. Earlier full Debug build and exact-source Linux native controls passed; 27 external workers from run40555053148 were verified before another push cancelled the last three.

Author: KSemenenko

.github/workflows/benchmarks.yml

@@ -12,7 +12,7 @@ on:
 permissions: {contents: read, actions: read}
 concurrency:
   group: keyload-benchmarks-${{ github.ref }}
-  cancel-in-progress: true
+  cancel-in-progress: ${{ github.event_name == 'workflow_dispatch' }}
 env:
   NUGET_PACKAGES: ${{ github.workspace }}/artifacts/benchmark-runtime/packages
 jobs:
@@ -177,7 +177,7 @@ jobs:
           KEYLOAD_WORKLOAD_OUTCOME: ${{ steps.workload.outcome }}
         run: node scripts/Features/BenchmarkComparisons/session-finalize.mjs
       - name: Save benchmark results
-        if: ${{ !cancelled() }}
+        if: ${{ always() }}
         uses: actions/upload-artifact@v7
         with:
           name: ${{ matrix.artifactName }}

docs/Features/BenchmarkComparisons/SessionExecution.md

@@ -174,8 +174,13 @@ matrix, restarts or separate benchmark scenario.
   other databases, workload fan-out or public figures. This focused check does not
   qualify the full eight-control gate or complete session suite; those gates stay
   unchanged for `mode=contracts`, `mode=full` and push runs.
-  A newer run of the same branch cancels the superseded run instead of waiting
-  behind its measurements, including when changing between manual modes and full runs.
+  All modes share one branch concurrency group. A manual dispatch cancels the
+  superseded run, including when changing between manual modes and full runs.
+  Automatic push runs preserve the active cohort and queue the newest pending
+  revision, so another change cannot discard hours of unfinished measurements.
+  Intentional replacement after a repair cancels the obsolete run first.
+  Workers retain partial original reports and logs on cancellation; those artifacts
+  remain incomplete evidence and cannot authorize publication.
 - Ingestion at 1/10/500 clients, the minimum measured operation counts, recall,
   replicated acknowledgement, resource bounds and all required workload families are
   kept. The number of jobs is not the number of measured scenarios.
@@ -250,9 +255,10 @@ its byte size remain reviewable when several database containers are running.
 
 Local execution retains the actual client binary/source/process identity with no
 invented GitHub provenance; every database server, KeyLoad included, is identified
-by its image reference and digest. Session and aggregation jobs declare the
-explicit `ubuntu-24.04` runner: the `ubuntu-latest` image moved to Ubuntu 26.04 with a
-7.0 kernel that the pinned MongoDB server refuses to start on. It cannot authorize public figures. External
+by its image reference and digest. Preparation, session and aggregation jobs use
+the common runner family documented in
+[Additional Databases](AdditionalDatabases.md#deployment-and-qualification-limits).
+Local execution cannot authorize public figures. External
 advertised-endpoint reachability remains open until actual execution proves it;
 measured workload reports do not close that independent gate.