ci(e2e): isolate concurrent integration rounds

Signed-off-by: Drew Newberry <anewberry@nvidia.com>

Author: drew

.github/workflows/branch-e2e.yml

@@ -242,6 +242,7 @@ jobs:
     uses: ./.github/workflows/integration-runner.yml
     with:
       category: conformance
+      stability-campaign: ${{ inputs.stability_campaign }}
       source-sha: ${{ needs.prepare-integration.outputs.source_sha }}
       integration-inputs-artifact-id: ${{ needs.prepare-integration.outputs.integration_inputs_artifact_id }}
       test-matrix: >-
@@ -261,6 +262,7 @@ jobs:
     uses: ./.github/workflows/integration-runner.yml
     with:
       category: feature-specific
+      stability-campaign: ${{ inputs.stability_campaign }}
       source-sha: ${{ needs.prepare-integration.outputs.source_sha }}
       integration-inputs-artifact-id: ${{ needs.prepare-integration.outputs.integration_inputs_artifact_id }}
       test-matrix: >-
@@ -279,6 +281,7 @@ jobs:
     uses: ./.github/workflows/integration-runner.yml
     with:
       category: driver-specific
+      stability-campaign: ${{ inputs.stability_campaign }}
       source-sha: ${{ needs.prepare-integration.outputs.source_sha }}
       integration-inputs-artifact-id: ${{ needs.prepare-integration.outputs.integration_inputs_artifact_id }}
       test-matrix: >-

.github/workflows/integration-runner.yml

@@ -18,6 +18,11 @@ on:
         description: Integration-test category, used for the job name and concurrency key
         required: false
         type: string
+      stability-campaign:
+        description: Keep concurrent stability rounds independent
+        required: false
+        type: boolean
+        default: false
       test-matrix:
         description: JSON array of environment, installer, and testsuite combinations
         required: false
@@ -35,8 +40,8 @@ permissions:
   packages: read
 
 concurrency:
-  group: ${{ github.workflow }}-integration-${{ inputs.category }}-${{ github.ref }}
-  cancel-in-progress: true
+  group: ${{ github.workflow }}-integration-${{ inputs.category }}-${{ github.ref }}-${{ inputs.stability-campaign && github.run_id || 'normal' }}
+  cancel-in-progress: ${{ !inputs.stability-campaign }}
 
 jobs:
   integration:

CI.md

@@ -42,7 +42,9 @@ database scenarios, TypeScript interactive exec, mechanistic policy checks, and
 the gateway-without-drivers contract. It excludes the experimental WSL GPU job. The standard
 PR and merge-queue E2E selection is unchanged. Each campaign dispatch gets a
 unique concurrency key and runtime-image tag, so up to five runs can execute
-at once without sharing mutable image tags. Dispatch ten runs in two waves of
+at once without sharing mutable image tags. The reusable integration runner
+also keys its conformance, feature, and driver jobs by run ID in stability mode;
+ordinary PR runs retain their existing cancellation behavior. Dispatch ten runs in two waves of
 five, retain each run ID and per-job result, and count failures and skips.
 Run `mise run test` once on the same commit before the E2E rounds.
 

architecture/build.md

@@ -106,7 +106,7 @@ The branch E2E workflow composes reusable integration, Docker, Podman, VM,
 Kubernetes, and GPU jobs from commit-addressed build artifacts. Its manual
 stability mode adds the normally optional Kubernetes HA and credential-driver
 lanes and limits GPU jobs to Linux runners. Each dispatch gets its own image
-tag and concurrency key, allowing bounded parallel runs against one immutable
+tag and concurrency key, including in the integration runner, allowing bounded parallel runs against one immutable
 candidate commit without changing PR or merge-queue gates. Stability mode adds
 the managed Podman driver path, Kubernetes database scenarios, and Docker-hosted
 SDK and policy-advisor E2E checks.