Make the daily soak exercise real workflows and report useful results

.github/workflows/server-perf-soak.yml

@@ -50,6 +50,7 @@ jobs:
           DW_PERF_CONCURRENCY: ${{ github.event_name == 'workflow_dispatch' && inputs.concurrency || '24' }}
           DW_PERF_NAMESPACES: "8"
           DW_PERF_TASK_QUEUES: "16"
+          DW_PERF_STANDARD_WORKFLOWS: "true"
           DW_PERF_REDIS_CACHE_DB: "1"
           DW_PERF_MAX_SERVER_MEMORY_MB: "1024"
           DW_PERF_MAX_POLLING_KEYS: "2048"
@@ -66,10 +67,20 @@ jobs:
           DW_PERF_RUNNER_ENVIRONMENT: "self-hosted"
         run: scripts/perf/run-vultr-soak.sh
 
+      - name: Summarize measured workload
+        if: always()
+        run: |
+          if [ -f build/perf/summary.md ]; then
+            cat build/perf/summary.md >> "$GITHUB_STEP_SUMMARY"
+          else
+            printf '## Server endurance soak\n\nNo complete measurement summary was produced. Inspect setup, timeout, and provisioning logs; this is not a pass.\n' >> "$GITHUB_STEP_SUMMARY"
+          fi
+
       - name: Upload perf artifacts
         if: always()
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7
         with:
           name: server-perf-soak
           path: build/perf/
           if-no-files-found: warn
+          retention-days: 14

benchmarks/capacity/v1/bindings/php/capacity_adapter.php

@@ -623,6 +623,10 @@ function capacityConformanceEvidence(array $fixtures): array
     ];
 }
 
+if (realpath($_SERVER['SCRIPT_FILENAME'] ?? '') !== __FILE__) {
+    return;
+}
+
 $mode = $argv[1] ?? '';
 if ($mode === 'describe') {
     echo json_encode(capacityAdapterDescriptor(), JSON_THROW_ON_ERROR | JSON_UNESCAPED_SLASHES).PHP_EOL;

docker-compose.yml

@@ -111,6 +111,7 @@ services:
         WORKFLOW_PACKAGE_REF: ${WORKFLOW_PACKAGE_REF:-}
         WORKFLOW_PACKAGE_COMMIT: ${WORKFLOW_PACKAGE_COMMIT:-}
     command: php artisan queue:work --sleep=1 --tries=3 --max-time=3600
+    restart: unless-stopped
     healthcheck:
       test: ["CMD", "server-process-healthcheck", "worker"]
       interval: 5s

docs/perf-runner.md

@@ -1,7 +1,49 @@
 # Server Perf Runner
 
 The server perf harness exercises the HTTP worker polling path for bounded
-memory growth and polling-cache cleanup.
+memory growth and polling-cache cleanup, with a low-rate PHP standard-workflow
+canary alongside it. The daily Action summary exposes both results directly.
+
+This is an endurance test, not a capacity claim: the canary uses the existing
+`capacity.v1.one_activity` binding (one 1-KiB echo activity), one workflow at a
+time, with at most one start every five seconds. It verifies each result,
+completed status, and ordered activity history. It skips missed start slots
+instead of accumulating a backlog. Its rate is not the capacity suite's
+concurrency/warmup measurement protocol. A slow correct completion is measured,
+not rejected against an invented throughput target; a result timeout is 30s.
+
+Health, readiness, and cluster-info probes run even without workflow-row growth.
+Artifacts include HTTP p50/p95/p99, backpressure separately from successful
+polls, sampled CPU/memory for Server, its queue worker/scheduler, MySQL, Redis,
+and the PHP SDK process container. The SDK client and worker share a 0.5-CPU,
+256-MiB container on the existing host. A 200 with an invalid health payload,
+missing probe results, or an entirely backpressured poll run fails validation.
+Compose restarts the Server queue worker after its configured hourly recycle;
+an exited queue worker must not disappear unnoticed for the second soak hour.
+Each synthetic polling thread owns and heartbeats one registration, distributed
+across the configured namespaces and queues (24 workers, not 128 Cartesian
+registrations, in the daily default). HTTP success carrying a stale/rejected
+worker status fails the run. Polling-cache activity must be present in at least
+the configured minimum sample fraction, not merely once at startup.
+
+The default post-load drain is 310 seconds. Real activity poll-result caches
+retain replay evidence through their five-minute lease plus five seconds;
+heartbeat retention throttles and wake signals also outlive the old 12-second
+idle-only drain. No cache flush, shorter production TTL, or relaxed final-key
+limit is used to obtain a pass. The load remains two hours; build/bootstrap,
+bounded drain, and teardown are additional host time.
+
+`DW_PERF_STANDARD_WORKFLOWS=false` disables only the SDK canary for a focused
+polling reproduction. The Compose wrapper enables it by default. Raw
+`server_soak.py` invocation does not build the SDK image; use the wrapper.
+
+The soak still does not exercise timer/signal/query execution, recovery fault
+injection, all SDK bindings, or HA. Those need the separate conformance and
+performance qualification experiments. Zero cache keys for an unexercised
+policy do not prove that feature works. Review failed setup and cleanup steps
+as well as workload results; neither is a successful experiment. Scheduled
+artifacts are retained for 14 days; record significant findings on the owning
+issue rather than committing generated run directories.
 
 ## Runner Shape
 

scripts/perf/Dockerfile.sdk

@@ -0,0 +1,14 @@
+FROM composer:2 AS dependencies
+WORKDIR /sdk
+COPY benchmarks/capacity/v1/bindings/php/composer.json benchmarks/capacity/v1/bindings/php/composer.lock ./
+RUN composer install --no-dev --no-interaction --no-progress --prefer-dist
+
+FROM php:8.3-cli
+RUN docker-php-ext-install pcntl
+WORKDIR /app
+COPY benchmarks/capacity/v1/bindings/php benchmarks/capacity/v1/bindings/php
+COPY --from=dependencies /sdk/vendor benchmarks/capacity/v1/bindings/php/vendor
+COPY scripts/perf/standard_workflow_soak.php scripts/perf/standard_workflow_soak.php
+USER 1000:1000
+ENTRYPOINT ["php"]
+CMD ["benchmarks/capacity/v1/bindings/php/capacity_adapter.php", "worker"]

scripts/perf/run-server-soak.sh

@@ -35,10 +35,11 @@ PROMETHEUS_CONTAINER="${PROJECT}-prometheus"
 PROMETHEUS_CONFIG_DIR=""
 
 mkdir -p "$ARTIFACT_DIR"
+export DW_PERF_DRAIN_SECONDS="${DW_PERF_DRAIN_SECONDS:-310}"
 
 if [ -z "$LOAD_TIMEOUT_SECONDS" ]; then
   DURATION_SECONDS="${DW_PERF_DURATION_SECONDS:-120}"
-  DRAIN_SECONDS="${DW_PERF_DRAIN_SECONDS:-12}"
+  DRAIN_SECONDS="$DW_PERF_DRAIN_SECONDS"
   LOAD_TIMEOUT_SECONDS=$((DURATION_SECONDS + DRAIN_SECONDS + 300))
   if [ "$LOAD_TIMEOUT_SECONDS" -lt 300 ]; then
     LOAD_TIMEOUT_SECONDS=300
@@ -57,6 +58,7 @@ export DW_WORKER_TOKEN="${DW_WORKER_TOKEN:-}"
 export DW_OPERATOR_TOKEN="${DW_OPERATOR_TOKEN:-}"
 export DW_ADMIN_TOKEN="${DW_ADMIN_TOKEN:-}"
 export DW_AUTH_BACKWARD_COMPATIBLE="${DW_AUTH_BACKWARD_COMPATIBLE:-true}"
+export DW_PERF_STANDARD_WORKFLOWS="${DW_PERF_STANDARD_WORKFLOWS:-true}"
 
 OVERRIDE_FILE="$ARTIFACT_DIR/docker-compose.perf.yml"
 cat > "$OVERRIDE_FILE" <<YAML
@@ -96,6 +98,8 @@ services:
     ports: !override []
 YAML
 
+compose=(docker compose -p "$PROJECT" -f "$ROOT_DIR/docker-compose.yml" -f "$OVERRIDE_FILE" -f "$ROOT_DIR/scripts/perf/standard-workflow.compose.yml")
+
 cleanup() {
   local status=$?
 
@@ -104,13 +108,14 @@ cleanup() {
   docker logs "${PROJECT}-scheduler-1" > "$ARTIFACT_DIR/scheduler.log" 2>&1 || true
   docker logs "${PROJECT}-mysql-1" > "$ARTIFACT_DIR/mysql.log" 2>&1 || true
   docker logs "${PROJECT}-redis-1" > "$ARTIFACT_DIR/redis.log" 2>&1 || true
+  docker logs "${PROJECT}-soak-sdk-1" > "$ARTIFACT_DIR/soak-sdk.log" 2>&1 || true
 
   docker rm -f "$PROMETHEUS_CONTAINER" >/dev/null 2>&1 || true
   if [ -n "$PROMETHEUS_CONFIG_DIR" ]; then
     rm -rf "$PROMETHEUS_CONFIG_DIR"
   fi
 
-  docker compose -p "$PROJECT" -f "$ROOT_DIR/docker-compose.yml" -f "$OVERRIDE_FILE" down -v --remove-orphans || true
+  "${compose[@]}" --profile standard-soak down -v --remove-orphans || true
   exit "$status"
 }
 trap cleanup EXIT
@@ -224,7 +229,10 @@ else
   echo "Starting perf stack with project ${PROJECT} on a dynamic host port"
 fi
 setup_status=0
-docker compose -p "$PROJECT" -f "$ROOT_DIR/docker-compose.yml" -f "$OVERRIDE_FILE" up -d --build --wait || setup_status=$?
+"${compose[@]}" up -d --build --wait || setup_status=$?
+if [ "$setup_status" -eq 0 ] && [ "$DW_PERF_STANDARD_WORKFLOWS" = true ]; then
+  "${compose[@]}" build soak-sdk || setup_status=$?
+fi
 if [ "$setup_status" -ne 0 ]; then
   echo "Perf environment setup failed before product smoke execution; docker compose could not build or start the stack." >&2
   write_environment_setup_failure "$setup_status" "docker_compose_up" "docker compose failed before server_soak.py started"

scripts/perf/run-vultr-soak.sh

@@ -266,6 +266,7 @@ for name in \
   DW_PERF_CONCURRENCY \
   DW_PERF_NAMESPACES \
   DW_PERF_TASK_QUEUES \
+  DW_PERF_STANDARD_WORKFLOWS \
   DW_PERF_REDIS_CACHE_DB \
   DW_PERF_MAX_SERVER_MEMORY_MB \
   DW_PERF_MAX_POLLING_KEYS \