docs(benchmarks): publish validated AWS baseline

Author: rahulbsw

ROADMAP.md

@@ -163,12 +163,17 @@ StreamForge aims to be the **fastest, most reliable, and easiest-to-use Kafka se
 - [x] Add single-destination produced accounting and focused regression tests
 - [x] Pass the loopback-only local Podman sustained validation with exact
   consumed, produced, delivered, output, and error counts
+- [x] Replace ad hoc AWS host provisioning with a cost-bounded private
+  Terraform and ECS-on-EC2 benchmark environment
+  - Keep the task and host in a private subnet with no public IP, internet
+    gateway, NAT gateway, load balancer, SSH access, or public ingress
+  - Gate billable runtime behind an explicit flag and hard expiry
+  - Validate a publication-eligible three-repetition sustained baseline and
+    destroy all provisioned resources
 - [ ] Run the corrected legacy/partition-ordered and
   acknowledged/queued live Kafka comparison matrix
 - [ ] Produce a clean-worktree, matched Java/Rust comparison before publishing
-  a public throughput claim
-- [ ] Replace ad hoc AWS host provisioning with cost-bounded Terraform and
-  ECS-on-EC2 benchmark jobs after the local comparison matrix passes
+  a comparative throughput claim
 - [ ] Implement rebalance-aware completed-offset coordination before supporting
   partition-ordered manual commits
 - [ ] Profile transform-heavy and aggregation-heavy workloads

docs/IMPLEMENTATION_STATUS.md

@@ -95,8 +95,8 @@ Implemented in the current source:
   rebalance-aware completed-offset coordination is implemented.
 
 These modes are implemented and unit-tested. The partition-ordered/queued
-combination has completed a valid sustained local Kafka run; the remaining
-mode comparison matrix and a new AWS run are still pending.
+combination has completed valid sustained local and private AWS Kafka runs; the
+remaining mode comparison matrix is still pending.
 
 Default compatibility values remain:
 
@@ -169,8 +169,32 @@ Verified for the current source on 2026-07-25 UTC:
 - Kafka was published only on `127.0.0.1:9092`; the benchmark network was
   internal; ingress and output runners exposed no ports. Final topic and disk
   reclamation checks passed.
-- No AWS rerun has been attempted after the valid local result. All resources
-  from the previous AWS attempt remain verified deleted.
+- The corrected private AWS harness passed three 120-second
+  partition-ordered/queued repetitions after a one-million-record untimed
+  warm-up per repetition. Each repetition reconciled exactly 16,200,000
+  records across ingress, input offsets, consumed, produced, broker-delivered,
+  output offsets, and the independent output validator, with zero errors.
+- Publication-eligible AWS aggregate at commit
+  `d848f118e62b41c7605250c69c9da087af688d0c`: median
+  `106,692.649362 msg/s`, minimum `106,651.683407`, maximum
+  `106,798.575491`, coefficient of variation `0.0580%`, median `1.124`
+  StreamForge mean cores, and median peak RSS `137.1 MiB`.
+- Environment: AWS `c7i.2xlarge`, Intel Xeon Platinum 8488C, 8 partitions,
+  8 threads, and target ingress of 135,000 messages/second. All repetitions
+  were classified as engine-saturated rather than ingress-limited.
+- The workload ran in a private subnet with no public IP, internet gateway,
+  NAT gateway, load balancer, SSH access, or public security-group rule.
+  PrivateLink endpoint ingress used security-group references only.
+- The ECR image scan completed with zero critical findings. The environment was
+  intentionally isolated and ephemeral; inherited high and medium base-image
+  findings remain a reason not to treat this benchmark image as a production
+  runtime image.
+- `terraform destroy` reported 45 resources destroyed. Direct AWS service
+  inventories then found zero live benchmark EC2 instances, volumes, ENIs,
+  VPCs, endpoints, S3 buckets, ECR repositories, CodeBuild projects, IAM roles,
+  active task definitions, or Auto Scaling groups. ECS retains only an
+  `INACTIVE` deleted-cluster record with zero registered, running, or pending
+  tasks.
 
 Previously verified on 2026-07-24:
 
@@ -214,9 +238,9 @@ from a blanket SIMD rewrite:
    partition-ordered/acknowledged, and partition-ordered/queued.
 2. Run a clean-worktree, matched Java/Rust comparison with identical payloads,
    partitions, acknowledgement semantics, warm-up, duration, and validation.
-3. Provision the later AWS benchmark with cost-bounded Terraform and
-   ECS-on-EC2 jobs only after the local matrix passes; keep all endpoints
-   private or restricted to the user's IP.
+3. Reuse the validated cost-bounded private Terraform/ECS environment for
+   matched reruns; keep the billable runtime disabled except during a bounded
+   measurement window.
 4. Keep raw/lazy envelope work behind the existing 30% parse/serialization
    threshold; the AWS passthrough profile measured about 16.5% parsing and 3.2%
    serialization on overlapping inclusive stacks.

docs/PERFORMANCE.md

@@ -10,10 +10,37 @@ StreamForge performance depends on payload size, partition count, broker and
 network latency, filter and transform complexity, destination fan-out, delivery
 semantics, and available CPU and memory.
 
-No headline throughput result is published here. A result belongs in public
-documentation only when it comes from a reproducible end-to-end comparison,
-uses the same workload and delivery guarantees as the comparison target, and
-improves the approved baseline without correctness regressions.
+A throughput result belongs in public documentation only when it comes from a
+reproducible end-to-end run and passes the count, error, warm-up, duration, and
+variance gates below. Comparisons with another implementation additionally
+require the same workload and delivery guarantees.
+
+## Validated sustained baseline
+
+On 2026-07-25, the private AWS harness measured a median output-delivery rate of
+**106,693 messages/second** across three 120-second repetitions. The range was
+106,652–106,799 messages/second with 0.058% coefficient of variation.
+
+| Workload property | Value |
+|---|---|
+| Mode | `partition_ordered` processing with queued delivery |
+| Host | AWS `c7i.2xlarge`, Intel Xeon Platinum 8488C |
+| Parallelism | 8 Kafka partitions, 8 StreamForge threads |
+| Ingress target | 135,000 messages/second |
+| Warm-up | 1,000,000 untimed records per repetition |
+| Timed validation | 16,200,000 records per repetition |
+| Correctness | Exact input, consumed, produced, delivered, and output counts; zero errors |
+| Resource use | 1.124 median mean CPU cores; 137.1 MiB median peak RSS |
+
+The workload was a deterministic passthrough test on commit
+`d848f118e62b41c7605250c69c9da087af688d0c`. It ran in a private subnet with no
+public IP, internet gateway, NAT gateway, load balancer, SSH access, or public
+security-group rule. Terraform destroyed all 45 resources after collection.
+
+This is a StreamForge baseline, not a Java comparison. A matched Java/Rust run
+remains required before making a relative implementation claim. The
+[full result and validation contract](https://github.com/rahulbsw/streamforge/blob/main/docs/benchmarks/results/BENCHMARK_RESULTS.md)
+are retained with the repository evidence.
 
 ## Runtime controls
 

docs/benchmarks/README.md

@@ -16,6 +16,8 @@ scripts/benchmarks/run_throughput_test.sh
 Criterion microbenchmarks remain under `benches/`.
 
 [`results/phase2-baseline-20260724.md`](results/phase2-baseline-20260724.md) is
-a historical coupled-harness record. No schema-version-3 sustained result is a
-public baseline yet; the first validated local run was intentionally withheld
-because its worktree and saturation gates were not publication-eligible.
+a historical coupled-harness record. The current publication-eligible
+schema-version-3 AWS sustained baseline is
+[`results/BENCHMARK_RESULTS.md`](results/BENCHMARK_RESULTS.md). The first
+corrected local run remains diagnostic because its worktree and saturation
+gates were not publication-eligible.