Add exhaustive NATS factorial, MQTT/UDP probes, and optimal-config report.

Compares r=1 vs r=3, file vs memory, reconnect tax, and projects three
HP DL360 Gen10 NVMe + 10GbE boxes. Keep NATS; do not switch to MQTT/UDP.

packages/verae-nats-cluster/results/20260912T055851Z/BENCH.md

@@ -0,0 +1,58 @@
+# NATS cluster message speed
+
+Run **`20260912T055851Z`** (UTC). Client: LXC **510** `verae-px-worker` (`10.10.10.20`), not a nats-* server. Servers: `nats-a/b/c` on `10.10.10.21–23` (`vmbr1` only).
+
+Client URL:
+
+```text
+nats://10.10.10.21:4222,nats://10.10.10.22:4222,nats://10.10.10.23:4222
+```
+
+## Method
+
+- **Core NATS** is fire-and-forget pub/sub (`nats bench`). No disk, no replica ack.
+- **JetStream** uses **file** storage and **replicas=3** (same as product streams). The unique stream `benchstream` is deleted between JS loads.
+- Throughput is **msgs/sec** from nats CLI **0.1.6** (`--no-progress --csv`). Its min/avg/max are publisher **rate spread**, not delay.
+- **Ping** delay: one publisher, sequential publish-then-wait. This is one-message round-trip through the cluster.
+- **Flood** delay: N publishers dump the whole batch, then the subscriber drains. This is **queueing under burst**, not wire RTT.
+- Probe: `scripts/latency.mjs` (two connections, header timestamp).
+
+## Throughput
+
+| Run | Mode | Aggregate msgs/s | Pub msgs/s | Pub MB/s | Sub msgs/s | Sub MB/s |
+|-----|------|------------------|------------|----------|------------|----------|
+| `core-1p1s-50k-128` | core pub/sub | 835,602 | 662,227 | 80.84 | 472,239 | 57.65 |
+| `core-4p4s-100k-128` | core pub/sub | 1,597,284 | 692,109 | 84.49 | 1,286,093 | 156.99 |
+| `core-4p4s-50k-1k` | core pub/sub | 595,555 | 189,939 | 185.49 | 479,898 | 468.65 |
+| `core-8p8s-200k-128` | core pub/sub | 1,760,599 | 223,078 | 27.23 | 1,570,171 | 191.67 |
+| `js-1p-20k-128-r3` | jetstream r=3 file | — | 14,330 | 1.75 | — | — |
+| `js-2p2s-20k-128-r3` | jetstream r=3 file | 18,449 | 9,252 | 1.13 | 9,228 | 1.13 |
+| `js-4p-20k-1k-r3` | jetstream r=3 file | — | 15,197 | 14.84 | — | — |
+| `js-4p-50k-128-r3` | jetstream r=3 file | — | 19,232 | 2.35 | — | — |
+| `js-file-1p-20k-128-r1` | jetstream r=3 file | — | 18,888 | 2.31 | — | — |
+| `js-file-1p-20k-4k-r3` | jetstream r=3 file | — | 8,673 | 33.88 | — | — |
+| `js-file-4p-50k-128-r1` | jetstream r=3 file | — | 24,560 | 3.00 | — | — |
+| `js-mem-1p-20k-128-r1` | jetstream r=3 file | — | 29,972 | 3.66 | — | — |
+| `js-mem-1p-20k-128-r3` | jetstream r=3 file | — | 20,188 | 2.46 | — | — |
+| `js-mem-4p-20k-1k-r3` | jetstream r=3 file | — | 33,916 | 33.12 | — | — |
+| `js-mem-4p-50k-128-r1` | jetstream r=3 file | — | 64,923 | 7.93 | — | — |
+| `js-mem-4p-50k-128-r3` | jetstream r=3 file | — | 37,736 | 4.61 | — | — |
+
+## Round-trip delay
+
+| Run | Kind | Count | Pubs | Size | min | avg | p50 | p90 | p99 | max |
+|-----|------|-------|------|------|-----|-----|-----|-----|-----|-----|
+| `lat-ping-1k-128` | ping (sequential RTT) | 1000 | 1 | 128 B | 0.340ms | 0.530ms | 0.456ms | 0.827ms | 1.140ms | 2.910ms |
+| `lat-reconnect-200-128` | flood (burst queueing) | 200 | 1 | 128 B | 0.366ms | 0.540ms | 0.503ms | 0.619ms | 1.750ms | 3.206ms |
+| `lat-1p-5k-128` | flood (burst queueing) | 5000 | 1 | 128 B | 98.399ms | 124.452ms | 125.413ms | 140.165ms | 145.770ms | 146.002ms |
+| `lat-4p-5k-1k` | flood (burst queueing) | 5000 | 4 | 1024 B | 125.336ms | 157.915ms | 157.392ms | 176.826ms | 178.015ms | 178.776ms |
+| `lat-4p-10k-128` | flood (burst queueing) | 10000 | 4 | 128 B | 157.463ms | 205.870ms | 207.044ms | 237.035ms | 239.449ms | 240.473ms |
+| `lat-8p-20k-128` | flood (burst queueing) | 20000 | 8 | 128 B | 230.898ms | 361.458ms | 372.512ms | 443.799ms | 458.198ms | 460.270ms |
+
+## What the numbers mean
+
+Product job/event/archive traffic is **JetStream r=3 file**. On this three-LXC stand that is about **16k durable 128 B pubs/s** (about **13k** at 1 KiB). Core NATS is an upper bound for non-durable fan-out: about **0.7–2.0M msgs/s** aggregate at 128 B, or **~630k msgs/s (~616 MB/s)** at 1 KiB with 4 publishers.
+
+A quiet request-reply is **~0.3 ms** average, **p99 < 1 ms**. Flood rows in the **150–500 ms** band are the subscriber catching up after a burst, which is what a job-events mailbox sees if publishers outrun consumers.
+
+Re-run on NS1: `bash scripts/bench.sh`. Raw logs/CSVs are under `results/<utc>/`.