Track 1. Cost of scaling (closed-loop POST /messages)
Fixed total API budget: 1x2.0/1G, 2x1.0/512M, 4x0.5/256M
Profile            reps  RPS median (min..max)   p95 ms median   errors
1x2.0cpu_1g           2   100.2 (98.5..101.9)      308.6        0
2x1.0cpu_512m         2    99.0 (95.2..103.0)      655.2        0
4x0.5cpu_256m         2    88.6 (85.6..91.6)     1199.7        0

Growth series: 4 replicas, per-replica CPU/RAM
Profile            reps  RPS median (min..max)   p95 ms median   errors
4x0.5cpu_256m         2    90.7 (89.7..91.7)     1205.1        0
4x1.0cpu_512m         2   200.7 (199.0..202.4)      422.8        0
4x2.0cpu_1g           2   279.1 (277.9..280.3)      214.8        0

Resource samples (max over sampled runs):
1x2.0cpu_1g        api_cpu=188.95% pg_cpu=32.8% lb_cpu=4.54% pg_conn=11 pg_active=1 pg_wait=0
1x2.0cpu_1g        api_cpu=181.64% pg_cpu=29.27% lb_cpu=4.29% pg_conn=11 pg_active=3 pg_wait=0
2x1.0cpu_512m      api_cpu=202.63% pg_cpu=40.47% lb_cpu=5.04% pg_conn=21 pg_active=2 pg_wait=0
2x1.0cpu_512m      api_cpu=202.01% pg_cpu=35.54% lb_cpu=5.54% pg_conn=21 pg_active=2 pg_wait=0
4x0.5cpu_256m      api_cpu=145.59% pg_cpu=31.42% lb_cpu=4.97% pg_conn=36 pg_active=1 pg_wait=0
4x0.5cpu_256m      api_cpu=128.44% pg_cpu=28.98% lb_cpu=5.56% pg_conn=36 pg_active=1 pg_wait=0
4x0.5cpu_256m      api_cpu=145.59% pg_cpu=31.42% lb_cpu=4.97% pg_conn=36 pg_active=1 pg_wait=0
4x0.5cpu_256m      api_cpu=128.44% pg_cpu=28.98% lb_cpu=5.56% pg_conn=36 pg_active=1 pg_wait=0
4x1.0cpu_512m      api_cpu=335.58% pg_cpu=101.78% lb_cpu=10.55% pg_conn=36 pg_active=2 pg_wait=0
4x1.0cpu_512m      api_cpu=336.77% pg_cpu=74.69% lb_cpu=9.84% pg_conn=41 pg_active=2 pg_wait=0
4x2.0cpu_1g        api_cpu=399.93% pg_cpu=106.4% lb_cpu=15.19% pg_conn=41 pg_active=6 pg_wait=0
4x2.0cpu_1g        api_cpu=526.87% pg_cpu=106.65% lb_cpu=16.2% pg_conn=41 pg_active=3 pg_wait=0

Interpretation:
- At a fixed total 2 CPU budget, splitting the budget across replicas does not
  increase throughput; the per-request CPU cost dominates and p95 grows.
- In the growth series, 4 replicas improve from ~90 to ~280 RPS as total API CPU
  grows from 2 to 8 cores, then the curve flattens: PostgreSQL (1 CPU) and the
  write/idempotency transaction become the shared bottleneck.
- This is a local comparative experiment, not a production capacity guarantee.
