Track 1 — cost of scaling
hypothesis: With a fixed total API CPU/RAM budget (2.0 CPU / 512M), two replicas outperform both one (a single uvicorn worker cannot spend 2.0 CPU) and four (each worker is capped at 0.5 CPU, p95 collapses). Growing per-replica CPU still raises RPS until that slice is no longer tiny. PostgreSQL and HAProxy stay below their limits; extra replicas past 2 only add pool connections to the same 1.0-CPU database. The confirmed improvement is 2×1.0 instead of 4×0.5 at the same budget.
profile: {"concurrency": 24, "warmup_seconds": 5, "measure_seconds": 20, "workload": "closed-loop POST /api/v1/chats/{id}/messages through HAProxy", "fixed_budget": "API total 2.0 CPU / 512M; postgres 1.0 CPU / 512M; pool 10 per replica"}

id              series    N  api_cpu  api_mem  pg_cpu  rps    p95   err   api_u%% pg_u%% lb_u%%  pool%%  backends  sat
fixed-1         fixed    1  2.0     512M    1.0     187.5 203.6    0   49.3  46.8  39.8  110.0      11.0  pool
fixed-2         fixed    2  1.0     256M    1.0     249.5 196.7    0   85.1  66.2  46.6  105.0      21.0  pool
fixed-4         fixed    4  0.5     128M    1.0     185.8 464.5    0   82.8  61.7  37.6   98.5     39.38  pool
grow-0.25       grow     2  0.25    256M    1.0      25.1 1916.3    0   96.0  26.4  11.6  105.0      21.0  pool
grow-0.50       grow     2  0.50    256M    1.0      93.8 593.2    0   97.3  37.6  24.1  105.0      21.0  pool
improve-pool20  improve  2  1.0     256M    1.0     140.7 396.9    0   89.2  63.3  45.2   87.8     35.14  api

finding: Fixed 2.0 CPU / 512M budget: 1/2/4 replicas → 187.5 / 249.5 / 185.8 RPS. One replica uses only 49% of a 2.0 limit (single uvicorn worker ≈ 1 core). Four replicas cap each worker at 0.5 CPU (p95 464.5 ms). Growing 2×CPU 0.25→0.50→1.0: 25.1 → 93.8 → 249.5 RPS; API util 96%/97% then 85%; postgres util 66%; LB util 47%. Confirmed improvement: same budget, 4×0.5 → 2×1.0 raises RPS 185.8 → 249.5 (+34.2%) and p95 464.5 → 196.7 ms. Raising DB_POOL_MAX 10→20 was tried and rejected (140.7 RPS, p95 396.9 ms).
