HTTP/2 same-connection fairness, October 17, 2026 (US Central).
One isolated AWS c7g.large, Ubuntu 24.04 ARM64, one Superweb worker.
Client and backend run on the same host over loopback; TLS is verified.
TCP_NODELAY is enabled on the test client, Superweb and asyncio backend.
Each round starts sixteen API requests with an 800 ms backend delay, then runs
10 fast API requests and 10 static GETs sequentially on the same H2 connection.
Five rounds; every response is checked. Results include client/backend overhead.
This is a fairness regression, not a throughput, CPU/RAM or server ranking test.
The previous binary did not support H2 proxying; no speedup is claimed against it.
The reported p95 uses the lower observed quantile: sorted[floor(.95 * (n - 1))].

Reproduce: place the three test_*.py files in test/ and the harness in bench/.
Install OpenSSL, libnghttp2, pkg-config and Python 3 (no Python packages required).
python3 bench/h2_proxy_fairness.py --binary /path/to/superweb --rounds 5 --output results.json

Normal and ASan/UBSan server suites were run on macOS ARM64 and Linux ARM64.
TLS and HTTP/2 are absent from the Intel macOS download.

Initial investigation: initial-client-nagle.json retains the first trial with
Nagle enabled in the client, yielding roughly 40 ms probe delays on Linux.
client-nodelay.json uses the identical server binary with only client TCP_NODELAY
changed. The 40 ms delay disappears. These are client-configuration diagnostics,
not a claimed server speedup. The final results use TCP_NODELAY and a later binary
that also corrects HTTP/2 access-log protocol and client-address reporting.