Validation

The Gateway-H3 validation flow reuses the same checks as Gateway-64 — same endpoints, same expected responses, same anti-cheat probes. The only differences:

• A separate compose.gateway-h3.yml is built and started (instead of compose.gateway.yml).
• Endpoint probes go to the same proxy on port 8443 but use curl --http2 rather than curl --http3.

Why curl --http2 for an h3 test? Most curl builds don’t ship HTTP/3 support by default, and requiring QUIC-enabled curl would break validate.sh on contributors’ machines. The compromise: h3-capable proxies (Caddy, nginx-quic, Envoy, HAProxy) advertise h2 on the same port anyway, so curl can reach the same endpoints for correctness validation. The h3-specific path is exercised at benchmark time by h2load-h3 with --alpn-list=h3 — if h3 is broken, the benchmark will show 0 req/s and you’ll see it immediately.

Checks performed

Same set as Gateway-64 validation:

1. HTTPS port responds within 30 seconds
2. HTTP/2 protocol negotiation (as a proxy-health sanity check)
3. /static/reset.css — Content-Type: text/css
4. /static/app.js — Content-Type: application/javascript
5. /static/app.js — non-zero response body
6. /static/nonexistent.txt — HTTP 404
7. /json/50 — returns 50 items with computed total field per item
8. /json/50 — Content-Type: application/json
9. /async-db?min=10&max=50&limit=50 — returns 1–50 items with nested rating, tags, boolean active
10. /async-db — Content-Type: application/json
11. /async-db?min=9999&max=9999 — returns count: 0 (anti-cheat)
12. /baseline2?a=13&b=42 — returns 55
13. /baseline2?a=<random>&b=<random> — returns correct sum (anti-cheat)

What validation does NOT check

• HTTP/3 protocol negotiation at the edge — requires QUIC-capable curl, not available in most builds. Covered implicitly by the benchmark: if the proxy can’t speak h3, h2load-h3 will report 0 req/s.
• QUIC connection migration, 0-RTT, or specific QUIC features — out of scope for correctness validation.
• Precompressed sidecar serving — validated at runtime by the benchmark’s Accept-Encoding: br;q=1, gzip;q=0.8 request header. If the proxy serves the raw file instead of the .br / .gz sidecar, response size will be larger and bandwidth numbers will reflect that.