Add echo-ws-limited: WebSocket echo with short-lived connections

[MDA2AV]

Closes #1001.

Both existing WebSocket profiles hold their connections open for the entire run, so the upgrade handshake is paid 512 times and amortized over millions of frames — its cost is invisible. This profile retires each connection after 10 messages, so the handshake is paid continuously and lands in the throughput number. It's the WebSocket counterpart to the HTTP/1.1 limited-conn profile.

[echo-ws-limited]="1|10|0-31,64-95|512,4096|ws-echo"

gcannon already supports this

No load-generator change was needed. Its per-connection quota check sits above the ws/http branch in worker.c, so -r caps frames exactly as it caps requests; once the quota drains the connection is retired, and every replacement connection runs fire_ws_upgrade() again. The adapter simply never passed -r for ws-echo — that's the one line added. The other two WebSocket profiles pass 0 and are byte-for-byte unaffected:

echo-ws           …/ws --ws -c 512 -t 64 -d 5s -p 1
echo-ws-pipeline  …/ws --ws -c 512 -t 64 -d 5s -p 16
echo-ws-limited   …/ws --ws -c 512 -t 64 -d 5s -p 1 -r 10

Measured, not assumed

Against a reference RFC 6455 echo server, 8 connections over 3s:

	responses	reconnects	server handshakes
no -r	335,916	—	8
-r 10	300,662	30,066	24,210

30,066 × 10 = 300,662 — exactly one connection retired per 10 messages, each replacement performing a real handshake, against 8 for the entire run without it.

The issue wondered whether this differs from the regular WebSocket tests: it does, though the reference server here is a single-threaded Python asyncio script, so the ~10% throughput delta is a floor rather than a prediction for real frameworks — the handshake share grows as per-message cost shrinks.

Nothing for framework authors to implement

The endpoint, framing and echo contract are identical to echo-ws, so any framework already subscribed to it satisfies this profile. The one thing it newly enforces is that Sec-WebSocket-Accept is computed per connection — a server returning a precomputed value answers the first handshake and fails every one after it.

Wiring

PROFILE_ORDER, the CATALOG entry and PROFILE_DOC mapping, the isolated-test list in benchmark.sh, ws-only readiness detection in framework.sh and validate.sh, and the WebSocket validation gate — which now fires for any of the three ws profiles instead of echo-ws alone, so subscribing to this one is enough to have /ws validated.

Impact until a round runs it: none

No framework subscribes yet (the issue is tagged next-round), and the generator only emits a profile once results exist — so the board and every composite score are unchanged. Docs added under test-profiles/ws/echo-limited/ with a card on the WebSocket index; verified all three pages render, type-rules widget included.

🤖 Generated with Claude Code