test(windows): assert what the server did, not what the client said

[EdmondDantes]

Closes #307.

The three tests still failing on WINDOWS_X64_ZTS_RELEASE each assert something
other than what the server did. The server's answers in all three artifacts are
correct.

compression/060 named two %Agarbage: 400 lines while the body prints one
per encoding the build has. SKIPIF skips only when neither brotli nor zstd is
built, so a build with exactly one — the Windows build has brotli — ran and
failed every time. The per-encoding lines stay, for diagnosis, and the verdict
is a single line that holds whatever the build contains.

chaos/002 floods 2000 CONTINUATION frames at a server that is entitled to
tear the connection down mid-flood; the test says so in its own comment and
wraps the flood in try/catch. H2TestClient::writeFrame ignored what fwrite
returned and kept writing into a reset socket, and on Windows each of those
raised a warning into the transcript — nine of them. fwrite warns rather than
throwing, so the catch never fired. The client latches the gone peer and
throws once.

tls/006 read keep-alive out of curl's verbose log, counting
Re-?using existing connection. That wording has changed between curl releases,
and the log cannot distinguish "curl opened three connections" from "curl said
it differently". The server sees the answer directly: three requests on one
kept-alive connection carry one peer port. HttpRequest::getRemotePort()
replaces the log scrape, and the output line is connections: 1 rather than
reuses: 2. Same move as #299 for static/013.

Evidence

Locally, with openssl on PATH so the TLS group executes: the whole suite is 496
tests, 259 executed, 0 failed, 0 warned — up from 245 executed, because the
TLS group runs. tls/006 and chaos/002 pass 4 of 4 runs each with the new
assertions.

chaos/002 and tls/006 pass on this machine before the change as well: what
they measured was the client's phrasing and the platform's warning, so the
reproduction is the Windows job, not this box. 060 skips here — neither
encoding is built — and its defect is arithmetic in the file: two expected lines
against one echo per built encoding, with the CI .out showing exactly one.

The Windows job is the proof for all three.

[github-actions]

Coverage

Total lines: 83.17% → 83.15% (-0.02 pp)

File	Baseline	Current	Δ	Touched
src/http3/http3_callbacks.c	83.54%	83.44%	-0.11 pp
src/http3/http3_listener.c	77.36%	77.96%	+0.60 pp
src/http3/http3_packet.c	90.43%	85.22%	-5.22 pp
src/websocket/php_websocket.c	78.61%	78.44%	-0.17 pp