test(fpc): keep-alive and WebSocket EAGAIN regression tests

[freitasjca]

Two standalone FPC regression tests for transports the DUnitX suite cannot reach.

Horse's own test suite is DUnitX and Delphi-only, so the FPC keep-alive path and
the FPC + epoll WebSocket transport have no way to be covered from it. Both of
these are plain FPC programs with no test-framework dependency, and each exits
with a non-zero code on failure so they gate directly.

Purely additive: three new files plus two .gitignore entries. Nothing existing
is modified, and neither test is wired into CI — that seemed better left to you
to do the way you prefer.

---

tests/src/FPCHttpKeepaliveTest.dpr — keep-alive latency (#562)

30 sequential GETs on one reused connection, asserting each completes within
35 ms. Verified against 3.3.4 on FPC trunk: max 1 ms, avg 0 ms.

Two details are load-bearing rather than incidental:

KeepConnection := True on the client. TFPHTTPClient.KeepConnection
defaults to False, and the stall only occurs on a reused connection — Linux
starts every socket in quickack mode, so the first ACK is immediate and Nagle
never engages. An earlier version of this test omitted it, opened a fresh
connection per request, and passed identically with and without the fix.

A compile-time refusal below FPC 3.3.1. The provider's keep-alive and
TCP_NODELAY handling sit behind {$IF FPC_FULLVERSION >= 30301}, so on 3.2.2
there is nothing to test: the server closes after each response and the run dies
with Error reading data from socket, an error that names nothing relevant.
Failing at compile time says so instead.

The header documents the cause as Nagle interacting with the peer's delayed ACK,
per the wire capture on #562 — the body segment leaves 11-14 µs after the ACK
arrives, across an ACK delay that itself varies 42.5-43.9 ms.

This complements tests/fpc_keepalive_regression.py rather than replacing it.
That one drives FPCHttpKeepAliveServer.dpr from Python; this one is
self-contained Pascal and runs where Python is not installed.

Build and run:

U=/usr/local/fpc-trunk/lib/fpc/3.3.1/units/x86_64-linux
fpc -MDelphi -Sh -B -n -Fu"$U/*" -Fu"$U" -Fu../../src FPCHttpKeepaliveTest.dpr
./FPCHttpKeepaliveTest

---

tests/fpc-websocket/ — WebSocket EAGAIN (#549)

Written during the review of #549 and kept afterwards. 5 checks, all passing
on FPC 3.2.2 against the reworked implementation now in 3.3.4 — errno captured
once and passed to both predicates, close deferred via FStateLock plus an
FActiveOperations refcount. That is code this test was not written for, which
is the useful part: it asserts a property rather than an implementation.

The load-bearing assertion is silence, not liveness. epoll sets O_NONBLOCK
on accepted sockets, so recv returns -1/EAGAIN whenever no frame has arrived
yet. The original defect read that as a disconnect, broke the upgrader's read
loop on its first iteration, and let the HTTP pipeline resume — writing a stray
HTTP/1.1 200 OK onto a socket already handed to WebSocket. An earlier draft
asserted only that the connection was still open, and passed against the broken
code.

The runner refuses to build a tree lacking WS_SOCKET_READ_TICK_MS, so it
cannot report a pass on a checkout predating the fix.

cd tests/fpc-websocket && ./run-ws-eagain-test.sh

Exit code is the number of failed checks.

---

.gitignore

Two entries for the FPC test binaries. Their Linux builds have no extension, so
the existing *.exe rule does not cover them. One of the two is
FPCHttpKeepAliveServer — your own file, which produces the same uncovered
artefact when built on Linux.

---

Happy to adjust naming, placement, or wire either into the Docker test image if
you would like them in CI.