fix(core): a half-closed peer keeps its response, and a gone one is still caught (#249)

[EdmondDantes]

Closes #249. Needs true-async/php-src#27 and true-async/php-async#266, both merged — the reactor's report is what makes the second half of this work.

What was wrong

http_connection_read_callback_fn latched conn->write_failed on every terminating read, and a clean EOF is what a half-close produces. The peer had only finished sending; every reader of that flag takes it as "output can no longer reach the peer", so a streaming handler lost the rest of its response.

Measured with a 4 MiB body written in 64 KiB chunks, the client sending one request, calling shutdown(SHUT_WR), then reading to EOF:

	half-closed	control
HTTP/1	110 wire bytes, headers only, no terminator	4194995, terminator
HTTP/2	65536–393216 DATA bytes of 4194304, no END_STREAM	4194304, END_STREAM

Why the read side alone cannot decide it

The kernel hands so_error to whichever syscall asks first. With a saturated outbound queue that is the write, so the read behind it returns a clean EOF and reports nothing. On bare sockets after an RST: write -> ECONNRESET, write -> EPIPE, read -> b''.

That is why dropping the latch on its own is not the fix. h1/032-h1-try-twins-peer-gone proves it: with the read-side change alone the handler ran all 100000 iterations of 4 KiB against a peer that had sent an RST and never saw an error, where the contract is HttpException 499.

What lands here

The read side latches on a read error alone. The verdict a fire-and-forget write could not deliver now comes from the handle: ZEND_ASYNC_IO_WRITE_FAILED, set by the reactor, read by http_connection_absorb_write_verdict from all four write completions — the three batched ones and the TLS cipher completion. A latched connection releases its outbound tail instead of chaining another refused write.

Guarded on ZEND_ASYNC_API_VERSION_NUMBER >= 0x001A00, the form the file already uses: built against an older reactor there is no other report to use, so the previous behaviour compiles in.

Evidence

h1/057 and h2/064 are new and fail against main: HTTP/1 gets 1 response, 0 terminators and 0 body bytes of 2097152; HTTP/2 gets 393216 of 1048576 with no END_STREAM. h1/032 holds the other half and fails on the read-side change alone. Full suite 351 of 351, ctest 16 of 16, on a PHP built from php-src true-async at ABI 0.26.0.

[github-actions]

Coverage

Total lines: 83.06% → 83.00% (-0.06 pp)

File	Baseline	Current	Δ	Touched
src/core/http_connection.c	81.26%	81.44%	+0.18 pp	●
src/core/http_connection.h	81.82%	81.82%	+0.00 pp	●
src/core/http_connection_tls.c	73.48%	73.55%	+0.07 pp	●
src/http1/http_parser.c	84.43%	84.05%	-0.38 pp
src/http3/http3_callbacks.c	83.37%	83.05%	-0.32 pp
src/http3/http3_listener.c	77.36%	76.91%	-0.45 pp
src/http3/http3_packet.c	90.43%	85.22%	-5.22 pp
src/http_server_config.c	92.48%	92.42%	-0.06 pp
src/websocket/php_websocket.c	78.44%	78.61%	+0.17 pp
src/websocket/ws_session.c	90.04%	89.48%	-0.55 pp