Report sendable() from the transport, not from a missing slot (#177)

[EdmondDantes]

Reduced from what this PR set out to do. tryWrite() is not here — two review passes showed it cannot be honest on HTTP/1 yet, and the reason is worth writing down.

What is left, and it is a real bug

sendable() is documented as backpressure advice. On a compressed response it answered a constant true: http_compression_maybe_install_stream_wrapper replaces stream_ops with compressing_stream_ops, which carried neither a sendable nor an is_alive slot, and a NULL slot is defined as "this transport has no queue of its own — report writable". The queue was still there, one layer down. HTTP/3 had no sendable either, for a stream that does queue. Both now answer from the transport.

Why tryWrite() was dropped

A non-blocking append needs a transport that can refuse. Threading a nonblocking flag into append_chunk works on HTTP/2, HTTP/3 and the worker path. On HTTP/1 it does not: the streaming writer is http_connection_send → send_raw, which awaits its own uv_write, so an accepted chunk parks regardless.

Routing the non-blocking case through http_connection_send_batched — the recipe WebSocket::trySend() uses — gives a second, unordered channel on a connection whose chunked framing assumes one: a chunk body waits in out_pending_buf behind an in-flight write while mark_ended's 0\r\n\r\n goes out directly and jumps ahead of it. That corrupts the body exactly when a refusal-and-retry happens, which is the case the feature exists for. On TLS it is worse: the high-water mark counts out_pending_len, which a TLS connection never fills, so the gate never fires and the failure becomes a spurious 499 instead of a refusal.

Filed as #179: one serialized outbound path per HTTP/1 connection, which tryWrite() needs and which also removes a park from the streaming hot path. #177 stays open behind it.

Suite: 276 passed, 0 failed.

[github-actions]

Coverage

Total lines: 82.22% → 82.10% (-0.12 pp)

File	Baseline	Current	Δ	Touched
src/compression/http_compression_response.c	80.14%	75.68%	-4.47 pp	●
src/core/http_connection.c	75.45%	75.26%	-0.19 pp
src/core/worker_dispatch.c	78.43%	76.63%	-1.80 pp	●
src/http1/http1_stream.c	69.57%	66.23%	-3.33 pp	●
src/http2/http2_strategy.c	83.33%	84.13%	+0.79 pp	●
src/http3/http3_callbacks.c	82.62%	83.17%	+0.55 pp	●
src/http3/http3_dispatch.c	82.77%	82.77%	+0.00 pp	●
src/http3/http3_listener.c	78.08%	76.88%	-1.20 pp
src/http3/http3_static_response.c	76.37%	76.37%	+0.00 pp	●
src/http_response.c	86.11%	85.37%	-0.74 pp	●
src/http_server_class.c	75.07%	75.03%	-0.04 pp
src/http_sse.c	85.11%	85.11%	+0.00 pp	●
src/websocket/php_websocket.c	78.44%	78.61%	+0.17 pp
src/websocket/ws_session.c	90.41%	89.85%	-0.55 pp

❌ Regression in touched files (> 1.0 pp drop)

• src/compression/http_compression_response.c dropped -4.47 pp
• src/core/worker_dispatch.c dropped -1.80 pp
• src/http1/http1_stream.c dropped -3.33 pp

Add [coverage-drop-ok] to a commit message in this PR to override.