feat(runtime): node:http/https streaming response bodies end-to-end + client/server parity fixes

[proggeramlug]

What this is

Working the node:http/http2/https parity mandate. The in-repo vectors turned out to be already green on main (node-suite http 18/18, http2 9/9, https 5/5; test_parity_http* inventory tests diff-clean vs node v26) — the mandate's "5/18, 13 crashes" numbers were stale. The real gap is against Node's own test/parallel corpus (scripts/node_core_subset.py radar): http 30%, https 20%, http2 7% at baseline, with 186 of 403 runtime failures being silent hangs.

Root cause of the silent-hang class: nothing streamed

• Server: the entire response was buffered into a HyperResponseShape and only handed to hyper via a oneshot at res.end(). res.flushHeaders() set a flag and sent nothing; res.write() buffered. Any client waiting on headers/chunks before response completion waited forever.
• Client: dispatch_request did response.bytes().await — the 'response' callback only fired after the complete body arrived.

Changes

Server streaming (perry-ext-http-server/response.rs, server.rs):

• HyperResponseShape.body is now ShapeBody::Full | ShapeBody::Stream (channel-backed hyper Body impl, Frame::data + Frame::trailers).
• res.flushHeaders() / first res.write() flush the head to the wire immediately (chunked unless Content-Length set — Node's wire behavior); res.end() sends the final chunk + trailers and closes the channel. Single-shot res.end(body) keeps the buffered Content-Length path byte-for-byte.
• Real backpressure: in-flight byte counter; write() returns false past the 16 KiB HWM; the pump emits 'drain' when it sinks below. Reaper treats a dead stream receiver as peer-gone and never synthesizes over a streaming response.
• Custom res.statusMessage now reaches the HTTP/1 status line via hyper's ReasonPhrase extension.
• req.httpVersion reflects the wire version (1.0/1.1/2.0); new httpVersionMajor/httpVersionMinor (HIR member lowering + codegen static table + dynamic dispatch tower — the dot-access path previously hit the pre-tagged-receiver 0.0 sentinel).

Client streaming (perry-ext-http/client_dispatch.rs, client_events.rs, lib.rs):

• New ResponseHead / ResponseChunk / ResponseEnd pump events: the response callback fires when headers arrive; body chunks deliver as 'data' events as they come off the socket (buffered until listeners attach, flushed at 'end').
• req.flushHeaders() dispatches body-less requests immediately (new js_http_client_request_flush_headers static route + stdlib twin); a later end() drains the write/finish/end callback ordering exactly once.
• req.abort() before end() now tears down properly: emits 'abort' + once-only 'close', suppresses dispatch (server never sees the request), no spurious 'error'.
• Dropped the perry/<version> default User-Agent (Node sends none) and send Node's default Connection: keep-alive — aligns server-side header observations with Node.

Results

Node-core radar (Node v22 own tests, --auto-optimize):

api	pass before → after
http	85 → 89 (of 283 judged)
https	10 → 9¹
http2	12 → 12

Corpus tests fixed: test-http-flush-headers, test-http-flush-response-headers, test-http-client-response-timeout, test-http-abort-before-end, test-http-client-agent-abort-close-event.
¹ the two "new" failures are verified flakes, not regressions: test-https-agent-sni hangs identically on the pre-change binary (TLS cert-chain resolution, env-dependent); test-http-dns-error passes on both builds when run idle (DNS/load flake during the sweep).

The modest pass-count delta understates the change: most remaining corpus failures sit behind additional blockers (client 'readable'/read() pull API, Expect: 100-continue, a real client-socket surface, Agent pool internals, keep-alive wire enforcement). The streaming foundation is what SSE / long-poll / chunked responses — i.e. real Express/Fastify/Hono workloads — need, and those follow-ups now have somewhere to land.

Regression guard

• in-repo node-suite http 18/18, http2 9/9, https 5/5 (apparent compile failures in early sweeps were cold ext-crate rebuild timeouts in the harness, re-verified with proper timeouts).
• neighbor modules net/events/stream: 827/883; all 49 failures re-run against a pristine origin/main build — every one fails there too (0 regressions).
• cargo test -p perry-ext-http -p perry-ext-http-server green; cargo fmt --all clean; no Cargo.toml-version / CLAUDE.md / CHANGELOG edits (maintainer folds metadata at merge).

Files

perry-ext-http-server/{response,server,request,handle_dispatch,http2_server}.rs, perry-ext-http/{lib,client_dispatch,client_events,client_request_surface,tls_client,agent,tests}.rs, perry-stdlib/src/{common/dispatch.rs,http/client_request_surface.rs}, perry-codegen/src/{ext_registry.rs,runtime_decls/stdlib_ffi.rs,lower_call/native_table/{http_client,http_server}.rs}, perry-hir/src/lower/expr_member.rs