fix(http): client write/end callbacks + backpressure, real client timeouts, dynamic listener registration (#4909)

[proggeramlug]

Sub-tickets (1) client-side OutgoingMessage write/end callbacks + backpressure and (2) client HTTP timeouts from the #4909 triage, mirroring the #4954 server-side fix onto the perry-ext-http client path — plus the cross-cutting dispatch bugs that kept the corpus shapes hanging.

Root causes fixed

Client req.write / req.end (perry-ext-http)

• The static dispatch table routed req.write/req.end to single-arg entry points: the (encoding?, callback?) tail was dropped and write() returned the always-truthy handle, so while (req.write(buf, cb)) producer loops never terminated (the dominant test-http-outgoing-* hang). New js_http_client_request_write_full / _end_full queue the callbacks, return a real NaN-boxed backpressure boolean (16 KiB HWM), and accept Buffer chunks (previously misread as StringHeaders — same layout confusion as node:http res.write(Buffer)/res.end(Buffer): binary contents zeroed on wire (length preserved) #1124).
• Flush ordering matches Node via a new PendingHttpEvent::Flushed: write callbacks → 'finish' → end callback.

Client timeouts

• req.setTimeout(ms, cb) only stored the delay; the deadline existed only once a request was dispatched. The canonical req.setTimeout(n); req.on('timeout', …) against a never-responding server (or before req.end()) hung forever. A real timer is now armed at setTimeout() / at request creation for options.timeout; 'timeout' fires at most once and never after completion; setTimeout(0) clears, per Node.
• req.destroy() on an in-flight request emits the coded ECONNRESET "socket hang up" then 'close' (once-only); response/error drains are suppressed for completed/destroyed requests; .on('response') listeners now fire alongside the factory callback.

Dynamic dispatch (untyped out receivers — function write(out) { … } corpus shape)

• out.on('finish', cb) silently dropped the listener: the stdlib dispatch_http.rs name-gate had no "on" in the method allowlist. Added on/once/addListener/prependListener/removeListener/off/removeAllListeners to the gate and the ext dispatcher.
• out.constructor.name now reads ClientRequest/ServerResponse/IncomingMessage (plain { name } stand-in object).

Server side (perry-ext-http-server)

• req.resume() consumed the one-shot 'end'/'data' emit flags even with no listener registered yet, so the canonical drain pattern req.resume(); req.on('end', cb) never fired 'end' and the server never responded — a second, independent cause of the silent hangs.
• Dynamic res.write(chunk, cb) (js_node_http_res_write_with_cb) always returned 1 — no backpressure (the fix(http): wire res.write/res.end callbacks + backpressure boolean into static dispatch (#4909) #4954 fix only covered the static path); now returns the same HWM boolean.
• Dynamic res.end(buf, cb) crashed with TypeError: value is not a function: closure_arg accepted any POINTER_TAG value, so the Buffer chunk was taken as the end(cb) callback form and invoked. Now validated via js_value_is_closure.
• Dynamic end_with_cb fired 'finish'/'close' before the callbacks; reordered to Node's write-cbs → 'finish' → end-cb → 'close' (matching js_node_http_res_end_full).

Validation

• node:http/https: 89 tests exit non-zero with no output (triage; likely server-hang/framing, partly cascades from #4903) #4909 bucket re-measured (108 tests, baseline origin/main build vs this branch, same raw-byte comparator, 8 s cutoff): +8 fail→pass, 0 regressions — test-http-outgoing-finish (byte-for-byte), test-http-client-timeout-event, test-http-agent-uninitialized(-with-handle), test-http-early-hints, plus test-http-client-race(-2) and test-http-timeout-client-warning which now differ only by Node's DEP0169-stderr/pid noise that node_core_subset.py's normalizer strips.
• Most remaining bucket entries now exit with a real assertion message instead of the silent (no output) hang this issue tracks — i.e. the triage acceptance ("capture the actual non-zero exit cause") holds for the residue.
• New e2e regression test crates/perry/tests/issue_4909_client_write_end_timeouts.rs (both green): flush-order + backpressure round-trip, and setTimeout→'timeout'→destroy→ECONNRESET→'close'.
• cargo test -p perry-ext-http -p perry-ext-http-server green; curated node-suite http/https and the in-repo test_parity_http(s)/gap http tests match baseline (no flips).
• test-http-outgoing-message-write-callback (fix(http): wire res.write/res.end callbacks + backpressure boolean into static dispatch (#4909) #4954's representative) still passes after the end_with_cb reorder.

Out of scope / follow-ups for #4909

• 100-continue (checkContinue/writeContinue), agent keep-alive socket reuse, streamed (multi-chunk) responses (res.flushHeaders + late chunks), server-side req/res/server.setTimeout timers (need a socket surface), raw-framing/smuggling cases.
• Discovered (pre-existing, unchanged): test-http-response-setheaders.js SIGSEGVs nondeterministically (~6/8 runs) on both baseline and this branch — worth its own ticket.

Code-only PR per convention — no version bump / changelog (maintainer folds at merge).

Part of #4909 / #2132.