feat(http): route the node:http client's simplest shape onto turnloop

[proggeramlug]

First lane of moving the node:http / node:https CLIENT off reqwest.
perry-ext-http -> reqwest is one of only four tokio edges reached ALWAYS
rather than as a fallback, and six of the repo's seventeen live in this crate.

No tokio edge is removed yet — tokio_inventory.py is unchanged at 17/7,
which is correct: reqwest still serves everything outside lane 1. Only the
annotations move.

What this does

client_turnloop.rs drives turnloop_http::client::Http1Connection over
perry_ffi::turnloop_net on the agent's own loop. It is offered every exchange
in dispatch_request_snapshot, immediately before dispatch_request — three
hunks in lib.rs, nothing else moved. PendingHttpEvent was already
transport-agnostic, so the lane emits the same
ResponseHead/ResponseChunk/ResponseEnd/TransportError and the existing
drain, agent-admission release included, is untouched.

Two behaviours fall out rather than being ported: Node's client never follows a
3xx (driving the codec directly means nothing follows anything), and
res.statusMessage is unchanged — http1::Head carries no reason phrase, so
the canonical one stands in, exactly as reqwest did.

Decline set

Narrow and named, per the pattern fetch established: a non-default agent
handle, a non-empty body, an explicit timeout, NODE_USE_ENV_PROXY=1,
TE: trailers / Connection: Upgrade / Expect: 100-continue, non-http://,
!available(), and whatever client::Request::new refuses.

The three bypass headers are declined on the same predicates those bypass
modules trigger on, so the routing cannot disagree with itself.

One decline that was not anticipated: an explicit Host header.
Request::head strips a caller's host and substitutes the URL authority
(a Fetch rule); reqwest sends what the caller set. A unit test pins that the
codec really does rewrite it.

Keep-alive is deliberately out of this lane. Releasing a connection before
Event::End is the framing-misattribution hazard, and it deserves its own
change. The cost is invisible to JS: req.reusedSocket and
agent.sockets/freeSockets are fed by agent.rs's facade pool, already
decoupled from the physical connection.

Verification

• cargo check --workspace --all-targets (UI excludes): exit 0, re-run at the
  final tree state
• cargo test -p perry-ext-http: 148 lib + 2 integration, 0 failed
• tests/turnloop_client_exchange.rs, new end-to-end binary in its own process,
  asserting liveness three ways — try_dispatch returned true,
  completed_total() moved, and a plain TcpListener saw a well-formed head.
  Sabotage-proven: forcing the lane to decline makes it fail with the
  intended message, so a green run means the lane actually ran
• Gap suite --filter test_gap_http_: 4/4 pass, 100% parity, three times
• Liveness in the real compiled binary (temporary probe, removed before commit):
  the lane accepted the request and output was byte-identical to node v26.5.1
• check_file_size.sh, addr_class_inventory.py, gc_runtime_root_holders.py,
  check_gc_env_knobs.py, check_node_version_consistency.py: OK. rustfmt
  clean, zero clippy and zero rustc warnings in the new files

scripts/run_lint_gates.sh was not run — it is broken at the base commit
(Install cargo-xwin extraction error, fixed separately in #11081), so the
individual gates were run instead.

Also fixed

A per-request reserved-id leak — free_handle_id missing from on_closed,
the #6441 exhaustion shape that ws, ext-net and http-server each already carry
an arm for.

Corrections to the roadmap this lane was scoped from

Several long-standing claims in tokio_inventory.json turn out to be wrong, and
the annotations are updated:

• agent.rs is not "a second connection pool layered over reqwest's own".
  It is a per-origin admission engine — FIFO waiter queue, maxSockets /
  maxTotalSockets / maxFreeSockets, socket facades — sitting above the
  transport, with its own unit tests. reqwest owns only the physical connection.
  Its reqwest coupling is 9 call sites in 3 functions. It is not the blocker
  the inventory describes and it barely shrinks. The per-name FIFO that the
  migration plan assumed would need building already exists.
• The "three raw tokio::net::TcpStream bypasses" list names the wrong
  third. Measured: plain_client.rs (TE: trailers), continue_client.rs
  (Expect), and client_upgrade.rs (Connection: Upgrade, node:http client never emits 'upgrade': a 101 Switching Protocols response is delivered as 'response' and the socket is lost #10468).
  client_connect_override.rs (agent.createConnection) reaches zero
  tokio — it runs on perry-ext-net's raw_net vtable. All three real ones are
  deletable, since Event::Upgrade exists too.
• agent.createConnection is still blocked, for a different reason than
  recorded. Detached::from_fd + Loop::attach are real in turnloop, but a
  binding crate cannot reach them: there is zero attach / adopt /
  from_fd / Detached surface in perry_ffi::turnloop_net or
  perry-runtime/src/turnloop_net/abi.rs. That lane needs a perry-ffi ABI
  addition first.
• test_gap_fetch_expect_continue_header is unaffected — fetch: shorthand { headers } drops a Headers instance before js_fetch_headers_to_json, silently sending no headers #11024 is a codegen
  defect (shorthand { headers } dropping a Headers instance), not transport,
  so its skip-list entry is not stale.

Remaining lanes

Lane	Scope	Removes
2	request bodies + Lifecycle deadlines	POST/PUT/PATCH
3	TLS via perry_tls_session, needs per-request config rather than fetch's process-wide OnceLock	all https — the biggest single surface
4	keep-alive + Agent: Pool, release only after End	closes #10328
5	proxy CONNECT via Route	small
6	fold in the three bypass modules	3 tokio TcpStream users
7	agent.createConnection	blocked on the ABI addition above

Lane 2 wants wants_step() / is_mid_message() / eof_is_clean(), which do
not exist in alpha.6 — this lane expresses the progress loop without them and
bounds zero-consumed steps so a non-advancing decoder cannot spin the agent's
loop. It should be sequenced behind the turnloop bump (#11083).

[coderabbitai]

Warning

Review limit reached

Next included review available in 50 seconds.

Check out review usage here.

View limit details

Limit details: You’ve used all 8 included reviews currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: a4c11b7b-8fc3-42d0-afbb-341e3a11f899

📥 Commits

Reviewing files that changed from the base of the PR and between e27f0a0 and 6063c9d.

⛔ Files ignored due to path filters (1)

• Cargo.lock is excluded by !**/*.lock

📒 Files selected for processing (8)

• CLAUDE.md
• Cargo.toml
• changelog.d/11091-node-http-client-turnloop-lane-1.md
• crates/perry-db-turnloop/src/lib.rs
• crates/perry-ext-http/src/client_turnloop.rs
• crates/perry-ext-http/src/lib.rs
• crates/perry-ext-http/tests/turnloop_client_exchange.rs
• scripts/tokio_inventory.json

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[proggeramlug]

Landed on main in merge train 265 (#11108), released as v0.5.1648 at 9d26936298.

Cherry-picked from this PR's head 6063c9d548 and validated as one tree with 15 other PRs — CI 22/22 green, all 6 gap-suite shards. A train rebase gives the commits new SHAs, so GitHub cannot auto-close the source PR; closing by hand.

Nothing needed from you. Thanks.