[robobun]

Problem

• A node:http2 server counts the open streams of SETTINGS_MAX_CONCURRENT_STREAMS in JS (#peerInitiatedStreams, src/js/node/http2.ts:4044). node:http2: refuse a stream over the advertised limit before it is opened, on maxSessionInvalidFrames #44337 moves that count to native code. No test says when a stream gives its slot back.
• Three pieces of code exist in copies: the charge against maxSessionInvalidFrames (2), the two writes of an error exit of request() (6) and the end of the local half of a stream (4), in connection.rs and h2_frame_parser.rs.

Fix

• Add test/js/node/http2/node-http2-max-concurrent-streams.test.ts with 33 tests, written for node:test so that node runs the same file. Each test pins one order in which a stream takes a slot or gives it back.
• Connection::count_invalid_frame, H2FrameParser::close_unsent and Stream::end_local replace the copies.
• The tests pass without the source change.
• Verified: the 33 tests pass on Bun 1.4.3. node --test on v26.3.0 passes 29 and skips 4 that are Bun only. The debug build of node:http2: refuse a stream over the advertised limit before it is opened, on maxSessionInvalidFrames #44337 has this source and passes test/js/node/http2/.

Behaviour change: none

Background

• A stream uses one slot from its HEADERS frame until both sides ended it or one side reset it. A request over the limit gets RST_STREAM(REFUSED_STREAM) and no 'stream' event.
• Every test asserts what node v26.3.0 does. Where Bun 1.4.3 differs from node, this PR has no test: node:http2: refuse a stream over the advertised limit before it is opened, on maxSessionInvalidFrames #44337 handles those cases.
• This PR is stacked on node:http2: count a refused stream where the session refuses it, not in rstStream #44248. The tests have a file of their own because the file of node:http2: count a refused stream where the session refuses it, not in rstStream #44248 is written for bun:test.

Downsides

• None found. Each helper has the body of the copies that it replaces, and the http2 suites pass on a build with them.

Notes

What the tests pin. Raw HTTP/2 client, limit 1 unless the test says another value.

Order	Node v26.3.0	Bun 1.4.3 and this branch
Both sides ended the stream, one request per round trip	next request served	same
Response of HEADERS only, sendTrailers(), compat res.end(), one request per round trip	served	same
The end of a response that waited for a WINDOW_UPDATE	served	same
Request END_STREAM after the response ended, in one write with the next HEADERS	served	same
RST_STREAM from the peer (CANCEL, NO_ERROR, INTERNAL_ERROR, REFUSED_STREAM), in one write with the next HEADERS	served	same
The same for a stream that waits for trailers, and for the compat API	served	same
stream.close(CANCEL) and stream.close(NO_ERROR) by the server	served	same
3 requests in one write, handler answers from setImmediate	1 served, 2 refused	same
3 requests in one write, handler ends the response in the same turn (4 shapes): no test here, #44337 has them	1 served, 2 refused	3 served
respond() that fails in Bun (weight, parent, options, maxSendHeaderBlockLength): 4 tests, Bun only	does not fail	next request served
session.settings() in 'session', in a handler (lower and raise), below the open count	applies from the call	same
server.updateSettings() before the connection	applies	same
Limit 0, trailers on the open stream, pushed streams, HPACK state after a refused block, frames behind a refused HEADERS	as the test says	same
A stream over a limit that the client ACKed: no test, see #44337	connection error (nghttp2_session.c:3889)	RST_STREAM(REFUSED_STREAM), session open
5 requests at limit 2 through the core API, the compat API, TLS, and performServerHandshake() over a Duplex	2 served, 3 refused	same

The node column comes from the same scenarios as plain scripts on node v26.3.0, and from the test file itself on node.

Why the tests are sequential. They run servers in the test process. describe.concurrent does not make them faster, and each test then waits for the others inside its own timeout.

Why a new file. The rule from #37588 and #41305 is that a test of node:http2 passes on Bun and on node. node-http2-rejected-streams.test.ts uses bun:test, so node needs a shim to run it. This file uses node:test and node:assert, like node-http2-continuation.test.ts: node --test test/js/node/http2/node-http2-max-concurrent-streams.test.ts. The 4 tests that node skips need a respond() that fails, and the calls that make it fail in Bun do not fail on node.

The three helpers. count_invalid_frame has the body that both call sites had: node's count++ > max, then terminated = true and on_too_many_invalid_frames(). close_unsent has the two writes (state = CLOSED, rst_code) that each of the six exits had. end_local has the transition that four places had after this side wrote END_STREAM: CLOSED and free_resources when the peer has ended, HALF_CLOSED_LOCAL when it has not. The dispatches stay where they are.

A reset that Bun can write twice. stream.close(code) schedules an RST_STREAM, and the destroy that follows schedules one more. When a read comes between the two, both go out. CI showed it on Windows 11 aarch64. The test of close(CANCEL) counts each reset once. #43539 removes the second reset.

Open PRs on the same lines: #40240 rewrites the six request() exits.

[robobun]

Status: closed. The tests and the two helpers of this PR are one commit of #44337 now, which is one PR on main.

[robobun]

Closing: this is folded into #44337. A review asked for one PR on main in place of a stack of three. The tests and the two helpers of this PR are one commit there.