quico 0.4.0 → 0.7.0

package/README.md

@@ -39,28 +39,37 @@
 6. [Design Choices](#-design-choices)
    - [The send model: range reconciliation](#the-send-model-range-reconciliation-not-retransmission)
    - [Congestion control](#congestion-control-bbr-style-not-loss-based)
+     - [The first round trips](#the-first-round-trips)
+     - [A returning client: Careful Resume](#a-returning-client-careful-resume)
    - [Stream scheduling](#stream-scheduling-byte-level-round-robin)
    - [Loss detection](#loss-detection-timers-and-how-we-treat-late-acks)
+   - [The handshake](#the-handshake-coalesced-datagrams-its-own-loss-detection)
    - [Flow control](#flow-control-consumption-based-windows)
 7. [Interoperability](#-interoperability)
 8. [API Reference](#-api-reference)
    - [`createServer()`](#quicocreateserveroptions-requestlistener)
-   - [`request()` / `get()`](#quicorequesturl-options-callback)
+   - [`request()` / `get()`](#quicorequesturloptions-options-callback)
    - [`IncomingMessage`](#incomingmessage) · [`ServerResponse`](#serverresponse) · [`ClientRequest`](#clientrequest)
    - [`Agent`](#agent) · [`createSecureContext()`](#quicocreatesecurecontextoptions)
+   - [0-RTT (early data)](#0-rtt-early-data)
    - [Framework compatibility](#framework-compatibility)
-   - [WebTransport](#webtransport)
+   - [WebTransport](#webtransport-1)
+   - [WebSocket](#websocket-1)
+   - [CONNECT-UDP (proxying UDP)](#connect-udp-proxying-udp)
+   - [Fingerprinting](#fingerprinting-1)
+   - [Protocol layers](#protocol-layers)
    - [Low-level QUIC API](#low-level-quic-api)
 9. [Configuration & Tuning](#-configuration--tuning)
 10. [Debugging](#-debugging)
 11. [Testing](#-testing)
 12. [Project Structure](#-project-structure)
-13. [Known Limitations](#-known-limitations)
-14. [Roadmap](#-roadmap)
-15. [Contributing](#-contributing)
-16. [Sponsors](#-sponsors)
-17. [Contact](#-contact)
-18. [License](#-license)
+13. [Security](#-security)
+14. [Known Limitations](#-known-limitations)
+15. [Roadmap](#-roadmap)
+16. [Contributing](#-contributing)
+17. [Sponsors](#-sponsors)
+18. [Contact](#-contact)
+19. [License](#-license)
 
 ---
 
@@ -146,6 +155,7 @@ stack trace ends in code you can open, not in a `.node` binary.
 |---|---|
 | **Node.js 18+** | ESM only (`import`). Tested on 18, 20 and 22. |
 | **No compiler** | Pure JavaScript — no native addons. |
+| **Dependencies** | [`lemon-tls`](https://github.com/colocohen/lemon-tls) (TLS 1.3) and [`flat-ranges`](https://www.npmjs.com/package/flat-ranges) (range arithmetic). Nothing else. |
 
 ```bash
 npm install quico
@@ -189,7 +199,8 @@ quico.request('https://www.google.com/', (res) => {
 }).end();
 ```
 
-> ⚠️ Don't forget `.end()` — like `node:https`, the request isn't sent until you call it.
+> GET and HEAD requests are ended for you on the next tick. For any method with a
+> body, call `.end()` — like `node:https`, the request isn't sent until you do.
 
 > 📂 More examples in [`examples/`](./examples)
 
@@ -198,30 +209,79 @@ quico.request('https://www.google.com/', (res) => {
 
 ## ✨ Features
 
-### HTTP/3 Server & Client
-- Full request/response cycle over QUIC (RFC 9114)
-- QPACK header compression, static and dynamic tables (RFC 9204)
-- Node.js-compatible API: `req.headers`, `res.writeHead()`, `res.end()`
-- Stream multiplexing — many concurrent requests on one connection
-- Connection pooling and coalescing: concurrent requests to the same origin share
-  a single QUIC connection
-- Automatic `Alt-Svc` advertisement for HTTP/3 discovery
-- Automatic H3 → H2 → H1.1 fallback when QUIC is unavailable
+### One HTTP stack, three versions
+QUICO does not import `node:http`, `node:https`, `node:http2` or `node:tls`. HTTP/1.1,
+HTTP/2 and HTTP/3 are all implemented here, and all three hand the application the
+same `IncomingMessage` / `ServerResponse` — so a framework written against `node:http`
+runs over any of them without knowing which. For the network and cryptography, all
+QUICO takes from Node is `node:net` (TCP), `node:dgram` (UDP) and `node:crypto`
+(primitives).
+
+- **HTTP/3** (RFC 9114) with QPACK, static and dynamic tables (RFC 9204); up to 100
+  blocked streams — a field section that references an entry whose insertion has not
+  arrived yet waits for it (`SETTINGS_QPACK_BLOCKED_STREAMS`)
+- **HTTP/2** (RFC 9113) with HPACK, including h2c by connection preface
+- **HTTP/1.1** with keep-alive, chunked transfer and trailers
+- One server listens on UDP and TCP and serves all three; `Alt-Svc` is advertised
+  on HTTPS HTTP/1.1 and HTTP/2 responses so browsers discover HTTP/3 (see `altSvc`)
+- Client with automatic H3 → H2 → H1.1 fallback, connection pooling and
+  per-origin coalescing
+- `103 Early Hints` (`res.writeEarlyHints()`) on all three versions
+- Extra protocols on the same TCP port through ALPN or byte sniffing
+  (`server.addProtocol()`)
 
 ### QUIC Transport
-- QUIC v1 (RFC 9000) — Initial, Handshake and 1-RTT packet spaces
+- QUIC v1 (RFC 9000) and **QUIC v2** (RFC 9369) — Initial, Handshake and 1-RTT packet
+  spaces; everything a version decides (salt, key labels, packet type bits, Retry key)
+  in one profile per version (`quic_version.js`), the rest shared
+- Compatible version negotiation (RFC 9368): a client that starts in v1 and offers v2 is
+  moved to v2 without a round trip when the server prefers it (`quicVersions`); Version
+  Negotiation and a second attempt when there is no compatible version
 - TLS 1.3 handshake via LemonTLS, no OpenSSL (RFC 9001)
 - Version Negotiation (RFC 9000 §6)
 - HelloRetryRequest — X25519, P-256, P-384
 - AES-128-GCM and AES-256-GCM
 - Key Update — bidirectional key phase rotation (RFC 9001 §6)
-- Connection ID management and packet coalescing
+- Session resumption on the server, with ticket keys shared across connections
+  (`ticketKeys`)
+- Retry and address validation, switchable at runtime (`validateAddress`)
+- Handshakes under way counted for the application (`getPendingHandshakes()`,
+  `'handshakeLoad'`), and capped per client address (`maxHandshakesPerAddress`)
+- Anti-amplification limit — 3× until the client's address is validated (RFC 9000 §8.1)
+- Connection migration (RFC 9000 §9): a client may move to another network (Wi-Fi to
+  cellular) and go on in the same connection — each side issues spare connection IDs
+  (NEW_CONNECTION_ID, RETIRE_CONNECTION_ID, Retire Prior To) and the new path is
+  addressed by one of its own (§9.5); NAT rebinding is followed too. Each new address is
+  validated with PATH_CHALLENGE first, and congestion/RTT state starts over on a new IP
+- NEW_TOKEN (RFC 9000 §8.1.3): a returning client's address needs no proof — no Retry,
+  no 3× limit; the agent keeps one token per server and uses it once
+- ACK frequency (draft-ietf-quic-ack-frequency): a receiver follows ACK_FREQUENCY and
+  IMMEDIATE_ACK; a sender with many packets in flight asks for an ACK every ~10 packets —
+  an 8 MB download, ~340 ACK packets instead of ~3,400
+- `grease_quic_bit` (RFC 9287): the fixed bit sent unpredictable to a peer that allows it
+  (not on a UDP port shared with other protocols)
+- Certificate compression (RFC 8879, in LemonTLS): the server's chain compressed for a
+  client that offers it (Chrome: brotli), only when it comes out smaller, and once per chain
+  (cached). The client offers brotli, zstd and zlib on every version. A client certificate
+  goes compressed when the server's CertificateRequest asks. Decompressing stops at the
+  declared length and never goes past 128 KB. With an RSA-4096 chain, the server's
+  Handshake flight over QUIC went from 9.6 KB to 6.0 KB. The server's algorithms are
+  `createServer({ certificateCompression })`; what went on the wire, each way, is
+  `socket.getNegotiationResult().certificateCompression` on every version. The chain is
+  compressed when the server is made, in the background, not by the first client's
+  handshake (brotli at quality 11: ~4 ms for a small chain, ~18 ms for 9 KB)
+- Encrypted Client Hello (RFC 9849, in LemonTLS) — HTTP/3, HTTP/2 and HTTP/1.1 alike: the
+  site name and ALPN travel encrypted, to a key published in DNS (`ech`; see
+  [Encrypted Client Hello](#encrypted-client-hello-ech))
+- Stateless reset with HMAC-derived tokens (`resetTokenSecret`)
 - Stream and connection-level flow control with sliding windows
+- Stream limits enforced in both directions, and raised with `MAX_STREAMS` as
+  streams close (RFC 9000 §4.6)
 - Loss detection with PTO and exponential backoff (RFC 9002)
 - BBR-style congestion control with pacing
+- Path MTU discovery (DPLPMTUD, RFC 8899)
 - Unreliable datagrams (RFC 9221)
-- Stateless reset, idle timeout, keep-alive
-- TLS keylog export for Wireshark decryption
+- Idle timeout, keep-alive, TLS keylog export for Wireshark
 
 ### WebTransport
 - Client and server (RFC 9297 + draft-ietf-webtrans-http3)
@@ -229,6 +289,24 @@ quico.request('https://www.google.com/', (res) => {
 - Unreliable datagrams
 - Natural API: `req.on('stream')`, `res.sendDatagram()`, `res.writeHead(200)` to accept
 
+### WebSocket
+- Server over all three versions: HTTP/1.1 Upgrade, HTTP/2 extended CONNECT
+  (RFC 8441) and HTTP/3 extended CONNECT (RFC 9220) — one `'upgrade'` event for all
+- `WebSocketServer` in the shape of the `ws` package, and a WHATWG-style
+  `WebSocket` client (HTTP/1.1)
+
+### CONNECT-UDP
+- A UDP proxy (RFC 9298, MASQUE) over HTTP/3 — QUIC datagrams — and HTTP/2 and
+  HTTP/1.1 — capsules: `server.on('connectUdp', (req, tunnel) => tunnel.accept())`
+- Its client: `quico.connectUdp(proxy, { host, port })`, a `dgram.Socket` that goes
+  through the proxy
+
+### Fingerprinting
+- JA4H (HTTP request), HTTP/2 and HTTP/3 protocol fingerprints, per request
+- TLS fingerprints (JA4, JA4S, JA5) through LemonTLS
+- `profileRequest()` — scores how well a request's protocol behaviour matches the
+  client it claims to be
+
 ---
 
 ## 🧬 Design Choices
@@ -251,13 +329,14 @@ recover its frames, and put them back in a retransmission queue. The sender is
 
 **What QUICO does instead.** There is no `sent_packets` table, no stored frame copies,
 and no retransmission queue — those concepts don't exist in the codebase. Each stream
-keeps only three things:
+keeps only this:
 
 ```js
 {
-  pending_data:     Uint8Array,     // the bytes the application wrote
+  q:                [Uint8Array, …],  // the bytes the application wrote, as written
   acked_ranges:     [from, to, …],  // byte ranges the peer has confirmed
-  in_flight_ranges: { 47: [from, to], 48: [from, to] }   // packet number → bytes it carries
+  in_flight_ranges: { 47: [from, to], 48: [from, to] },  // packet number → bytes it carries
+  missing_ranges:   [from, to, …]   // the answer to the question below, kept (see below)
 }
 ```
 
@@ -268,14 +347,54 @@ missing = invert(acked_ranges ∪ in_flight_ranges, 0, total_size)
 ```
 
 Whatever's missing goes out. That's the whole algorithm. The sender doesn't remember
-what it *did* — it recomputes what the peer *still needs*, from scratch, every time.
-It's a declarative model: continuously reconcile "what the peer has" against "what the
-peer should have", and emit the difference.
+what it *did* — it asks what the peer *still needs*, every time. It's a declarative
+model: continuously reconcile "what the peer has" against "what the peer should have",
+and emit the difference.
+
+**The answer is kept, not recomputed.** `acked_ranges ∪ in_flight_ranges` was computed
+from every packet in flight, three or four times per send pass — with thousands of
+packets in flight on a large transfer, that was a fifth of the CPU. Each stream now holds
+the answer itself, as [flat-ranges](https://github.com/colocohen/flat-ranges)' three
+states of what the peer has:
+
+| flat-ranges | QUICO | the peer… |
+|---|---|---|
+| `have` | `acked_ranges` | has these bytes — acknowledged |
+| `not_have` | `missing_ranges` | surely lacks them — never sent, or given up on |
+| `unknown` | `in_flight_ranges` (by packet) | may have them — in flight |
+
+`have` and `not_have` never overlap, and `missing_ranges` is what the sender sends from.
+It changes at exactly five points ([`nothingMissing`](./src/quic_connection.js)
+documents them):
 
-**Loss detection, therefore, is one line:**
+| Event | `missing_ranges` |
+|---|---|
+| the application writes bytes | added |
+| bytes yielded into a packet (`get_stream_chunks`) | removed |
+| an ACK credits bytes — a late one included | removed |
+| the expiry timer gives a packet up | its bytes added, except any acknowledged meanwhile (`add_not_have`) |
+| a packet that never left (encryption failed) | its bytes added |
+
+Adding a packet's bytes back is exact because each byte is in flight in one packet at
+most: bytes are yielded only while missing, and become missing again only when the
+packet that carried them expires. `sendstate_test` checks `missing_ranges` against the
+complement of `acked_ranges ∪ in_flight_ranges`, rebuilt from scratch every few
+milliseconds under loss, reordering and late ACKs, and that it never overlaps
+`acked_ranges`.
+
+Congestion control counts something else: **packets, by their size on the wire** — header,
+frames, padding and AEAD tag, whatever they carry, as RFC 9002 counts bytes in flight.
+Each packet the peer must acknowledge enters `pkt_in_flight` when it leaves and goes when
+it is acknowledged or given up on; `app_bytes_in_flight` is their sum, kept as they come
+and go, and an ACK delivers what its packets weighed. The window, the pacing rate and the
+bandwidth BBR measures are one unit. (They used to be STREAM bytes, which left out the
+packets' overhead, packets of control frames, and most of the handshake.)
+
+**Loss detection, therefore, is still one step:**
 
 ```js
-delete st.in_flight_ranges[pn];   // → those bytes are "missing" again
+delete st.in_flight_ranges[pn];                    // not in flight any more
+flat_ranges.add_not_have(acked, missing, lostRanges); // → missing again, unless acked
 ```
 
 Deleting the mapping is the entire recovery action. Nothing is queued, scheduled, or
@@ -312,15 +431,20 @@ reassembles by offset and genuinely does not care how the bytes were packetized.
   isn't a contradiction to reconcile — it just adds a range to `acked_ranges`, and the
   next pass naturally stops sending those bytes.
 
-**What it costs.** Computing `invert(acked ∪ in_flight)` per stream per send
-opportunity is O(number of ranges) work, where a queue pop would be O(1). With
-coalesced flat range arrays this stays cheap in practice, but it is real CPU that a
-queue-based sender doesn't spend — a deliberate trade of cycles for correctness.
-
-**The receive side is symmetric.** Incoming data goes into `chunks` keyed by offset
-plus a `ranges` array, and delivery walks the contiguous prefix. Same range algebra,
-mirrored — which is exactly why the receiver must tolerate chunks whose boundaries
-don't line up with anything it saw before: on the sending side, they never had to.
+**What it costs.** A send opportunity reads a stream's few `missing_ranges` — usually
+one or two — and keeping them is a range insert or removal per write, and per packet
+sent, acknowledged or expired. The expiry walk visits a
+stream's packets oldest first and stops at the first one still within the timeout, so
+it costs the packets that expired, not every packet in flight.
+
+**The receive side is symmetric.** Incoming data is recorded in a `ranges` array, and
+delivery walks the contiguous prefix. Same range algebra, mirrored — which is exactly
+why the receiver must tolerate chunks whose boundaries don't line up with anything it
+saw before: on the sending side, they never had to. It does so at insertion: only the
+bytes not already held are stored, so stored pieces never overlap, tile `ranges`
+exactly, and delivery is an exact-key walk. A peer cannot make reassembly expensive
+either — the number of gaps per stream is capped (1000, as in quic-go), and pieces are
+compacted into one buffer per contiguous run when they outnumber the gaps.
 
 ### Congestion control: BBR-style, not loss-based
 
@@ -371,21 +495,103 @@ That trace is a real run through a 10 Mbps bottleneck with a 25-packet queue: st
 finds the link in ~7 rounds, drain empties the queue, probe_bw settles within 5% of
 line rate.
 
+#### The first round trips
+
+A page's first view is decided before any of this has measured anything: by the window
+the connection starts with, and how fast it grows. QUICO starts at **20 packets (24 KB)**
+— `initialCongestionWindow` — where RFC 9002 suggests 10. That is where TCP effectively
+is when its response goes out: a 10-segment TCP window has grown on the acknowledged
+handshake segments by then, and a QUIC connection at 10 left a 60 KB page a round trip
+behind HTTP/2 on the same path. In `startup` the window then grows by every byte the
+peer acknowledges — the handshake's packets as well as the application's (RFC 9002
+§7.3.1, TCP's slow start). Through an emulated 450 ms, 9 Mbps link, Chromium loading a
+61 KB page that references a 61 KB script:
+
+| | HTML after its first byte | the script | `load` |
+|---|---|---|---|
+| a 10-packet start | 934 ms (2 round trips) | 1384 ms | 2326 ms |
+| 20 packets (default) | 490 ms (1) | 942 ms | 1890 ms |
+
+The window's ceiling is **8 MB** — what a long fat path needs at the pacing ceiling
+(14 MB/s × 450 ms ≈ 6.3 MB); a 4 MB download through the same link takes 5.4 s (it took
+8.7 s under the earlier 300 KB ceiling). The window reaches it only where BBR measured
+that much: on a short path, the BDP keeps it small.
+
+#### A returning client: Careful Resume
+
+A client back within the hour has already shown this server its path. QUICO keeps what
+BBR measured — the bottleneck bandwidth and the minimum RTT — in the client's session
+ticket, and a connection that resumes with it starts from there, as **RFC 9959 (Careful
+Resume)** lays out, with its BBR appendix — when turned on (`carefulResume: true`; off by
+default, for reasons measured at the end of this section):
+
+- **Saved** once BBR has a measurement worth keeping (a capacity of at least four initial
+  windows): during Startup at each round it grows by a quarter, then when it moves by a
+  quarter either way — in a NewSessionTicket sent after the handshake (lemon-tls
+  `sendTicket()`), inside the ticket's encryption: the client carries it, cannot read or
+  forge it, and the server keeps nothing. A resumed connection carries it forward.
+- **Not used** past an hour, from another network (the client's IPv4 /24 or IPv6 /48
+  changed: a laptop moved from home to a café), or when the first flight shows another
+  RTT — at most half, or over ten times, the saved one.
+- **Reconnaissance → jump → validation.** The first flight goes out at the initial window,
+  as on any connection. Once it is acknowledged without loss, the first time the sender
+  has more than its window allows, the window jumps to **half** the saved capacity, paced
+  over one round trip; slow start goes on from what that delivers.
+- **Safe Retreat.** A loss while the jump is out is taken for congestion: the saved state is
+  deleted, the window drops to half of what was actually delivered and holds until the jump
+  is acknowledged, and BBR goes on in ProbeBW from there.
+
+Chromium through the emulated link, the second download on a new connection resuming with
+the first one's ticket:
+
+| Path | File | First visit | Returning, Careful Resume | Returning, without |
+|---|---|---|---|---|
+| 450 ms, 9 Mbps | 512 KB | 2365 ms | **1736 ms** | 2384 ms |
+| 450 ms, 9 Mbps | 2 MB | 4107 ms | **3353 ms** | 4152 ms |
+| 100 ms, 50 Mbps | 4 MB | 1876 ms | **1428 ms** | 1634 ms |
+| 450 ms, 9 Mbps, 2 % random loss on the return | 2 MB | 4026 ms | 5357 ms | **4608 ms** |
+
+The last row is the RFC's price: a random loss during the jump looks like congestion, and
+the retreat costs more than QUICO's BBR loses without it — it does not react to loss at all.
+On a path that did get slower (the same RTT, a third of the rate), the jump is what the
+retreat is for; measured there, the returning download was still faster with it (6.8 s
+against 7.8 s), the queue having absorbed it.
+
+**Off by default.** Those measurements were of one file, downloaded twice — so the first
+visit had measured the path at full rate. A website's first visit is a page: a document and
+its resources, which measure a lower rate, so the jump on the next visit is smaller. Measured
+again that way (Chromium, the second visit on a new connection), and with the handshake
+and loss recovery improved since, which made connections without it faster too:
+
+| Path | Returning visit | With Careful Resume | Without |
+|---|---|---|---|
+| 100 ms, 50 Mbps | 4 MB download | 1314 ms | 1320 ms |
+| 450 ms, 9 Mbps | 512 KB download | 2833 ms | 2918 ms |
+| 100 ms, 50 Mbps | page (60 KB + 20 scripts) | 731 ms | **678 ms** |
+| 250 ms, 20 Mbps | page | 1769 ms | **1660 ms** |
+| 250 ms, 20 Mbps, 3 % random loss | page | 3583 ms | **2150 ms** |
+
+Almost nothing on clean paths, and a lot lost on a lossy one: with a hundred packets in the
+jump and a few percent of random loss, one is nearly always lost, Safe Retreat takes it for
+congestion, and the connection goes on at a fraction of the link. `carefulResume: true`
+turns it on — for an application whose returning clients download large files over clean
+paths, as in the first table; `getStats().carefulResume` gives a connection's phase.
+
 **Honest scope:** this is BBR-*style*, not a certified BBRv2. ProbeRTT — the periodic
 cwnd collapse that re-measures RTprop on a permanently saturated path — is not
 implemented, and there is no explicit loss response. The windowed filters bound the
 damage of stale measurements in the meantime.
 
-### Stream scheduling: byte-level round-robin
+### Stream scheduling: byte-level round-robin, by priority
 
 **The choice:** when several streams have data ready and one packet is going out, who
 gets the space? Implementations differ sharply — many drain streams in priority or
 creation order, filling each one before moving on.
 
 QUICO interleaves at the **byte level, inside a single packet**
-([`execute_quic_burst`](./src/quic_connection.js)). The scheduler cycles through all
+([`execute_quic_burst`](./src/quic_connection.js)). The scheduler cycles through the
 active streams, taking a chunk from each until the packet is full, then cycles again —
-so a 1350-byte packet routinely carries STREAM frames from three different streams:
+so a 1200-byte packet can carry STREAM frames from three different streams:
 
 ```
 ┌─────────────────────────────── one QUIC packet ───────────────────────────────┐
@@ -397,19 +603,104 @@ The loop repeats while any stream still yields data and space remains (`progress
 flag), so a stream that runs dry doesn't hold a slot — the remaining bandwidth
 redistributes to the others within the same packet.
 
-**Why it matters:** three parallel downloads progress at genuinely equal rates instead
-of finishing sequentially, and no stream can be starved by a large one ahead of it.
-The cost is fragmentation — more STREAM frame headers per packet (~24 bytes each) than
-a drain-one-stream-at-a-time scheduler. QUICO accepts that trade for fairness.
-Priority signalling (RFC 9218) would let applications override the equal split; it's
-on the roadmap.
+**Each packet starts where the last one left off.** A stream takes what it has, up to
+the room left — so a stream with plenty to send fills a whole packet on its own. If
+every packet started from the lowest stream number, that stream would fill every
+packet and the others would wait for it to finish: three 6 MB downloads ran one after
+another. The scheduler remembers who opened the previous packet (`rr_cursor`) and starts
+the next round just after it, so with three busy streams the packets go 0, 4, 8, 0, 4,
+8… and the downloads progress at equal rates. The cost is fragmentation — more STREAM
+frame headers per packet (~24 bytes each) than a drain-one-stream-at-a-time scheduler.
+QUICO accepts that trade for fairness.
+
+**Priorities decide who takes part in a round.** The streams with data are arranged in
+tiers, and a packet takes from the first tier with something to send; only if that
+runs dry does the next tier fill the rest of the same packet:
+
+- by **urgency**, most urgent first — RFC 9218's `u=0..7`, default 3; urgency −1 is
+  HTTP/3's own control and QPACK streams, ahead of everything (a header block that
+  refers to a QPACK entry stuck behind response bytes would block);
+- within an urgency, **non-incremental** streams go one at a time, by stream number,
+  each a tier of its own — a stylesheet whole before the next;
+- then the **incremental** ones, taking turns as above;
+- a **group** (any object, given to `setStreamPriority`) takes *one* turn in its round,
+  and passes it to its member with the highest `order` that has something to send;
+  equal orders take turns. This is WebTransport's `sendGroup` / `sendOrder`, mapped in
+  the WebTransport layer; QUIC only knows "streams that share one turn".
+
+A stream nobody set a priority for takes turns at urgency 3 — as every stream did
+before. The mechanism lives in QUIC (`setStreamPriority(id, { urgency, incremental,
+group, order })`); the policy comes from above: HTTP/3 reads the client's `priority`
+header and `PRIORITY_UPDATE` frames, the application can override them
+(`res.setPriority`) — the client's later signals do not undo the application's decision
+— and WebTransport maps its streams' `sendGroup` / `sendOrder`. A client asks for a
+priority with the `priority` header on its request (`'u=1, i'`) — QUICO's client: the
+`priority` option, `req.setPriority()`. HTTP/2 has the same
+rules in its own scheduler (`h2.js`), for the DATA that waits for a window or for the
+socket; what the socket already holds leaves in the order it was written. An HTTP/2
+server sends `SETTINGS_NO_RFC7540_PRIORITIES = 1`: it reads RFC 9218's signals, not
+RFC 7540's priority tree.
+
+A scheduler can only order what is waiting. A stream fed in small writes with
+backpressure has little queued at any moment, and while its queue is empty the next
+tier sends — so an application that wants a stream strictly first should hand it its
+data ahead of time, not a chunk per `'drain'`. Likewise a stream blocked by the peer's
+flow control lets the next one go, rather than leaving the packet empty.
 
 **The subtle part** falls straight out of the reconciliation model: the sender decides
 what to send by asking which bytes aren't acked or in flight — but chunks selected
 *earlier in the same burst* have no packet number yet, so they're in neither set, and
-the next iteration would select them again. QUICO parks them under a temporary
-`_burst` key in the in-flight map, which real packet-number keys replace the moment
-the packet is encrypted and sent. The set arithmetic stays honest even mid-burst.
+the next iteration would select them again. `get_stream_chunks` takes what it yields out
+of `missing_ranges` at once, so the next iteration sees it as taken; the packet's number
+records it in `in_flight_ranges` when it is sent. A packet that does not leave hands its
+bytes back to `missing`. One packet can carry two pieces of one stream — new data and a
+repaired gap — so a packet's entry in `in_flight_ranges` is a range list, read whole by
+the ACK, the expiry timer and a stream reset.
+
+**Many streams on one connection.** What the sender does per packet and per ACK does not
+walk every stream: each packet in flight records the streams it carries, so an ACK (and
+the expiry timer) visits those; the streams that hit flow control, and those whose FIN
+is still to go, are kept apart the same way. Walked, those loops were most of the CPU of
+a connection with many streams — a thousand concurrent responses on one connection ran
+at 25 Mbps on loopback, two thousand at 10. Now 1000 run at ~60 and 2000 at ~55, against
+~70 with 200 (one core, client and server in one process). What remains per stream is
+the scheduler's arrangement of the streams with data, once per burst.
+
+**Writes go out once per tick.** A write to a stream does not send: it asks for one send
+pass at the end of the current tick, and every write until then rides in it. A response
+written as headers, ten small writes and an end is one datagram, not fourteen. A pass
+for any other reason — an ACK, a timer — still runs at once and takes what is waiting.
+
+**Datagrams** (RFC 9221) are never retransmitted, but they go under the same congestion
+control: they wait in a queue, are taken into packets ahead of stream data, and the
+packets that carry them count in flight (with nothing to resend). A datagram sent with
+a priority (WebTransport's `sendOrder` / `sendGroup`) is not first: it takes its place
+in its group, as a member at its order. While the path cannot
+keep up the oldest are dropped — any that have waited 250 ms (configurable), and past
+1 MB queued — since a late datagram is worth nothing. The peer's
+`max_datagram_frame_size` is honoured.
+
+### The send queue: the application's buffers, not copies
+
+What a stream holds for sending is the list of buffers the application wrote, by
+reference, each with the stream offset it starts at. A packet takes a view of one of
+them, or — where it straddles two — a copy of its own ~1200 bytes. Writes of up to
+4 KB are the exception: they are gathered into a 16 KB buffer of the stream's own, so
+ten thousand one-byte writes are a handful of pieces. A buffer leaves the queue the
+moment the peer has acknowledged all of it.
+
+The effect is on memory: `res.end(file)` holds the file once, not once per layer, and
+any number of concurrent responses of the same `Buffer` share it. As with
+`net.Socket`, **a buffer passed to `write()` / `end()` must not be modified afterwards**
+— it is referenced until acknowledged, and a retransmission reads it again.
+
+**Backpressure on the sending side** works as on a TCP socket. A write is complete —
+its callback runs — once the stream has *sent* all but `writableHighWaterMark` of what
+was written (bytes in flight are bounded by the congestion window, as the kernel
+bounds them for TCP). So on HTTP/3 `res.write()` returns `false` and `'drain'` follows,
+and `readable.pipe(res)` reads at the client's pace: a 30 MB file piped to a browser
+holds ~5 MB of buffers on the server, not the file. The same applies to WebTransport
+streams and to tunnels (WebSocket over HTTP/3, CONNECT).
 
 ### Within a stream: new data leads, gaps follow
 
@@ -448,10 +739,54 @@ decisions worth calling out:
   and if those ACKs are then barred from producing RTT samples, the timer never learns
   that it's too short. The loop is self-sealing. Accepting the sample breaks it.
 
+- **Optimistic ACKs are caught** (RFC 9000 §21.4). Every 256–1023 packets, at random,
+  one 1-RTT packet number is skipped; an ACK that covers it acknowledges a packet that
+  was never sent, and closes the connection (PROTOCOL_VIOLATION). A peer acknowledging
+  ahead of what arrived would otherwise inflate the congestion window at will. The send
+  history keeps a zero-length placeholder for the skipped number, so its index math
+  stays contiguous.
+
 The PTO itself is `srtt + max(4·rttvar, 1ms) + max_ack_delay`. That last term is not
 optional: peers may legitimately sit on an ACK for their full advertised delay (25 ms
 is common), and a timer without it expires packets moments before their ACKs arrive.
 
+The timer is the backstop, not the first word. On a new connection rttvar is half the
+first RTT sample, so the timer comes ~3 RTTs after a packet left — and a lost packet in
+a page's first response is exactly where those RTTs are paid. Two things come first:
+
+- **Loss by ACK: the time threshold** (RFC 9002 §6.1.2). A packet is given up on once a
+  packet sent after it is acknowledged and `max(9/8·RTT, RTT + 5 ms)` has passed since
+  it was sent (RTT: the larger of srtt and the latest sample) — not the timer. Then
+  exactly the timer's path: resent, parked for a late ACK, counted lost. The RFC's
+  packet threshold (three later packets acknowledged) is left out on purpose: a packet
+  that arrives out of order is not resent for that, only for being later than the
+  threshold; the 5 ms floor keeps that so on short paths, where an eighth of the RTT is
+  less than Wi-Fi's own retransmissions reorder by. A 60 KB response through a 200 ms
+  path with one packet lost: ~10 ms longer than with none, where it was ~185 ms.
+- **The tail loss probe** (RFC 8985). The last packet of a response — or a request in
+  one packet — has no later one to be acknowledged: after `2·srtt + max_ack_delay`
+  without an ACK, a copy of its data goes as a new packet. Nothing is given up on (the
+  original stays in flight; whichever arrives counts), one probe until an ACK comes,
+  only once the handshake is confirmed, and only while the timer is further off.
+
+Chromium through 250 ms with 3 % loss (a 60 KB page and 20 scripts): page load p75
+2.41 s → 2.02 s; through 20 ms with 1 % loss and 5 ms jitter: median 474 → 419 ms.
+
+**Our own ACKs are delayed and batched** (RFC 9000 §13.2). In the 1-RTT space an ACK
+goes out once two ack-eliciting packets are owed — decided at the end of the
+event-loop turn, so datagrams that arrived together get one ACK — at once for a
+packet out of order (that is what lets the sender detect loss quickly), and
+otherwise within `max_ack_delay`. Any packet we send anyway carries the ACK that is
+owed, and the ACK Delay field reports how long it was held, so the peer can take it
+out of its RTT. Initial and Handshake packets are acknowledged immediately, as the
+RFC requires — in the handshake's flush, with whatever else those spaces send (see
+[the handshake](#the-handshake-coalesced-datagrams-its-own-loss-detection)). Before this, every ack-eliciting packet got an ACK packet of its own:
+a 20 MB download had the receiver send ~18,000 of them; now ~1,200.
+
+The ACK timer runs at `max_ack_delay − 15 ms` (10 ms against the 25 ms we
+advertise): timers fire late by up to their granularity, ~15.6 ms on Windows, and
+the peer's loss timer counts on our ACK arriving within the advertised value.
+
 **ACK ranges are capped at 32** (RFC 9000 §13.2.4 permits acking a subset). Under
 loss, packet-number gaps never close — a lost PN is gone forever, its data returns
 under a new PN — so an uncapped ACK frame grows with the cumulative loss count until
@@ -459,18 +794,128 @@ it no longer fits in a packet. We measured exactly that against a line-rate send
 ACK packets inflating 325 B → 1017 B second by second, until they crossed the MTU and
 stopped being sendable at all.
 
+
+### The handshake: coalesced datagrams, its own loss detection
+
+The handshake is the one place where every datagram must arrive before anything can
+go on — so it is built to send as few as it can (RFC 9000 §12.2):
+
+- **One flush, coalesced datagrams.** Initial and Handshake packets are not sent as
+  TLS produces each message. What the two spaces have to send — CRYPTO data neither
+  acknowledged nor in flight, an ACK they owe — goes out at the end of the turn, in
+  packets as full as the datagram allows, the packets in as few datagrams as they fit.
+  The datagram is padded to 1200 bytes, not an Initial packet with zeros ahead of the
+  Handshake packets that could have filled it. A server's whole flight (ServerHello,
+  EncryptedExtensions, Certificate, CertificateVerify, Finished, and the ACK of the
+  ClientHello) is **one datagram** with a small chain, where it was seven; a client's
+  answer one, where it was up to eight.
+- **RTT from the first round trip.** ACKs of Initial and Handshake packets give RTT
+  samples, as 1-RTT ones do (RFC 9002 §4.1: RTT measurement is "unified across packet
+  number spaces"); the client has one when its handshake completes, not only after its
+  first 1-RTT packet is acknowledged.
+- **ACK-based loss detection, in these two spaces only.** A packet is lost once a later
+  one is acknowledged and three more are, or 9/8 of an RTT has passed (RFC 9002 §6.1);
+  its range is sent again at once. The retransmission timer stays as the backstop. The
+  1-RTT space keeps the recovery described above.
+- **The client's anti-deadlock probe** (RFC 9002 §6.2.2.1). Until one of its Handshake
+  packets is acknowledged, a client keeps its timer even with nothing in flight — and
+  sends a PING when it fires: a server held by its anti-amplification limit has no
+  timer of its own, and waits for any datagram from the client.
+- **The first 1-RTT packet rides in the last handshake datagram.** When 1-RTT keys
+  exist, the flush's last datagram waits for the end of the turn, and the first 1-RTT
+  packet made in it goes in it, last (a short header has no length): the server's
+  SETTINGS with its flight, the client's request — with its control streams — with its
+  Finished, in room the Initial's padding took anyway. A connection with one request:
+  8 datagrams where it was 13; before the response, one datagram each way after the
+  ClientHello. Not where it would cost the server a round trip: a datagram the
+  anti-amplification budget would have let through alone is not made too large for it.
+- **The peer's retransmission is news** (RFC 9002 §6.2.3). Handshake data received
+  again means the peer has not heard from us: what we sent in answer and is still
+  unacknowledged goes again at once, with the ACK the packet is owed — a server seeing
+  the ClientHello again, a client seeing the server's flight again — not a
+  retransmission timeout (three RTTs on the first sample) later. Twice per connection
+  at most. A 1-RTT packet that rode in a datagram found lost this way, or by the
+  ACK-based detection above, is lost with it and sent again in the new datagram — the
+  request does not wait for the 1-RTT space's own timeout because it travelled with the
+  Finished. A late ACK for either still counts.
+
+Chromium against a QUICO server through a 250 ms path with 3 % loss: time to first byte
+went from a median of 1182 ms to 516 ms (two round trips — the floor), and the page from
+2.8 s to 2.0 s. QUICO's client through the same path (100 connections): the 90th
+percentile of time to first byte from ~1150 ms to ~850 ms, mostly the lost
+Finished-and-request datagram answered on the server's retransmission.
+
+
+### Path MTU discovery
+
+Packets start at what every QUIC path carries — 1200 bytes of frames — and most paths
+carry more: ~1450 on Ethernet. Every byte more in a packet is a packet less, and a packet
+is an encryption, a system call and its bookkeeping — where a JavaScript sender spends its
+time: a 4 MB download goes out in ~16 % fewer packets at 1452.
+
+A round trip after the handshake is confirmed — the first responses go first; a probe is
+1.4 KB of padding a slow link would put ahead of them — the sender probes (DPLPMTUD,
+RFC 8899 / RFC 9000 §14.3), in a send pass of its own if nothing else would run one (a
+client that only downloads may have nothing to send), and after the data of the pass it
+shares:
+a PING padded to a size, sent in the paced flow of the connection, and acknowledged — by
+its own packet number — only if that size crossed the path. Sizes are whole UDP datagrams.
+The first probe is at **the largest datagram the peer has sent on this path** (the limit is
+most often the peer's own link — a PPPoE line, a VPN — and the same both ways), then at the
+ceiling (1452, and the peer's `max_udp_payload_size`), and past a failure a binary search
+down to 16 bytes. A size fails after three unacknowledged probes; the search runs again ten
+minutes after it ends. A probe carries no stream data: a lost one — and a probe too large is
+lost by design — holds nothing back, and is no loss to congestion control (§14.4).
+
+**Black holes.** The MTU can shrink mid-connection. Packets above the floor lost in a row,
+with none acknowledged since, are a suspicion, not a verdict — congestion loses packets in
+runs too (Startup's overshoot does, every connection). A probe of the current size settles
+it: acknowledged, it was congestion; lost three times, packets go back to 1200 and the
+search starts again later. A new path starts from 1200, as its congestion state does.
+
+| Path (largest datagram) | Found | Probes dropped |
+|---|---|---|
+| Ethernet (1472) | 1452 — the ceiling | 0 |
+| a VPN (1392) | 1387, within the search's 16 bytes | ~9 |
+| 1452, cut to 1300 mid-download | back to 1200 after 3 confirming probes | — |
+
 ### Flow control: consumption-based windows
 
 **The choice:** when do you tell the peer it may send more? The naive trigger is
-"bytes arrived"; QUICO slides its receive windows based on **application
-consumption** — a `MAX_STREAM_DATA` / `MAX_DATA` update goes out once the app has
-actually *read* a threshold of data, not merely when packets landed.
-
-The difference shows up under backpressure: a slow consumer causes the window to stop
-advancing, which slows the sender, instead of letting unbounded data accumulate in the
-receive buffer. Node stream semantics (`write()` return values, `highWaterMark`)
-propagate through H3 into these windows, so ordinary `.pipe()` backpressure reaches
-all the way to the peer's congestion controller.
+"bytes arrived"; QUICO slides its receive windows on **consumption** — a
+`MAX_STREAM_DATA` / `MAX_DATA` update goes out once in-order data has been delivered
+up the stack and half a window has been used, never for bytes still sitting out of
+order in the reassembly buffer.
+
+**Where "consumed" is measured.** For what the application reads — a request body on
+the server, a response body on the client, a WebTransport stream on either side — it is
+the application's read: the
+bytes that have left the Node stream's buffer, counted after each `read()` (which
+flowing mode, `pipe()` and async iteration all go through). A reader that reads
+nothing holds the peer to one stream window; one that reads slowly sets the peer's
+pace. Everything else — HTTP/3's own streams — is consumed at the hand-off from QUIC to the HTTP/3 layer, and bounded there by its
+own limits (a HEADERS frame by the field-section size we advertise, control frames by
+a fixed cap).
+
+These bytes count in the connection's window too (`initial_max_data`, 1 MB by
+default): a few streams that nobody reads can hold the whole connection. An
+application with many parallel streams — MoQT's stream per group — reads or stops
+(`stopSending`) each one, and sizes the windows for it with `transportParams`.
+
+**Autotuning.** A window caps the peer at one window a round trip: 256 KB a stream is
+10 Mbit/s at 200 ms. So, as Chromium and quiche do, a window the application consumed
+half of within two round trips of the last update — the window is what limits, not the
+path — doubles, up to **16 MB** a stream and **24 MB** the connection (a larger
+`transportParams` value is kept as given); the connection's grows with its streams, to
+1.5× the largest. It grows only with what is consumed, so a reader that falls behind
+keeps the window it has. Measured through an emulated 20 Mbit/s, 200 ms link, a 4 MB
+download by QUICO's client: 7.8 s before, 4.2 s now (Chromium: 3.1 s).
+
+**HTTP/2, over TCP.** The client's stream window is **6 MB** (Chrome's), not
+HTTP/2's 64 KB, and the connection's 16 MB. Credit still goes back as the
+application reads, so a response nobody reads stops at 6 MB. The 64 KB default held a
+response to 64 KB a round trip: 16 MB through a 100 ms round trip took 52 s, and takes
+under 1 s now. An application that wants less sets `initialWindowSize` on the request.
 
 ### Reading the source
 
@@ -506,6 +951,11 @@ delay, bandwidth limits and packet loss.
 - **http3** — three concurrent HTTP/3 requests, which must share **one** QUIC
   connection
 
+0-RTT over TCP is verified with OpenSSL both ways (`s_client -early_data`, `s_server
+-early_data`: taken, refused and sent again, and after a HelloRetryRequest). Over QUIC,
+0-RTT is verified with Chromium (its second visit's request in its first flight, and its
+retry after a `425`), and with aioquic and quic-go in both directions.
+
 Additionally verified against production servers: **Google**, **Cloudflare**,
 **Facebook**, **Microsoft**, **LiteSpeed** and **nginx** — including a 12.7 MB
 download from `cloudflaremirrors.com` byte-for-byte.
@@ -522,37 +972,278 @@ and finally the low-level QUIC primitives for non-HTTP protocols.
 
 ### `quico.createServer([options][, requestListener])`
 
-Creates an HTTP/3 server. Returns a server object with `listen()`, `close()` and
-`on()`.
+Creates a server for HTTP/3, HTTP/2 and HTTP/1.1. Returns a server object with
+`listen()`, `close()` and `on()`.
 
 **Options**
 
 | Option | Type | Default | Description |
 |---|---|---|---|
-| `key` | `Buffer \| string` | — | Private key in PEM format. Required unless `SNICallback` is used. |
-| `cert` | `Buffer \| string` | — | Certificate chain in PEM format. |
+| `key` | `Buffer \| string \| Array` | — | Private key in PEM format (a string without `-----` is read as a file path). Required unless `SNICallback` is used. An array — `node:tls`'s form, keys or `{ pem, passphrase }` — for several certificates of different kinds, each key paired with the chain in `cert` that carries its public key (see [Post-quantum TLS](#post-quantum-tls)). |
+| `cert` | `Buffer \| string \| Array` | — | Certificate chain in PEM format (or a file path, as for `key`); an array, one chain per key. |
 | `ca` | `Buffer \| string \| Array` | `null` | Additional trusted CAs. |
-| `SNICallback` | `Function` | `null` | `(servername, cb) => cb(null, secureContext)` for multi-domain setups. |
+| `SNICallback` | `Function` | `null` | `(servername, cb, info) => cb(null, secureContext)` for multi-domain setups. `info.transport` is `'tcp'` or `'quic'`. |
+| `http3` | `boolean` | `true` | Listen for HTTP/3 over UDP. |
 | `http2` | `boolean` | `true` | Also listen for HTTP/2 over TCP (fallback for non-QUIC clients). |
 | `http1` | `boolean` | `true` | Also listen for HTTP/1.1 over TCP. |
+| `ALPNProtocols` | `string[]` | — | ALPN names to accept, in preference order, as in `node:tls`. Served HTTP versions missing from the list are put first; naming a version that is disabled throws. |
 | `maxConnections` | `number` | `10000` | Maximum concurrent QUIC connections. |
+| `ticketKeys` | `Buffer` (48 bytes) | random | Session-ticket key, shared by TCP and QUIC. Pass a stable one to keep resumption working across restarts. |
+| `ech` | `{ keys }` | — | Encrypted Client Hello keys, TCP and QUIC alike: `{ keys: [{ config, privateKey, retry }] }`, from `quico.generateEchKey()`. A key that does not match its config throws here. See [Encrypted Client Hello](#encrypted-client-hello-ech). |
+| `earlyData` | `boolean` | `false` | Accept 0-RTT — a returning client's request in its first flight, answered one round trip sooner: over HTTP/3, and over TLS 1.3 on TCP (HTTP/2, HTTP/1.1). Off by default: early data can be replayed. See [0-RTT](#0-rtt-early-data). |
+| `earlyDataMaxAge` | `number` (ms) | `10000` | 0-RTT: how far the age a client states for its ticket may be from what the server measures (RFC 8446 §8.3) — a captured first flight is good for this long at most. 1–600000. |
+| `earlyDataReplay` | `(id, ttl, cb) => void` | in memory | 0-RTT's anti-replay store: `cb(null, true)` if `id` is new (and keep it `ttl` ms), `cb(null, false)` if seen. Share one (Redis, memcached) between servers that share `ticketKeys`. No answer in 100 ms, or an error: no 0-RTT, the handshake goes on. |
+| `earlyDataFilter` | `(req, cb) => void` | `GET`, `HEAD`, `OPTIONS` | 0-RTT: whether a request that came early may be handled before the handshake completes — `cb(null, true)`; otherwise the client gets `425 Too Early` and sends it again after the handshake (RFC 8470). No answer in 100 ms: `425`. |
+| `sessionTimeout` | `number` | `7200` | Seconds a session ticket stays good, TCP and QUIC alike (`node:tls`'s option; its default is 300). `0`–`604800` (RFC 8446's 7 days); `0` issues tickets that never resume. |
+| `requestCert`, `rejectUnauthorized` | `boolean` | `false`, `true` | Client certificates, as in `node:tls`, on every version — HTTP/3 included — verified against `ca`. With `rejectUnauthorized` (the default) a client without a valid one is refused; with `false` it gets in, and `req.socket.authorized` says which. `req.socket.getPeerCertificate()` / `authorizationError` as on a `TLSSocket`; on a resumed connection, the certificate it proved on its first. |
+| `minVersion`, `maxVersion` | `string` | `'TLSv1.2'`, `'TLSv1.3'` | `'TLSv1.2'` or `'TLSv1.3'`. HTTP/3 is TLS 1.3 only: `maxVersion: 'TLSv1.2'` needs `http3: false`. |
+| `certificateCompression` | `string[]` \| `boolean` | `true` | The algorithms the server's certificate may go compressed in (RFC 8879), in order: `['brotli', 'zstd', 'zlib']` — the default, `true`, is all of them that Node has (zstd from 22.15). It goes compressed only to a client that offered one, and only when that comes out smaller. `false`: always plain. TCP and QUIC alike; a name unknown throws from `createServer`. |
+| `honorCipherOrder` | `boolean` | `false` | As `node:tls`: the cipher suite by the server's order (`ciphers`). By default the client's — a client that asks for AES-256 first gets it. |
+| `ciphers`, `ecdhCurve`, `sigalgs` | `string` | — | As `node:tls` takes them (OpenSSL's colon-separated names; `ecdhCurve: 'X25519:P-256'`). The server's default groups: every post-quantum one Node has, then X25519, P-256, P-384 (see [Post-quantum TLS](#post-quantum-tls)). HTTP/3 uses the TLS 1.3 suites of `ciphers` (AES-GCM and ChaCha20-Poly1305); a list without one needs `http3: false`. A value naming nothing usable throws from `createServer`. |
+| `handshakeTimeout` | `number` | TCP `120000`, QUIC `10000` | Milliseconds a TLS handshake may take. Over QUIC, counted from the connection's first packet, and on a server until the client's Finished is verified — not only until the server's own Finished has gone: a client that sends a ClientHello and nothing more holds the connection for this long, not for the idle timeout. |
+| `passphrase` | `string` | — | For an encrypted `key`. |
+| `validateAddress` | `boolean` | `false` | Validate client addresses with a QUIC Retry before any handshake work. |
+| `handshakeLoadThreshold` | `number` | `500` | HTTP/3 handshakes under way at which `'handshakeLoad'` fires (`high: true`), and again under half of it (`high: false`). `0`: off. |
+| `maxHandshakesPerAddress` | `number` | `64` | HTTP/3 handshakes under way from one client address; past it an Initial from there makes no connection. `0`: no cap. |
+| `resetTokenSecret` | `Buffer` (≥ 16 bytes) | random | Secret for QUIC stateless-reset tokens. Keep it stable across restarts so clients of a crashed instance get a reset instead of a timeout. |
 | `socket` | `dgram.Socket` | `null` | Use an externally-bound IPv4 UDP socket instead of creating one. Must already be bound. |
 | `socket6` | `dgram.Socket` | `null` | Same, for IPv6. |
+| `greaseQuicBit` | `boolean` | `true`; `false` with `socket` / `socket6` | QUIC: invite clients to send the fixed bit unpredictable (RFC 9287), so the bit stays free to change. Off by default on an external socket, whose port may be shared with STUN, TURN or DTLS — told apart from QUIC by that bit (RFC 9443). `true` when that port carries QUIC alone. |
+| `port` | `number` | `443` | The port advertised in `Alt-Svc` when `listen()` is not used (externally fed server). |
+| `altSvc` | `string \| false` | — | The `Alt-Svc` header on HTTPS HTTP/1.1 and HTTP/2 responses. By default `h3=":<port>"; ma=86400` when this server runs HTTP/3, and none with `http3: false`. A string is sent as is — for HTTP/3 served by something else (`'h3=":443"; ma=86400'`); `false` sends none. A response's own `alt-svc` header wins. |
+| `connectUdp` | `object` | — | CONNECT-UDP's policy: `{ template, allowPrivate, maxTunnelsPerConnection, idleTimeout, lookup, proxyName }` — see [CONNECT-UDP](#connect-udp-proxying-udp). |
+| `shouldUpgradeCallback` | `(req) => boolean` | listens for `'upgrade'`? | As in `node:http`: whether an upgrade request (HTTP/1.1 `Upgrade`, or extended CONNECT over HTTP/2 and HTTP/3) is taken up as `'upgrade'`, or handled as an ordinary `'request'`. |
+| `headersTimeout` | `number` | `60000` | How long a request's head may take to arrive. HTTP/1.1: `408` and close. HTTP/2: a header block left open (HEADERS without END_HEADERS) blocks the whole connection, so the connection goes (GOAWAY). HTTP/3: the request stream is reset, `H3_REQUEST_INCOMPLETE`. `0` disables it. |
+| `requestTimeout` | `number` | `300000` | As `node:http`: the whole request, body included, must arrive within it, counted from the end of its head. HTTP/1.1: `408` and close. HTTP/2 and HTTP/3: that request's stream is reset (`CANCEL` / `H3_REQUEST_INCOMPLETE`) and the request gets `'aborted'`; the connection serves on. Not applied to `CONNECT` (WebSocket, WebTransport), whose stream is meant to stay open. The response is not timed. `0` disables it. |
+| `keepAliveTimeout` | `number` | `5000` | HTTP/1.1 only: an idle connection between requests. |
+| `streamResetBurst`, `streamResetRate` | `number` | `1000`, `33` | Rapid Reset (CVE-2023-44487) protection, `node:http2`'s options and defaults: a client may cancel `streamResetBurst` requests at once, then `streamResetRate` a second. Past it the connection is closed — HTTP/2 `GOAWAY` `ENHANCE_YOUR_CALM` (a `'sessionError'`, `ERR_HTTP2_STREAM_RESET_RATE`), HTTP/3 `H3_EXCESSIVE_LOAD`. HTTP/3 counts cancelled request streams (`RESET_STREAM` or `STOP_SENDING`, once each), not a WebTransport session's own streams. `Infinity` for either: no limit. The `'stream_reset'` event remains for a policy of your own. |
+| `maxRequestsPerSocket` | `number` | `0` (no limit) | `node:http`'s: requests one connection may carry. HTTP/1.1: the Nth response says `Connection: close` and the connection ends (a request past it, pipelined, gets `503`). HTTP/2 and HTTP/3: `GOAWAY` after the Nth — requests in flight finish, the client takes the next to a new connection (nginx's `keepalive_requests`). |
+| `maxHeaderSize` | `number` | `16384` | HTTP/1.1: request line plus headers, in bytes (`node:http`'s option and default). Over it: `431`. |
+| `maxHeaderCount` | `number` | `2000` | HTTP/1.1: header lines. Over it: `431` (`node:http` drops the extra ones instead). |
+| `noDelay`, `keepAlive`, `keepAliveInitialDelay` | | `true`, `false`, `0` | TCP socket options. |
+| `transportParams` | `object` | — | QUIC's windows and stream limits, as sent to every client: `initialMaxData`, `initialMaxStreamDataBidiLocal`, `initialMaxStreamDataBidiRemote`, `initialMaxStreamDataUni`, `initialMaxStreamsBidi`, `initialMaxStreamsUni`, `maxIdleTimeout`. See [Transport parameters](#transport-parameters). |
+| `initialCongestionWindow` | `number` | `20` | QUIC: the congestion window a connection starts with, in full-size packets (1200 B), `2`–`100`. See [The first round trips](#the-first-round-trips). |
+| `carefulResume` | `boolean` | `false` | QUIC: a returning client's connection starts from what the server measured on its last one (RFC 9959). Off by default: for page loads it gains almost nothing and costs a lot with random loss. Worth it where returning clients download large files on clean paths. See [Careful Resume](#a-returning-client-careful-resume). |
+| `startupCap` | `'stall'` \| `'always'` \| `'off'` | `'stall'` | QUIC: Startup's cap on the congestion window, 2 × BDP plus ACK aggregation (BBR v3). `'stall'`: from the first round in which bandwidth stopped growing — fewer drops at a router's queue where the window used to overshoot, no cost to a page's ramp-up. `'always'`: the whole of Startup — fewest drops, a slower ramp on long lossy paths (for bulk transfers). `'off'`: none. Also per connection (`connectionOptions`, `conn.setOptions`), at runtime (`setOptions`), and a client's own option for what it sends. |
+| `pmtuDiscovery` | `boolean` | `true` (`false` on Windows) | QUIC: path MTU discovery. Off on Windows by default — Node cannot set the Don't Fragment bit there, so a probe may cross fragmented and be taken for a fit. Off, packets stay at 1200 bytes of frames (or what was found before it was turned off). |
+| `maxUdpPayloadSize` | `number` | `1452` | QUIC: the largest UDP payload sent (1200–65527): where path MTU discovery stops. Lower it for a VPN or a tunnel; raise it for jumbo frames on a LAN. |
+| `initialRtt` | `number` (ms) | `333` | QUIC: the RTT assumed before there is a sample (RFC 9002's kInitialRtt) — how soon a lost handshake packet is sent again. Lower for clients known to be near (one country, a CDN edge); too low against far clients, handshake packets are resent needlessly. 10–10000. |
+| `reorderWindow` | `number` (ms) | `5` | QUIC: the least a packet is allowed to arrive out of order before a later one's ACK gives it up (the floor of the time threshold, see [Loss detection](#loss-detection-timers-and-how-we-treat-late-acks)). Raise it on paths that reorder more (some mobile networks, per-packet load balancing) — `getStats().spuriousLosses` counting up says so. 0–1000. |
+| `quicVersions` | `Array` | `['v1', 'v2']` | QUIC versions accepted, in preference (`'v1'`, `'v2'`, or the numbers). A client that offers both through `version_information` gets the first of these it can be moved to without a round trip (RFC 9368) — v1 by default; `['v2', 'v1']` moves such clients to v2. A client whose version is not listed gets Version Negotiation. `req.socket.quicVersion` says which. |
+| `connectionOptions` | `(info, callback) => void` | — | QUIC: each connection's own `initialCongestionWindow`, `transportParams`, `carefulResume` and `startupCap`, decided when its ClientHello arrives. See [Per-connection settings](#per-connection-settings-connectionoptions). |
+| `maxWebTransportSessions` | `number` | no limit | WebTransport sessions a connection may have open at once — sent as `SETTINGS_WT_MAX_SESSIONS`; a CONNECT past it is refused with `H3_REQUEST_REJECTED` before the application hears of it. |
+| `maxConcurrentStreams`, `headerTableSize`, `initialWindowSize`, `maxFrameSize`, `maxHeaderListSize`, `connectionWindowSize`, `enableConnectProtocol` | | `100`, `4096`, `65535`, `16384`, `131072`, 16 MB, `true` | HTTP/2 settings. |
+| `maxSessionMemory` | `number` | `10` | HTTP/2: megabytes a connection may have queued for a peer that is not reading (`node:http2`'s option). DATA waits for the socket instead of queueing; past this, the connection is closed. |
+
+#### Per-connection settings: `connectionOptions`
+
+The server's settings are everyone's. `connectionOptions` decides them per QUIC connection —
+from the client's address, the name it asked for, and whether it has been here before:
+
+```js
+createServer({
+  key, cert,
+  connectionOptions(info, callback) {
+    // info: { remoteAddress, remotePort, servername, resumed,
+    //         previous: { rtt, bandwidth, age } | null }
+    lookupNetwork(info.remoteAddress, (err, net) => {
+      if (err) return callback(err);                       // → the server's settings
+      callback(null, {
+        initialCongestionWindow: net.satellite ? 40 : 20,  // packets
+        transportParams: { initialMaxData: 32 * 1024 * 1024 },
+        carefulResume: !net.shared,
+        startupCap: net.bulk ? 'always' : 'stall',
+      });
+    });
+  },
+}, handler);
+```
+
+It is asked once the ClientHello is in and before anything is answered: the handshake waits
+for it, as it waits for an asynchronous `SNICallback` (whose turn it takes first), and
+`transportParams` — merged over the server's — go out in that same handshake. `previous` is
+what the client's last connection from the same network measured, within the hour: what
+[Careful Resume](#a-returning-client-careful-resume) would start from (`rtt` in ms,
+`bandwidth` in bits/s, `age` in seconds); `null` on a first visit, from another network, or
+past its lifetime. Every field of the answer is optional: `transportParams`, and any of
+the connection options — `initialCongestionWindow`, `carefulResume`, `startupCap`,
+`pmtuDiscovery`, `maxUdpPayloadSize`, `initialRtt`, `reorderWindow`. What a connection
+has is the answer's, else what the server had when it was made (`createServer`, or
+`setOptions` since), else the default; `conn.setOptions` changes it afterwards.
+
+It cannot hold a connection hostage: no answer within **1 s**, an error, a throw, or an
+answer that does not check out (checked whole — nothing of it is taken) leaves the
+connection on the server's settings, and `process.emitWarning` says why (code
+`QUICO_CONNECTION_OPTIONS`). It runs for every Initial that opens a connection — spoofed
+ones included — so anything it calls out to should be cached; `validateAddress` screens
+spoofed addresses before it is asked.
 
 Setting `http2: false` and `http1: false` makes the server **HTTP/3 only** — no TCP
-listener is opened.
+listener is opened. Without `key`/`cert`/`SNICallback`, the TCP side serves cleartext
+HTTP/1.1 and h2c.
 
 **Methods**
 
-- `server.listen(port[, host][, callback])` — start listening. Binds UDP (and TCP
-  unless both fallbacks are disabled).
+- `server.listen(port[, host][, callback])` — start listening. Binds UDP (unless
+  `http3: false` or external sockets were given) and TCP (unless both fallbacks are
+  disabled). Port `0` picks a free port.
 - `server.close([callback])` — stop listening and close all connections.
+- `server.address()`
+- `server.handleConnection(socket)` — serve a TCP socket accepted elsewhere.
+- `server.setTimeout([msecs][, callback])` / `server.timeout` — as in `node:http`:
+  a connection idle for `msecs` emits `'timeout'` with its socket, and is destroyed
+  if nothing listens. `0` (default) disables it; changes apply to new connections.
+  TCP connections (HTTP/1.1, HTTP/2) — an HTTP/3 connection has QUIC's idle timeout.
+- `server.handlePacket(msg, rinfo)` / `server.hasConnection(rinfo)` — feed UDP
+  datagrams yourself when sharing a port (external socket mode).
+- `server.addProtocol(name, handler, { sniff? })` — serve another protocol on the same
+  TCP port: by ALPN `name`, or for TLS connections without ALPN, by
+  `sniff(firstBytes) → true | false | null`. `handler(socket, protocol)`.
+- `server.getTicketKeys()` / `setTicketKeys(keys)`,
+  `getEchKeys()` / `setEchKeys(ech)`,
+  `getValidateAddress()` / `setValidateAddress(on)`,
+  `getResetTokenSecret()` / `setResetTokenSecret(secret)` — change these at runtime.
+- `server.setOptions(options)` / `server.getOptions()` — every other option that can
+  change while the server runs, in one call, by the names `createServer` takes. See
+  [Changing options while running](#changing-options-while-running).
+- `server.getPendingHandshakes()` — HTTP/3 handshakes under way: connections made whose
+  handshake is not complete (the client's Finished verified) and has not failed.
+  What load is, and what to do about it, is the application's — see `'handshakeLoad'`:
+
+  ```js
+  // Retry (one round trip more) for new clients only while handshakes pile up.
+  server.on('handshakeLoad', ({ high }) => server.setOptions({ validateAddress: high }));
+  ```
+
+  `maxHandshakesPerAddress` is a cap, not a judgement of load: an address with that
+  many handshakes under way gets no more connections until some end. High enough for
+  many clients behind one address (a carrier NAT); with Retry on, it counts only
+  addresses that came back with a token.
+- `server.getStats()` — the server's counters since it was made (a copy). Counts only:
+  no addresses, names, headers or payloads. Codes are the protocol's, as hex strings,
+  each mapped to how many times it was seen:
+
+  | Field | Counts |
+  |---|---|
+  | `connections.quic` / `.tcp` | Connections taken: a QUIC Initial that began a handshake; a TCP socket accepted |
+  | `handshakes.quic` / `.tcp` | `{ completed, failed }` — failed: an alert either way, a timeout, the client gone first |
+  | `handshakes.resumed` | Completed with a session ticket |
+  | `protocols` | `{ h3, h2, 'http/1.1', other }` — connections by the protocol they spoke |
+  | `earlyData` | `{ accepted, rejected, tooEarly }` — 0-RTT the clients offered, and requests answered 425 |
+  | `closes.quic` | `{ sent, received, silent }` — QUIC's CONNECTION_CLOSE by code (`'0x12f'`: a TLS alert, illegal_parameter); silent: ended with none (idle timeout, stateless reset) |
+  | `closes.h3` | `{ sent, received }` — HTTP/3's CONNECTION_CLOSE by code (`'0x100'`: H3_NO_ERROR) |
+  | `closes.h2` | `{ sent, received }` — HTTP/2 GOAWAY by code |
+  | `closes.http1` | `{ sent }` — HTTP/1.1 requests refused before the application saw them, by status (`'400'`) |
+  | `streamResets.h3` / `.h2` | `{ sent, received }` — RESET_STREAM / RST_STREAM by code |
+  | `amplificationBlocked` | Datagrams held at the anti-amplification limit (RFC 9000 §8.1) until the client's address was validated |
+  | `internalErrors` | Exceptions of ours caught in the QUIC and HTTP/3 layers — a bug, never the client's |
+  | `since`, `uptime` | When counting began; milliseconds since |
+
+  ```js
+  setInterval(() => {
+    const s = server.getStats();
+    metrics.gauge('tls_failed', s.handshakes.quic.failed + s.handshakes.tcp.failed);
+    metrics.gauge('h3_errors', Object.entries(s.closes.h3.sent).filter(([c]) => c !== '0x100').reduce((n, [, k]) => n + k, 0));
+  }, 10000);
+  ```
+
+#### Changing options while running
+
+`server.setOptions(partial)` changes what can change, by `createServer`'s names;
+`server.getOptions()` returns all of them as they are now (a copy).
+
+```js
+const server = createServer({ key, cert, startupCap: 'stall', initialRtt: 100 }, app);
+
+server.on('handshakeLoad', ({ high }) => {
+  server.setOptions({ validateAddress: high, maxHandshakesPerAddress: high ? 16 : 64 });
+});
+
+server.setOptions({ startupCap: 'always' });          // returns server: chainable
+const { startupCap, validateAddress } = server.getOptions();
+
+try {
+  server.setOptions({ startupCap: 'off', initialRtt: -5 });
+} catch (e) {
+  e.code;      // 'ERR_OUT_OF_RANGE' — "initialRtt must be between 10 and 10000 (ms), got -5"
+  server.getOptions().startupCap;   // unchanged: one bad value, and nothing changes
+}
+server.setOptions({ cert });        // TypeError: "cert cannot be changed with setOptions(): use setSecureContext()"
+```
+
+| Option | Applies |
+|---|---|
+| `validateAddress` | from the next new client |
+| `handshakeLoadThreshold` | at once (the count as it stands) |
+| `maxHandshakesPerAddress` | from the next Initial |
+| `earlyData`, `earlyDataMaxAge` | from the next handshake (`earlyData: false` refuses early data on tickets already given out too) |
+| `startupCap`, `initialCongestionWindow`, `carefulResume`, `pmtuDiscovery`, `maxUdpPayloadSize`, `initialRtt`, `reorderWindow` | to HTTP/3 connections made from then on — those open keep theirs |
+
+**One connection's.** The same two calls on a connection — the handle of `'connection'`
+(its first packet) and `'secureConnection'`, and an HTTP/3 request's `req.socket`; on the
+client, an HTTP/3 response's `res.socket`. Every request on that connection shares it.
+
+```js
+server.on('connection', (conn) => {
+  if (conn.setOptions) conn.setOptions({ initialRtt: 80 });        // QUIC: before the handshake
+});
+
+app.get('/downloads/*', (req, res) => {
+  if (req.socket.setOptions) req.socket.setOptions({ startupCap: 'always' });   // HTTP/3 only
+  // …
+});
+```
+
+Over TCP (HTTP/1.1, HTTP/2) there is no `setOptions`: these options are QUIC's. Some are
+about how a connection begins, and past that point the call throws `ERR_INVALID_STATE`
+rather than take a value that would do nothing:
+
+| Option | On a running connection |
+|---|---|
+| `reorderWindow`, `startupCap` | any time (`startupCap` matters only while in Startup) |
+| `pmtuDiscovery` | any time: off — probes stop, the size found stays; on — the search starts |
+| `maxUdpPayloadSize` | any time: lower — packets shrink at once; higher — the search may go there |
+| `initialRtt` | before the first RTT sample |
+| `initialCongestionWindow` | before the first 1-RTT packet |
+| `carefulResume` | before the handshake is confirmed |
+| `transportParams` | not: they went out in the handshake (`connectionOptions` decides them) |
+| `validateAddress`, `handshakeLoadThreshold`, `maxHandshakesPerAddress` | not: the server's |
+
+A client sets them in its request options (`request`, `connect`, `WebTransport`,
+`connectUdp`) for what it sends: `startupCap`, `initialCongestionWindow`, `pmtuDiscovery`,
+`maxUdpPayloadSize`, `initialRtt`, `reorderWindow`.
 
 **Events**
 
-- `'request'` — `(req, res)` — same as the `requestListener` argument
-- `'listening'`
-- `'error'` — `(err)`
+| Event | Arguments | |
+|---|---|---|
+| `'request'` | `(req, res)` | Same as the `requestListener` argument. |
+| `'upgrade'` | `(req, socket, head)` | WebSocket and other upgrades, on all three versions — see [WebSocket](#websocket-1). With no listener, an upgrade request arrives as an ordinary `'request'`, as in Node (see `shouldUpgradeCallback`). |
+| `'connectUdp'` | `(req, tunnel)` | A request to proxy UDP (RFC 9298), on all three versions — see [CONNECT-UDP](#connect-udp-proxying-udp). With no listener, such a request arrives as before (`'upgrade'` or `'request'`). |
+| `'timeout'` | `(socket)` | A TCP connection was idle for `server.timeout` ms. |
+| `'connection'` | TCP: the socket, before TLS. QUIC: a connection handle, on the first packet. | Destroy it here to refuse a peer before any crypto is done. |
+| `'secureConnection'` | TCP: the TLS socket. QUIC: the connection handle. | After the handshake. |
+| `'stream_open'` | `(conn, streamId)` | HTTP/2 and HTTP/3: a request stream announced itself, before its headers are decoded. |
+| `'handshakeLoad'` | `({ pending, high })` | HTTP/3 handshakes under way reached `handshakeLoadThreshold` (`high: true`), or are back under half of it (`high: false`). |
+| `'stream_reset'` | `(conn, streamId, code)` | HTTP/2 RST_STREAM; HTTP/3 RESET_STREAM or STOP_SENDING — once per stream. A default limit applies (`streamResetBurst`, `streamResetRate`); this is for a policy of your own on top, for both versions. |
+| `'OCSPRequest'` | `(certificate, issuer, callback)` | As `node:tls`: a client asked for the certificate's OCSP status, TCP or QUIC. `certificate` and `issuer` are DER `Buffer`s (`issuer` `null` when the chain has none); `callback(null, response)` staples a DER OCSP response, `callback(null, null)` none, `callback(err)` fails the handshake. The handshake waits for the answer — fetch and cache responses ahead of time. Without a listener nothing is stapled. (Chromium asks over TCP, not over QUIC.) |
+| `'tlsClientError'` | `(err, socket \| conn)` | A handshake failed, as in `node:tls`: over TCP the TLS socket, over QUIC the connection handle. Either side may have ended it — plain HTTP sent to the TLS port (`ERR_SSL_WRONG_VERSION_NUMBER`), no common ALPN, or a client refusing the certificate (`ERR_TLS_ALERT_RECEIVED`, `err.alert` = 42 for `bad_certificate`). |
+| `'clientError'` | `(err, socket)` | HTTP/1.1, once the connection is up (parse errors, the socket's errors). TLS handshake failures are `'tlsClientError'`. |
+| `'sessionError'` | `(err, session)` | HTTP/2: the connection failed (a protocol error — GOAWAY already sent —, the socket's error, a peer not reading past `maxSessionMemory`). `session` is the connection handle of `'stream_open'`. As `node:http2`. |
+| `'listening'` | | |
+| `'error'` | `(err)` | |
+
+The `conn` handle given to the HTTP/2 and HTTP/3 stream events has `remoteAddress`,
+`remotePort`, `fingerprint()`, `sendRstStream(streamId, 'refused' | 'cancelled' | code)`,
+`goaway([code])` and `destroy([code])` — the same on both versions, so one rate limiter
+serves both:
+
+- `goaway()` stops new requests; those already in flight finish. HTTP/2 sends GOAWAY;
+  HTTP/3 sends GOAWAY too, refuses any later request with `H3_REQUEST_REJECTED` (safe for
+  the client to retry elsewhere), and closes the connection once the last response has
+  been delivered. `goaway('overloaded')` says why: `ENHANCE_YOUR_CALM` / `H3_EXCESSIVE_LOAD`.
+- `destroy()` cuts the connection now, in-flight requests included (`'overloaded'` or a
+  number sets the HTTP/3 close code).
 
 ```js
 const server = quico.createServer({ key, cert, http2: false, http1: false });
@@ -569,8 +1260,9 @@ server.listen(4433, '0.0.0.0');
 
 ### `quico.request(url|options[, options], callback)`
 
-Issues an HTTP/3 request. Returns a [`ClientRequest`](#clientrequest). Accepts a URL
-string, an options object, or both (options override the URL).
+Issues an HTTP request — HTTP/3 first, falling back to HTTP/2 and HTTP/1.1. Returns a
+[`ClientRequest`](#clientrequest). Accepts a URL string, an options object, or both
+(options override the URL). Requests are always made over TLS.
 
 **Options**
 
@@ -581,9 +1273,21 @@ string, an options object, or both (options override the URL).
 | `path` | `string` | `'/'` | Request path including query string. |
 | `method` | `string` | `'GET'` | HTTP method. |
 | `headers` | `object` | `{}` | Request headers. |
+| `priority` | `object` | — | `{ urgency: 0..7, incremental }` — the request's priority (RFC 9218), as the `priority` header; `req.setPriority()` changes it later. |
+| `ech` | `Buffer \| string` | — | The server's ECH key list (an ECHConfigList; base64 — the `ech=` of its HTTPS record — or bytes): the ClientHello goes encrypted, on every version. See [ECH](#encrypted-client-hello-ech). |
+| `echGrease` | `boolean` | `true` | Without `ech`: send GREASE ECH, as browsers do. |
+| `ecdhCurve` | `string` | `'auto'` | The key-exchange groups to offer, in order — `node:tls`'s option, on every version. `'auto'` is what `node:tls` offers and the browsers do: the post-quantum hybrid X25519MLKEM768 first, with an X25519 share beside it (a server without it needs no retry), then X25519, P-256, P-384 — the hybrid where Node has ML-KEM (24.7+), the classical groups alone before that. `'X25519:P-256'` and the like for classical only; `'MLKEM1024'`, `'SecP384r1MLKEM1024'` and the other post-quantum groups by name (see [Post-quantum TLS](#post-quantum-tls)). A name not implemented, or the hybrid in a Node without ML-KEM (it needs 24.7+ on OpenSSL 3.5), throws from `request()`. A request shares only connections opened with the same groups. |
+| `startupCap` | `'stall'` \| `'always'` \| `'off'` | `'stall'` | HTTP/3: Startup's window cap on what the client sends (an upload) — as the server's option. |
+| `pmtuDiscovery`, `maxUdpPayloadSize`, `initialRtt`, `reorderWindow`, `initialCongestionWindow` | | | HTTP/3: this client's connection options — as the server's. `res.socket.setOptions()` changes them while it runs. |
 | `rejectUnauthorized` | `boolean` | `true` | Verify the server certificate. Set `false` for self-signed certs. |
-| `ca` | `Buffer \| string \| Array` | `null` | Trust these CAs in addition to the defaults. |
-| `agent` | `Agent \| false` | `globalAgent` | Connection pool to use. `false` disables pooling. |
+| `ca` | `Buffer \| string \| Array` | `null` | Trust these CAs instead of the platform roots (as in Node). |
+| `sigalgs` | `string` | lemon-tls's list | The signature schemes to offer, in order — `node:tls`'s option. `'mldsa65:mldsa87:ecdsa_secp256r1_sha256'` asks a server that holds an ML-DSA certificate for it; the default offers ML-DSA after the classical schemes. A request shares only connections opened with the same list. |
+| `key`, `cert`, `passphrase` | | — | A client certificate, for a server that asks for one — on HTTP/3, HTTP/2 and HTTP/1.1. A request with one shares only connections opened with the same certificate (see `Agent`). |
+| `earlyData` | `boolean` | `false` | With a session whose ticket allows it, send a `GET`, `HEAD` or `OPTIONS` in the first flight (0-RTT) — on every version. Refused or answered `425`, it is sent again after the handshake by itself; `res.earlyData` says what happened. See [0-RTT](#0-rtt-early-data). |
+| `session` | `Buffer` | the agent's | A TLS session to resume with, from an earlier `'session'` event (as `https.request`). Without it the agent offers its own (see `Agent`). |
+| `requestOCSP` | `boolean` | `false` | Ask the server for its stapled OCSP response (`tls.connect`'s option), on every version; it comes as the request's `'OCSPResponse'` (a DER `Buffer`). Only a full handshake carries one — a resumed session has no certificate. |
+| `quicVersions` | `Array` | `['v1', 'v2']` | HTTP/3: the QUIC version of the first flight, then the others it offers (RFC 9368) — a server that prefers v2 moves the connection there. `['v2']`: v2 from the first packet. |
+| `agent` | `Agent \| false` | `globalAgent` | Connection pool to use. `false`, as in Node: a connection of this request's own, closed when it is done (over HTTP/2 and HTTP/1.1 it went on a pool shared by every `agent: false` request), and no sessions kept — pass `session` to resume. |
 | `http2` | `boolean` | `true` | Allow HTTP/2 fallback. |
 | `http1` | `boolean` | `true` | Allow HTTP/1.1 fallback. |
 | `h3Timeout` | `number` | `3000` | Milliseconds to wait for QUIC before falling back. |
@@ -609,7 +1313,8 @@ req.end();
 
 ### `quico.get(url|options[, options], callback)`
 
-Same as `request()` with `method: 'GET'`, and `end()` is called automatically.
+Same as `request()` with `method: 'GET'`. (`request()` itself ends GET and HEAD
+requests on the next tick.)
 
 ---
 
@@ -625,11 +1330,28 @@ response.
 | `method` | Request method (server side) |
 | `url` | Request path (server side) |
 | `statusCode` | Response status (client side) |
+| `earlyData` | Server: `true` if the request came in 0-RTT (it could be a replay). Client: `'accepted'` (sent early, taken), `'rejected'` (sent again after the handshake), `null` (not sent early). |
 | `socket` / `connection` | Compatibility shim for frameworks |
 
 Standard `Readable` behaviour applies: `'data'`, `'end'`, `'error'`, `pipe()`,
 `for await (const chunk of req)`, and `setTimeout(ms, cb)`.
 
+A body cut short — on the server, a request the client cancels or whose connection
+drops; on the client, a response the server resets or cuts off — gets `aborted = true`,
+an `'aborted'` event, an `'error'` (`ECONNRESET`, or the cause) only if something listens
+for one, and `'close'`; never `'end'`, which would say the body arrived whole, and never
+an unhandled `'error'`, because a browser cancelling a request is ordinary traffic. This
+is node:http's behaviour, on all three versions. HTTP/2 requests have
+`req.h2Fingerprint()`, HTTP/3 requests `req.h3Fingerprint()`; over HTTP/2 and HTTP/3,
+`req.headers` also keeps the pseudo-headers (`:method`, `:path`, `:authority`, …), as
+`node:http2` does, plus `host` taken from `:authority`.
+
+A request handler that throws ends only its own exchange — HTTP/1.1 answers `500` and
+closes the connection, HTTP/2 and HTTP/3 reset that one stream (`INTERNAL_ERROR`,
+`H3_INTERNAL_ERROR`), and other requests on the connection go on — and the exception is
+then rethrown on the next tick, so it reaches `process.on('uncaughtException')` (or ends
+the process) exactly as in Node. A response already ended before the throw is delivered.
+
 ---
 
 ### `ServerResponse`
@@ -640,12 +1362,35 @@ A Node.js `Writable` stream.
 - `res.setHeader(name, value)` / `getHeader` / `getHeaders` / `getHeaderNames` /
   `hasHeader` / `removeHeader`
 - `res.flushHeaders()`
-- `res.write(chunk)` / `res.end([chunk])`
+- `res.write(chunk)` / `res.end([chunk])` — over HTTP/1.1, a body given whole to `end()`
+  (nothing written before it) is sent with `Content-Length`, otherwise chunked, as
+  `node:http` does
 - `res.statusCode`, `res.headersSent`
+- `res.writeEarlyHints(hints[, callback])` — `103 Early Hints`, on all three versions
+- `res.destroy([err])` / `req.destroy([err])` — give up on the exchange, as in Node: the
+  stream is reset, the connection serves on. HTTP/3: RESET_STREAM and STOP_SENDING with
+  `H3_REQUEST_CANCELLED` (`H3_INTERNAL_ERROR` with an `err`; RFC 9114 §4.1.1); after a
+  finished response, only STOP_SENDING with `H3_NO_ERROR` for a request body still
+  coming. HTTP/2: RST_STREAM `CANCEL` (`INTERNAL_ERROR` with an `err`). HTTP/1.1: the
+  socket. Both went unsent: a destroyed exchange held its stream until the idle timeout.
+- `res.addTrailers({ … })` before `end()` — response trailers, on all three versions
+  (HTTP/1.1 chunked; HTTP/2 and HTTP/3 a HEADERS frame after the body). What gRPC
+  answers with: `grpc-status` / `grpc-message` go only there. A request's trailers are in
+  `req.trailers` / `req.rawTrailers` by `'end'`; on the client, `res.trailers`
+- `res.setPriority({ urgency, incremental })` — HTTP/2 and HTTP/3 (over HTTP/1.1, one
+  response at a time, it does nothing — the same handler runs on all three): the response's
+  priority (RFC 9218: urgency 0..7, 0 first, default 3; incremental: shares the
+  connection with others of its urgency rather than going whole, one at a time). Without
+  it, the client's `priority` header and `PRIORITY_UPDATE` frames decide; set, it is the
+  server's decision and later signals from the client do not override it. The client's
+  header is in `req.headers.priority`. See
+  [Stream scheduling](#stream-scheduling-byte-level-round-robin-by-priority).
 
 For WebTransport sessions, `ServerResponse` additionally exposes
-`sendDatagram()`, `createBidirectionalStream()` and `createUnidirectionalStream()` —
-see [WebTransport](#-webtransport).
+`sendDatagram()`, `maxDatagramSize()`, `datagrams`, `getStats()`, `createBidirectionalStream()`,
+`createUnidirectionalStream()`, `canOpenStream()`, `createSendGroup()`, `setWebTransportProtocol()`,
+`closeSession()` and `drainSession()` — see
+[WebTransport](#webtransport-1).
 
 ---
 
@@ -653,10 +1398,45 @@ see [WebTransport](#-webtransport).
 
 A Node.js `Writable` stream.
 
-- `req.write(chunk)` / `req.end([chunk])`
-- `req.abort()` / `req.destroy()`
+- `req.write(chunk)` / `req.end([chunk])` — the body streams, as `node:http`'s does: written
+  while the connection is being made it is held (past 64 KB `write()` returns `false`, then
+  `'drain'`), and from then on it goes as it is written, with the transport's backpressure —
+  the QUIC stream's, HTTP/2's window, the socket's. Over HTTP/1.1 a body not yet complete goes
+  chunked, unless a `content-length` was set: the bytes must then add up to it (too many or too
+  few fails the request with `ERR_HTTP_CONTENT_LENGTH_MISMATCH`; on a pooled connection the
+  difference would be read as another request). The first protocol to be ready takes the
+  request, and only it sends: a request whose answer takes HTTP/3 longer than `h3Timeout` is not
+  sent over TCP as well. The first 64 KB of a body are kept, so that a request refused
+  unprocessed (`REFUSED_STREAM`, `H3_REQUEST_REJECTED`, past a GOAWAY), or whose protocol failed
+  before it was sent, goes again whole; a larger one that has started cannot, and fails.
+- `req.addTrailers({ … })` before `end()` — request trailers (`node:http`'s), on all three
+  versions: a HEADERS frame after the body over HTTP/2 and HTTP/3, after the last chunk over
+  HTTP/1.1 (so only chunked, as in `node:http`). Connection-specific and forbidden fields are
+  dropped, as for responses.
+- `req.setPriority({ urgency, incremental })`, or the `priority` option of `request()` — the
+  request's priority (RFC 9218: urgency 0..7, 0 first, default 3; incremental: its response
+  may share the connection with others of its urgency). Before the request goes it is the
+  `priority` header (`'u=1, i'`; the default sends none), on all three versions; after, a
+  `PRIORITY_UPDATE` frame over HTTP/2 and HTTP/3 (HTTP/1.1 has none — nothing is sent). It
+  is the client's signal: a server whose application set the response's priority
+  (`res.setPriority`) keeps its own.
+- `req.abort()` / `req.destroy([err])` — stop the exchange where it is carried: the
+  request's stream over HTTP/3 and HTTP/2 (a pooled connection serves on), the socket over
+  HTTP/1.1. `res.destroy()` on an unfinished response does the same. A request the server
+  reset after it was sent — a non-idempotent one (RFC 9110 §9.2.2) — fails with the reset,
+  and is not sent again over another version; one it refused is. As in Node, the response then sees `'aborted'` and `'close'` (an `'error'` only
+  if it listens for one), never `'end'`; `destroy()` before any response fails the request
+  with `ECONNRESET` (`socket hang up`), `abort()` emits `'abort'` and no error.
 - `req.protocol` — which transport was actually used: `'h3'`, `'h2'` or `'https'`
-- Events: `'response'`, `'error'`
+- Events: `'response'`, `'error'`, `'timeout'`, `'session'` (a `Buffer`: a TLS session the
+  server sent, to resume with — as a `TLSSocket`'s, once per ticket, on the request that
+  opened the connection), `'OCSPResponse'` (with `requestOCSP`: the server's stapled OCSP
+  response, DER, on the request that opened the connection)
+- A request the server did not process — refused (HTTP/2 `REFUSED_STREAM`, HTTP/3
+  `H3_REQUEST_REJECTED`) or past its `GOAWAY` (a server that closes its connections after
+  `maxRequestsPerSocket` requests does that to the ones in flight) — is sent again once,
+  on a new connection (RFC 9113 §8.7, RFC 9114 §4.1.1, §5.2). When every protocol fails,
+  the `'error'` has the last one's reason as its `cause`.
 
 ---
 
@@ -668,22 +1448,66 @@ Connection pool and protocol cache, analogous to `https.Agent`.
 import { Agent } from 'quico';
 
 const agent = new Agent({
-  http2: true,        // allow H2 in the fallback chain
-  http1: true,        // allow H1.1 in the fallback chain
-  h3Timeout: 3000,    // ms before falling back from H3
-  maxSockets: 256,
-  timeout: 30000
+  http2: true,          // allow H2 in the fallback chain
+  http1: true,          // allow H1.1 in the fallback chain
+  h3Timeout: 3000,      // ms before falling back from H3
+  maxSockets: 256,      // HTTP/1.1 sockets per origin; more requests wait
+  maxCachedSessions: 100, // TLS sessions kept to resume with
+  ecdhCurve: 'X25519:P-256'  // its requests' key exchange (default 'auto': post-quantum first), unless one names its own
 });
 
 quico.request({ hostname: 'example.com', agent }, handler).end();
 ```
 
-The agent caches which protocol works per host (including `Alt-Svc` advertisements
-with their `ma=` lifetime), pools live QUIC connections for multiplexing, and
-coalesces concurrent requests to the same origin onto a single connection.
+HTTP/2 and HTTP/3 multiplex every request to an origin over one connection, which
+the agent keeps and shares. HTTP/1.1 needs a socket per request, and for those the
+agent takes `node:http.Agent`'s options, with Node's meanings:
+
+| Option | Default | |
+|---|---|---|
+| `keepAlive` | `true` | Keep a socket for the next request (Node's `globalAgent` default; a plain `new http.Agent()` defaults to `false`). |
+| `keepAliveMsecs` | `1000` | TCP keep-alive initial delay on kept sockets. |
+| `maxSockets` | `256` | Sockets per origin at once. Requests beyond it wait for a socket. |
+| `maxFreeSockets` | `256` | Idle sockets kept per origin. |
+| `timeout` | `5000` | How long an idle socket is kept. |
+
+**Protocol discovery.** The agent remembers which protocol worked for each origin,
+and separately what each origin advertises in `Alt-Svc` (RFC 7838), for its `ma=`
+lifetime. When an origin advertises HTTP/3 — on its own port or another — the next
+request tries HTTP/3 there first, and falls back to the origin over TCP (or its own
+HTTP/3) if that fails; a failed alternative is left alone for five minutes, as
+browsers do. Concurrent requests to an origin share one connection.
+
+**Which connections a request shares.** As `https.Agent` names its connections
+(`agent.getName(options)`), a pooled connection is keyed by its origin and its TLS
+identity — `ca`, `cert` / `key`, `rejectUnauthorized`, `servername` (hashed). A
+connection opened with `rejectUnauthorized: false` never carries a request that
+verifies, nor one verified against a private `ca` a request that trusts the public
+roots; before, any connection to the origin was reused. The same goes for the
+standalone `h1Request` / `h2Request` pools.
+
+**Session resumption.** As `https.Agent`, the agent keeps the TLS sessions each
+origin's server sent — up to 4 per origin, newest last; `maxCachedSessions` (default
+`100`, the oldest dropped; `0` keeps none) bounds the origins — and a new connection
+resumes with one — no certificate exchange — over HTTP/3, HTTP/2 and HTTP/1.1, and for
+`WebSocket` (`wss://`) and `connectUdp` too. Each session is used **once** (RFC 8446
+§C.4): two connections opened at once take two, so they cannot be linked, and both can
+send 0-RTT — with one kept per origin, the second's early data was always refused as a
+replay. A QUICO server sends two tickets per full handshake (one per resumption), as
+OpenSSL and BoringSSL do. An expired session is not offered (lemon-tls `sessionInfo`).
+TCP sessions are kept per ALPN (`h2`, `http/1.1`): a ticket carries 0-RTT only for the
+protocol it was issued under. Before, every connection was a full handshake, and a QUICO server
+never saw its own ticket again: what it saves there about the path (Careful Resume,
+above) is only used by a client that resumes. A session is offered only where it was
+made: the same transport (a TCP session is not offered to QUIC), origin, server name and
+verification — one from a connection with `rejectUnauthorized: false`, or another `ca`, is
+not offered to one that verifies. A session the handshake fails with is dropped. A
+resumed connection reports the server's certificate and `authorized` from the session, as
+Node does. For QUIC it also keeps the server's NEW_TOKEN token (RFC 9000 §8.1.3) — used
+once, by the next connection, which then needs no proof of address.
 
 `quico.globalAgent` is the default instance used when no `agent` option is given.
-Call `agent.destroy()` to close all pooled connections.
+Call `agent.destroy()` to close everything it pools.
 
 ---
 
@@ -694,6 +1518,199 @@ Re-exported from LemonTLS.
 
 ---
 
+### 0-RTT (early data)
+
+A returning visitor's request can go in its first flight, with the ClientHello, under a
+key from its last session — the answer comes one round trip sooner. On a phone at
+150 ms RTT, that is 150 ms off the first byte of every page that opens a new connection.
+It works the same on every version: over HTTP/3 in QUIC's 0-RTT packets, and over TLS 1.3 on
+TCP (HTTP/2, HTTP/1.1) as TLS early data — with Encrypted Client Hello as well (the early
+data is the inner hello's, RFC 9849 §6.1). Chromium and Firefox send 0-RTT over HTTP/3.
+
+The price: early data has no protection against replay. Someone who captured the first
+flight can send it again, and the server would run the request again. So it is **off by
+default**, and turned on it is held to four rules:
+
+1. **A ticket carries early data once.** The server remembers each ticket's early data
+   (`earlyDataReplay`, in memory by default) for twice `earlyDataMaxAge`; the second time,
+   the handshake goes on without it.
+2. **Only for a short while.** The age the client states for its ticket must be within
+   `earlyDataMaxAge` (10 s) of what the server measures — a captured flight goes stale.
+3. **Only under the same settings.** A ticket from before a change of transport
+   parameters or of the HTTP/3 SETTINGS this server sends (or a restart with other ones),
+   or one a connection with lower limits (`connectionOptions`) would not honour, resumes
+   the session without early data: a client sends its early data under what it remembered
+   (RFC 9000 §7.4.1, RFC 9114 §7.2.4.2).
+4. **Only safe requests.** `earlyDataFilter` decides which requests may run before the
+   handshake completes — by default `GET`, `HEAD` and `OPTIONS`; the others get
+   `425 Too Early` and the client sends them again once the handshake is done (RFC 8470),
+   by itself — a browser does, and so does `quico.request`.
+
+Refused early data costs nothing but the round trip saved: the request is served as it
+would have been without it.
+
+```js
+const server = quico.createServer({
+  key, cert,
+  ticketKeys,                       // stable keys, or tickets die with the process
+  earlyData: true,
+  earlyDataFilter(req, cb) {        // asynchronous: a database, a rule engine
+    if (req.url.startsWith('/api/')) return cb(null, false);   // → 425, after the handshake
+    cb(null, true);
+  }
+}, (req, res) => {
+  // req.earlyData: this request may be a replay — never act on it (a payment,
+  // a login) without the handshake; the filter is where to say so.
+  res.end('hello');
+});
+```
+
+**A restart.** The default store is in memory and starts empty: a ticket issued before a
+restart (with `ticketKeys` kept) gets no early data for twice `earlyDataMaxAge` after it —
+a flight captured just before the restart cannot be sent again just after it. One 1-RTT
+handshake for such a client, and its new ticket is fine.
+
+**More than one server.** Each server remembers in its own memory, so a flight captured
+at one could be replayed at another that shares `ticketKeys`. Give them one store:
+
+```js
+const server = quico.createServer({
+  key, cert, ticketKeys, earlyData: true,
+  // id: the ticket (a SHA-256, hex); ttl: how long to remember it, ms.
+  earlyDataReplay(id, ttl, cb) {
+    redis.set('0rtt:' + id, '1', 'PX', ttl, 'NX', (err, ok) => cb(err, ok === 'OK'));
+  }
+});
+```
+
+`server.setOptions({ earlyData: false })` turns it off at once — tickets already given out
+included.
+
+**What the client remembers.** Each session keeps, with the ticket, the server's transport
+parameters and its HTTP/3 SETTINGS; early data goes within them (RFC 9000 §7.4.1, RFC 9114
+§7.2.4.2). A server that accepts the 0-RTT and then gives less — a transport parameter
+lower than remembered, a setting omitted or lower — breaks the connection it took the
+early data on: the client closes it, `PROTOCOL_VIOLATION` or `H3_SETTINGS_ERROR`, naming
+what changed.
+
+**The client.** `earlyData: true` on a request with a session (its own, or the agent's):
+
+```js
+const req = quico.request('https://example.com/', { earlyData: true }, (res) => {
+  console.log(res.earlyData);   // 'accepted', 'rejected' or null
+});
+req.end();
+```
+
+Only `GET`, `HEAD` and `OPTIONS` are sent early; anything else waits for the
+handshake. Refused — by the server's TLS, or with a `425` — the request is sent again after
+the handshake, and the response is the one to that.
+
+**Over TCP.** The same options, the same checks (one use per ticket, the age, the context —
+a ticket from the HTTP/3 side carries no early data here, nor the other way round), the same
+`earlyDataFilter` and `425`. A ticket allows up to 16 KB of early data — room for a request,
+not an upload; what does not fit goes after the handshake. The server reads the early
+request as it comes and may answer as soon as its own handshake flight is out, so the
+client has the answer one round trip after its first flight.
+
+---
+
+### Encrypted Client Hello (ECH)
+
+Without ECH, the first message of every TLS connection — the ClientHello, with the
+site's name (SNI) and the ALPN list — crosses the network in the clear: anyone on the
+path sees which site is being visited, HTTPS or not. With ECH (RFC 9849) the browser
+sends the real ClientHello encrypted, inside an outer one that names only a shared
+*public name*, to a key the server publishes in its DNS HTTPS record. The server
+decrypts it and the handshake goes on as usual — `SNICallback` gets the real name.
+
+ECH hides names that **share an address**: when many sites sit behind one IP and one
+public name (a CDN, a host with many domains), the network sees only which group. A
+single site on its own IP is still known by the IP.
+
+**1. A key, kept across restarts** — like DNSSEC keys:
+
+```js
+import quico from 'quico';
+const k = quico.generateEchKey({ publicName: 'ech.example.com' });
+// save k.config and k.privateKey (bytes); publish k.configListBase64 in DNS
+```
+
+**2. The DNS record** — the `ech` key of the domain's HTTPS record (RFC 9460), e.g.
+with [dnssec-server](https://github.com/colocohen/dnssec-server):
+
+```
+www  IN HTTPS 1 . alpn="h3,h2" ech="AEb+DQBC…"
+```
+
+**3. The server** — with a certificate for the public name as well, served when a
+client's key could not be used:
+
+```js
+const server = quico.createServer({
+  SNICallback: (name, cb) => cb(null, contexts[name]),   // incl. ech.example.com
+  ech: { keys: [{ config: saved.config, privateKey: saved.privateKey }] },
+}, (req, res) => {
+  res.end(req.socket.servername + ' ' + req.socket.echStatus);
+});
+```
+
+`req.socket.echStatus` is `'accepted'` (the hidden name was used), `'rejected'` (the
+client's key was not one of ours: the handshake ran as the public name, and the client
+was sent our current keys — `retry_configs` — to connect again with), or `null` (no ECH
+offered). `req.socket.echAccepted` is the boolean.
+
+**Rotation** — new key in DNS first, then on the server, the old one kept with
+`retry: false` until the old record's TTL has passed:
+
+```js
+server.setEchKeys({ keys: [
+  { config: next.config, privateKey: next.privateKey },
+  { config: prev.config, privateKey: prev.privateKey, retry: false },
+] });
+```
+
+The key itself is not certified by anyone: its trust comes from DNS (DNSSEC, and
+DNS-over-HTTPS on the client). A substituted key could reveal which name was asked
+for, never let anyone impersonate the site — the certificate is still checked against
+that name. Tested against Go's `crypto/tls` client (accept, retry, HelloRetryRequest,
+resumption); DTLS is not covered.
+
+#### The client
+
+`quico.request`, `WebTransport` and `connectUdp` take the server's key list as `ech` —
+the `ech=` value of its HTTPS record (base64), or the bytes. The real ClientHello goes
+encrypted, over HTTP/3, HTTP/2 and HTTP/1.1 alike:
+
+```js
+const req = quico.request({ hostname: 'www.example.com', path: '/', ech: 'AEb+DQBC…' }, (res) => {
+  console.log(res.socket.echStatus);   // 'accepted'
+});
+req.end();
+```
+
+`res.socket.echStatus` (and `echAccepted`) on every version: `'accepted'`, `'grease'`, or
+`null`. When the server cannot use the key (it rotated its keys, say), it proves itself
+for the public name and sends its current keys; the request is tried once more with
+them, automatically. The agent keeps those keys (RFC 9849 §6.1.8) and uses them first on
+the next requests given the same `ech` — `quico.request` and `connectUdp` — so a server
+that rotated its keys costs one retry, not one per connection until your DNS has the new
+ones. They are kept as TLS sessions are: per origin and TLS settings, a day at most, none
+with `agent: false` or `maxCachedSessions: 0`. Rejected in their turn, they are forgotten
+and your `ech` is tried again. If it sends none — its ECH is off — the request fails with
+`ERR_ECH_REJECTED` (`err.retryConfigs` null) instead of falling back to a ClientHello
+in the clear: going on without ECH is the caller's decision.
+
+Without `ech`, the client sends **GREASE ECH** (RFC 9849 §6.2), as Chrome and Firefox
+do: an ECH extension no server can decrypt, so that clients that do use ECH do not stand
+out. `echGrease: false` turns it off. QUICO does not look up HTTPS records itself (no
+DNS-over-HTTPS here): the application passes the key list it resolved.
+
+ECH is TLS 1.3 only. Tested against Go's `crypto/tls` server (accept, retry,
+HelloRetryRequest, GREASE) and QUICO's own server, over QUIC and TCP.
+
+---
+
 ### Framework compatibility
 
 A core design goal is a **familiar API**. If you know `node:https`, QUICO should feel
@@ -763,29 +1780,181 @@ quico.createServer({ key: KEY, cert: CERT }, (req, res) => {
 }).listen(4433);
 ```
 
+**Datagrams** (`res.sendDatagram(data)`) are never retransmitted, but they are sent under
+the connection's congestion control (RFC 9221 §5.4): they wait for room in a packet and
+go ahead of stream data. While the path cannot keep up, the oldest are dropped — any that
+have waited 250 ms, and past 1 MB queued. `sendDatagram` returns `true` when the datagram
+is queued, `false` when it cannot be sent at all (larger than `res.maxDatagramSize()`, or
+the peer takes no datagrams). Datagrams sent right after accepting the session are held
+until the `200` has left: a browser drops datagrams for a session it does not know yet.
+
+`res.datagrams` is the browser's datagrams object too — the client's `wt.datagrams`,
+the same code: `readable`, `writable`, `createWritable({ sendOrder, sendGroup })`,
+`maxDatagramSize`, `incomingMaxAge`, `outgoingMaxAge` and the high-water marks. Its
+`outgoingMaxAge` is the session's own: sessions on one connection share its queue, each
+with its own age limit and its own count of datagrams dropped.
+
+**Statistics** — `res.getStats()`, as the client's `wt.getStats()` (the same function):
+a Promise of the browser's `WebTransportConnectionStats` — `bytesSent`, `packetsSent`,
+`packetsLost`, `bytesReceived`, `packetsReceived` (whole packets), `smoothedRtt`,
+`rttVariation`, `minRtt` (ms), `estimatedSendRate` (bits/s, the congestion control's
+bandwidth estimate; `null` until measured), `atSendCapacity` (the congestion window is
+full) — with QUICO's `congestionWindow` and `bytesInFlight` besides, and `datagrams`:
+`{ droppedIncoming, expiredIncoming, expiredOutgoing, lostOutgoing }`, this session's.
+The connection's figures are shared by every session on it, as in the browser. A sender
+that has to choose what to drop for a slow receiver reads `atSendCapacity`, or
+`bytesInFlight` against `congestionWindow`: the congestion is the connection's, not one
+stream's.
+
+**Priorities** are the browser's: a stream's `sendOrder` and `sendGroup`
+(`stream.sendGroup = res.createSendGroup()`), and `res.sendDatagram(data, { sendOrder,
+sendGroup })` as the browser's `datagrams.createWritable({ sendOrder, sendGroup })`. The
+groups of a session share its sending equally; within a group the highest `sendOrder`
+with something to send goes first, equal orders taking turns; a stream in no group is in
+the session's own. Datagrams sent without them go ahead of every stream. The session's
+urgency — the CONNECT request's `priority`, or `res.setPriority()` — is its streams'.
+
+**Streams** are Node streams: a `Duplex` for a bidirectional stream, a `Readable` for the
+client's unidirectional one (`'unidirectionalStream'`), a `Writable` for ours. What the
+client sends is credited back to it as the application reads, so a stream nobody reads
+holds one stream window (256 KB bidirectional, 128 KB unidirectional, or as set with
+`transportParams`), not the whole upload. Each direction can be given up with an
+application error code (32-bit, mapped onto HTTP/3's codes as draft-ietf-webtrans-http3
+§4.4 defines), on its own — as in the browser, the two directions of a bidirectional
+stream are independent:
+
+| | |
+|---|---|
+| `stream.resetStream(code[, { reliableSize }])` | Stop sending: what was written and not yet sent is dropped. A browser's reader fails with `streamErrorCode`. With `reliableSize`, the first that many bytes written still arrive whole before the reset (RESET_STREAM_AT, draft-ietf-quic-reliable-stream-reset) — a MOQT subgroup's header, say; the stream's own WebTransport header always does (draft-ietf-webtrans-http3 §4.4), so the peer knows its session. Needs a peer that takes RESET_STREAM_AT (`stream.reliableResetSupported`; QUICO advertises it); otherwise a RESET_STREAM. Also after `end()`, until all of it is acknowledged (RFC 9000 §3.1) — a MOQT subgroup past its delivery timeout is dropped, not sent to the end. Returns `true` when a reset went, `false` when there was nothing left to reset. |
+| `stream.stopSending(code)` | Stop reading (STOP_SENDING): the peer stops. A browser's writer fails with `streamErrorCode`. |
+| `stream.destroy()` | Both, with code 0, for whatever direction is unfinished. |
+| `'peerReset'` (`code`) | The peer stopped sending (RESET_STREAM): our reading is over — what was read stands, the readable side ends, and its WHATWG `readable` fails with the code. Our writing goes on. |
+| `'peerStopSending'` (`code`) | The peer stopped reading (STOP_SENDING): our writing is over — writes from now on are dropped, the WHATWG `writable` fails. Our reading goes on. |
+| `stream.streamErrorCode` | The peer's code from either (`null` if it sent none). |
+| `'committed'` | Our sending part is "all data committed" (draft-ietf-webtrans-overview §4.3; "Data Recvd", RFC 9000 §3.1): every byte written and the FIN acknowledged by the peer. Once; never for a stream reset first. `'finish'` (and `end()`'s callback) only says the bytes were handed to QUIC — a delivery deadline (MOQT's SUBGROUP_DELIVERY_TIMEOUT) runs from the FIN to here, and `resetStream()` before it drops what is left. `sendState` is then `'committed'`. |
+| `stream.id`, `stream.sessionId` | Its QUIC stream ID, and its session's — read-only. |
+| `stream.sendState`, `stream.recvState` | Each direction, read-only: `'open'`, `'finished'` (the FIN sent / received), `'committed'` (ours: all of it and the FIN acknowledged), `'reset'` (we gave up sending / the peer did), `'stopped'` (the peer asked us to stop / we stopped reading); `null` for the direction a unidirectional stream does not have. A `'drain'` says the writer may write; `sendState` says whether what it writes still goes anywhere. |
+| `'close'` | Once both directions are over — by FIN, reset or stop, from either side, in any order. When the peer's reset or stop is what ends the last one, an `Error` (`ERR_WEBTRANSPORT_STREAM_RESET` / `ERR_WEBTRANSPORT_STOP_SENDING`) goes to an `'error'` listener — only if there is one. |
+
+**Opening streams** is the browser's: `res.createBidirectionalStream(options[, callback])`
+and `res.createUnidirectionalStream(options[, callback])` take `{ sendOrder, sendGroup,
+waitUntilAvailable }` and return the stream at once, in Node's way (`net.connect`'s).
+Past the client's `MAX_STREAMS` without `waitUntilAvailable` they return `null`
+(`res.canOpenStream({ bidirectional })` says beforehand); with it, the stream is opened
+and its bytes wait for the client to allow more. Until then `stream.pending` is `true`;
+when it is open, `'ready'` and `callback(null, stream)` — or `callback(err)` (code
+`ERR_WEBTRANSPORT_STREAM_NOT_OPEN`) if it was given up first, its session closed.
+
+```js
+const st = res.createUnidirectionalStream({ waitUntilAvailable: true, sendOrder: 5 }, (err) => {
+  if (err) return;          // never opened: the session went first
+  // open: the client has allowed it
+});
+st.write(object);           // may be written at once; it waits with the stream
+```
+
+**The subprotocol** — the browser's `new WebTransport(url, { protocols: [...] })`: the
+client's offer is `req.webTransportProtocols`, in its order of preference (parsed from
+WT-Available-Protocols), and `res.setWebTransportProtocol(name)` before `writeHead(200)`
+picks one — sent as WT-Protocol, what the client's `wt.protocol` then says. A name the
+client did not offer throws: the client would fail the session over it.
+
+```js
+const chosen = req.webTransportProtocols.find(p => SUPPORTED.includes(p));
+if (chosen) res.setWebTransportProtocol(chosen);
+res.writeHead(200);
+```
+
+**Streams that arrive early** — opened by the client right after its CONNECT, before
+the application accepts the session, or even read before the CONNECT itself — are
+held (one stream window each, up to 64 per connection) and delivered as
+`'stream'` / `'unidirectionalStream'` once `res.writeHead(200)` accepts it. If the
+application answers otherwise, or the CONNECT never comes, they are reset with
+`WEBTRANSPORT_BUFFERED_STREAM_REJECTED`.
+
+**The session** ends with `'close'` on `req`. `res.closeSession(code, reason)` closes it with
+CLOSE_WEBTRANSPORT_SESSION — a 32-bit code and a reason of up to 1024 bytes, what a
+browser's `wt.closed` resolves with — and resets its streams. A client's close arrives the
+same way: `req.closeInfo` is `{ closeCode, reason }` when it sent one. `res.drainSession()`
+sends DRAIN_WEBTRANSPORT_SESSION (finish up, start nothing new); a client's DRAIN is
+`'draining'` on `req`.
+
 #### Client (Node.js)
 
+QUICO's `WebTransport` is the browser's API — the same class, promises, streams and
+options — so a client written for the browser runs in Node unchanged:
+
 ```js
-import { WebTransport } from 'quico';
+import { WebTransport, WebTransportError } from 'quico';   // in a browser: globals
 
-const wt = new WebTransport('https://localhost:4433/echo');
+const wt = new WebTransport('https://example.com:4433/live', {
+  serverCertificateHashes: [{ algorithm: 'sha-256', value: certHash }],   // or a CA-signed cert
+});
 await wt.ready;
 
-const stream = await wt.createBidirectionalStream();
-stream.write('hello');
-stream.on('data', (chunk) => console.log('echo:', chunk.toString()));
-stream.end();
+const { readable, writable } = await wt.createBidirectionalStream();
+const writer = writable.getWriter();
+await writer.write(new TextEncoder().encode('hello'));
+await writer.close();
+for (const reader = readable.getReader();;) { const { value, done } = await reader.read(); if (done) break; /* … */ }
 
-const uni = await wt.createUnidirectionalStream();
-uni.write('fire and forget');
-uni.end();
+const uni = await wt.createUnidirectionalStream();             // a WritableStream
+await uni.getWriter().abort(new WebTransportError({ streamErrorCode: 7 }));   // RESET_STREAM, code 7
 
-wt.sendDatagram(Buffer.from('ping'));
-wt.on('datagram', (data) => console.log('pong:', data.toString()));
+const incoming = wt.incomingUnidirectionalStreams.getReader(); // ReadableStream of ReadableStreams
+const dgWriter = wt.datagrams.writable.getWriter();            // and wt.datagrams.readable
 
-wt.close();
+wt.close({ closeCode: 0, reason: 'done' });
+const { closeCode, reason } = await wt.closed;                 // the server's, if it closed first
 ```
 
+The whole browser API is there: `ready`, `closed` (resolving with `{ closeCode, reason }`),
+`draining`, `close()`, `createBidirectionalStream()` / `createUnidirectionalStream()` with
+`sendOrder`, `sendGroup` and `waitUntilAvailable`, `incomingBidirectionalStreams`,
+`incomingUnidirectionalStreams`, `datagrams` (`readable`, `writable`, `createWritable()`,
+`maxDatagramSize`, `incomingMaxAge`, `outgoingMaxAge`, `incomingHighWaterMark`,
+`outgoingHighWaterMark`), `createSendGroup()`, `getStats()`, `reliability`,
+`congestionControl`, `protocol`, and stream `getStats()`; the options
+`serverCertificateHashes` (the leaf's SHA-256, a certificate valid for at most 14 days, as
+the browser requires), `protocols`, `requireUnreliable`, `congestionControl`,
+`transportParams` (QUICO's own: the windows and stream limits, as the server's),
+`key` / `cert` / `passphrase` (a client certificate, as `https.request` takes it),
+`quicVersions` and `ecdhCurve` (as `quico.request`'s),
+`session` (a TLS session from an earlier `'session'` event, to resume — not with
+`serverCertificateHashes`: a resumed handshake has no certificate to pin),
+`initialCongestionWindow` (as the server's),
+`allowPooling` and the `anticipatedConcurrent…` hints. `WebTransportError` takes the browser's
+`{ streamErrorCode }` form as well as the spec's `(message, options)`. A test runs the same
+client function in Chromium and in Node against a QUICO server and checks both see the same.
+`sendOrder` and `sendGroup` order what the client sends, as on the server (see above).
+
+The same object also has a Node-style API, for Node code:
+
+| | |
+|---|---|
+| `wt.on('stream', duplex)`, `wt.on('unidirectionalStream', readable)` | The server's streams as Node streams. A stream goes to a Node listener if there is one, otherwise to `incoming…Streams` once that is read, and waits for either. |
+| `wt.on('datagram', buf)`, `wt.sendDatagram(buf[, { sendOrder, sendGroup }])` → `true`/`false`, `wt.maxDatagramSize()` | Datagrams; with a `sendOrder` / `sendGroup` one takes its place among the streams, as the server's `res.sendDatagram`. |
+| `'ready'`, `'draining'`, `'close'` | The session's events. |
+| `'session'` (`Buffer`) | A TLS session the server sent, to pass back as `session` next time and resume. |
+| `stream.write()`, `stream.on('data')`, `stream.pipe()` | Every stream is a Node stream too: `createBidirectionalStream()` gives a `Duplex` that has the browser's `readable` / `writable`; `createUnidirectionalStream()` gives the browser's `WritableStream` that is Node's `Writable` too — `write()` / `'drain'`, `end()`, `resetStream()`, its events, `writableLength` and the rest of its state, and `readable.pipe()` into it. |
+| `stream.resetStream(code)`, `stream.stopSending(code)`, `stream.streamErrorCode` | Give up on a direction with a code; the peer's code when it gave up. |
+| `stream.id`, `stream.sessionId`, `stream.sendState`, `stream.recvState`, `stream.pending` | As on the server. `create…Stream()`'s Promise is settled once the stream is open — at once, or with `waitUntilAvailable` when the server allows it — as the browser's is. |
+| `wt.canOpenStream({ bidirectional })`, `wt.getStats()` | As the server's `res.canOpenStream()` and `res.getStats()`. |
+| `'peerReset'` / `'peerStopSending'` (`code`), `'close'` | As on the server: the direction the peer gave up on ends, the other goes on; `'close'` once both are over. |
+| `stream.sendOrder`, `stream.sendGroup` | Its place among the session's streams (see [Priorities](#webtransport-1) on the server — the same rules). |
+
+The client's streams are the server's kind: the server's are recognised by their
+WebTransport prefix, not guessed from stream numbers, and what the server sends is credited
+back as the application reads — a stream nobody reads holds one stream window.
+
+Besides the browser's options, the second argument takes `ca` and `rejectUnauthorized`
+(default `true`), as in `https.request`: the server's certificate is verified against the
+platform roots unless `ca` is given or `serverCertificateHashes` pins it. If the session
+cannot be established, `ready` and `closed` reject with a `WebTransportError` (source
+`'session'`, keeping node:tls's `err.code`, e.g. `DEPTH_ZERO_SELF_SIGNED_CERT`, when TLS is
+the reason), `state` becomes `'failed'` and the connection is closed, as in the browser.
+`'error'` is emitted too, but only when a listener is attached.
+
 #### Client (Browser)
 
 A QUICO server works with the **native browser WebTransport API** — no client library
@@ -805,6 +1974,257 @@ console.log('Received:', new TextDecoder().decode(value));
 
 ---
 
+### WebSocket
+
+#### Server
+
+One handler serves WebSocket over HTTP/1.1, HTTP/2 (RFC 8441) and HTTP/3 (RFC 9220):
+all three arrive as the server's `'upgrade'` event.
+
+```js
+import quico, { WebSocketServer } from 'quico';
+
+const server = quico.createServer({ key: KEY, cert: CERT });
+const wss = new WebSocketServer({ server, path: '/chat', pingInterval: 30000 });
+
+wss.on('connection', (ws, req) => {
+  ws.on('message', (data, isBinary) => ws.send(data, { binary: isBinary }));
+});
+
+server.listen(4433);
+```
+
+`WebSocketServer(options)` takes `server` (or `noServer: true` and call
+`wss.handleUpgrade(req, socket, head, cb)` yourself), `path`, `pingInterval` (clients
+that miss a pong are dropped) and the options of `upgradeWebSocket`:
+
+- `origin` — a string, an array, or `fn(origin, req)`; a mismatch is answered with 403
+- `verifyClient(req)` — return `false` (401) or `{ code, message, headers }`
+- `protocol` (string or `fn(offered, req)`) / `protocols` (preference list)
+- `headers` — extra response headers
+- `maxPayload` — default 100 MB
+
+`wss.clients` is the set of open connections; `wss.close([code][, reason])` closes them.
+For a single upgrade without a server object, `upgradeWebSocket(req, socket, head,
+options)` returns the socket, or `null` after answering the client itself.
+
+Sockets have both surfaces: WHATWG (`onmessage`, `addEventListener`, `binaryType`) and
+the `ws` package's (`on('message', (data, isBinary))`, `on('close', (code, reason))`,
+`ping()`, `on('pong')`). Pings are answered automatically.
+
+#### Client
+
+```js
+import { WebSocket } from 'quico';
+
+const ws = new WebSocket('wss://localhost:4433/chat', [], { rejectUnauthorized: false });
+ws.onopen = () => ws.send('hello');
+ws.onmessage = (e) => console.log(e.data);
+```
+
+The client connects over HTTP/1.1. Options beyond the WHATWG constructor: `origin`,
+`headers`, `ca`, `rejectUnauthorized`, `maxPayload`, `agent` (whose TLS sessions a
+`wss://` connection resumes with and keeps — `globalAgent` by default) and `session` (one
+to resume with, from an earlier `'session'`). `ws.on('session', buf)`: each one the server
+sends.
+
+---
+
+### CONNECT-UDP (proxying UDP)
+
+A QUICO server can be a UDP proxy (RFC 9298, the MASQUE working group's): a client
+asks it for a UDP "socket" to one target, a host and a port, and from then on sends
+the target UDP payloads through it and receives the target's. This is what a browser's
+or an OS's privacy relay does for QUIC traffic.
+
+```js
+const server = quico.createServer({ key: KEY, cert: CERT });
+
+server.on('connectUdp', (req, tunnel) => {
+  if (req.headers.authorization !== 'Bearer ' + TOKEN) return tunnel.reject(407);
+  tunnel.accept();   // to tunnel.target, from the request
+  tunnel.on('close', (stats) => console.log(tunnel.target, stats.bytesToClient, tunnel.closeReason));
+});
+
+server.listen(443);
+```
+
+**How it travels.** The request names the target in its path, by a URI template the
+clients are configured with: by default `/.well-known/masque/udp/{target_host}/{target_port}/`
+(an IPv6 address with its colons percent-encoded). Over HTTP/3 it is an extended CONNECT
+(`:protocol: connect-udp`), and each UDP payload is an HTTP Datagram (RFC 9297) in a QUIC
+DATAGRAM frame — unreliable, as UDP is. A client that did not negotiate datagrams,
+HTTP/2 (extended CONNECT) and HTTP/1.1 (`Upgrade: connect-udp`, answered 101) carry them
+as DATAGRAM capsules on the request stream. The proxy relays from a UDP socket of its own,
+connected to the target, so only the target's packets come back through it. A payload
+larger than a QUIC datagram can carry is dropped, not moved to the stream (§5: a
+reliable, ordered copy of a UDP packet is not one).
+
+**What the application decides** — in `'connectUdp'`, now or later (the client may
+send before the answer; the first 32 payloads wait):
+
+| | |
+|---|---|
+| `tunnel.accept()` | Relay to `tunnel.target`. The host is resolved (`lookup`), the socket connected, then the 2xx goes. |
+| `tunnel.accept({ target: { host, port } })` | Relay to another target. |
+| `tunnel.accept({ socket })` | Relay through the application's own `dgram` socket (bound where it wants, say). Packets from senders other than the target are not relayed; the socket is not closed with the tunnel. |
+| `tunnel.accept({ relay: false })` | No relay: the application gets each payload from the client as `'message'` (a `Buffer`) and answers with `tunnel.send(buf)`. |
+| `tunnel.accept({ headers, allowPrivate })` | Extra response fields; the private-address rule for this tunnel. |
+| `tunnel.reject(status[, headers])` | Refuse it. |
+| `tunnel.send(buf)` | A payload to the client, in any mode. `false`: not sent (too large, or the stream is full — dropped, as UDP drops). |
+| `tunnel.close()` | End it (the request stream is ended). |
+
+`tunnel.req` is the request (an `IncomingMessage`), `tunnel.target` `{ host, port }`,
+`tunnel.state` `'pending'` / `'open'` / `'closed'`. Events: `'open'`, `'message'` (with
+`relay: false`), `'close'` (with the stats; `tunnel.closeReason` says why: `'client'`,
+`'local'`, `'idle'`, `'aborted'` — the stream was reset or the connection went —,
+`'socket'` — the target's socket failed, an ICMP unreachable, say — or `'rejected'`;
+`tunnel.error` holds the error when there was one).
+
+**Figures.** `tunnel.stats`: `{ target, address, port, httpVersion, state, transport
+('datagram' | 'capsule'), openedAt, closedAt, durationMs, packetsFromClient,
+bytesFromClient, packetsToClient, bytesToClient, droppedFromClient, droppedToClient }`.
+`server.getConnectUdpStats()`: `{ active, accepted, rejected, closed }` and the same
+counters summed over every tunnel, open ones included.
+
+**Refused on its own** — before `'connectUdp'`, or at `accept()`:
+
+| Case | Answer |
+|---|---|
+| A path that is not the template's | 404 |
+| An empty host, a port outside 1..65535 | 400 |
+| HTTP/1.1, not a GET | 400 |
+| More tunnels on one connection than `maxTunnelsPerConnection` | 429 |
+| The host does not resolve | 502, `Proxy-Status: quico; error=dns_error` |
+| The host resolves only to private addresses | 403, `Proxy-Status: quico; error=destination_ip_prohibited` |
+| No `'connectUdp'` listener | Not a tunnel: the request goes to `'upgrade'` or `'request'`, as before |
+
+**Options** (`createServer({ connectUdp: { … } })`):
+
+| Option | Default | |
+|---|---|---|
+| `template` | `/.well-known/masque/udp/{target_host}/{target_port}/` | The URI template — variables in the path, or a query (`/masque{?target_host,target_port}`). |
+| `allowPrivate` | `false` | Allow targets in private, loopback, link-local, CGNAT, multicast and reserved ranges (IPv4 and IPv6, IPv4-mapped included). Off, a proxy cannot be used to reach its own network (RFC 9298 §7). |
+| `maxTunnelsPerConnection` | `16` | |
+| `idleTimeout` | `120000` | ms with nothing either way before a tunnel is closed; never under two minutes (RFC 9298 §3.1), `0` for none. |
+| `lookup` | `dns.lookup` | `(host, { all: true }, cb)`, as `dns.lookup`. |
+| `proxyName` | `'quico'` | The name in `Proxy-Status`. |
+
+RFC 9298 asks that a UDP proxy serve authenticated users only — the check is the
+application's, in `'connectUdp'`.
+
+#### Client: `quico.connectUdp(proxy, target[, options][, callback])`
+
+A UDP socket to `target` (`{ host, port }`) through a proxy. What it returns behaves as a
+connected `dgram.Socket`, so code written for `dgram` moves to a proxied path by changing
+the line that makes the socket:
+
+```js
+import { connectUdp } from 'quico';
+
+const sock = connectUdp('https://proxy.example', { host: '1.1.1.1', port: 53 },
+                        { headers: { 'proxy-authorization': 'Bearer ' + TOKEN } });
+sock.on('connect', () => sock.send(dnsQuery));
+sock.on('message', (msg, rinfo) => { /* rinfo: the target, as dgram gives it */ });
+sock.on('error', console.error);
+```
+
+`proxy` is the proxy's URL — and, if it contains `{target_host}`, the URI template too
+(`'https://proxy.example/masque{?target_host,target_port}'`); without one, the default
+template's path — or `{ hostname, port, template }`.
+
+HTTP/3 first, each payload a QUIC datagram. If QUIC has not connected within `h3Timeout`
+(UDP blocked on the way, a proxy without HTTP/3), TCP: HTTP/2 extended CONNECT or HTTP/1.1
+Upgrade, as ALPN decides, the payloads in capsules.
+
+Over HTTP/3 the tunnels to one proxy share a connection (RFC 9298's tunnels are requests;
+their datagrams are told apart by Quarter Stream ID): a tunnel joins the connection opened
+with the same TLS settings, or waits for the one being made — no handshake of its own, one
+congestion controller, one UDP port — and the connection closes with its last tunnel. A
+tunnel the proxy refuses, or one closed, leaves the others as they are; a connection lost
+ends each tunnel on it with `ERR_CONNECT_UDP_CLOSED`. `shareConnection: false` gives a
+tunnel its own. Over TCP each tunnel has its own connection.
+
+| | |
+|---|---|
+| `send(msg[, offset, length][, callback])` | A connected socket's forms: a `Buffer`, a typed array, a string, or an array of them (one payload). Before `'connect'`, up to 64 are held and go when the tunnel opens. Larger than `maxPayloadSize`: `EMSGSIZE`. When the stream already holds too much (capsules), dropped as UDP drops — no error. With a port and address: `ERR_SOCKET_DGRAM_IS_CONNECTED`. |
+| `close([callback])` | Ends the tunnel: the request stream's end, and the connection's when it was the last on it. |
+| `remoteAddress()` | The target `{ address, family, port }`. |
+| `address()` | The local end of the connection to the proxy. |
+| `maxPayloadSize` | The largest payload `send()` takes now: over HTTP/3, what a QUIC datagram carries on this path (~1195 bytes at the start; more after path MTU discovery); over capsules, 65527. |
+| `protocol` | `'h3'`, `'h2'` or `'http/1.1'`, once open. |
+| `statusCode`, `responseHeaders` | The proxy's answer. |
+| `closeReason` | `'local'`, `'proxy'` (it ended the tunnel) or `'error'`. |
+| `stats` | `{ protocol, transport ('datagram' \| 'capsule'), target, openedAt, closedAt, durationMs, packetsSent, bytesSent, packetsReceived, bytesReceived, dropped }` |
+
+TLS sessions to the proxy are kept and resumed with as `quico.request`'s are: the
+`agent` option's (`globalAgent` by default; `false`: none), each used once, or `session`;
+each the proxy sends is the socket's `'session'` event.
+
+Events: `'connect'`, `'message'` `(msg, rinfo)`, `'error'`, `'close'`, `'session'`. A refusal is an
+`'error'` with `code: 'ERR_CONNECT_UDP_REFUSED'`, `statusCode` and `proxyStatus` (the
+proxy's `Proxy-Status`), then `'close'`.
+
+| Option | Default | |
+|---|---|---|
+| `protocol` | `'auto'` | `'h3'` (no fallback), `'h2'` or `'http/1.1'` to use only that. |
+| `h3Timeout` | `3000` | ms to wait for QUIC before TCP. |
+| `fallback` | `true` | `false`: HTTP/3 only, even with `'auto'`. |
+| `shareConnection` | `true` | `false`: an HTTP/3 connection of its own, not shared with the proxy's other tunnels. |
+| `headers` | — | Fields for the request (authorization). |
+| `ech`, `echGrease` | | Encrypted Client Hello to the proxy, as `quico.request`'s. |
+| `ecdhCurve` | | The key-exchange groups to offer the proxy, as `quico.request`'s. |
+| `template` | the default | When the proxy URL carries none. |
+| `rejectUnauthorized`, `ca`, `key`, `cert`, `passphrase`, `servername`, `quicVersions` | | As `quico.request`. |
+
+Tested against quic-go's masque-go in both directions — its client through QUICO's proxy
+(an HTTP/3 request inside the tunnel: QUIC in QUIC), QUICO's client through its proxy.
+
+---
+
+### Fingerprinting
+
+Who is really on the other end? A browser and a script sending the same User-Agent
+still differ in how they speak the protocols underneath. QUICO exposes those
+differences, on every layer it implements.
+
+| Function | Input | Returns |
+|---|---|---|
+| `ja4h(req)` | a request | JA4H: `{ hash, raw, a, b, c, d }` — method, version, header names and order, cookie names (never values) |
+| `req.h2Fingerprint()` | HTTP/2 request | `{ hash, raw, … }` from the client's SETTINGS, WINDOW_UPDATE, PRIORITY and pseudo-header order |
+| `req.h3Fingerprint()` | HTTP/3 request | `{ hash, raw, … }` from the client's QUIC transport parameters, HTTP/3 SETTINGS and pseudo-header order |
+| `h2Fingerprint(observed)` / `h3Fingerprint(observed)` | raw observations | the same, computed from your own data |
+| `req.socket.getJA4()` / `getJA4S()` / `getJA5()` | TLS | TLS fingerprints, from LemonTLS — `null` when the installed LemonTLS does not provide them |
+| `req.socket.getNegotiationResult()` / `res.socket.getNegotiationResult()` | TLS | What the handshake settled, on every version: `{ versionName, cipherName, groupName, alpn, resumed, helloRetried, certificateCompression: { sent, received }, … }` — `sent` and `received` the algorithm each side's certificate went in (`'brotli'`, `'zlib'`, `'zstd'`), `null` where it went plain or not at all (a resumed connection) |
+| `profileRequest(req)` | a request | `{ claimed, protocol, fingerprints: { ja4, ja4s, h2, h2raw, h3, h3raw }, signals, suspicion }` — signals where the request's TLS and HTTP/2 behaviour contradicts the client it claims to be, and a 0–100 score. HTTP/3 fingerprints are reported, not yet checked against browsers. |
+
+```js
+server.on('request', (req, res) => {
+  const p = quico.profileRequest(req);
+  if (p.suspicion >= 60) { res.writeHead(403); return res.end(); }
+  // …
+});
+```
+
+---
+
+### Protocol layers
+
+The unified server and client are built from layers you can use on their own — for a
+custom transport, a proxy, or protocols other than HTTP.
+
+| Export | What it is |
+|---|---|
+| `createH1Server(options, handler)` | HTTP/1.1 server on its own |
+| `createH2Server(options, handler)` | HTTP/2 server on its own |
+| `createH3Server(options, handler)` | HTTP/3 + WebTransport server on its own (over `createQuicServer`) |
+| `h1Request(options, cb)` / `h2Request(options, cb)` / `h3Request(options, cb)` | Single-protocol clients, no fallback. Each takes `session` and emits `'session'`, as `https.request`. `h2Request` streams its body: `write()` sends as it is called and returns `false` while the stream's window or the socket is full, then `'drain'` — an upload is never held whole |
+| `createTcpListener(options)` | TCP accept + TLS + ALPN dispatch: `register(alpn, handler, { sniff })`, `listen()` |
+| `createQuicServer(options)` | QUIC transport server — see below |
+| `createQuicClientSocket(opts)` | A UDP socket driving one client `QUICConnection` (`remoteIp`, `remotePort`, `hostname`, `alpn`, `ecdhCurve`, `onConnect(quic)`, `onSocket(quic)` before the handshake — the place to attach `'keylog'`) |
+
+---
+
 ### Low-level QUIC API
 
 For non-HTTP protocols over QUIC, or full control of the connection.
@@ -817,9 +2237,23 @@ For non-HTTP protocols over QUIC, or full control of the connection.
 | `SNICallback` | `Function` | `null` | Per-hostname contexts |
 | `alpn` | `string[]` | `['h3']` | ALPN protocols to accept |
 | `maxConnections` | `number` | `10000` | Connection cap |
+| `ticketKeys` | `Buffer` | random | Session-ticket key |
+| `sessionTimeout` | `number` | `7200` | Seconds a session ticket stays good |
+| `requestCert`, `rejectUnauthorized`, `ciphers`, `ecdhCurve`, `sigalgs`, `handshakeTimeout`, `passphrase`, `certificateCompression` | | | As `createServer` (TLS 1.3) |
+| `initialCongestionWindow` | `number` | `20` | Initial congestion window, in packets |
+| `carefulResume` | `boolean` | `false` | Careful Resume (RFC 9959) for returning clients |
+| `startupCap` | `string` | `'stall'` | Startup's window cap: `'stall'`, `'always'`, `'off'` (as `createServer`) |
+| `quicVersions` | `Array` | `['v1', 'v2']` | QUIC versions accepted, in preference (RFC 9368, RFC 9369) |
+| `connectionOptions` | `Function` | — | Per-connection settings (see above) |
+| `validateAddress` | `boolean` | `false` | Address validation with Retry |
+| `earlyData`, `earlyDataMaxAge`, `earlyDataReplay` | | `false`, `10000`, in memory | 0-RTT, as `createServer` (`quic.isEarlyStream(id)`: whether a stream came early) |
+| `handshakeLoadThreshold`, `maxHandshakesPerAddress` | `number` | `500`, `64` | As `createServer`; `getPendingHandshakes()`, `setOptions()` / `getOptions()` too |
+| `pmtuDiscovery`, `maxUdpPayloadSize`, `initialRtt`, `reorderWindow` | | | As `createServer` |
+| `resetTokenSecret` | `Buffer` | random | Stateless-reset secret |
 | `socket` / `socket6` | `dgram.Socket` | `null` | External, already-bound UDP sockets |
 
-Returns `{ listen, close, on, handlePacket, hasConnection }`. The `'connection'`
+Returns `{ listen, close, on, handlePacket, hasConnection }` plus the runtime getters and
+setters for ticket keys, address validation and the reset secret. The `'connection'`
 event fires with `(quicConnection, peer)`.
 
 ```js
@@ -849,22 +2283,33 @@ and you route datagrams yourself.
 | Member | Description |
 |---|---|
 | `quic.state` | `'idle'`, `'handshaking'`, `'connected'`, `'closing'`, `'draining'`, `'closed'` |
-| `quic.sendStream(id, data, fin)` | Write to a stream |
-| `quic.sendDatagram(data)` | Send an unreliable datagram |
-| `quic.maxDatagramSize()` | Largest datagram payload that fits the current path |
-| `quic.close(errorCode, reason)` | Send CONNECTION_CLOSE |
+| `quic.sendStream(id, data, fin)` | Write to a stream. A new stream beyond the peer's `MAX_STREAMS` waits until the peer raises the limit. |
+| `quic.whenStreamCommitted(id, cb)` | `cb()` once every byte and the FIN of the stream are acknowledged ("Data Recvd"); never if it is reset first |
+| `quic.resetStream(id, code[, reliableSize])` | Abandon the sending part of a stream (RESET_STREAM); with `reliableSize`, its first bytes still delivered (RESET_STREAM_AT, when the peer takes it: `quic.resetStreamAtSupported`) |
+| `quic.stopSending(id, code)` | Ask the peer to stop sending on a stream (STOP_SENDING) |
+| `quic.sendDatagram(data[, prio])` | Send an unreliable datagram; `false` if not connected or too large. With `prio` (`{ urgency, group, order }`) it is scheduled among the streams instead of ahead of them |
+| `quic.setStreamPriority(id, { urgency, incremental, group, order }[, fromPeer])` | A stream's place in the scheduler — see [Stream scheduling](#stream-scheduling-byte-level-round-robin-by-priority). `fromPeer`: the peer's signal, which does not replace what the application set |
+| `quic.getStreamPriority(id)` | What is set for a stream, or `null` |
+| `quic.canOpenStream(bidi)` | Whether a stream of ours could be opened now, within the peer's `MAX_STREAMS` |
+| `quic.isReceiving(id)` | Whether a stream's receiving part is open (opened, and neither finished nor reset) |
+| `quic.maxDatagramSize()` | Largest datagram payload that fits one packet (a local estimate) |
+| `quic.close(errorCode, reason[, application])` | Send CONNECTION_CLOSE; `application: true` for application error codes |
+| `quic.remoteAddress` / `remotePort` | The peer's current address — follows NAT rebinding |
+| `quic.earlyDataStatus` | 0-RTT: `null` (none), `'offered'` (a client's, undecided), `'accepted'`, `'rejected'` |
+| `quic.isEarlyStream(id)` | Server: whether the stream was opened by 0-RTT data |
+| `quic.whenConfirmed(cb)` | `cb()` once the handshake is confirmed (RFC 9001 §4.1.2), at once if it is |
+| `quic.alpnProtocol`, `servername`, `getCertificate()`, `getTicketInfo()`, `isSessionReused()` | What the handshake decided, with `TLSSocket`'s names |
 | `quic.on(event, fn)` / `off` | Event subscription |
 
-**Events:** `'connect'`, `'stream'` (`id, data, fin`), `'datagram'`, `'close'`,
-`'error'`, `'keylog'`.
+**Events:** `'connect'`, `'stream'` (`id, data, fin`), `'stream_reset'` (`id, code,
+finalSize`), `'stop_sending'` (`id, code`), `'datagram'`, `'path'` (`address, port`),
+`'peer_transport_params'`, `'close'`, `'error'`, `'keylog'`; a client's 0-RTT: `'early'` (it may
+send application data now), `'early_accepted'`, `'early_rejected'` (what went early is forgotten — send it again).
 
 **Connection options** (when constructing directly): `idleTimeout` (default `30000`),
 `handshakeTimeout` (default `10000`), `keepAlive` (`true` → half the idle timeout, or
 a number of ms), `alpn`, `hostname`, `rejectUnauthorized`, `ca`.
 
----
-
-
 
 ---
 
@@ -872,35 +2317,49 @@ a number of ms), `alpn`, `hostname`, `rejectUnauthorized`, `ca`.
 
 ### Transport parameters
 
-These are the values QUICO advertises to peers. They are currently compiled-in
-defaults; making them configurable is on the roadmap.
+These are the values QUICO advertises to peers. The windows, stream limits and idle timeout can be
+set — `createServer({ transportParams })`, `new WebTransport(url, { transportParams })`
+— by RFC 9000's names in camelCase; a value out of range throws when the server or
+client is made.
 
-| Parameter | Default | Meaning |
-|---|---|---|
-| `initial_max_data` | 1 MB | Connection-level receive window |
-| `initial_max_stream_data_bidi_local/remote` | 256 KB | Per-stream receive window |
-| `initial_max_stream_data_uni` | 128 KB | Per-unidirectional-stream window |
-| `initial_max_streams_bidi` | 100 | Concurrent bidirectional streams |
-| `initial_max_streams_uni` | 3 | Concurrent unidirectional streams |
-| `max_idle_timeout` | 30 s | Idle before the connection is dropped |
-| `max_ack_delay` | 25 ms | Maximum ACK delay we advertise |
-| `max_datagram_frame_size` | 65527 | Datagram support (RFC 9221) |
-| `active_connection_id_limit` | 4 | Connection IDs we will track |
-
-Receive windows slide automatically as the application consumes data, so large
-downloads are not limited by the initial values.
+| Parameter | Option | Default | Meaning |
+|---|---|---|---|
+| `initial_max_data` | `initialMaxData` | 1 MB | Connection-level receive window |
+| `initial_max_stream_data_bidi_local` | `initialMaxStreamDataBidiLocal` | 256 KB | Receive window of a bidirectional stream we open |
+| `initial_max_stream_data_bidi_remote` | `initialMaxStreamDataBidiRemote` | 256 KB | Receive window of one the peer opens |
+| `initial_max_stream_data_uni` | `initialMaxStreamDataUni` | 128 KB | Receive window of a unidirectional stream the peer opens |
+| `initial_max_streams_bidi` | `initialMaxStreamsBidi` | 100 | Bidirectional streams the peer may have open at once |
+| `initial_max_streams_uni` | `initialMaxStreamsUni` | 100 | Unidirectional streams the peer may have open at once (HTTP/3 keeps three for itself; at least 3) |
+| `max_idle_timeout` | `maxIdleTimeout` | 30 s | Milliseconds without a packet from the peer before the connection is closed — silently, with no CONNECTION_CLOSE (RFC 9000 §10.1); `0`: no limit. The one in force is the **smaller** of the two sides' (RFC 9000 §10.1), and never under three PTOs; `keepAlive: true` pings at half of it |
+| `disable_active_migration` | `disableActiveMigration` | not sent | `true`: clients may not move to another address (connection migration, RFC 9000 §9). Unset, they may — a phone from Wi-Fi to cellular goes on in the same connection |
+| `max_ack_delay` | | 25 ms | Maximum ACK delay we advertise |
+| `max_datagram_frame_size` | | 65527 | Datagram support (RFC 9221) |
+| `active_connection_id_limit` | | 4 | Connection IDs we will track (and issue to the peer, up to its own) |
+| `version_information` | `quicVersions` | v1, v2 | The versions offered or chosen (RFC 9368) |
+| `grease_quic_bit` | | sent | The peer may send the fixed bit as 0 (RFC 9287); not on a shared UDP port |
+| `reset_stream_at` | | sent | RESET_STREAM_AT is taken (draft-ietf-quic-reliable-stream-reset): a peer's reset can keep its stream's first bytes delivered |
+| `min_ack_delay` | | 1 ms (16 ms on Windows) | ACK_FREQUENCY and IMMEDIATE_ACK are taken (draft-ietf-quic-ack-frequency) |
+
+Receive windows slide automatically as data is consumed, and grow (autotuning) while
+the application keeps up, so large downloads are not limited by the initial values. Stream limits work the same way: as
+the peer's streams close, `MAX_STREAMS` gives them back, keeping the number open at
+once, so a long-lived connection is not capped at 100 requests.
 
 ### Per-connection timers
 
 ```js
 // via the low-level API
 new QUICConnection({
-  idleTimeout: 60000,      // default 30000
+  idleTimeout: 60000,      // default 30000 — what we advertise (transportParams.maxIdleTimeout)
   handshakeTimeout: 10000, // default 10000
-  keepAlive: true          // PING at idleTimeout/2; or a number of ms
+  keepAlive: true          // PING at half the idle timeout in force; or a number of ms
 });
 ```
 
+The idle timeout in force is the smaller of ours and the peer's, as RFC 9000 §10.1 has
+it — the peer's arrives with its transport parameters; `createQuicClientSocket` takes
+`idleTimeout` and `keepAlive` too.
+
 ---
 
 ## 🐛 Debugging
@@ -911,6 +2370,7 @@ new QUICConnection({
 |---|---|
 | `QUICO_DEBUG=1` | Full per-packet trace: every datagram, packet, frame, stream write, and TLS transition. **Very verbose** — it can become the bottleneck it observes, so don't use it for throughput measurements. |
 | `QUICO_DEBUG_BBR=1` | One line per round-trip with congestion-control state: `BtlBw`, `RTprop`, `BDP`, `cwnd`, pacing rate. Cheap enough to leave on. |
+| `QUICO_TRACE_ACK=1` | One line per packet received and per ACK sent — enough to line up with a peer's log when it rejects an acknowledgement. |
 
 ```bash
 QUICO_DEBUG_BBR=1 node server.js
@@ -978,41 +2438,189 @@ quico/
 ├── interop/               — QUIC Interop Runner endpoint (Dockerfile + shim)
 ├── examples/              — Runnable examples and test suites
 └── src/
-    ├── server.js            — Unified server (drop-in for node:https)
+    ├── server.js            — Unified server: HTTP/3 + HTTP/2 + HTTP/1.1 (drop-in for node:https)
+    ├── server_stats.js      — The server's counters (server.getStats())
     ├── client.js            — Unified client with H3 → H2 → H1 fallback
     ├── agent.js             — Connection pooling, coalescing, Alt-Svc cache
     ├── streams.js           — IncomingMessage, ServerResponse, ClientRequest
+    ├── http_common.js       — Shared header and message helpers
+    │
     ├── h3.js                — HTTP/3 framing, QPACK, WebTransport detection
     ├── h3_server.js         — HTTP/3 + WebTransport server
-    ├── h3_client.js         — HTTP/3 client (DNS, UDP, QUIC, H3 pipeline)
-    ├── quic_server.js       — UDP listener, connection routing, retry/VN
+    ├── h3_client.js         — HTTP/3 client
+    ├── qpack.js             — QPACK encoder and decoder tables
+    ├── quic_version.js      — QUIC versions: v1 and v2 (RFC 9369), one profile each
+    ├── webtransport.js      — WebTransport client
+    ├── wt_stream.js         — WebTransport streams (both sides)
+    ├── connect_udp.js       — CONNECT-UDP proxy (RFC 9298): templates, capsules, the relay
+    ├── connect_udp_client.js — quico.connectUdp(): CONNECT-UDP's client
+    │
+    ├── h2.js                — HTTP/2 connection: frames, streams, flow control
+    ├── h2_frames.js         — HTTP/2 frame encoding/decoding
+    ├── h2_server.js         — HTTP/2 server
+    ├── h2_client.js         — HTTP/2 client and session pool
+    ├── hpack.js             — HPACK
+    ├── hpack_primitives.js  — Prefixed integers, string literals, Huffman (shared by HPACK and QPACK)
+    │
+    ├── h1.js                — HTTP/1.1 parser and serializer
+    ├── h1_server.js         — HTTP/1.1 server
+    ├── h1_client.js         — HTTP/1.1 client and socket pool
+    ├── tcp_listener.js      — TCP accept, TLS, ALPN dispatch and sniffing
+    │
+    ├── ws.js                — WebSocket framing, and the WHATWG and ws-style event surfaces
+    ├── ws_server.js         — WebSocketServer, upgradeWebSocket
+    ├── ws_client.js         — WebSocket client
+    │
+    ├── ja4h.js              — JA4H request fingerprint
+    ├── h2_fingerprint.js    — HTTP/2 and HTTP/3 fingerprints
+    ├── client_profile.js    — profileRequest()
+    │
+    ├── quic_server.js       — UDP listener, connection routing, Retry, stateless reset
     ├── quic_socket.js       — Client-side UDP socket + connection bootstrap
     ├── quic_connection.js   — QUIC state machine, loss detection, CC, flow control
     ├── transport.js         — Packet parsing, frame encoding/decoding
     ├── crypto.js            — AEAD, header protection, key derivation
     ├── tls_bridge.js        — LemonTLS integration, transport parameters
-    ├── utils.js             — VarInt, ACK ranges, binary helpers
-    └── webtransport.js      — WebTransport client + stream classes
+    └── utils.js             — VarInt, binary helpers, debug flags
 ```
 
 ---
 
+## 🔐 Security
+
+### Post-quantum TLS
+
+With Node 24.7+ (ML-KEM and ML-DSA in `node:crypto`, OpenSSL 3.5), on every version —
+HTTP/3, HTTP/2, HTTP/1.1 — and checked against OpenSSL 3.5 in both directions:
+
+- **Key exchange** — the hybrids X25519MLKEM768, SecP256r1MLKEM768 and SecP384r1MLKEM1024
+  (draft-ietf-tls-ecdhe-mlkem), and ML-KEM alone: MLKEM512, MLKEM768, MLKEM1024
+  (draft-ietf-tls-mlkem; MLKEM1024 is CNSA 2.0's). A server takes any of them, post-quantum
+  before classical: the client's first post-quantum group, by the client's order — a
+  HelloRetryRequest for it when the client sent no key share for it (a browser sends one
+  for X25519MLKEM768, its first: one round trip; a CNSA 2.0 client listing MLKEM1024 first
+  gets MLKEM1024). The client
+  offers X25519MLKEM768 by default, as `node:tls` and the browsers do; the others by
+  `ecdhCurve`. Each costs 1–1.7 KB in a ClientHello, so they are not offered unasked.
+- **Authentication** — ML-DSA-44, -65 and -87 certificates (FIPS 204,
+  draft-ietf-tls-mldsa), and chains signed with ML-DSA, verified and presented. A server
+  can hold an ML-DSA and a classical certificate at once — `key` and `cert` arrays, as in
+  `node:tls` — and presents the one the client's `signature_algorithms` ask for first:
+  a client that puts ML-DSA first gets ML-DSA, a browser (which offers none) gets ECDSA or
+  RSA. The client offers ML-DSA after the classical schemes (an ML-DSA chain is several
+  KB more in every handshake); `sigalgs: 'mldsa65:…'` asks for it first.
+
+  ```js
+  createServer({
+    key:  [fs.readFileSync('ecdsa.key'), fs.readFileSync('mldsa.key')],
+    cert: [fs.readFileSync('ecdsa.pem'), fs.readFileSync('mldsa.pem')],   // in any order
+    minVersion: 'TLSv1.3',      // TLS 1.2 has no post-quantum key exchange
+  }, app);
+  ```
+- **What was negotiated** — `socket.getNegotiationResult()`: `groupName`
+  (`'MLKEM1024'`, …) and `signatureAlgorithmName` (`'mldsa65'`, …).
+- **TLS 1.2** has no post-quantum key exchange; a server that still accepts it is open to
+  a client pushed onto it. `minVersion: 'TLSv1.3'` closes it (CNSA 2.0 admits TLS 1.3
+  alone). The default stays `node:tls`'s, TLS 1.2 and up; TLS 1.3's downgrade sentinel
+  (RFC 8446 §4.1.3) is sent and checked either way.
+- **The cipher suite** is the client's choice among ours, as in `node:tls` — AES-256 for a
+  client that asks for it first; `honorCipherOrder: true` for the server's.
+- An **empty `key_share`** (RFC 8446 §4.2.8 — the client asks the server to choose; how
+  scanners enumerate a server's groups) is answered with a HelloRetryRequest.
+
+QUICO parses everything a peer sends, so every buffer a peer can make it hold has a
+bound — and when a peer breaks the rules, the connection (or the stream) is closed
+with the error the RFC names, not left to grow.
+
+- **Stream limits are enforced** before any state is allocated: a stream ID beyond the
+  advertised `MAX_STREAMS` is `STREAM_LIMIT_ERROR`; a frame for a stream that cannot
+  exist is `STREAM_STATE_ERROR`; a closed stream cannot be reopened.
+- **Every way a stream ends** — FIN, RESET_STREAM, STOP_SENDING, a WebTransport
+  session closing, the connection closing — goes through one release path per layer,
+  so nothing a stream allocated outlives it.
+- **Reassembly is bounded**: gaps per stream are capped, and out-of-order data costs
+  linear, not quadratic, work.
+- **The handshake's CRYPTO data** is held to 64 KB past what was read and 128 pieces
+  ahead of a gap (`CRYPTO_BUFFER_EXCEEDED`, RFC 9000 §7.5), each byte copied once; pieces
+  sent again in other boundaries, overlapping, join as they should. A ClientHello split
+  over several Initial packets — a post-quantum one is — is joined the same way. Over TCP,
+  a handshake message declaring more than 128 KB is refused on sight (`illegal_parameter`),
+  before its bytes are kept.
+- **Transport parameters are checked** before any is used (RFC 9000 §7.4, §18.2): out of
+  range (`ack_delay_exponent` above 20, `max_ack_delay` of 2^14 or more,
+  `active_connection_id_limit` below 2, `max_udp_payload_size` below 1200, a stream
+  count above 2^60), malformed (truncated, an integer that does not fill its value, a
+  connection ID over 20 bytes, a reset token not 16 bytes), sent twice, or — from a client
+  — one only a server sends: `TRANSPORT_PARAMETER_ERROR`, on both sides.
+- **QUIC frames**: an unknown frame type, or a frame cut short, is
+  `FRAME_ENCODING_ERROR`; a packet with no frames, a frame its packet type may not
+  carry, or a DATAGRAM larger than we allowed, `PROTOCOL_VIOLATION` (§12.4, RFC 9221 §3).
+- **Connection IDs are bounded**: a peer's retired IDs are not kept, the
+  `RETIRE_CONNECTION_ID`s it has not acknowledged are capped at twice our
+  `active_connection_id_limit` (`CONNECTION_ID_LIMIT_ERROR`, §5.1.2), as are its active IDs; a
+  server issues at most 32 IDs per connection, and every route to a connection goes with it.
+- **HTTP/3 and QPACK buffers are bounded**: a HEADERS frame larger than the field
+  section size we advertise is refused before it is buffered; unknown frames are
+  skipped as they arrive; a QPACK capacity or literal larger than we advertised is
+  `QPACK_ENCODER_STREAM_ERROR`; bytes already parsed are dropped, not re-scanned.
+- **0-RTT is off by default**, and on, held to one use per ticket, a freshness window,
+  the same transport parameters, and safe requests only — see [0-RTT](#0-rtt-early-data).
+- **HTTP/3 framing rules** are enforced: DATA before HEADERS is
+  `H3_FRAME_UNEXPECTED`, a request stream without HEADERS is `H3_REQUEST_INCOMPLETE`;
+  a control-stream frame on a request stream, or one of HTTP/2's reserved frame types
+  anywhere, is `H3_FRAME_UNEXPECTED`; a payload that is not exactly its fields,
+  `H3_FRAME_ERROR`.
+- **HTTP/3's direction rules** (RFC 9114 §4.6, §5.2, §7.2): neither side takes part in
+  server push, and each holds the other to that — at the server, `PUSH_PROMISE` is
+  `H3_FRAME_UNEXPECTED`, a push stream `H3_STREAM_CREATION_ERROR`, `CANCEL_PUSH`
+  `H3_ID_ERROR`; at the client, `MAX_PUSH_ID` is `H3_FRAME_UNEXPECTED`, a push stream,
+  `PUSH_PROMISE` or `CANCEL_PUSH` `H3_ID_ERROR`. `MAX_PUSH_ID` that goes down, `GOAWAY`
+  that goes up, or a server's `GOAWAY` naming anything but a request stream:
+  `H3_ID_ERROR`. `SETTINGS` with one of HTTP/2's identifiers, one twice, or
+  `SETTINGS_H3_DATAGRAM` / `SETTINGS_ENABLE_CONNECT_PROTOCOL` other than 0 or 1:
+  `H3_SETTINGS_ERROR`.
+- **HTTP/1.1 line endings**: a request must end its lines with CRLF — a bare LF is `400`,
+  since a server that took one where a hop in front did not would read another message
+  (request smuggling). A response may end them with LF alone, as RFC 9112 §2.2 allows and
+  browsers and curl accept: the client is its last reader, with no one behind it to
+  disagree. A bare CR is refused either way.
+- **Key shares** (RFC 8446 §4.2.8): one of the wrong length, an ML-KEM key that is not
+  one, a share for a group the client did not list, two for one group, or `key_share`
+  without `supported_groups` — a fatal alert, never a handshake in a group the client did
+  not agree to.
+- **Our own failures are ours**: an exception inside the HTTP/3 code closes with
+  `H3_INTERNAL_ERROR`, not a code that blames the peer; one inside the QUIC packet path
+  drops that packet and is counted (`quic.context.internal_errors`).
+- **Fuzzed**: every parser in front of a peer's bytes — QUIC frames, transport
+  parameters, packet headers, HTTP/1.1, WebSocket, HPACK and HTTP/2 connections, QPACK,
+  capsules — and HTTP/3 and QUIC frames on live connections, with mutated valid input and
+  noise; none ends in an exception of its own, a hang, or an internal-error close.
+- **Closing says so**: `server.close()` sends every connection its `CONNECTION_CLOSE`
+  and closes the socket only once those have left, so clients learn at once rather than
+  by their idle timeout.
+
+To report a vulnerability, please see [SECURITY.md](./SECURITY.md) — or write to
+**support@quicojs.dev** — rather than opening a public issue.
+
+---
+
 ## ⚠️ Known Limitations
 
 Being explicit about what is **not** implemented yet:
 
 | Area | Status |
 |---|---|
-| **0-RTT / session resumption** | Not implemented — every connection is a full handshake. |
-| **ChaCha20-Poly1305** | Not implemented. AES-128-GCM and AES-256-GCM only. On hardware without AES acceleration this costs performance. |
-| **Connection migration** | Not implemented; QUICO advertises `disable_active_migration`. Connections do not survive a client IP change. |
+| **0-RTT** | HTTP/3, HTTP/2 and HTTP/1.1 (TLS 1.3); not DTLS. The client sends early only `GET`, `HEAD` and `OPTIONS` requests. Over HTTP/3, what it sends early assumes the default SETTINGS (no QPACK dynamic table), not the ones it was given last time — the safe choice, at the cost of a few header bytes. |
+| **ECH** | Server and client (see [Encrypted Client Hello](#encrypted-client-hello-ech)). The client takes the key list from the application — it does not resolve HTTPS records itself. TLS only, not DTLS. |
+| **Client-initiated migration** | The server takes a client that moves; the QUICO client itself has no API to move to another local address. |
 | **ECN** | Not implemented — Node's `dgram` does not expose the required socket options. |
-| **Retry / address validation** | Not implemented (client does not consume Retry tokens; server does not issue them). |
-| **Anti-amplification limit** | Not enforced (RFC 9000 §8.1). |
-| **Stream priority** (RFC 9218) | Not implemented. |
-| **Delayed ACK** | Not implemented — an ACK is sent per ack-eliciting packet. Correct, but more packets than necessary. |
-| **`MAX_STREAMS` issuance** | Streams limits are honoured but not proactively raised, capping very-high-concurrency workloads. |
-| **TypeScript types** | Not shipped yet. |
+| **Path MTU discovery** | DPLPMTUD up to 1452-byte datagrams (see [Path MTU discovery](#path-mtu-discovery)). Its probes rely on the Don't Fragment bit, which Node's `dgram` cannot set: Linux sets it on UDP by default, Windows does not — there a probe can cross fragmented and be taken for a fit. |
+| **Request bodies (`quico.request`)** | Streamed; only the first 64 KB are kept for sending again, so a larger body that started over one protocol cannot move to another. |
+| **Server push** | Not implemented, on either side: the server sends no `PUSH_PROMISE`, the client allows none (it sends no `MAX_PUSH_ID`). Browsers have dropped it too. |
+| **Post-quantum TLS** | Needs ML-KEM and ML-DSA in `node:crypto` (Node 24.7+, OpenSSL 3.5). Before that, classical only — `node:tls` there offers the hybrid through OpenSSL, which lemon-tls cannot reach. The hybrid ClientHello is ~1.4 KB: two QUIC Initial packets rather than one. Public CAs do not issue ML-DSA certificates yet. |
+| **WebSocket** | No permessage-deflate. The client connects over HTTP/1.1 only. |
+| **CONNECT-UDP** | Context ID 0 only (UDP payloads) — no CONNECT-IP, no extensions that register contexts. The client opens a connection per socket (no pooling of tunnels on one connection). |
+| **Other transport parameters** | `max_ack_delay` and the rest are compiled in (see [Configuration](#-configuration--tuning)); the windows, stream limits, idle timeout and `disableActiveMigration` can be set. |
 
 Performance has not been formally benchmarked. QUICO is pure JavaScript; expect it to
 be slower than native implementations, especially for bulk transfer.
@@ -1029,28 +2637,33 @@ be slower than native implementations, especially for bulk transfer.
 - Loss detection, PTO with exponential backoff (RFC 9002)
 - BBR-style congestion control with pacing
 - Bidirectional flow control with sliding windows
+- Stream limits enforced in both directions, raised with `MAX_STREAMS`
+- Delayed and batched ACKs, piggybacked on outgoing packets (RFC 9000 §13.2)
+- Retry / address validation and the anti-amplification limit
+- NAT rebinding with path validation
+- Stateless reset
+- Session resumption, server and client
+- 0-RTT, server and client, over QUIC (RFC 9001 §4.6) and TLS 1.3 on TCP (RFC 8446 §4.2.10), with RFC 8470's 425; off by default
 - HTTP/3 server and client (RFC 9114)
 - QPACK with static and dynamic tables (RFC 9204)
+- HTTP/2 and HTTP/1.1, implemented in QUICO — one stack for all three versions
 - WebTransport client and server — streams + datagrams (RFC 9297)
+- WebSocket over HTTP/1.1, HTTP/2 (RFC 8441) and HTTP/3 (RFC 9220)
 - Unreliable datagrams (RFC 9221)
+- CONNECT-UDP proxy and client (RFC 9298)
+- Stream priorities (RFC 9218) for HTTP/2 and HTTP/3, and WebTransport's `sendOrder` / `sendGroup`
+- GOAWAY, 103 Early Hints
+- JA4H, HTTP/2 and HTTP/3 fingerprints, `profileRequest()`
 - `node:https`-style API — `createServer`, `request`, `get`, `Agent`
 - Express / Fastify / Koa support
 - Connection pooling, coalescing and multiplexing
 - TLS keylog export
 - **Interop: 6/6 implementations green** (handshake, transfer, http3, both directions)
 
-### 🔄 In Progress
-- Delayed ACK and ACK piggybacking
-- 0-RTT and session resumption
-- Configurable transport parameters
-
 ### ⏳ Planned
-- Connection migration (PATH_CHALLENGE / PATH_RESPONSE)
 - ChaCha20-Poly1305 (0x1303)
-- Retry / address validation + anti-amplification limit
-- GOAWAY and graceful shutdown
-- Stream priority (RFC 9218)
-- TypeScript type definitions
+- Backpressure from a slow application reader into QUIC flow control
+- TypeScript type definitions, kept in step with the API
 - Performance benchmarks
 - Fuzz testing