Add Hyperliquid WebSocket client

[v9ai]

• Add WebSocket client, codec, and exchange modules
• Implement message handling and subscription management

[cjdsellers]

Thanks @NicolaD, good progress here!

[cjdsellers]

It's interesting tokio-stream is required, but not for the other adapters. This suggests a different pattern somewhere?

[v9ai]

Aspect	Other adapters: channel + tokio::select! (no tokio-stream)	This adapter: stream-first (uses tokio-stream)	What it says about the dep
Direct answer	Doesn’t need tokio-stream because it pulls from channels via recv().await and coordinates with select!.	Needs tokio-stream when it wraps Tokio primitives as Streams or uses stream combinators.	Yes—different pattern. Presence of tokio-stream signals a stream-centric composition somewhere.
Primary abstraction	Futures + channels (mpsc, broadcast) + select!.	Stream<Item = T> (from futures-core) + tokio_stream::wrappers + StreamExt combinators.	Stream API → extra crate.
What you’ll see in code	while let Some(x) = rx.recv().await { ... }, tokio::select! { ... }.	ReceiverStream::new(rx).map(...).merge(...).next().await, StreamMap.	Wrappers/combinators imply tokio-stream.
Fan-in / merging sources	Manual select! over multiple recv() branches; verbose for many/dynamic sources.	merge, SelectAll, or StreamMap to combine many streams succinctly.	tokio-stream added to get these tools.
Timers / heartbeats	tokio::time::interval() polled inside loops.	IntervalStream + combinators (timeout, throttle, chunks_timeout).	Time-aware ops come from tokio-stream features.
Public API shape	Pull style: async fn next_event() -> Option<T>.	Stream style: fn into_stream(self) -> impl Stream<Item = T>.	Exposing a stream almost always pulls in tokio-stream.
Composability (map/filter/batch)	Hand-rolled; more boilerplate/state machines.	One-liners with StreamExt (map, filter, chunks_timeout, etc.).	Crate enables declarative pipelines.
Backpressure semantics	Channel capacity controls it; explicit.	Same underlying channels; wrappers don’t change capacity.	Neutral—dependency is about ergonomics.
Fairness/ordering	select! has randomized fairness; ordering is imperative.	Per-stream order preserved; cross-stream order depends on combinator.	Must document semantics either way.
Testability	Drive loops manually; custom harnesses.	collect::<Vec<_>>(), take(n), etc. on streams simplify tests.	Stream API often easier to test.
Cost / deps	Fewer deps; minimal indirection.	Adds tokio-stream; thin wrapper overhead.	Small runtime impact; main cost is dep surface.
If you want to remove it	Keep pull-style API; replace any wrappers with direct recv()/timeout().	N/A.	Run cargo udeps; drop tokio-stream if unused.
If you want to keep it (justify)	N/A.	You’re wrapping mpsc/broadcast/interval as streams, merging sources, or exposing a stream API.	Consider feature-gating: stream-api = ["tokio-stream"].
Reply	“The other adapters use channel + select!, so they don’t need tokio-stream.”	“This adapter embraces a Stream-based composition (wrapping channels/timers and/or using combinators), which is why tokio-stream is required.”

[v9ai]

more details here: https://www.vadim.blog/tokio-vs-tokio-stream-websocket-adapters