Add new connectrpc-health crate

[torkelrogstad]

Inspired by connectrpc.com/grpchealth and the equivalent Tonic implementation. Breaking up into four different commits felt kind of awkward, but did it per CONTRIBUTING.md guidelines.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[torkelrogstad]

I have read the CLA Document and I hereby sign the CLA

[iainmcgin]

[claude code] Thanks — this is a strong contribution and we'd like to take it. We ran it through both a correctness/concurrency review and an API-ergonomics review; the hand-written code came out of both in very good shape: locks are block-scoped and never held across awaits, send_if_modified suppresses no-op wake-ups, dropping a Watch stream releases the subscriber (with a regression test for exactly that), mutex poisoning is recovered rather than propagated, peer-supplied service names are only ever used as lookup keys (no insertion path a caller can drive), and every documented behaviour of StaticChecker has a colocated test. The deliberate deviation on Watch for unknown services (NOT_FOUND rather than the spec's open stream with SERVICE_UNKNOWN, matching connect-go's grpchealth) is documented consistently in all four places it matters — that's fine by us. The buf/Taskfile wiring matches the other generated directories and the generated output follows the fully-qualified-paths convention.

The asks below are about the public API surface, since this becomes a published crate.

Before merge

1. Feature-gate the client half. connectrpc-health/Cargo.toml:16 unconditionally enables connectrpc's server and client features, so a server that only exposes the health service still pulls in the whole client transport stack. Please add a client feature on this crate (off by default or default — your call) that forwards to connectrpc/client, and gate the HealthClient/HealthExt re-exports behind it.
2. Make the common case one line and leave the user holding a usable handle. Today the static, always-SERVING case is two Arc::news, an Arc::clone, from_arc, and register — and the discoverable constructor (HealthService::new) seals the checker inside the service so there's nothing left to call set_status on. A crate-level convenience along the lines of install_static(router, [SERVICE_NAME]) -> (Router, Arc<StaticChecker>) (keeping from_arc for custom checkers) would close most of the gap to grpchealth/tonic-health. Related: checker(&self) -> &Arc<C> is an unusual return type — either return Arc<C> (clone internally) or have the doc spell out Arc::clone(svc.checker()), since "Clone it" currently steers users into clippy::clone_on_ref_ptr territory.
3. Decide and document the set_status register-on-miss semantics. StaticChecker::set_status silently creates an entry for any name (static_checker.rs:61-76), so a typo'd service name "works" while probes for the real FQN keep reporting the old status, and the map grows if the name ever comes from non-static input. connect-go's StaticChecker.SetStatus errors on unknown names. Either split registration from updates (or add a strict try_set_status), or keep the current behaviour but say loudly on set_status itself that it registers on miss and must only be fed the generated *_SERVICE_NAME constants; a remove_service would round the API out either way.
4. Guide fixes. The example at docs/guide.md:1001 doesn't compile: Status is used but not imported, and the Cargo snippet above it omits the connectrpc = { version = "0.6", features = ["server"] } dependency the example needs. Marking the block rust,no_run (like the lib.rs Quick Start) would let CI catch future drift.
5. Crate metadata. Please add a readme (the sibling crates point at the workspace README) and bump the crate version to match the workspace now that 0.6.1 is out.

Fine as follow-ups (or in this PR if you prefer)

• From<ServingStatus> for Status (or TryFrom) for probe-loop code decoding raw responses; the forward direction already exists.
• The default Checker::watch returning Unimplemented is spec-legal, but it's easy for a custom-Checker user to ship without noticing Watch is dead — consider a doc callout right on the trait method (or a tracing warning) rather than only under # Errors.
• A with_services_status(...) form so a service can be registered as NotServing from the start.
• The #[allow(clippy::upper_case_acronyms)] on the generated-module wrappers appears unused.
• An end-to-end test that a client disconnect mid-Watch drops the server-side subscriber (the WatchStream-drop unit test is convincing; the hyper-level version would tie it together).

On process: CI for first-time contributors needs a maintainer to approve the workflow run — we'll do that once the above lands, and the new "Check generated code" job that just merged should pass given the Taskfile/buf wiring you've already included. Happy to re-review quickly after the changes.

[torkelrogstad]

Thanks for your review on this @iainmcgin. I've added a new commit at the start that modifies the codegen to achieves the feature-gate you requested. The other commits have been amended to incorporate your feedback.

While working on the codegen changes I realized it would make sense to expose codegen knobs for disabling client/server generation all together, if you're only interested in one of them. Skipped this here, as that seemed like a bigger change. But curious what you think about that.

[torkelrogstad]

@iainmcgin friendly ping