Summary

Fixes #303.

StdioMcpSessionTransport#sendMessage gates the actual write on inboundReady and outboundReady (two Sinks.One<Void>) both completing before proceeding:

return Mono.zip(inboundReady.asMono(), outboundReady.asMono()).then(Mono.defer(() -> {
    ...
}));

Mono.zip is the wrong operator here. It's built to combine values, and it completes (empty) as soon as the first of the zipped sources completes without a value — it does not wait for the others. Since inboundReady/outboundReady are Mono<Void> (they only ever complete, never emit a value via tryEmitValue(null)), Mono.zip here only ever waits for whichever of the two signals fires first, silently breaking the intended "wait for both" barrier.

I verified this empirically before touching anything (both with tryEmitEmpty() and with the exact tryEmitValue(null) this code uses):

Mono.zip:  fired after only ONE of two Mono<Void> signals? true   <- bug
Mono.when: fired after only ONE of two Mono<Void> signals? false
Mono.when: fired after BOTH signals?                        true

Mono.when is the correct operator for waiting on multiple completion-only signals — it waits for all of them regardless of whether they emit a value.

Fix

One-line change: Mono.zip(...) → Mono.when(...).

Test

Added monoZipDoesNotWaitForBothVoidSignals_monoWhenDoes to StdioServerTransportProviderTests, which reproduces the exact combinator pattern used by sendMessage and asserts:

• Mono.zip incorrectly proceeds after only one of the two Mono<Void> signals fires.
• Mono.when correctly waits for both.

I want to be upfront about scope: this is a targeted unit test of the reactive operator semantics, not a full black-box integration test driving the real threading race through sendMessage end-to-end — inboundReady/outboundReady are private fields on a private inner class with no injectable scheduler, so reproducing the exact race deterministically through the public API would need either reflection or new test seams, which felt like more than a one-line fix warrants. Happy to extend it if a maintainer would rather see that.

Verification

• ./mvnw -pl mcp-core,mcp-test -am test — 1007 tests, 0 failures attributable to this change (one pre-existing flaky timeout in HttpServletStatelessIntegrationTests, unrelated HTTP transport module, confirmed to pass in isolation and on main without this change).
• ./mvnw -pl mcp-core,mcp-test -am spring-javaformat:validate — clean.

Related PRs

A few prior PRs proposed the identical one-line fix but appear to have gone stale without review: #846, #981, #987. This PR adds a regression test on top of the same fix.