fix(transport): handle SIGINT/SIGTERM in stdio transport and gracefully shut down LSP servers

[bug-ops]

Summary

• run_stdio, the default transport used by every stdio-based MCP client (Claude Desktop, Claude Code, etc.), installed no signal handler at all — an uncaught SIGINT/SIGTERM bypassed kill_on_drop and orphaned every spawned LSP child process. It now shares the same signal-handling logic run_http already had.
• LspServer::shutdown() (the graceful LSP shutdown/exit handshake) was previously dead code outside tests. serve_with now calls a new Translator::shutdown_servers() after the transport future resolves — on signal, stdin EOF, or HTTP's own shutdown — which drains and gracefully shuts down every registered LspServer concurrently, with a bounded per-server grace period before falling back to kill_on_drop.
• run_http's own graceful shutdown wait is now bounded too, so a stuck in-flight connection can't block LSP cleanup indefinitely.

Known limitation

Process termination via an uncaught panic under panic = "abort" ([profile.release]) still bypasses this cleanup, since no Drop runs on that path. A real fix needs process-group isolation, which is out of scope here and documented as a # Limitations note on Translator::shutdown_servers. Recommend filing a narrower follow-up issue for that vector before/independent of closing this one.

Test plan

• cargo +nightly fmt --all -- --check
• cargo clippy --all-targets --all-features --workspace -- -D warnings
• cargo nextest run --workspace --all-features --lib --bins (489/489)
• RUSTDOCFLAGS="-D warnings" cargo doc --no-deps --all-features --workspace
• New unit tests: empty-registry shutdown, concurrent drain via a fake LSP server, serve_with's shutdown path draining a registered server, run_stdio's prompt-return behavior, and a paused-clock regression test proving run_http no longer self-terminates on ordinary uptime
• Manually verified against the built binary: stdio SIGTERM/SIGINT no longer orphans a spawned LSP server; HTTP transport stays up over multi-minute uptime with no signal, shuts down cleanly within ~0.02s on SIGTERM, and correctly force-times-out ~30s after a signal if a connection is stuck mid-drain

Closes #241

[Copilot]

Pull request overview

This PR addresses process hygiene in mcpls-core by ensuring both stdio and HTTP transports react to SIGINT/SIGTERM and trigger an orderly shutdown path that drains and shuts down spawned LSP servers instead of orphaning them.

Changes:

• Added shared SIGINT/SIGTERM handling and applied it to both run_stdio and run_http, including a bounded HTTP graceful shutdown wait.
• Introduced a post-transport shutdown sequence in serve_with that signals background tasks to stop and drains registered LSP servers via Translator::shutdown_servers().
• Extended LspServer::shutdown() to attempt the LSP shutdown/exit handshake and then wait a bounded grace period for the child process to exit before relying on kill_on_drop.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
crates/mcpls-core/src/transport.rs	Adds shared shutdown-signal handling; stdio returns on signal; HTTP graceful shutdown is bounded post-signal.
crates/mcpls-core/src/lsp/mod.rs	Re-exports a test-only fake LSP server helper for shutdown-path tests.
crates/mcpls-core/src/lsp/lifecycle.rs	Adds bounded child-exit grace in LspServer::shutdown() and introduces a shared test helper for mock servers.
crates/mcpls-core/src/lib.rs	Ensures transport completion triggers a centralized shutdown sequence that drains LSP servers.
crates/mcpls-core/src/bridge/translator.rs	Implements shutdown_servers() to concurrently drain and shut down registered servers with per-server timeouts.
CHANGELOG.md	Documents the fix and its limitations under Unreleased/Fixed.