feat: lifecycle ordering preserved under concurrency (#4)

[meymchen]

Closes #4.

Holds two LSP-spec lifecycle invariants under the concurrent dispatcher (issue #1), plus the structural changes they required.

Behavior

1. Initialize precedence (gate the spawn step)

• Any request arriving while state == Uninitialized with method != "initialize" is answered ServerNotInitialized (-32002) without spawning a handler.
• Any notification arriving while Uninitialized with method ∉ {initialized, exit} is dropped with a debug! log. This matches LSP §Initialize ("notifications should be dropped, except for the exit notification"); $/cancelRequest is dropped pre-init too.

2. Exit aborts in-flight work

• All spawned handlers now live in a single tokio::task::JoinSet. On exit, the read-loop shutdown()s the set — aborting every in-flight handler (releasing their outgoing-sender clones) — then drops the master sender so the send-loop drains queued output and exits cleanly, and hands back the exit code.
• This fixes the real bug: previously send_handle.await blocked until the slowest handler finished (handlers held out_tx clones), so exit effectively awaited a slow didOpen instead of aborting it.

Structural changes (steered to align with the LSP spec + ADR 0003)

• process::exit moved to the binary boundary. dispatcher::run now returns Outcome::{TransportClosed, Exit(code)}; StdioBuilder::serve maps Exit → process::exit(code), while lspf::serve returns. A library no longer kills your process, and exit is testable in-process.
• initialize runs inline (ADR 0003's documented design) so the state → Running transition is synchronous before the next message is dispatched — the prior spawned initialize is what made gating racy. initialize is therefore no longer cancellable.
• Registry shrinks to RequestId → CancellationToken (cancel routing only); handles live in the JoinSet.

Tests

• New tests/lifecycle_ordering.rs: two table-driven init-precedence tests (requests → ServerNotInitialized; feature/unknown notifications dropped) + the exit-abort smoke test (slow didOpen in flight → serve returns < 500ms, no publishDiagnostics).
• cancellation.rs reworked to cancel a post-init shutdown (spec-legal client timing) instead of an in-flight initialize, which is now dropped pre-init per the spec.
• concurrency_cap.rs assertion corrected: inline initialize takes no concurrency permit → 2 acquire spans, not 3.
• concurrency.rs and the stdio smoke tests pass unchanged.

Full suite green (20 tests), clippy clean, fmt applied.

Acceptance criteria

• Pre-init request (method ≠ initialize) → ServerNotInitialized, no spawn
• Pre-init notification (∉ {initialized, exit}) → dropped with debug!
• On exit: all in-flight handles aborted; sender dropped; send-loop drains cleanly
• Table-driven init-precedence test
• Smoke test: slow didOpen in flight at exit is aborted, not awaited
• Existing smoke tests still pass

Note for review

ADR 0003 also lists shutdown as inline, but it stays spawned here so it remains cancellable (the only request the current trait surface can cancel). The Outcome/exit-termination-boundary is a new decision with no ADR yet — happy to record both as a follow-up if wanted.

🤖 Generated with Claude Code