fix(server): correct process-handler and shutdown lifecycle

[lesquel]

Summary

Resolves the process/shutdown lifecycle items of #33 (Tier-1/2). The plugin factory runs once per workspace/worktree in one process, so these defects compound per invocation. Verified by hand.

#	Defect	Fix
D1	process.on for uncaught/unhandled → listeners accumulate per invocation (dup emits, MaxListenersWarning, feedback loop)	Module-scoped install-once guard (extracted to server/lifecycle.ts). Deliberately not process.once — once removes the listener after one fire, leaving a 2nd rejection unhandled. Handlers route through shared singletons.
D2	shutdown() never closed SSE clients → keepalive setIntervals leak after server.stop()	shutdown() calls eventBus.closeAll() before server.stop(); closeAll() now clears each client's pingInterval explicitly (stored on SSEClient) rather than relying on controller.close() → stream cancel() (a Bun impl detail)
D3	No re-entrancy guard → SIGINT+SIGTERM/exit ran cleanup twice	Closure-scoped guard; original shutdown step order preserved exactly (only closeAll prepended)
D4	process.once("exit", () => void shutdown()) dropped all async cleanup (fires during event-loop drain)	Listener removed; SIGINT/SIGTERM (intact) are the correct graceful hooks

Design notes

• Install-once + shutdown-guard extracted to server/lifecycle.ts (testable seam; server/ may compose core/ — no dependency-rule violation; documented in server/AGENTS.md).
• The D2 robustness hardening (explicit pingInterval clear in closeAll) was added in response to the fresh review's nit — it makes the interval-leak fix self-contained instead of depending on undocumented Bun ReadableStream behavior.

Testing

• +9 behavioral tests (listener-count-stable-across-invocations; idempotent/concurrent shutdown; closeAll invoked; no exit listener). Non-flaky, no leaked process listeners.
• bun test 633/0 · bunx tsc --noEmit clean (bus.ts keepalive change did not regress existing bus tests)
• Independent fresh-context adversarial review: APPROVE, no blocking issues. Shutdown step-ordering exhaustively diffed (original vs new, side by side) — every step/condition preserved. All 3 review nits fixed in this PR.

Pre-existing, out of scope (noted, not fixed): SIGINT/SIGTERM use process.once, so in the multi-instance model only the first invocation's shutdown runs on a signal — a separate multi-instance design question.

Refs #33 (Tier-1/2, partial — does not close).