fix(wa-sqlite): make direct and dedicated teardown terminal

[marcus-pousette]

Summary

• Mark direct clients closed before database cleanup can fail.
• Always terminate a dedicated worker and remove its listeners when the init, close, or drop RPC rejects.
• Preserve existing error behavior: close() is best-effort, drop() reports the first teardown error, and repeated teardown does not retry the backend.
• Add focused regressions for failed direct close/drop and dedicated-worker init/close/drop.

Why

cybersemics/em#4325 uses persistent OPFS with dedicated-worker, plus direct for unit tests and some Puppeteer profiles. Its integration currently falls back from a rejected drop() to close() because wa-sqlite can leave the lower handle live after teardown fails.

em also treats initialization as retryable. If a dedicated worker rejects its init RPC, createTreecrdtClient() returns no client handle, so em cannot clean up the worker itself. Terminating that failed endpoint must therefore be guaranteed by wa-sqlite.

This extracts only the relevant single-client lifecycle slice from #208. It intentionally does not include #209 SharedWorker peer invalidation, protocol/server changes, cross-tab behavior, OPFS memory fallback, or a generic lifecycle-hook abstraction.

With this behavior, em can discard the terminal client after drop() settles, remove its drop() → close() fallback, and safely retry after a dedicated-worker initialization rejection. em still owns WebSocket shutdown, materialization unsubscribe, provider reset, and lifecycle serialization.

Validation

• focused single-client teardown suite — 5/5
• wa-sqlite TypeScript build
• Prettier and git diff --check
• full CI build and test matrix