Drain TreeCRDT work before teardown

[marcus-pousette]

Stack

This is the focused teardown follow-up after #108, #109, and #106. Review and merge that stack first.

Summary

• register persistence synchronously and bind it to the exact session ordered before the call
• keep the real latest lifecycle promise separate from the failure-tolerant scheduling tail, so persistence receives init/drop failures instead of leaking into a later session
• retain settled persistence failures until waitForIdle or teardown consumes them
• stop ingress, drain captured persistence plus the shared write/materialization pipeline to a stable version, stop sync again, then release the client
• verify failed-init ordering, drop → persist → init, idle waiting, error-safe draining, and failed cleanup ownership
• strengthen the delayed-startup browser test by verifying the text entered before initialization is actually persisted

The shared write/materialization queues intentionally represent em’s one Redux integration pipeline; em creates one active app thoughtspace per browser realm. Client, provider session, persistence accounting, and WebSocket state remain factory-local.

Validation

• the three runtime/teardown file patches match the previously validated version exactly by stable patch ID; the startup assertion was rebased onto Simplify window.em test access #109’s explicit preload type
• yarn lint:tsc
• all TreeCRDT unit suites — 10 files, 42 tests passed, including all 5 runtime lifecycle tests
• focused post-restack runtime/lifecycle check — 20 tests passed
• prior cumulative validation: full yarn lint, full unit suite, and Puppeteer startup plus treecrdt-single-tab passed