daemon supervisor: catch-up is never retried, adoption blocks readiness, one rejection kills all

[Dmatut7]

Three supervisor-side defects we hit in production and verified against main (fb2db8ee1). Fixes + tests exist on our downstream line; happy to turn any of them into PRs.

1. A failed client catch-up is never retried or requeued. daemon-supervisor.ts ~:1505 — catchUpClient(client).catch(log); one transient failure and the client keeps a permanently partial view with no signal. We logged 708 of these in production; 86 client/session pairs failed once and stayed stale forever. Our fix mirrors the worker-side retry (bounded backoff, then a terminal session_snapshot_failed so the client re-pulls).

2. Startup readiness awaits adopting ALL workers (daemon-supervisor.ts start()), including create() calls with 24h timeouts — one wedged worker blocks the daemon, and clients attaching during the window see an empty session list. We moved adoption off the readiness path and seed roster rows as recovering so list is honest immediately.

3. unhandledRejection → process.exit(1) (daemon-mode.ts ~:646): one stray rejection (a single failed fs write in bookkeeping) kills the daemon and every session. Suggest log-and-isolate with a counter and a degraded state, keeping uncaughtException fatal.

Full context: https://github.com/Dmatut7/prime-agent-rlm/blob/merge/repl-kernel/docs/fork/upstream-feedback-20260911.md (§2.13 restates the three from the auto-closed #2238).