Retry timed-out provider initialization once

[brsbl]

Summary

• classify JSON-RPC request timeouts with a typed error
• retry provider startup once, with a fresh process and adapter, only when initialize times out
• suppress the retry during shutdown and cover the recovery path with a lifecycle regression test

Why

A cold Codex app-server process can occasionally miss the 30-second initialize deadline. The PartySocket upgrade protects the server-to-daemon transport, but this failure is inside the daemon-to-provider JSON-RPC startup path.

Verification

• pnpm exec turbo run test --filter=@bb/agent-runtime --force (795 tests)
• pnpm exec turbo run typecheck --filter=@bb/agent-runtime
• independent review-loop re-review: CORRECT

[brsbl]

Closing as stale. The branch is retained, so this can be reopened if the retry work is still wanted.