fix: reconnect backoff and safe retry lifecycle in the CLI service

[R44VC0RP]

Summary

The daemon reconnect loop previously slept fixed delays inside the command loop, logged full stack traces every 2 seconds during outages (my daemon log reached 41 MB with 2,228 identical failures), and could strand retries or leak in-flight attempts across reload.

• Capped, jittered exponential backoff (250 ms to 30 s from BridgeTiming), reset after a stable connection.
• Connection attempts run in a supervised fiber, so stop/reload interrupt an in-flight attachment instead of waiting out the sleep.
• Retries carry a generation number; a retry scheduled by a canceled attempt's cleanup cannot tear down its replacement.
• Scopes are forked from the daemon scope and failed attempts close immediately, releasing partial resources.
• Errors log as one timestamped JSON line, deduplicated while the message repeats; unexpected defects are reported and retried instead of leaving the daemon idle-but-"ok".

Independent of the client changes (#10); works against the unmodified client.

Verification

• packages/cli: bun test test/service.test.mjs — 5 tests covering resource release and growing delays under repeated failure, partial route-registration retry, interruption of in-flight attachments, stale-retry cancellation across reload, and defect recovery.
• bun run build (typecheck + bundle).

Split out of #7.