Add exponential backoff for Amp status monitoring

[dinsley]

Summary

Priority: P0

Replace the fixed three-second amp top --stream-jsonl restart loop with capped exponential backoff and jitter.

The current loop recovers quickly, but a persistent configuration, authentication, or compatibility failure can cause the plugin to launch and log a new Amp process every three seconds indefinitely.

Proposed behavior

• Keep the initial status launch immediate.
• Increase the retry delay after consecutive failures.
• Apply jitter so multiple plugin instances do not retry in lockstep.
• Cap the maximum delay so recovery remains automatic.
• Reset the failure count only after a valid live snapshot, or after a clearly defined stable period.
• Cancel pending retries when monitoring is paused or the plugin shuts down.
• Preserve the current fail-closed behavior while disconnected or receiving incompatible snapshots.
• Surface retry attempt and next-delay information to diagnostics without making normal logs noisy.

Suggested defaults to validate during implementation:

• initial delay: 3 seconds;
• multiplier: 2;
• maximum delay: 60 seconds; and
• bounded/full jitter.

Acceptance criteria

• Consecutive launch failures and unexpected exits increase the retry delay up to a cap.
• A valid live connection resets the backoff state.
• stop() reliably cancels every pending retry.
• Only connection-state transitions and useful retry summaries are logged at normal production levels.
• Retry scheduling is deterministic in tests through injected time and randomness.
• Tests cover spawn errors, immediate exits, schema mismatch, recovery, stable reset, and shutdown during backoff.
• Existing users still recover automatically when Amp becomes available again.