fix(sync): retain backoff after partial progress failures

[branarakic]

Summary

• preserve useful sync progress from mixed progress-and-failure rounds
• retain and grow the ordinary per-peer retry backoff when that same round hits transport or integrity pressure
• prevent transient connection-open events from replaying the full subscribed graph scope after only the base catch-up cooldown

Live evidence

A testnet edge peer repeatedly replayed its full subscribed-user graph fanout every 1-3 minutes. The round successfully integrated verified triples, then hit a graph-scoped durable authorization deadline. Progress accounting cleared the peer failure history, so the next transient connection-open bypassed exponential retry damping and repeated the full scan.

Correctness boundaries

• clean progress still clears backoff
• pure local backpressure behavior is unchanged
• mixed progress plus a real failure records progress without marking the peer fresh and grows peer-scoped backoff
• topology or protocol probe changes can still bypass stale backoff
• no timing constants or deployment configuration change

Validation

• git diff --check
• focused sync-on-connect churn and retry tests: 2 files, 69 passed
• agent build: TypeScript, type tests, package-root tests passed
• full agent unit lane: 139 files, 1,789 passed, 5 skipped, 0 failed

Integration note

This is a standalone patch based on exact testnet-canary commit 1dba8f5e7b12680a267356dae9e5eb944248436a. Draft only; no node was restarted or deployed.