[GEN-08] `topological_sort()` runs O(N) on every `SyncRequest`, then truncates to 500

[intendednull]

Commit: 2f26d91 · Finding: GEN-08

Problem

crates/client/src/listeners.rs:298-307 calls dag.topological_sort() over the entire DAG on every WireMessage::SyncRequest, then takes only the first 500 entries. topological_sort (crates/state/src/dag.rs:316) iterates every event, builds in-degree + dependents maps, and allocates a Vec of size self.events.len().

For a server with 50k events every sync request pays the full sort cost even though only the first 500 are sent. An N+1 amplifier as peers increase. The in-source TODO acknowledges the temporary nature; the worker has a heads-based sync that should be the migration target (see #65, #207).

Distinct from #65 (protocol change) and #207 (500 cap) — this finding is specifically about the wasted work per request.

Fix

• Near-term: cache the truncated snapshot on the listener side and invalidate on new event insertion. Even a simple "remember the last sorted Vec<_> and the tip hash" saves 99% of the cost for idle channels.
• Long-term: migrate clients to WorkerRequest::Sync { heads } — see Worker sync protocol uses bulk SyncBatch instead of per-author Advertise/Request/Response #65.