diff --git a/HANDOFF.md b/HANDOFF.md
index 4db0293..3917177 100644
--- a/HANDOFF.md
+++ b/HANDOFF.md
@@ -26,7 +26,14 @@ closed and chronicled in git history 08-11..08-12.
   (relay-control-probe unit there; monitor in the 08-15 session). Verdict
   decides: per-IP shaping -> pace retry + port PDS-direct backfill;
   both-dip -> network-wide; neither -> theory dies. Mitigation proposed to
-  operator: retry interval 24h + host-workers 1, compaction 24h.
+  operator: retry interval 24h + host-workers 1, compaction 24h —
+  APPLIED 08-15 ~05:55Z (config-only restart). THE FIX, found upstream:
+  their retry fetches the KNOWN PDS first and uses the relay 302 only as
+  fallback (retry.go:425-438); our resync.zig routes everything through
+  bsky.network. Port PDS-direct-first (task #11) — parity, not invention —
+  then consider restoring the 4h interval. Shaping theory stays unproven
+  (mitigation restart reset the retry timer, cancelling the experiment);
+  the mitigation covers it either way.
 - Prior deploy `56dbf2e` (receipt 21/21; 401 contract byte-identical to
   Bluesky's edge, probed side-by-side). Prod archive IS key-gated:
   STREAM_ARCHIVE_API_KEY in /opt/stream-experiment/.env (never in git),