Animica 7.2.0 is a sync-hardening release built around one stuck-node case. On 2026-07-07 mainnet had a natural one-block fork at height 38,728. Nodes that had accepted the losing sibling could never import the winning one, because the headers pipeline discarded any header at or below the local head at every layer:
- the canonical-overlap trimmer binned the whole batch as overlap (observed live: 73,896 headers seen, 0 accepted), and a full-window overlap batch was credited as progress whenever the sync target was unknown, so the stall watchdog never fired;
- the header processor treated a batch beginning with the sibling as an anchor mismatch and dropped it;
- a re-served sibling was never “actionable”, and blocks at height ≤ head were never fetched;
- the one recovery keyed to that signature returned false because the network-best height was
Noneon every stuck node.
The fix. A bounded fork-sibling predicate — height ≤ head, within reorg depth, hash different from the canonical block at that height, parent on our canonical chain — lets the sibling through all four layers so fork choice finally sees the competing branch; queued descendants then cascade the reorg. Full-window overlap batches no longer count as progress, and the reset recovery falls back to pipeline-vetted evidence when no peer height survives.
Pinned checkpoint at 38,728. As with the earlier 28,167 outage, a node whose block at 38,728 is the orphan sibling rolls back to 38,727 at boot, prunes the dead-fork index and re-pulls the canonical chain. Verified end-to-end on a clone of live data surgically wedged on the orphan. P2P identity is unaffected because the chain-1 parameters hash is pinned precisely so source edits cannot fragment the handshake.
Review hardening. The adversarial review of the wedge fixes narrowed the snapshot genesis recovery to the one known-benign case (a height-0 entry whose block-0 matches the pinned genesis hash), capped distinct sibling fetches at four per height because compact headers carry no verifiable PoW, made the reset-recovery fallback use only queued-block heights, and disabled sibling ingest entirely when _max_reorg_depth is 0.
8.0.3 (2026-07-15) later confirmed by four independent code traces and an end-to-end reproduction that any node on ≥ 7.2.0 self-heals a natural one-block fork without a pin; the per-fork pins remain as a fast-converge safety net for older nodes.