Tracking: kad-dht query convergence & routing-table reliability

[tabcat]

Tracking issue for a set of internal @libp2p/kad-dht reliability fixes found while hardening the IPNS-over-DHT write path. All are internal (no public API changes); each lands as its own PR, linked here as it opens. Ordered roughly by impact.

• getClosestPeers should only add peers that responded
  • Every contacted peer is added to the closest set even if it never answered, so dead or stale entries fill the closest-K and receive wasted PUT_VALUEs. Mirrors go-libp2p, which excludes PeerUnreachable peers from the converged set.
  • fix(kad-dht): only keep peers that responded in getClosestPeers #3537
• Terminate a query path once the closest-K set is full
  • Lets a path stop instead of draining every seeded peer. Only honest once the live-peer fix lands, otherwise the set can fill with dead peers.
  • feat(kad-dht): terminate query paths once the closest-K set is full #3539
• Run the periodic routing-table refresh
  • Only the startup refresh fires today; the interval refresh is throttled out because the per-CPL refresh timestamp is never recorded.
  • fix(kad-dht): run the periodic routing table refresh #3538
• Evict unresponsive peers from the routing table on query failure
  • A peer that fails a lookup is never removed (eviction only happens via bucket-full ping-replace), so dead entries are re-queried on every lookup. Ping-confirm before evicting. (needs design + test)
  • Scratched: a maintained routing table stays healthy on its own, so proactive eviction adds little. A 45-min live run on the fix(kad-dht): run the periodic routing table refresh #3538 refresh branch (passive node, no manual queries) held a stable table at ~5-8% unreachable with no upward drift (~46 peers, ~20 connections kept open to the KEEP_ALIVE closest-peer set).
• Unbiased Fisher-Yates shuffle for the disjoint-path seed
  • Scratched: shuffling is not required for S/Kademlia disjoint-path lookups (the spec only says to distribute the k closest into d buckets); the bias is harmless to the lookup result and go-libp2p doesn't shuffle at all.
• Fix alpha option doc
  • The JSDoc says @default 3 but the code uses ALPHA = 10.
  • docs(kad-dht): correct the alpha option default in the JSDoc #3541
• Defer the dialability check until after the distance checks
  • So it only runs for peers that pass the distance filter, not every candidate.
  • refactor(kad-dht): only check dialability for closer peers #3528

Evaluated, not pursuing: per-request timeout cap (reach is ceiling-independent once the live-peer fix lands), backfilling failed PUTs to peers outside the closest-K (breaks Kademlia write semantics, periodic republish covers it), streaming PUTs before the lookup completes (closest-K is not known until convergence).