p2p/discover/topicindex: fix search deadlock and early close when empty buckets

[srene]

Summary

Fix two related stalls in Search:

1. IsDone deadlock (closes topicindex/search: IsDone can miss termination — searcher hangs until Close #27): treat no unasked nodes AND no buffered results as the terminating condition. Drop the queriesWithoutNewNodes >= 4 gate and the counter that backed it.

2. QueryTarget empty-bucket stall (closes topicSearch: QueryTarget stalls when farthest bucket lacks unasked nodes and numRequests==0 #65): move the numRequests == 0 break inside the len(new) > 0 block so the warm-up frontier only halts on a populated unqueried bucket. Empty unqueried buckets are now invisible to the gate.

Supersedes #45.

The IsDone bug

IsDone returned true only if queriesWithoutNewNodes >= 4, after verifying the result buffer and every bucket's new set were empty. The counter only advances when AddNodes runs without finding novel nodes.

But once every bucket's new set is empty:

• QueryTarget returns nil.
• No query is dispatched.
• No response arrives.
• AddNodes is never called.
• queriesWithoutNewNodes can never reach 4.

The search loop in v5_topic.go: topicSearch.run() then sat with only <-s.quit progressable:

select {
case <-s.quit:                  // only live case
case queryCh <- nextQuery:      // queryCh nil (no target)
case resp := <-s.queryRespCh:   // no response coming
case resultCh <- result:        // no result to emit
}

iter.Next() blocks indefinitely waiting on the (now silent) resultCh.

The QueryTarget bug

QueryTarget walks the search buckets far → close, collecting buckets with unasked nodes into a withnew pool. After each bucket is added, the original code applied a "warm-up frontier" gate: if numRequests == 0 { break } halted the scan so closer-to-topic buckets weren't probed until the current bucket had received at least one response. This avoids spamming nodes close to the topic before farther rings have been sampled.

The bug was that this gate fired regardless of whether the bucket had any candidates. An empty unqueried farthest bucket — natural state when the routing-table feed is slow to deliver maximally-distant peers — broke the scan at i=0 with withnew = [], returning nil and stalling the search even though closer buckets held unasked candidates.

Fix: move the break inside the len(new) > 0 block. The frontier still halts on populated unqueried buckets (so warm-up semantics and load-spreading intent are preserved), but empty buckets are now invisible to the gate and the scan continues past them.

Change

• p2p/discover/topicindex/search.go
  • IsDone: returns true as soon as resultBuffer is empty and all b.new are empty.
  • QueryTarget: warm-up frontier gate (break on numRequests == 0) is now nested inside the len(new) > 0 check. Cross-bucket randomization, numRequests field, and math/rand usage are unchanged from upstream.
• p2p/discover/topicindex/search_test.go
  • TestSearchIsDoneNoUnaskedNodes, TestSearchIsDoneBufferedResults: cover the IsDone fix.
  • TestSearchQueryTargetSkipsEmptyFarBucket: empty bucket 0 + populated bucket 5 → picks from bucket 5.
  • TestSearchQueryTargetPrefersFarthest: bucket 0 with unqueried nodes locks the pool to itself; closer buckets only join after bucket 0 drains.
  • TestSearchQueryTargetWidensFrontierAfterResponse: once bucket 0 has been queried, refilled bucket 0 + populated bucket 5 both appear in picks across many calls.
  • TestSearchQueryTargetRestartsAfterRefill: bucket 0 refill takes priority over remaining bucket 5 candidate.

Test plan

• go test ./p2p/discover/topicindex/ -run TestSearch passes
• go test -run TestTopic ./p2p/discover/ passes
• go vet ./p2p/discover/... clean