flaky: dht_pair fixture races on routing-table population ("Node A should know about Node B")

[yashksaini-coder]

The dht_pair fixture in tests/core/kad_dht/test_kad_dht.py intermittently fails during setup with:

AssertionError: Node A should know about Node B

at test_kad_dht.py:177. Because it's a fixture, any test that uses it fails when this trips (currently ~5 tests: test_register_validator, test_find_node_reply_does_not_prepend_unknown_target, and others).

Evidence it's a flake, not a real bug

Hit on two unrelated PRs this week, neither of which touches kad_dht:

• refactor(webrtc): deduplicate _varint.py — use libp2p.utils.varint #1395 (webrtc varint dedup) — tox (3.11, core), failed on test_find_node_reply_does_not_prepend_unknown_target
• test(timed_cache): de-flake test_expiry_removal (#1408) #1409 (timed_cache de-flake) — tox (3.13, core), failed on test_register_validator

Both bottom out in the same fixture assertion. Passes on re-run / locally.

Why it races

The fixture does:

await dht_a.peer_routing.routing_table.add_peer(peer_b_info)
await dht_b.peer_routing.routing_table.add_peer(peer_a_info)
async with background_trio_service(dht_a), background_trio_service(dht_b):
    await trio.sleep(0.1)              # fixed wait
    try:
        await dht_a.find_peer(...)     # best-effort, exceptions only logged
        await dht_b.find_peer(...)
    except Exception:
        ...
    assert dht_a.routing_table.peer_in_table(...)   # hard assert

The hard assert depends on the peer still being resident after a fixed 0.1s sleep and a best-effort find_peer. Under CI load that window isn't always enough (the added peer may not have survived early service maintenance / discovery hasn't converged), so peer_in_table is False and setup blows up.

Suggested fix

Replace the fixed sleep + one-shot assert with a bounded poll until both routing tables are populated, e.g.:

with trio.fail_after(10):
    while not (
        dht_a.routing_table.peer_in_table(host_b.get_id())
        and dht_b.routing_table.peer_in_table(host_a.get_id())
    ):
        await trio.sleep(0.05)

Same idea used for the other CI de-flakes (#1401, #1408): wait for the condition, don't guess a duration. Happy to send a PR if that direction looks right.