test(live): ask redis who is subscribed, not the relay's own map

[houko]

The failure

TestAChangeOnOneServerReachesTheOther fails on CI roughly one run in eight, always the same way:

--- FAIL: TestAChangeOnOneServerReachesTheOther (10.01s)
    relay_test.go:131: the change never crossed to the other server: 0 sync frame(s) arrived in ten seconds, each rejected as []

It hit #254 and #258 during this batch. Neither touches internal/live; #258 does not touch Go at all.

Why

Relay.Subscribe records a subscription in r.subscriptions and only afterwards waits for Redis to confirm it, because client.Subscribe merely queues the command — Receive is what sends it:

subscription := r.client.Subscribe(ctx, documentChannel(name))
r.subscriptions[name] = subscription          // relay.go:140 — visible here
r.mu.Unlock()

if _, err := subscription.Receive(ctx); err != nil {   // relay.go:144 — confirmed only here

waitForSubscription polled that map, so it returned during the window between those two statements. Pub/sub keeps no backlog, so the change the test publishes in that window is not delayed, it is dropped, and the test then waits out its full ten seconds for a frame that no longer exists.

The relay's own ordering is correct — it confirms before it publishes — so nothing in relay.go changes here.

0 frames is what points at this rather than at a slow machine: every server publishes SyncStep1 and QueryAwareness from Subscribe itself, so a server that were subscribed would have received those too. Receiving nothing at all means it was not subscribed when the other side published.

The change

Ask Redis instead, with PUBSUB NUMSUB, which only counts a subscriber once the SUBSCRIBE has actually landed. Both servers share one Redis, so one question covers the pair.

Verification

• The internal/live suite passes, including under CPU contention and -race.
• Not verified: the original failure did not reproduce locally in 60 runs under load, on the old code or the new. This closes a window that is certain from reading the code, not one demonstrated by a local repro. If the test still flakes after this lands, the root cause is elsewhere and this reasoning should be discarded rather than patched over.