fix(scheduler): atomic-read counters in control loop

[singchia]

Summary

control() reads countIncomingReqs and countProcessedReqs with plain loads, but worker goroutines and PublishRequest update them via atomic.AddInt64. Under the Go race detector this consistently trips a data-race warning — surfaced in singchia/geminio's race-enabled CI jobs.

Fix

Snapshot each counter with a single atomic.LoadInt64 so the diff and the baseline update see the same value, and no plain read races with a concurrent atomic write.

incoming := atomic.LoadInt64(&s.countIncomingReqs)
processed := atomic.LoadInt64(&s.countProcessedReqs)
incomingReqsDiff := incoming - s.countIncomingReqsL
processedReqsDiff := processed - s.countProcessedReqsL
s.countIncomingReqsL = incoming
s.countProcessedReqsL = processed

countIncomingReqsL / countProcessedReqsL are only touched by control() itself, so they stay plain.

Test plan

• go test -race ./... in this repo — clean
• go test -race ./test/integration/... in singchia/geminio with a local replace pointing at this branch — clean (previously raced)
• Gemino unit/regression/e2e race suites — all clean with this fix

Impact

• No behavioral change on non-race builds.
• Safe to tag as v2.2.2 and bump downstream.

🤖 Generated with Claude Code