Fix sentinel error aliasing, hot-path allocations, and resource leaks

- Deduplicate sentinel errors: httpx.ErrNoHealthy, ErrCircuitOpen, and
  ErrRetryExhausted are now aliases to the canonical sub-package values
  so errors.Is works across package boundaries
- Retry transport returns ErrRetryExhausted only when all attempts are
  actually exhausted, not on early policy exit
- Balancer: pre-parse endpoint URLs at construction, replace req.Clone
  with cheap shallow struct copy to avoid per-request allocations
- Circuit breaker: Load before LoadOrStore to avoid allocating a Breaker
  on every request for known hosts
- Health checker: drain response body before close for connection reuse,
  probe endpoints concurrently, run initial probe synchronously in Start
- Client: add Close() to shut down health checker goroutine, propagate
  URL resolution errors instead of silently discarding them
- MockClock: fix lock ordering in Reset (clock.mu before t.mu), fix
  timer slice compaction to avoid backing-array aliasing, extract
  fireExpired to deduplicate Advance/Set

balancer/health.go

@@ -2,6 +2,7 @@ package balancer
 
 import (
 	"context"
+	"io"
 	"net/http"
 	"sync"
 	"time"
@@ -70,8 +71,10 @@ func newHealthChecker(opts ...HealthOption) *HealthChecker {
 }
 
 // Start begins the background health checking loop for the given endpoints.
-// All endpoints are initially considered healthy.
+// An initial probe is run synchronously so that unhealthy endpoints are
+// detected before the first request.
 func (h *HealthChecker) Start(endpoints []Endpoint) {
+	// Mark all healthy as a safe default, then immediately probe.
 	h.mu.Lock()
 	for _, ep := range endpoints {
 		h.status[ep.URL] = true
@@ -82,6 +85,9 @@ func (h *HealthChecker) Start(endpoints []Endpoint) {
 	h.cancel = cancel
 	h.stopped = make(chan struct{})
 
+	// Run initial probe synchronously so callers don't hit stale state.
+	h.probe(ctx, endpoints)
+
 	go h.loop(ctx, endpoints)
 }
 
@@ -111,7 +117,7 @@ func (h *HealthChecker) Healthy(endpoints []Endpoint) []Endpoint {
 	h.mu.RLock()
 	defer h.mu.RUnlock()
 
-	var result []Endpoint
+	result := make([]Endpoint, 0, len(endpoints))
 	for _, ep := range endpoints {
 		if h.status[ep.URL] {
 			result = append(result, ep)
@@ -137,13 +143,18 @@ func (h *HealthChecker) loop(ctx context.Context, endpoints []Endpoint) {
 }
 
 func (h *HealthChecker) probe(ctx context.Context, endpoints []Endpoint) {
+	var wg sync.WaitGroup
+	wg.Add(len(endpoints))
 	for _, ep := range endpoints {
-		healthy := h.check(ctx, ep)
-
-		h.mu.Lock()
-		h.status[ep.URL] = healthy
-		h.mu.Unlock()
+		go func() {
+			defer wg.Done()
+			healthy := h.check(ctx, ep)
+			h.mu.Lock()
+			h.status[ep.URL] = healthy
+			h.mu.Unlock()
+		}()
 	}
+	wg.Wait()
 }
 
 func (h *HealthChecker) check(ctx context.Context, ep Endpoint) bool {
@@ -156,6 +167,7 @@ func (h *HealthChecker) check(ctx context.Context, ep Endpoint) bool {
 	if err != nil {
 		return false
 	}
+	io.Copy(io.Discard, resp.Body)
 	resp.Body.Close()
 
 	return resp.StatusCode >= 200 && resp.StatusCode < 300