Routes may share one lease (affinity = "route") and skip crossbar's queue (queue = false)

route.go gains Route.Affinity/Queue with PerRoute(), Queues() and affinity
validation (checkRoutes moved here; config.go calls it once). limiter.Track
counts a request without holding or refusing it; a release hands the slot to a
waiter only while in flight <= parallel. The proxy leases a PerRoute() route
under an empty fingerprint (the row keeps the real one) and uses Track when
Queues() is false.

Implemented by Ornith (OpenCode); owner review removed a release-on-first-flush
workaround for a race in the owner's given test (see implementer log).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

internal/proxy/proxy.go

@@ -217,6 +217,12 @@ func (p *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 	}
 	model = resolveModel(model, r, routeCfg)
 	fp := fingerprint.Of(body)
+	// A route-affinity route puts every request (chat or control) on one lease per model, so the
+	// lease key's fingerprint is "" for all of them; the real fingerprint is kept for the row below.
+	leaseFP := fp
+	if routeCfg.PerRoute() {
+		leaseFP = ""
+	}
 	started := time.Now()
 
 	// v0 compatibility path: no lease table, no limiter, no recording.
@@ -231,14 +237,14 @@ func (p *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 	}
 
 	// Lease. The route's ordered host list is the candidate set.
-	host, reused, err := p.leases.Acquire(lease.Key{Route: route, FP: fp, Model: model}, routeCfg.Hosts, time.Now())
+	host, reused, err := p.leases.Acquire(lease.Key{Route: route, FP: leaseFP, Model: model}, routeCfg.Hosts, time.Now())
 	if err != nil {
 		switch {
 		case errors.Is(err, lease.ErrNoHost):
 			// No host healthy. Ask a waker to rouse a sleeping one; it answers
 			// (served or 503) when it has had a turn, else falls through to the
 			// plain 503.
-			if p.waker != nil && p.wakeOnErrNoHost(w, r, route, routeCfg, rest, model, fp, started, lease.Key{Route: route, FP: fp, Model: model}) {
+			if p.waker != nil && p.wakeOnErrNoHost(w, r, route, routeCfg, rest, model, fp, started, lease.Key{Route: route, FP: leaseFP, Model: model}) {
 				return
 			}
 			p.writeError(w, http.StatusServiceUnavailable, "no healthy host")
@@ -265,10 +271,30 @@ func (p *Handler) serveLeased(w http.ResponseWriter, r *http.Request, route stri
 		p.forward(w, r, route, host, leaseState(reused), rest, fp, model, started, 0, 0, "", true)
 		return
 	}
-	// Slot. A full queue is a 503; a context done while waiting means the client left.
-	release, waited, err := p.lim.Acquire(r.Context(), host, model)
-	if err != nil {
-		if errors.Is(err, limiter.ErrQueueFull) {
+	// Slot or track. A queue = false route leaves queueing to the client's own
+	// llama-server slot: crossbar only counts the request on the host, never
+	// holding it or refusing it.
+	var waited time.Duration
+	var release func()
+	if routeCfg.Queues() {
+		var err error
+		release, waited, err = p.lim.Acquire(r.Context(), host, model)
+		if err != nil {
+			if errors.Is(err, limiter.ErrQueueFull) {
+				p.writeRecord(store.Request{
+					Route:   route,
+					FP:      fp,
+					Model:   model,
+					Host:    host,
+					Started: started,
+					TotalMs: time.Since(started).Milliseconds(),
+					Status:  http.StatusServiceUnavailable,
+					Err:     "queue full",
+				})
+				p.writeError(w, http.StatusServiceUnavailable, "queue full")
+				return
+			}
+			p.log.Warn("request", "route", route, "host", host, "method", r.Method, "path", rest, "status", 499)
 			p.writeRecord(store.Request{
 				Route:   route,
 				FP:      fp,
@@ -276,24 +302,13 @@ func (p *Handler) serveLeased(w http.ResponseWriter, r *http.Request, route stri
 				Host:    host,
 				Started: started,
 				TotalMs: time.Since(started).Milliseconds(),
-				Status:  http.StatusServiceUnavailable,
-				Err:     "queue full",
+				Status:  499,
+				Err:     "client cancelled while queued",
 			})
-			p.writeError(w, http.StatusServiceUnavailable, "queue full")
 			return
 		}
-		p.log.Warn("request", "route", route, "host", host, "method", r.Method, "path", rest, "status", 499)
-		p.writeRecord(store.Request{
-			Route:   route,
-			FP:      fp,
-			Model:   model,
-			Host:    host,
-			Started: started,
-			TotalMs: time.Since(started).Milliseconds(),
-			Status:  499,
-			Err:     "client cancelled while queued",
-		})
-		return
+	} else {
+		release = p.lim.Track(host, model)
 	}
 	defer release()