Wire wake and identity into crossbar; v2 smoke and README

Implemented-By: OpenCode session (model recorded in docs/implementer-log.md)

internal/proxy/ctxguard.go

@@ -69,6 +69,35 @@ func (p *Handler) guard(w http.ResponseWriter, r *http.Request, hosts []string,
 		return newHost, movedHeader(host, newHost), estimate, false
 	}
 
+	// Rule 4a: no host fits. Before refusing, ask a waker to rouse a candidate
+	// whose context may grow when it comes up; a woken host takes the lease.
+	if p.waker != nil {
+		for _, name := range hosts {
+			s, ok := p.health.Get(name)
+			if !ok {
+				continue
+			}
+			// A healthy host does not need waking; only a down host might grow
+			// a larger context when it comes up.
+			if s.Healthy {
+				continue
+			}
+			// A host with a known per-slot context smaller than the estimate
+			// cannot serve it no matter how it wakes.
+			if psc := s.PerSlotCtx(); psc != 0 && psc < estimate {
+				continue
+			}
+			if p.cfg.Hosts[name].Wake == nil || !p.waker.Wake(r.Context(), name) {
+				continue
+			}
+			if err := p.leases.Move(lease.Key{Route: route, FP: fp, Model: model}, name, time.Now()); err != nil {
+				p.writeError(w, http.StatusBadGateway, "upstream failed")
+				return host, "", estimate, true
+			}
+			return name, movedHeader(host, name), estimate, false
+		}
+	}
+
 	// Rule 4: no host fits. Answer 400 with the estimate and the largest
 	// available per-slot context, and record the row.
 	p.refuseCtx(w, host, route, model, fp, started, estimate, largestSlotCtx(hosts, p.health))