perf(graph): warm schema cache on startup to kill cold-start spikes

Following the schema cache PR, warm pods serve from cache (~24/25 hits
on a long-running pod). New pods, however, start cold: the first
LatestSchema query per (orgId, ref) still runs the wgc router compose
subprocess, which costs 100-300m CPU per call.

That cold-start cost is what kept tripping the HPA into TooManyReplicas:
HPA scales up → new pod added → new pod runs wgc on first query →
metrics spike → HPA scales up further → cycle repeats. Even after the
caching PR landed, observed pods cycling 2→4→2→4 in production, with
fresh pods showing 2 'Fetching latest schema' (cold) entries and 0
cache hits within their first minute.

Add Cache.AllOrgRefs() exposing every tracked (orgId, ref) pair, and
Resolver.WarmCache(ctx) which iterates them after the event-sourced
caches have been populated. For each ref it fetches the subgraphs,
runs sdlmerge, runs CosmoGenerator.Generate, and stores both results
in the cache. Errors per ref are logged and skipped so a single bad
ref does not block warming the rest.

Service startup calls WarmCache right after the Resolver is wired,
before the HTTP server starts accepting traffic, so the first
LatestSchema query a pod receives is already a cache hit.

cache/cache.go

@@ -102,6 +102,30 @@ func (c *Cache) Services(orgId, ref, lastUpdate string) ([]string, string) {
 	return services, c.lastUpdate[key]
 }
 
+// OrgRef identifies a single (organizationId, ref) pair that the cache
+// tracks subgraphs for.
+type OrgRef struct {
+	OrgId string
+	Ref   string
+}
+
+// AllOrgRefs returns every (orgId, ref) pair that currently has at least
+// one subgraph in the cache. Used by startup warmup to pre-compute the
+// merged SDL and SchemaUpdate for every known ref before the pod starts
+// serving traffic.
+func (c *Cache) AllOrgRefs() []OrgRef {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+
+	var out []OrgRef
+	for orgId, refs := range c.services {
+		for ref := range refs {
+			out = append(out, OrgRef{OrgId: orgId, Ref: ref})
+		}
+	}
+	return out
+}
+
 func (c *Cache) SubGraphId(orgId, ref, service string) string {
 	c.mu.RLock()
 	defer c.mu.RUnlock()

cmd/service/service.go

@@ -210,6 +210,8 @@ func start(closeEvents chan error, logger *slog.Logger, connectToAmqpFunc func(u
 			Debouncer:      graph.NewDebouncer(500 * time.Millisecond),
 		}
 
+		resolver.WarmCache(rootCtx)
+
 		config := generated.Config{
 			Resolvers:  resolver,
 			Complexity: generated.ComplexityRoot{},

graph/resolver.go

@@ -8,7 +8,9 @@ import (
 	"gitlab.com/unboundsoftware/eventsourced/eventsourced"
 
 	"gitea.unbound.se/unboundsoftware/schemas/cache"
+	"gitea.unbound.se/unboundsoftware/schemas/graph/model"
 	"gitea.unbound.se/unboundsoftware/schemas/middleware"
+	"gitea.unbound.se/unboundsoftware/schemas/sdlmerge"
 )
 
 //go:generate go run github.com/99designs/gqlgen
@@ -60,3 +62,60 @@ func (r *Resolver) handler(ctx context.Context, aggregate eventsourced.Aggregate
 func apiKeyId(orgId, name string) string {
 	return fmt.Sprintf("%s-%s", orgId, name)
 }
+
+// WarmCache precomputes the merged SDL and SchemaUpdate (cosmo router
+// config) for every (orgId, ref) tracked in the cache. Intended to run
+// once at startup, after the event-sourced caches have been populated
+// but before the pod accepts traffic, so the first request per ref does
+// not pay the cold-start cost of running sdlmerge + wgc compose.
+//
+// Errors per ref are logged and skipped rather than aborting the whole
+// warmup: a single bad ref must not block the pod from serving the
+// remaining refs.
+func (r *Resolver) WarmCache(ctx context.Context) {
+	refs := r.Cache.AllOrgRefs()
+	r.Logger.Info("Warming schema cache on startup", "refCount", len(refs))
+
+	for _, or := range refs {
+		services, lastUpdate := r.Cache.Services(or.OrgId, or.Ref, "")
+		if len(services) == 0 {
+			continue
+		}
+
+		subGraphs := make([]*model.SubGraph, len(services))
+		serviceSDLs := make([]string, len(services))
+		for i, id := range services {
+			sg, err := r.fetchSubGraph(ctx, id)
+			if err != nil {
+				r.Logger.Error("warmup: fetch subgraph", "error", err, "orgId", or.OrgId, "ref", or.Ref, "id", id)
+				subGraphs = nil
+				break
+			}
+			subGraphs[i] = r.toGqlSubGraph(sg)
+			serviceSDLs[i] = sg.Sdl
+		}
+		if subGraphs == nil {
+			continue
+		}
+
+		if sdl, err := sdlmerge.MergeSDLs(serviceSDLs...); err != nil {
+			r.Logger.Error("warmup: merge SDLs", "error", err, "orgId", or.OrgId, "ref", or.Ref)
+		} else {
+			r.Cache.SetMergedSDL(or.OrgId, or.Ref, lastUpdate, sdl, subGraphs)
+		}
+
+		cosmoConfig, err := r.CosmoGenerator.Generate(ctx, subGraphs)
+		if err != nil {
+			r.Logger.Error("warmup: generate cosmo config", "error", err, "orgId", or.OrgId, "ref", or.Ref)
+			continue
+		}
+		r.Cache.SetSchemaUpdate(or.OrgId, or.Ref, &model.SchemaUpdate{
+			Ref:               or.Ref,
+			ID:                lastUpdate,
+			SubGraphs:         subGraphs,
+			CosmoRouterConfig: &cosmoConfig,
+		})
+	}
+
+	r.Logger.Info("Schema cache warmup complete", "refCount", len(refs))
+}