Compute from a build of build-agent only once a toolchain is read, so a restart drops no fingerprint; fixtures as live (issue 1727)

modules/mesh-delivery-indexer/cmd/mesh-delivery-indexer/main.go

@@ -271,7 +271,7 @@ func run(ctx context.Context, getenv func(string) string, logOut io.Writer) erro
 			known.Delete(e.Key())
 		}
 		if ix := started.Load(); ix != nil && strings.HasPrefix(e.Key(), index.BuilderBuiltBy+".build.") {
-			if err := ix.Recompute(ctx); err != nil {
+			if err := ix.RecomputeForBuild(ctx); err != nil {
 				logger.Printf("computing the new-form fingerprints again after %s: %v", e.Key(), err)
 			}
 		}
@@ -284,10 +284,6 @@ func run(ctx context.Context, getenv func(string) string, logOut io.Writer) erro
 	go func() { errs <- follow(ctx, w, each) }()
 
 	ix := &index.Indexer{Store: kvStore{candidates}, Modules: kvStore{modules}, Manifests: kvStore{manifests}, Builds: known, Logf: logger.Printf}
-	started.Store(ix)
-	if err := ix.Recompute(ctx); err != nil { // the builds heard before this, which moved no one
-		logger.Printf("computing the new-form fingerprints at start: %v", err)
-	}
 	// The gate's source bucket (mesh-delivery-gate_source) comes with the gate (release 3e, novox/hq ADR 0450 §1): its key
 	// main is the mesh's source revision, which stamps the modules (issue 1344). Where it does not exist yet the indexer does
 	// not wait for it (that needs a timer, ADR 0437 §8): it says so, and is started again with the gate.
@@ -313,6 +309,12 @@ func run(ctx context.Context, getenv func(string) string, logOut io.Writer) erro
 	if err := tools.start(); err != nil {
 		return err
 	}
+	// Only now, with the toolchains read, are the builds heard to compute from (a build before it would write every Go
+	// module as having no toolchain reported, and back); the builds heard until now are computed once.
+	started.Store(ix)
+	if err := ix.RecomputeForBuild(ctx); err != nil {
+		logger.Printf("computing the new-form fingerprints at start: %v", err)
+	}
 	// A build module says each time it wrote its toolchains key: its stream, and the indexer's durable consumer on it, are
 	// made by the controller as the forge's are. Where the stream is not there yet the indexer says so and is started again
 	// with the build module, as with the gate's source.

modules/mesh-delivery-indexer/internal/index/toolchains.go

@@ -397,3 +397,18 @@ func (ix *Indexer) builderInput() (src, none string) {
 	}
 	return "", fmt.Sprintf("%s, which builds it, has no successful trunk build that recorded a source fingerprint", BuilderBuiltBy)
 }
+
+// RecomputeForBuild is Recompute for a build of BuilderBuiltBy heard on the builds bucket, or the one at start (issue 1727).
+// While the indexer has read no toolchain at all (it has just started and the toolchains buckets are not settled), every
+// Go module would be written as having no toolchain reported and then back: it computes nothing then, and the settling of
+// the toolchains computes them.
+func (ix *Indexer) RecomputeForBuild(ctx context.Context) error {
+	ix.mu.Lock()
+	none := len(ix.toolchains) == 0
+	ix.mu.Unlock()
+	if none {
+		ix.logf("the builds changed before any toolchain was read: the new-form fingerprints are computed when they are")
+		return nil
+	}
+	return ix.Recompute(ctx)
+}

modules/mesh-delivery-indexer/internal/index/toolchains_test.go

@@ -46,7 +46,11 @@ const (
 )
 
 func agentBuild(id, src, at string) [2]any {
-	raw := `{"id":"` + id + `","repository":"` + agentCloneURL + `","ref":"main","module":"build-agent","commit":"cada7e3c",` +
+	commit := "cada7e3c"
+	if id == "build-1792587792086548572" {
+		commit = "f10fbda1"
+	}
+	raw := `{"id":"` + id + `","repository":"` + agentCloneURL + `","ref":"main","module":"build-agent","commit":"` + commit + `",` +
 		`"path":"modules/build-agent","at":"` + at + `","source-fingerprint":"` + src + `","source-inputs":` + aInputs + `}`
 	return [2]any{delivery.BuildKey("build-agent", id), []byte(raw)}
 }
@@ -63,10 +67,7 @@ func newFormIndexer(t *testing.T, mods ...mod) (*Indexer, *memStore) {
 	if err := ix.Apply(context.Background(), f); err != nil {
 		t.Fatal(err)
 	}
-	// build-agent's value of today's form, from a merge of the controller (modules.build-agent holds its fingerprint).
-	if err := ix.Apply(context.Background(), mergedMsg(buildAgentRepo, map[string]string{"modules/build-agent": "agent-tree-1"})); err != nil {
-		t.Fatal(err)
-	}
+	// build-agent holds no fingerprint of today's form, as live (no merge of the controller is applied; issue 1727).
 	return ix, store
 }
 
@@ -189,6 +190,25 @@ func TestBasesAndTheBuilderAreComputedBeforeTheModulesTakingThem(t *testing.T) {
 	}
 }
 
+func TestABuildHeardBeforeAnyToolchainIsReadDropsNoFingerprint(t *testing.T) { // 1727 (review: a restart flapped every FP2)
+	ix, mods := newFormIndexer(t, builderModules(t)...)
+	reportToolchain(t, ix, "mesh-build-go", "novox", "go", "go1.27.1")
+	before, _ := moduleOf(t, mods, "mesh-tools")
+	if before.FP2 == "" {
+		t.Fatalf("mesh-tools has no fp2 to begin with: %+v", before)
+	}
+	// a restarted indexer has read no toolchains bucket yet
+	ix.mu.Lock()
+	ix.toolchains = nil
+	ix.mu.Unlock()
+	if err := ix.RecomputeForBuild(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if m, _ := moduleOf(t, mods, "mesh-tools"); m.FP2 != before.FP2 || m.None2 != "" {
+		t.Errorf("mesh-tools: fp2 = %q none2 = %q, want fp2 %q kept", m.FP2, m.None2, before.FP2)
+	}
+}
+
 func TestTheBuilderHasNoFingerprintWithoutATrunkBuildOfBuildAgentAndSaysSo(t *testing.T) { // 1727
 	ix, mods := newFormIndexer(t, builderModules(t)...)
 	reportToolchain(t, ix, "mesh-build-go", "novox", "go", "go1.27.1")