Make aggregate ONNX gates execute for real

Reference-count the process-global ONNX Runtime across embedder and routing-head sessions, make close idempotent, and require named proof that both aggregate routing gates executed rather than self-skipped (V-716). Owner explicitly requested direct commits to master.

JOURNAL.md

@@ -210,3 +210,20 @@ pops only the active top before resuming the flow underneath.
 passes (162.310s for mavend; every package green). Focused structural
 possession, repair-pointer, nested-stack, and repaired-clarify tests pass under
 the race detector.
+
+### ONNX test/runtime lifecycle
+
+V-716 found that each embedder constructor tried to initialize ONNX Runtime,
+while `Close` destroyed only its session. In one package process the first
+model-aware test ran and later tests converted “already initialized” into a
+green skip. The router now owns the process-global environment through
+reference-counted leases held by each embedder and routing-head session; the
+last owned lease performs cleanup, and close is idempotent.
+
+The router and mavend test packages hold a lease across their model gates.
+`make eval-router` additionally requires proof that both named aggregate gates
+actually executed. In one process the classifier baseline scored 72/96 and the
+routing heads 93/96; destination was 11/33 and 25/33 respectively, and ecosystem
+reach remained 28/30. The lifecycle reacquire test, focused race suite, full
+aggregate command, and portable no-runtime packages all pass. Measurement:
+`docs/evals/2026-08-13-onnx-runtime-lifecycle.md`.

Makefile

@@ -228,11 +228,15 @@ t:
 # Verbose so the report tables land in the terminal. MAVEN_ONNX_LIB points the
 # prod-representative baseline at the vendored runtime; override it or set it
 # empty to run only the deterministic hash ratchet. This is the measurement
-# Vikunja #319 compares before #320 flips the route decider.
+# Vikunja #319 compares before #320 flips the route decider. With a non-empty
+# runtime path the package must prove that at least two model gates executed;
+# a constructor skip after the first process-global initialization is a failure.
 MAVEN_ONNX_LIB ?= $(shell pwd)/deps/onnxruntime-linux-x64-1.26.0/lib/libonnxruntime.so
 
 eval-router:
-	MAVEN_ONNX_LIB="$(MAVEN_ONNX_LIB)" $(GO) test -v -count=1 ./internal/router/eval/
+	MAVEN_ONNX_LIB="$(MAVEN_ONNX_LIB)" \
+	MAVEN_ONNX_REQUIRED_GATES="$(if $(MAVEN_ONNX_LIB),2,0)" \
+		$(GO) test -v -count=1 ./internal/router/eval/
 
 # eval-reach — score the held-out ecosystem reach fixture (internal/router/eval,
 # ru_ecosystem_v1.json). Answers "does a real Russian utterance actually arrive

cmd/mavend/onnx_testmain_test.go

@@ -0,0 +1,36 @@
+package main
+
+import (
+	"fmt"
+	"os"
+	"testing"
+
+	"github.com/kami/maven/internal/router"
+)
+
+// TestMain holds one ONNX Runtime lease across the model-aware topic and
+// personal-boundary gates. Each test still owns and closes its model session;
+// the process-global environment is released only after the final test.
+func TestMain(m *testing.M) {
+	var lease *router.ONNXRuntimeLease
+	lib := os.Getenv("MAVEN_ONNX_LIB")
+	if lib != "" {
+		if _, err := os.Stat(lib); err == nil {
+			var acquireErr error
+			lease, acquireErr = router.AcquireONNXRuntime(lib)
+			if acquireErr != nil {
+				fmt.Fprintf(os.Stderr, "initialize shared ONNX test runtime: %v\n", acquireErr)
+				os.Exit(2)
+			}
+		}
+	}
+
+	code := m.Run()
+	if lease != nil {
+		if err := lease.Close(); err != nil {
+			fmt.Fprintf(os.Stderr, "ONNX test runtime cleanup: %v\n", err)
+			code = 1
+		}
+	}
+	os.Exit(code)
+}

docs/evals/2026-08-13-onnx-runtime-lifecycle.md

@@ -0,0 +1,65 @@
+# ONNX aggregate gate lifecycle — 2026-08-13
+
+Vikunja: V-716.
+
+## Finding
+
+`router.NewONNXEmbedder` initialized `onnxruntime_go` unconditionally. Its
+`Close` method destroyed the model session but left the package-global runtime
+environment alive. In one Go process the first `TestONNX*` therefore ran and
+every later constructor returned `The onnxruntime has already been initialized`.
+Those tests converted the constructor error into `t.Skip`, so the aggregate
+package still reported `PASS`.
+
+Reproduction before the fix:
+
+```text
+TestONNXBaseline     PASS  classifier+onnx 72/96
+TestONNXRoutingHeads SKIP  onnx: init environment: already initialized
+package              PASS
+```
+
+## Contract
+
+The router now owns ONNX Runtime through reference-counted leases. Every model
+session holds one lease; routing heads hold their own, so daemon shutdown order
+cannot unload the library underneath a live graph. The last lease destroys an
+environment Maven initialized. Package `TestMain` holds one extra lease across
+all model-aware tests in `internal/router/eval` and `cmd/mavend`.
+
+The aggregate router command additionally requires two named gates to record
+execution only after all of their model dependencies loaded. A second-test
+skip can no longer satisfy the command.
+
+## Measurement
+
+Runtime and model:
+
+```text
+deps/onnxruntime-linux-x64-1.26.0/lib/libonnxruntime.so.1.26.0
+models/embedder/multilingual-e5-small/model_quantized.onnx
+models/embedder/router-heads/router_heads.onnx
+```
+
+Focused aggregate, one process:
+
+```text
+TestONNXBaseline     PASS  72/96 full (75.0%), destination 11/33
+TestONNXRoutingHeads PASS  93/96 full (96.9%), destination 25/33
+ONNX aggregate proof: verified 2 required model gates:
+  [TestONNXBaseline TestONNXRoutingHeads]
+```
+
+The separate lifecycle test released the last lease and successfully acquired
+the runtime again. A second same-process package run also executed all three
+selected mavend gates without a skip:
+
+```text
+TestONNXPersonalBoundary         PASS  29/29
+TestONNXPersonalBoundaryFourFold PASS  99/104
+TestONNXTopics                   PASS  43/43
+```
+
+Finally, `make eval-router` completed the full package. In addition to the two
+required proof gates, `TestONNXClaimConfidenceDistribution` and
+`TestReachBaselineONNX` executed; ecosystem reach remained 28/30.

docs/evals/CLAUDE.md

@@ -110,6 +110,7 @@ A pair in `docs/routing.md` went stale unnoticed. Its source predated the
 
 | measurement | state |
 | --- | --- |
+| [Both ONNX routing gates execute in one process](2026-08-13-onnx-runtime-lifecycle.md) | live |
 | [Where the resident model's 7.9GB of RSS goes](2026-08-03-llama-prompt-cache.md) | live |
 | [Does one sqlite connection make reads queue? No](2026-08-07-store-connection-cap.md) | live |
 

docs/qa.md

@@ -301,6 +301,23 @@ make eval-recall
 
 A large miss against 72.7% means the deploy differs from the bench harness.
 
+`make eval-router` is also the aggregate ONNX lifecycle gate. When the default
+runtime path is non-empty it requires both `TestONNXBaseline` and
+`TestONNXRoutingHeads` to load their models and execute in the same Go process.
+The final line names both verified gates. A later test that skips because the
+first one consumed the process-global runtime now makes the command fail rather
+than leaving a green package result. To exercise the proof directly:
+
+```sh
+MAVEN_ONNX_LIB="$PWD/deps/onnxruntime-linux-x64-1.26.0/lib/libonnxruntime.so" \
+MAVEN_ONNX_REQUIRED_GATES=2 \
+  go test -v -count=1 \
+  -run '^(TestONNXBaseline|TestONNXRoutingHeads)$' ./internal/router/eval/
+```
+
+Set `MAVEN_ONNX_LIB=` explicitly when the intended measurement is the portable
+hash floor; that also disables the aggregate ONNX requirement.
+
 **Run on 02-08-2026 @ af9d213. The deploy matches the bench.** `eval-models`
 scored 56 of 77: 72.7% full, 77.9% intent-only, 2 false clarifies and 1 missed.
 That is the recorded figure to the decimal, and calendar sat at 2 of 2, so the

internal/router/eval/eval_test.go

@@ -197,6 +197,7 @@ func TestONNXBaseline(t *testing.T) {
 		t.Skipf("onnx embedder unavailable: %v", err)
 	}
 	defer emb.Close()
+	recordONNXGateExecuted(t)
 
 	f, err := Load()
 	if err != nil {

internal/router/eval/heads_test.go

@@ -54,6 +54,7 @@ func TestONNXRoutingHeads(t *testing.T) {
 		t.Skipf("routing heads unavailable: %v", err)
 	}
 	defer h.Close()
+	recordONNXGateExecuted(t)
 
 	f, err := Load()
 	if err != nil {

internal/router/eval/onnx_testmain_test.go

@@ -0,0 +1,99 @@
+package eval
+
+import (
+	"fmt"
+	"os"
+	"sort"
+	"strconv"
+	"sync"
+	"testing"
+
+	"github.com/kami/maven/internal/router"
+)
+
+const requiredONNXGatesEnv = "MAVEN_ONNX_REQUIRED_GATES"
+
+var executedONNXGates struct {
+	sync.Mutex
+	names []string
+}
+
+// recordONNXGateExecuted is called only after a gate has loaded every model it
+// needs. TestMain uses the names as proof that an aggregate command did not
+// turn a second initialization error into a green SKIP.
+func recordONNXGateExecuted(t *testing.T) {
+	t.Helper()
+	executedONNXGates.Lock()
+	executedONNXGates.names = append(executedONNXGates.names, t.Name())
+	executedONNXGates.Unlock()
+}
+
+func TestMain(m *testing.M) {
+	required, err := requiredONNXGateCount()
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(2)
+	}
+
+	lease, err := configuredONNXTestRuntime(required > 0)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(2)
+	}
+
+	code := m.Run()
+	if required > 0 {
+		executedONNXGates.Lock()
+		names := append([]string(nil), executedONNXGates.names...)
+		executedONNXGates.Unlock()
+		sort.Strings(names)
+		fmt.Fprintf(os.Stderr, "ONNX aggregate proof: verified %d required model gates: %v\n", len(names), names)
+		if len(names) < required {
+			fmt.Fprintf(os.Stderr, "ONNX aggregate gate failed: executed %d, require at least %d\n", len(names), required)
+			code = 1
+		}
+	}
+
+	if lease != nil {
+		if err := lease.Close(); err != nil {
+			fmt.Fprintf(os.Stderr, "ONNX test runtime cleanup: %v\n", err)
+			code = 1
+		}
+	}
+	os.Exit(code)
+}
+
+func requiredONNXGateCount() (int, error) {
+	raw := os.Getenv(requiredONNXGatesEnv)
+	if raw == "" {
+		return 0, nil
+	}
+	n, err := strconv.Atoi(raw)
+	if err != nil || n < 0 {
+		return 0, fmt.Errorf("%s must be a non-negative integer, got %q", requiredONNXGatesEnv, raw)
+	}
+	return n, nil
+}
+
+func configuredONNXTestRuntime(required bool) (*router.ONNXRuntimeLease, error) {
+	lib := os.Getenv("MAVEN_ONNX_LIB")
+	if lib == "" {
+		if required {
+			return nil, fmt.Errorf("ONNX aggregate gate requires MAVEN_ONNX_LIB")
+		}
+		return nil, nil
+	}
+	if _, err := os.Stat(lib); err != nil {
+		if required {
+			return nil, fmt.Errorf("ONNX aggregate gate runtime %s: %w", lib, err)
+		}
+		// Preserve the ordinary portable test path: individual model tests
+		// report a skip when optional local model dependencies are absent.
+		return nil, nil
+	}
+	lease, err := router.AcquireONNXRuntime(lib)
+	if err != nil {
+		return nil, fmt.Errorf("initialize shared ONNX test runtime: %w", err)
+	}
+	return lease, nil
+}

internal/router/heads.go

@@ -7,6 +7,7 @@ import (
 	"math"
 	"os"
 	"path/filepath"
+	"sync"
 
 	ort "github.com/yalue/onnxruntime_go"
 )
@@ -54,9 +55,12 @@ const (
 type RouterHeads struct {
 	tokenizer *unigramTokenizer
 	session   *ort.DynamicSession[int64, float32]
+	runtime   *ONNXRuntimeLease
 	intents   []Intent
 	sources   []Source
 	threshold float64
+	closeOnce sync.Once
+	closeErr  error
 }
 
 // headsMeta — router_heads.json, written beside the weights by the exporter.
@@ -70,8 +74,9 @@ type headsMeta struct {
 // NewRouterHeads loads the graph and its label order. modelPath points at the
 // .onnx; the external weights and router_heads.json sit beside it.
 //
-// It assumes the ONNX environment is already initialised, because the embedder
-// does that at startup and the runtime allows it once.
+// It shares the process-global ONNX environment with the embedder. The runtime
+// lease is independent so shutdown order cannot unload the library while this
+// graph's session is still alive.
 func NewRouterHeads(modelPath, tokenizerPath string) (*RouterHeads, error) {
 	metaPath := filepath.Join(filepath.Dir(modelPath), "router_heads.json")
 	raw, err := os.ReadFile(metaPath)
@@ -106,18 +111,24 @@ func NewRouterHeads(modelPath, tokenizerPath string) (*RouterHeads, error) {
 	if err != nil {
 		return nil, fmt.Errorf("heads: tokenizer: %w", err)
 	}
+	runtime, err := AcquireONNXRuntime("")
+	if err != nil {
+		return nil, fmt.Errorf("heads: %w", err)
+	}
 	session, err := ort.NewDynamicSession[int64, float32](
 		modelPath,
 		[]string{"input_ids", "attention_mask"},
 		[]string{"intent", "source", "slots", "clarify"},
 	)
 	if err != nil {
+		_ = runtime.Close()
 		return nil, fmt.Errorf("heads: create session: %w", err)
 	}
 
 	return &RouterHeads{
 		tokenizer: tok,
 		session:   session,
+		runtime:   runtime,
 		intents:   intents,
 		sources:   sources,
 		threshold: headsThreshold,
@@ -128,8 +139,14 @@ func (h *RouterHeads) Close() error {
 	if h == nil {
 		return nil
 	}
+	h.closeOnce.Do(func() {
+		if h.session != nil {
 			h.session.Destroy()
-	return nil
+			h.session = nil
+		}
+		h.closeErr = h.runtime.Close()
+	})
+	return h.closeErr
 }
 
 // headsResult — one forward pass, read back.

internal/router/onnxembedder.go

@@ -7,6 +7,7 @@ import (
 	"math"
 	"os"
 	"strings"
+	"sync"
 
 	ort "github.com/yalue/onnxruntime_go"
 	"golang.org/x/text/unicode/norm"
@@ -33,17 +34,21 @@ const (
 type onnxEmbedder struct {
 	tokenizer *unigramTokenizer
 	session   *ort.DynamicSession[int64, float32]
+	runtime   *ONNXRuntimeLease
 	id        string
+	closeOnce sync.Once
+	closeErr  error
 }
 
 func NewONNXEmbedder(modelPath, tokenizerPath, libPath string) (*onnxEmbedder, error) {
-	ort.SetSharedLibraryPath(libPath)
-	if err := ort.InitializeEnvironment(); err != nil {
-		return nil, fmt.Errorf("onnx: init environment: %w", err)
+	runtime, err := AcquireONNXRuntime(libPath)
+	if err != nil {
+		return nil, err
 	}
 
 	tok, err := newUnigramTokenizer(tokenizerPath)
 	if err != nil {
+		_ = runtime.Close()
 		return nil, fmt.Errorf("tokenizer: %w", err)
 	}
 
@@ -53,12 +58,14 @@ func NewONNXEmbedder(modelPath, tokenizerPath, libPath string) (*onnxEmbedder, e
 		[]string{"last_hidden_state"},
 	)
 	if err != nil {
+		_ = runtime.Close()
 		return nil, fmt.Errorf("onnx: create session: %w", err)
 	}
 
 	return &onnxEmbedder{
 		tokenizer: tok,
 		session:   session,
+		runtime:   runtime,
 		id:        modelIDFromPath(modelPath),
 	}, nil
 }
@@ -148,8 +155,17 @@ func (e *onnxEmbedder) embed(ctx context.Context, text string) ([]float32, error
 }
 
 func (e *onnxEmbedder) Close() error {
-	e.session.Destroy()
+	if e == nil {
 		return nil
+	}
+	e.closeOnce.Do(func() {
+		if e.session != nil {
+			e.session.Destroy()
+			e.session = nil
+		}
+		e.closeErr = e.runtime.Close()
+	})
+	return e.closeErr
 }
 
 func meanPool(hidden []float32, mask []int64, seqLen, dim int) []float32 {

internal/router/onnxruntime.go

@@ -0,0 +1,124 @@
+package router
+
+import (
+	"fmt"
+	"path/filepath"
+	"sync"
+
+	ort "github.com/yalue/onnxruntime_go"
+)
+
+// ONNXRuntimeLease keeps the process-global ONNX Runtime environment alive.
+//
+// onnxruntime_go exposes the environment as package-global state: it may be
+// initialized only once, and destroying it while any session is alive is
+// unsafe. A lease makes that ownership explicit. Model wrappers hold one for
+// the lifetime of their session; TestMain may hold an additional lease so a
+// package's model-aware gates share one environment instead of repeatedly
+// loading and unloading the shared library.
+type ONNXRuntimeLease struct {
+	once sync.Once
+	err  error
+}
+
+var sharedONNXRuntime = struct {
+	sync.Mutex
+	refs    int
+	libPath string
+	managed bool
+}{}
+
+// AcquireONNXRuntime acquires the process-global ONNX Runtime environment.
+// The first caller initializes it; later callers share it. A non-empty library
+// path must agree with the path that initialized the active environment.
+//
+// The external-environment branch is intentional. It keeps this package safe
+// when an executable initialized onnxruntime_go directly before constructing
+// a router model; in that case the lease never destroys state it does not own.
+func AcquireONNXRuntime(libPath string) (*ONNXRuntimeLease, error) {
+	libPath = cleanONNXLibraryPath(libPath)
+
+	sharedONNXRuntime.Lock()
+	defer sharedONNXRuntime.Unlock()
+
+	if sharedONNXRuntime.refs > 0 {
+		if err := compatibleONNXLibraryPath(sharedONNXRuntime.libPath, libPath); err != nil {
+			return nil, err
+		}
+		sharedONNXRuntime.refs++
+		return &ONNXRuntimeLease{}, nil
+	}
+
+	if ort.IsInitialized() {
+		// Some other component owns the already-live environment. Attach to it,
+		// but never unload its shared library when our last lease closes.
+		sharedONNXRuntime.refs = 1
+		sharedONNXRuntime.libPath = libPath
+		sharedONNXRuntime.managed = false
+		return &ONNXRuntimeLease{}, nil
+	}
+
+	if libPath != "" {
+		ort.SetSharedLibraryPath(libPath)
+	}
+	if err := ort.InitializeEnvironment(); err != nil {
+		return nil, fmt.Errorf("onnx: init environment: %w", err)
+	}
+	sharedONNXRuntime.refs = 1
+	sharedONNXRuntime.libPath = libPath
+	sharedONNXRuntime.managed = true
+	return &ONNXRuntimeLease{}, nil
+}
+
+// Close releases one environment lease. The last lease destroys an
+// environment initialized by this package, after every model session that
+// held a lease has already been destroyed.
+func (l *ONNXRuntimeLease) Close() error {
+	if l == nil {
+		return nil
+	}
+	l.once.Do(func() {
+		l.err = releaseONNXRuntime()
+	})
+	return l.err
+}
+
+func releaseONNXRuntime() error {
+	sharedONNXRuntime.Lock()
+	defer sharedONNXRuntime.Unlock()
+
+	if sharedONNXRuntime.refs == 0 {
+		return fmt.Errorf("onnx: release environment without a lease")
+	}
+	sharedONNXRuntime.refs--
+	if sharedONNXRuntime.refs != 0 {
+		return nil
+	}
+
+	managed := sharedONNXRuntime.managed
+	sharedONNXRuntime.libPath = ""
+	sharedONNXRuntime.managed = false
+	if !managed {
+		return nil
+	}
+	if err := ort.DestroyEnvironment(); err != nil {
+		return fmt.Errorf("onnx: destroy environment: %w", err)
+	}
+	return nil
+}
+
+func cleanONNXLibraryPath(path string) string {
+	if path == "" {
+		return ""
+	}
+	return filepath.Clean(path)
+}
+
+func compatibleONNXLibraryPath(active, requested string) error {
+	// Empty means "use the environment that is already active". RouterHeads
+	// intentionally has no second copy of the embedder's library-path config.
+	if active == "" || requested == "" || active == requested {
+		return nil
+	}
+	return fmt.Errorf("onnx: runtime already uses shared library %q, cannot also use %q", active, requested)
+}

internal/router/onnxruntime_test.go

@@ -0,0 +1,39 @@
+package router
+
+import (
+	"os"
+	"testing"
+)
+
+// TestONNXRuntimeCanBeReacquired pins the process-global lifecycle itself.
+// The model gates exercise live sessions; this smaller check proves that the
+// last lease unloads a managed environment and that a later gate can acquire
+// it again. It is opt-in for the same reason as every model measurement.
+func TestONNXRuntimeCanBeReacquired(t *testing.T) {
+	lib := os.Getenv("MAVEN_ONNX_LIB")
+	if lib == "" {
+		t.Skip("MAVEN_ONNX_LIB unset — see AGENTS.md § Embedder model for intent routing")
+	}
+	if _, err := os.Stat(lib); err != nil {
+		t.Skipf("missing %s: %v", lib, err)
+	}
+
+	first, err := AcquireONNXRuntime(lib)
+	if err != nil {
+		t.Fatalf("first acquire: %v", err)
+	}
+	if err := first.Close(); err != nil {
+		t.Fatalf("first close: %v", err)
+	}
+	if err := first.Close(); err != nil {
+		t.Fatalf("idempotent close: %v", err)
+	}
+
+	second, err := AcquireONNXRuntime(lib)
+	if err != nil {
+		t.Fatalf("reacquire after last close: %v", err)
+	}
+	if err := second.Close(); err != nil {
+		t.Fatalf("second close: %v", err)
+	}
+}