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.

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