fix(serving): confirm ONNX session before status

Add a status reconciliation guard to the ONNX serving adapter. After the Rust ONNX session manager reports a successful load, the adapter now lists sessions and verifies the requested model is available before recording backend served status.

If the session list does not include the model or cannot be read, serving returns a non-critical provider-load failure instead of publishing stale loaded state. This keeps durable serving state behind the provider-owned session truth.

Update the ONNX Runtime serving plan with M4 status, verification evidence, and remaining serving gaps.

Verification: cargo fmt --manifest-path rust/Cargo.toml --all -- --check; cargo test --manifest-path rust/crates/pumas-rpc/Cargo.toml serving; cargo test --manifest-path rust/crates/pumas-core/Cargo.toml serving

Author: MrScripty

docs/plans/onnx-runtime-embedding-serving/execution-and-coordination.md

@@ -454,6 +454,15 @@ Update during implementation:
   secrets, or full model paths. Verification passed:
   `cargo fmt --manifest-path rust/Cargo.toml --all -- --check` and
   `cargo test --manifest-path rust/crates/pumas-rpc/Cargo.toml openai_gateway`.
+- 2026-05-11: Added an ONNX serving status reconciliation guard before durable
+  served-state updates. After the Rust ONNX session manager reports a successful
+  load, `serving_onnx.rs` now lists sessions and verifies the requested model is
+  present before calling `record_served_model`; a mismatch returns a
+  non-critical provider-load failure instead of publishing stale loaded status.
+  Verification passed:
+  `cargo fmt --manifest-path rust/Cargo.toml --all -- --check`,
+  `cargo test --manifest-path rust/crates/pumas-rpc/Cargo.toml serving`, and
+  `cargo test --manifest-path rust/crates/pumas-core/Cargo.toml serving`.
 
 ## Commit Cadence Notes
 

docs/plans/onnx-runtime-embedding-serving/milestones.md

@@ -578,7 +578,7 @@ state.
 - [ ] Move existing Ollama and llama.cpp serving paths behind provider serving
       adapters before adding ONNX load/unload so the RPC handler only performs
       boundary parsing, validation orchestration, and response shaping.
-- [ ] Confirm the Rust ONNX provider status/list includes the model before
+- [x] Confirm the Rust ONNX provider status/list includes the model before
       recording loaded status.
 - [x] Add unload support through the Rust ONNX session manager and served
       status removal.
@@ -608,9 +608,11 @@ selected ONNX profile is running and the primary executable artifact is `.onnx`.
 Provider behavior drives ONNX artifact compatibility, and ONNX rejects
 llama.cpp-specific placement overrides with non-critical domain errors. The RPC
 serving boundary now loads/unloads ONNX through the Rust fake session manager
-and records/removes backend served status. Real ONNX Runtime execution,
-duplicate load/unload idempotency, session status reconciliation before record,
-and gateway embedding routing remain open.
+and records/removes backend served status. The ONNX serving adapter now
+confirms the Rust session manager lists the loaded model before recording
+backend served status. Real ONNX Runtime execution, duplicate load/unload
+idempotency, and route/profile fallback cleanup remain open. Gateway embedding
+routing has started under Milestone 5.
 
 ### Milestone 5: Pumas Gateway Routing
 

rust/crates/pumas-rpc/src/handlers/serving_onnx.rs

@@ -50,6 +50,7 @@ pub(super) async fn serve_onnx_model(
         }
     };
 
+    let onnx_model_id = load_request.model_id.clone();
     let session = match state.onnx_session_manager.load(load_request).await {
         Ok(session) => session,
         Err(error) => {
@@ -65,6 +66,18 @@ pub(super) async fn serve_onnx_model(
             .await;
         }
     };
+    if let Err(error) = confirm_onnx_session_loaded(state, &onnx_model_id).await {
+        warn!("ONNX fake session status confirmation failed: {}", error);
+        return non_critical_failure_response(
+            state,
+            serving_error(
+                ModelServeErrorCode::ProviderLoadFailed,
+                "ONNX Runtime loaded the selected model but did not report it as available",
+                &request,
+            ),
+        )
+        .await;
+    }
 
     let status = ServedModelStatus {
         model_id: request.model_id.clone(),
@@ -176,3 +189,24 @@ async fn resolve_onnx_model_path(
     }
     Ok(Some(onnx_path))
 }
+
+async fn confirm_onnx_session_loaded(
+    state: &AppState,
+    model_id: &OnnxModelId,
+) -> Result<(), String> {
+    let sessions = state
+        .onnx_session_manager
+        .list()
+        .await
+        .map_err(|error| error.to_string())?;
+    if sessions
+        .iter()
+        .any(|session| session.model_id.as_str() == model_id.as_str())
+    {
+        return Ok(());
+    }
+    Err(format!(
+        "ONNX model '{}' was absent from session list after load",
+        model_id.as_str()
+    ))
+}