[Router] Scope signal evaluation to authz-eligible decisions

[duchangyong]

Closes #2647

Purpose

• Evaluate the authz signal before the remaining routing signals.
• Conservatively partially evaluate candidate decision rule trees with known authz facts and remove only decisions that are definitely false.
• Derive non-authz signal usage from the retained candidates, then pass the same candidate set to the final decision engine so computation and result scopes stay aligned.
• Preserve existing all-signal behavior when forceEvaluateAll is enabled and preserve the default-model fallback when authz removes every auto-routing candidate.
• Affects: Router.

Test Plan

• Run focused classifier and extproc package tests:
  • go test ./pkg/classification ./pkg/extproc -count=1
• Run the complete Semantic Router Go suite:
  • make test-semantic-router
• Run repository structural lint, config-contract checks, and the fast test gate:
  • make agent-ci-gate ENV=cpu AGENT_BASE_REF=origin/main AGENT_SKIP_PRECOMMIT_BASELINE=1 ...
• Review formatting and patch integrity:
  • git diff --check
• Build a branch image, load it into the existing Podman-backed Kind cluster, and run a streaming-shaped ExtProc request against the existing model PVC:
  • verify the running image ID and /app/router-candle SHA
  • verify fixed-general route selection for authz group maas-route-9120
  • verify signal metrics contain only the authz-eligible decision's authz, jailbreak, and pii classifiers

The unit tests cover conservative AND/OR/NOT partial evaluation, malformed-rule retention, case-sensitive authz role matching, decision-scoped signal discovery, skipped irrelevant classifier work, force-all compatibility, extproc candidate propagation, and empty-candidate auto-model fallback.

Test Result

• go test ./pkg/classification ./pkg/extproc -count=1: passed.
• make test-semantic-router: passed.
• Agent CI gate: passed Go structural lint, reference-config contract checks, structure checks, and the full fast test command.
• git diff --check: passed.
• Existing-cluster runtime smoke: passed with branch image localhost/semantic-router-extproc:issue-2647-10826745 (image ID sha256:1875b27085bc4e4565f5924aef48f707daceeb49b9a36a87d82a66fade9fc954) and verified /app/router-candle SHA 7c5c8e1e8e375bd055d34963c2252e2d89c476ceb7b65098f3d75313e5d6f0fc.
• A streaming-shaped request (stream: true) for authz group maas-route-9120 selected route maas-route-9120, scene general, and model maas-model-rel-91201.
• Runtime metrics reported only authz, jailbreak, and pii signal extraction; unrelated configured signals such as keyword, embedding, domain, context, complexity, modality, and fact-check were not evaluated.
• The smoke test reused the existing kind-cluster, Deployment, and mounted model PVC without downloading models. Its cleanup restored the original localhost/semantic-router-extproc:route-scope-20260723 image and Ready replica.

---

Semantic Router PR Checklist

• PR title uses module-aligned prefixes such as [Router], [CLI], [Dashboard], [Operator], [Fleet-Sim], [Bindings], [Training], [E2E], [Docs], or [CI/Build]
• If the PR spans multiple modules, the title includes all relevant prefixes
• Commits in this PR are signed off with git commit -s
• The Purpose, Test Plan, and Test Result sections reflect the actual scope, commands, and blockers for this change

[netlify]

✅ Deploy Preview for vllm-semantic-router ready!

Name	Link
🔨 Latest commit	3ad3673
🔍 Latest deploy log	https://app.netlify.com/projects/vllm-semantic-router/deploys/6a6e0d74b3815200088b4385
😎 Deploy Preview	https://deploy-preview-2648--vllm-semantic-router.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.
🤖 Make changes	Run an agent on this branch

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[github-actions]

👥 vLLM Semantic Team Notification

The following members have been identified for the changed files in this PR and have been automatically assigned when their GitHub accounts are assignable in this repository:

📁 e2e

Owners: @FAUST-BENCHOU, @shraderdm, @drivebyer, @ramkrishs, @WUKUNTAI-0211, @AayushSaini101, @siloteemu
Files changed:

• e2e/config/config.authz-rbac.yaml
• e2e/profiles/authz-rbac/README.md
• e2e/profiles/authz-rbac/profile.go
• e2e/profiles/authz-rbac/values.yaml
• e2e/testcases/authz_signal_scope.go
• e2e/testcases/authz_signal_scope_test.go

📁 src/semantic-router

Owners: @FAUST-BENCHOU, @shraderdm, @drivebyer, @ramkrishs, @WUKUNTAI-0211, @AayushSaini101, @siloteemu
Files changed:

• src/semantic-router/pkg/classification/classifier_authz_scope.go
• src/semantic-router/pkg/classification/classifier_authz_scope_test.go
• src/semantic-router/pkg/classification/classifier_signal_authz.go
• src/semantic-router/pkg/classification/classifier_signal_context.go
• src/semantic-router/pkg/classification/classifier_signal_context_test.go
• src/semantic-router/pkg/classification/classifier_signal_usage.go
• src/semantic-router/pkg/extproc/req_filter_classification.go
• src/semantic-router/pkg/extproc/req_filter_classification_runtime.go
• src/semantic-router/pkg/extproc/req_filter_classification_signal_test.go

---

🎉 Thanks for your contributions!

This comment was automatically generated based on the OWNER files in the repository.

[github-actions]

✅ Supply Chain Security Report — All Clear

Scanner	Status	Findings
AST Codebase Scan (Py, Go, JS/TS, Rust)	✅	31 finding(s) — MEDIUM: 24 · LOW: 7
AST PR Diff Scan	✅	No issues detected
Regex Fallback Scan	✅	No issues detected

---

Scanned at 2026-08-01T15:16:17.604Z · View full workflow logs

[FAUST-BENCHOU]

The unit tests cover the scoped evaluation paths, but this is a behavior-visible routing change and the repository requires E2E coverage for it. The manual cluster smoke demonstrates the behavior once, but it won’t protect against future regressions. Could we extend the existing authz-rbac profile to verify that routing stays unchanged and a signal used only by an ineligible decision is not evaluated?

[duchangyong]

Hi @rootfs @Xunzhuo @FAUST-BENCHOU, the requested authz-rbac E2E coverage has been added: it verifies that an authz-ineligible decision does not evaluate its signal while fallback routing remains unchanged. All PR checks are now green (29 successful). Could you please take another look when you have a moment? Thank you!

[Xunzhuo]

[P2] Preserve all recipe decisions for the compatibility wrapper. EvaluateAllSignalsWithContext has no request scope, and before this change getUsedSignals used Config.AllRoutingDecisions(). Passing only Config.Decisions here drops signals referenced exclusively by non-default recipes for callers such as ClassificationService.ClassifyIntent. Please keep AllRoutingDecisions() in unscoped wrappers and reserve the narrower list for the explicit request-scoped APIs.

[duchangyong]

Thanks, addressed in e9ebd9b: EvaluateAllSignalsWithContext now uses Config.AllRoutingDecisions() again, while the explicit request-scoped APIs keep the narrower decision list. I also added regression coverage for a signal referenced only by a non-default recipe.

Validation:

• go test ./pkg/classification
• go test ./e2e/testcases
• full local Podman/Kind authz-rbac profile: 4/4 passed, including authz-signal-scope (admin_only_marker metric delta 0 and fallback model Qwen/Qwen2.5-7B-Instruct)
• all GitHub CI checks on 3ad3673 completed successfully, including both required checks and the Kubernetes integration test; conditional jobs were skipped as expected