[Test] Migrate the dashboard profile values to canonical v0.3

[rishabhsinha17]

Related #3233

Standalone prerequisite requested in #3233 (comment): migrate e2e/profiles/dashboard/values.yaml off the pre-v0.3 schema before the next dashboard journey lands, so tests stop accumulating per-test modelCards workarounds around a stale shared fixture.

Legacy vs effective

The old file declared model_config/categories/decisions/strategy/default_model/bert_model/classifier/prompt_guard/response_cache at the top level of the config block. The v0.3 loader drops every one of those keys, so what the router actually ran was the chart defaults: replace-with-your-model placeholder providers and modelCards, the general domain, the default-route decision, plus the file's one surviving key (global.model_catalog.kbs: []). The file documented a router that never existed.

What changed

e2e/profiles/dashboard/values.yaml is rewritten in canonical v0.3 (version/listeners/providers/routing/global, mirroring e2e/profiles/ai-gateway/values.yaml), encoding what the legacy keys meant where that is discernible:

• model_config."base-model" (qwen3, LoRA general-expert) -> providers.models[base-model] + routing.modelCards with the same LoRA and description. The backend ref points at vllm-llama3-8b-instruct.default.svc.cluster.local:8000, which this profile already deploys through the shared gateway stack resource manifests.
• categories: [other] -> routing.signals.domains: [other].
• top-level decisions (other_decision) -> routing.decisions, unchanged in content.
• strategy: priority -> global.router.strategy: priority.
• default_model: general-expert -> providers.defaults.model (the post-[Feature] Add unified model catalog and model hub #3489 canonical spelling; the merge with current main also respells reasoning.family and adds provider to backend_refs), with a bare general-expert modelCard so the reference resolves, same pattern as the ai-gateway profile.
• response_cache.enabled: false -> global.stores.response_cache.enabled: false. This is an intent-encoding change of effective behavior: the dropped legacy key meant disabled, while the chart default had it enabled.
• prompt_guard.enabled: false -> no jailbreak signal declared, which is how v0.3 spells disabled.
• bert_model/classifier model IDs and thresholds are dropped rather than translated: they have no canonical top-level home, they duplicated loader/registry defaults, and the effective config never contained them anyway.
• listeners: [] follows the convention of every migrated e2e profile on this gateway stack. This drops the chart-default router-owned listeners from the rendered config; the ExtProc/API servers are not configured through listeners, which twenty-odd profiles on the same stack already demonstrate.

Follow-through edits keep the assertions honest:

• e2e/testcases/dashboard_deploy_safe_failure.go: the fragment no longer carries its own routing.modelCards. The fixture now declares the model surface, merge mode leaves that node untouched, so the unknown-model check fires against the shared fixture instead of a per-test workaround. Verified locally: the merged document is rejected with routing.decisions[e2e-safe-failure].modelRefs[e2e-nonexistent-model] references unknown model "e2e-nonexistent-model".
• e2e/testcases/dashboard_config_read.go (review follow-up): config/all and config/yaml must each declare base-model under providers.models and other_decision under routing.decisions, decoded into JSON-shaped maps and checked by one walker, instead of passing on any nonempty payload. A Dashboard serving chart placeholders now fails both reads.
• src/semantic-router/pkg/config/maintained_asset_contract_test.go: the file is registered in maintainedValuesConfigAssets, so TestMaintainedConfigAssetsUseCanonicalV03Contract now parses it with the real loader and rejects any regression to legacy keys.

Verification

• go test ./pkg/config/ -run TestMaintainedConfigAssetsUseCanonicalV03Contract passes with the new asset registered.
• helm template over deploy/helm/semantic-router with the new values renders a config with no legacy keys; the rendered document parses via ParseYAMLBytes, and the safe-failure merge is rejected by the unknown-model validator as described above.
• gofmt, go build ./cmd/e2e, go vet, yamllint clean on the touched files.
• The dashboard profile e2e (selection: pr in the domain registry for e2e/profiles/dashboard/** and e2e/testcases/dashboard_*.go) is the real proof and triggers on this PR.

[netlify]

✅ Deploy Preview for vllm-semantic-router ready!

Name	Link
🔨 Latest commit	01f9180
🔍 Latest deploy log	https://app.netlify.com/projects/vllm-semantic-router/deploys/6a9ef1664286b400075a0e80
😎 Deploy Preview	https://deploy-preview-3505--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.

[Xunzhuo]

Thanks for migrating the shared fixture. The requested config/read/deploy follow-through is incomplete: dashboard-config-read still only checks a nonempty JSON object and nonempty YAML body. Both assertions pass if the Dashboard continues serving chart placeholders instead of the new base-model/other_decision fixture, so this prerequisite still does not prove the user-facing read surfaces reflect the migrated config. Please assert canonical identifying fields on both JSON and YAML reads while retaining the deploy safe-failure assertion.

[rishabhsinha17]

Reworked per the review. dashboard-config-read now asserts the canonical identifying fields on both read paths: config/all and config/yaml must each declare base-model under providers.models and other_decision under routing.decisions, checked by one walker over the JSON-shaped decode of each response. A Dashboard serving chart placeholders now fails both reads instead of passing on nonempty output. The deploy safe-failure assertion is retained unchanged.

The branch also merges current main to pick up the model catalog schema change from #3489, respelling the fixture to the new canonical fields (defaults.model, reasoning.family, backend_refs provider). The 09-05 dashboard go mod tidy failure does not reproduce after the merge. Verified locally: the maintained-asset contract suite over the migrated file, dashboard frontend tests and type-check, both go mod tidy checks, and the changed-file pre-commit hooks.

[rishabhsinha17]

@Xunzhuo checks are green on the rework, ready for another look. Review re-request via gh needs triage permission on this repo, so pinging here instead.