fix(evals): switch llm and grader targets to OpenRouter

GH Models rate limits (429) were failing most LLM evals. OpenRouter
has higher rate limits and built-in provider fallback.

Also excluded code-grader-sdk from CI (needs Azure keys in its
per-example targets.yaml).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.agentv/targets.yaml

@@ -19,17 +19,15 @@ targets:
 
   # ── LLM target (text generation, no agent binary needed) ────────────
   - name: llm
-    provider: openai
-    base_url: https://models.github.ai/inference/v1
-    api_key: ${{ GH_MODELS_TOKEN }}
-    model: ${{ GH_MODELS_MODEL }}
+    provider: openrouter
+    api_key: ${{ OPENROUTER_API_KEY }}
+    model: ${{ OPENROUTER_MODEL }}
 
   # ── Grader (LLM-as-judge) ──────────────────────────────────────────
   - name: grader
-    provider: openai
-    base_url: https://models.github.ai/inference/v1
-    api_key: ${{ GH_MODELS_TOKEN }}
-    model: ${{ GH_MODELS_MODEL }}
+    provider: openrouter
+    api_key: ${{ OPENROUTER_API_KEY }}
+    model: ${{ OPENROUTER_MODEL }}
 
   # ── Named agent targets ───────────────────────────────────────────
   - name: copilot-cli

.github/workflows/evals.yml

@@ -60,7 +60,7 @@ jobs:
           DEFAULT_PATTERNS: "evals/**/*.eval.yaml,examples/**/*.eval.yaml,examples/**/*.EVAL.yaml,examples/**/EVAL.yaml"
           # Exclude evals that need local scripts or multiple agent targets.
           # Negation patterns (!glob) are supported by the CLI.
-          EXCLUDE_PATTERNS: "!examples/showcase/multi-model-benchmark/**"
+          EXCLUDE_PATTERNS: "!examples/showcase/multi-model-benchmark/**,!examples/features/code-grader-sdk/**"
         run: |
           PATTERNS="${{ github.event.inputs.suite_filter || vars.EVAL_PATTERNS || env.DEFAULT_PATTERNS }}"
           EXCLUDES="${{ vars.EVAL_EXCLUDE_PATTERNS || env.EXCLUDE_PATTERNS }}"