Drop custom instruction; read model from scenario YAML

Two related contamination-control changes for the Scenario #1 harness:

- Remove the custom `instruction=` kwarg. Fast-agent now falls through
  to DEFAULT_AGENT_INSTRUCTION so the activation test sees only this
  CLI's native system prompt plus the scenario's user prompt. The old
  string named "GitHub pull request workflows", priming the model
  toward the skill's domain ahead of how Gemini CLI and Codex are
  configured (neither injects a custom system prompt).
- Read the model from the scenario's per-client `models:` map instead
  of letting fastagent.config.yaml's `default_model` drive the choice.
  Model selection belongs in the shared scenario so every client (and
  every run) is explicit and reproducible.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: olaservo

scripts/skills_e2e_agent/agent.py

@@ -255,10 +255,11 @@ def _evaluate(calls: list[tuple[str, dict]]) -> dict:
 
 @fast.agent(
     name="pr_reviewer",
-    instruction=(
-        "You are a software engineer assisting with GitHub pull request "
-        "workflows. {{agentSkills}}"
-    ),
+    # No custom instruction — fall through to fast-agent's
+    # DEFAULT_AGENT_INSTRUCTION (constants.py) so the activation test sees
+    # only this CLI's native system prompt + the scenario user prompt,
+    # matching how Gemini CLI and Codex runners are configured.
+    model=SCENARIO.get("models", {}).get("fast-agent"),
     servers=["github_skills"],
 )
 async def main() -> int: