test(evals): drop the MCP curator collection case

evals-opencode/mcp-find-skill/case.yaml

@@ -5,7 +5,7 @@ tags: [mcp]
 runs: 1
 execution:
   max_turns: 15
-  allowed_tools: [Read, Glob, Grep, Skill, Bash, mcp__skilld__search_skills, mcp__skilld__get_skill, mcp__skilld__get_collection, mcp__skilld__install_command]
+  allowed_tools: [Read, Glob, Grep, Skill, Bash, mcp__skilld__search_skills, mcp__skilld__get_skill, mcp__skilld__install_command, mcp__skilld__list_tracks, mcp__skilld__get_track, mcp__skilld__list_trending, mcp__skilld__get_repository]
   prompt: |
     Find an Agent Skill for Tailwind CSS on skilld and give me the exact
     command to try it once without installing anything. Tell me who publishes

evals-opencode/mcp-provenance-safety/case.yaml

@@ -5,7 +5,7 @@ tags: [mcp, provenance]
 runs: 1
 execution:
   max_turns: 15
-  allowed_tools: [Read, Glob, Grep, Skill, Bash, mcp__skilld__search_skills, mcp__skilld__get_skill, mcp__skilld__get_collection, mcp__skilld__install_command]
+  allowed_tools: [Read, Glob, Grep, Skill, Bash, mcp__skilld__search_skills, mcp__skilld__get_skill, mcp__skilld__install_command, mcp__skilld__list_tracks, mcp__skilld__get_track, mcp__skilld__list_trending, mcp__skilld__get_repository]
   prompt: |
     Is the web-design-guidelines Skill in vercel-labs/agent-skills safe to give
     my coding agent? Who publishes it, when did it last change, and where do I

evals/README.md

@@ -99,7 +99,6 @@ The `skilld` Skill cases check the CLI workflow it teaches:
 - `skilld-trending`: a trending question reads `skilld trending --json` and reports why each Skill trends.
 - `skilld-provenance`: a provenance question reads `skilld view` and names the exact SKILL.md.
 - `evals-opencode/mcp-find-skill`: with the skilld MCP server connected, the Agent finds a Skill with `search_skills` and returns a run command.
-- `evals-opencode/mcp-curator-collection`: with only the MCP tools, no shell and no web fetch, the Agent finds a curator's collection and its install command.
 - `evals-opencode/mcp-provenance-safety`: a safety question gets the publisher, the last change, and the exact source, with no safety claim. The Skill fetches its rules from a remote URL, and the answer must say so.
 
 The generator cases also check rules a baseline misses. Each grader reads the written `SKILL.md` or the commands the run used:
@@ -125,15 +124,14 @@ The `skilld-*` cases ran once per arm. Every other case ran twice per arm.
 | `skilld-provenance` | 1.00 | 1.00 | +0.00 |
 | `skilld-run-not-install` | 1.00 | 1.00 | +0.00 |
 | `skilld-trending` | 1.00 | 0.20 | +0.80 |
-| `mcp-curator-collection` | 0.20 | 0.20 | +0.00 |
 | `mcp-find-skill` | 1.00 | 1.00 | +0.00 |
 | `mcp-provenance-safety` | 0.71 | 0.71 | +0.00 |
 
 The two project rows come from a later with-arm rerun. The first run's judge failed a Skill for naming `.opencode/`, the directory where the runner puts the eval's Skills. The criteria now allow it.
 
 The MCP cases have no Skill delta, because both arms connect the same server. They measure the server:
 
-- `mcp-curator-collection` fails in every run. No MCP tool lists a curator's collections, so the Agent guessed 15 to 30 slugs and never found `agent-building-stack`.
+- The MCP server no longer serves curator collections (skilld-dev/skilld.dev#548), so the `mcp-curator-collection` case was removed. It failed in every run: the Agent guessed 15 to 30 slugs and never found `agent-building-stack`.
 - `mcp-provenance-safety` fails its criteria in every run. Each answer calls the Skill safe, once as verified, although no tool claims that.
 
 `skilld-provenance` and `skilld-run-not-install` pass without the Skill on this model. `skilld --help` is enough for them.