fix(agent-context): tune prompts and codegen skill triggers with eval set (#438)

* feat(agent-context): trigger evaluation for the prompts and codegen skill descriptions

Add should-trigger and should-not-trigger cases (12 each) for metaobjects-prompts and
metaobjects-codegen, a runner that wraps the skills in a throwaway plugin for
claude plugin eval, and a CI test that the cases parse and name existing skills.

First run: codegen 36/36, prompts 32/36 (one run of an entities-only near-miss
fired the skill). Narrow the prompts description to projects that declare
template.prompt; codegen needed no change. Rerun: 36/36 and 36/36.

* no-mistakes(review): Scope prompts skill gate; add adoption eval case

* no-mistakes(test): Tune codegen skill description to trigger on dialect config

* chore(agent-context): drop the parse-only trigger-case test; the evaluation is manual-only

A test that only parses the case files proves nothing about trigger behaviour, so the cases
and the documented run command are the evidence. README no longer claims a CI check.

Author: dmealing

agent-context/evals/README.md

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Run the skill trigger evaluation (agent-context/evals/skill-triggers) with `claude plugin eval`.
+# Manual, not CI: it drives real model sessions on your own Claude login (subscription quota).
+#
+# usage: agent-context/evals/run-skill-triggers.sh [OUT_DIR] [extra `claude plugin eval` flags]
+#   OUT_DIR defaults to a fresh temp dir; the scored result is written to OUT_DIR/result.json.
+#   e.g. agent-context/evals/run-skill-triggers.sh /tmp/triggers --case 'metaobjects-codegen/*'
+set -euo pipefail
+here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+skills="$here/../skills"
+out="${1:-$(mktemp -d)}"; [ $# -gt 0 ] && shift
+mkdir -p "$out"
+plugin="$out/plugin"
+rm -rf "$plugin"
+mkdir -p "$plugin/.claude-plugin"
+# `claude plugin eval` refuses an evals/ link that points outside the plugin, so copy.
+echo '{"name":"mo-skills","version":"0.0.0","description":"MetaObjects skills, wrapped for trigger evaluation"}' >"$plugin/.claude-plugin/plugin.json"
+cp -R "$skills" "$plugin/skills"
+cp -R "$here/skill-triggers" "$plugin/evals"
+cd "$plugin"
+# --ablation none: one arm, so the `tool_used: Skill` graders count toward the score.
+exec claude plugin eval . --ablation none --no-publish --trust-plugin -j 4 --threshold 0 \
+  --output-dir "$out/runs" --json "$out/result.json" "$@"