SCRUM-1066-issue 216

[emnarahmouni-exmergo]

Closes #216

• Adds evals/corpus/ade_bench_triggering.json: 30 real analytics-engineering requests sourced from dbt-labs/ade-bench (Apache-2.0), hand-labeled with the skill each should fire (or none), with provenance and license recorded on the file itself.
• Unlike each skill's own evals.json positive/negative list (written by whoever wrote the description it tests, checked with siblings disabled), this corpus runs every skill available at once and is scored as per-skill precision/recall/F1 (Corpus/CorpusCase/load_corpus in suite.py, Classifier protocol/run_corpus/SkillPrecisionRecall in runner.py, ClaudeCliClassifier in claude_agent.py, --corpus flag in __main__.py).
• python -m evals --corpus evals/corpus/ade_bench_triggering.json always exits 0 — it's a measurement against externally authored prompts, not a release gate, since the initial pass rate is expected to be low.
• Classifier.classify reports every skill marker a call finds via a ClassifyResult(fired_skills, error), not a single winner: a prompt two skills both claim shows up as real cross-skill contamination (a false positive for the unwanted one) instead of being silently hidden behind whichever name came first.
• A per-call failure is caught by the classifier itself and recorded per case, excluded from precision/recall. A setup failure (the claude binary missing) is left to propagate out of run_corpus immediately instead, so it aborts the whole run once with the existing prerequisite message and exit code 2, rather than being recorded as N separate case errors while the command still reports a clean exit.
• Adds a short mention of the corpus to references/evaluation.md's Tier 2 section.
• Removes a stray changelog entry that had leaked into the already-released 1.9.1 section via an unrelated PR (SCRUM-1123-explore inventory: cap --rank's payload by default #379), claiming this feature shipped when none of its code had actually merged.

Test plan

• uvx ruff check evals/ / uvx ruff format --check evals/ — clean
• uvx pytest evals — 46 passed (16/16 in test_runner.py, including new regression tests for the cross-skill-contamination fix and the setup-failure-propagation fix); 6 pre-existing test_wrapper.py failures are a Windows/PosixPath environment quirk that also fails on a clean main, unrelated to this diff
• Zero em-dashes across every touched file
• Manually verified the corpus JSON's provenance/license fields and spot-checked several hand-labels against their source ADE-bench task prompts