Summary

Two related defects across the LLM-based evaluators were identified through end-to-end testing against the np-canary Foundry project (using minimal_repro_dataset.jsonl and minimal_repro_dataset_empty.jsonl):

1. Skip-handling fragility — coherence / fluency / retrieval post-process the dict returned by _the_super_do_eval and then run math.isnan(result.get(self._result_key, 0)). PR Fix TypeError when evaluator returns None score for skipped/not-applicable results #5107 added the _score is not None guard that prevents the production TypeError: must be real number, not NoneType, but the structure is fragile: any future override that forgets the guard re-introduces the bug. task_completion (and relevance) avoid this entire class of defect by checking status == "skipped" before any numeric coercion runs.

2. No input pre-validation on retrieval — empty/missing query/context is silently sent to the LLM, which returns status: skipped, demoting the row to not_applicable instead of raising a clean USER_ERROR. (groundedness/coherence/fluency/relevance already have this via self._validator.)

Changes (3 files, +43 lines)

The status-first pattern (the canonical example lives in task_completion._parse_prompty_output):

`python
result = await self._the_super_do_eval(eval_input)

Honor explicit skip status before any numeric validation. Structurally safer

than relying solely on the None-guard below.

if result.get(f"{self._result_key}_status") == "skipped":
return result
_score = result.get(self._result_key, 0)
if _score is not None and math.isnan(_score):
raise EvaluationException(...)
`

The None-guard from #5107 is retained as a backstop. The skipped path can no longer reach math.isnan at all.

Out of scope

• relevance — already has both patterns inline. No change required.
• groundedness — has the same outer post-check shape as coherence/fluency/retrieval and would benefit from the same W4 refactor; left out per agreed scope (this PR is the four highest-risk evaluators only).
• The remaining ~14 LLM evaluators in the catalog need pre-validation as well; tracked separately.
• Republishing fixed evaluators to the azureml registry (operational, not a code change).
• Math-based evaluators (bleu/gleu/meteor/rouge/f1/task_navigation_efficiency) need typed input validation; tracked separately.

Validation

• python scripts/validation/code_health.py -i assets/evaluators/builtin — clean.
• pytest on test_coherence_evaluator_behavior.py + test_fluency_evaluator_behavior.py + test_relevance_evaluator_behavior.py — 199 passed.