Structural check: require eval results in testing directory

[dacharyc]

Context

From a PR review patterns audit across 7 MongoDB Agent Skills, testing/evidence gaps appeared in 8 instances across 4 PRs. The most common review comment was some form of "have you tested agent performance with and without this file?" — a question that should be answered by eval results, not reviewer intuition.

Issue #39 added support for recognizing the evals/ directory. This issue proposes going further: checking that eval results actually exist.

Proposed check

Add an opt-in warning-level check that verifies a skill's evals/ directory contains result files, not just test definitions.

Behavior

• Severity: warning (opt-in via flag, e.g., --require-eval-results)
• Check: If evals/ directory exists, verify it contains at least one file matching a results pattern (e.g., *results*, *output*, *report*, *.json with results-like structure)
• If evals/ exists but contains only definition/config files and no results: warn
• If evals/ does not exist: warn (when opt-in flag is set)

Message

• No eval results found in evals/ directory. Consider running evals and including results to demonstrate skill effectiveness.

Why opt-in

Not all workflows will have eval infrastructure set up. Making this opt-in (and eventually default in CI mode) allows gradual adoption.

Examples from PR reviews

PR	Issue
6	"Have we tested agent performance with and without this file?" (asked for 3 reference files)
5	"Have you found agents unable to adhere to patterns without this guidance?"
7	"Did you find in testing that the skill needed this?" (asked twice)
7	Meta: "I wonder if there is a way to indicate what information a skill needs for efficacy"

Related

Builds on #39 (recognize evals/ directory). Part of a series of checks derived from PR review pattern analysis.