test(docs-hygiene): add eval cases for the write-for-humans grounding gap #12023

Parent #7591 decides that "/docs-hygiene:write-for-humans grounding rule only if an eval shows the gap" (Q526), and its done-when needs the write-for-humans eval slice to carry a Probe result: comment, but that slice was never filed (comment 6556369883); this slice authors the eval cases and slice S3 runs them on a host. The rule under test, from the accepted Q526 list, is "introduce each concept before a sentence relies on it, or name it as assumed reader knowledge", and plugins/docs-hygiene/skills/write-for-humans/ has no such rule on main.

Done when

• plugins/docs-hygiene/evals/ gains at least two wfh-grounding-* cases. Each case has case.yaml, prompt.md, graders/skill-fired.md (same form as wfh-tutorial-is-reference/graders/skill-fired.md) and a grounding grader.
• Each grader has a samples/<grader-name>.json with at least one pass and one fail entry (same shape as wfh-tutorial-is-reference/samples/skill-fired.json).
• Each prompt asks for a human-facing document (README, design doc or how-to) from source material that uses at least one domain term before it defines it.
• The grounding grader is type: llm. Its text says PASS only when every such term is introduced before a sentence relies on it or is named as assumed reader knowledge, and FAIL otherwise.
• Each case.yaml sets runs: 3 or more.
• The diff edits neither plugins/docs-hygiene/skills/write-for-humans/SKILL.md nor reference/sentence-rules.md: the rule ships only if S3 shows the gap.
• A .changes/docs-hygiene/*.md fragment with bump: none describes the new cases.
• The PR's CI is green.

Out of scope

• Writing the grounding rule.
• Running the eval (slice S3).

Parent: #7591