docs: audit-derivability rubric/evals invert the context-fork inheritance claim

[kyle-sexton]

Problem

Two docs-hygiene artifacts state that skill-level `context: fork` inherits the invoking conversation:

• `plugins/docs-hygiene/skills/audit-derivability/context/rubric.md` (~L132-133): "must NOT be a `context: fork` (a fork inherits this context's contaminated history)"
• `plugins/docs-hygiene/skills/audit-derivability/evals/evals.json` (~L88): "Explicitly uses a non-fork subagent and rejects `context: fork` (fork inherits the contaminated context)"

Per the official skills doc (code.claude.com/docs/en/skills, "Run skills in a subagent"), skill-level `context: fork` starts BLANK: "It won't have access to your conversation history." The claim conflates the Agent-tool `subagent_type: "fork"` (which DOES inherit the conversation and ignores model overrides) with skill-level `context: fork` (which does not inherit).

Surfaced by the independent review of PR #1050 (re-anchor `sweep-all-disciplines`), whose fork-mechanism claims were verified against the official docs and now contradict these two artifacts.

Fix

Reconcile the rubric + evals wording to distinguish the two fork mechanisms; the conclusion (use a fresh non-fork subagent for uncontaminated review) may well survive — the stated reason is inverted.

References

• Official: https://code.claude.com/docs/en/skills (context: fork), https://code.claude.com/docs/en/sub-agents (Agent-tool fork)
• Internal: PLUGIN-PHILOSOPHY.md fresh-eyes section (Agent-tool fork wording is correct)

[kyle-sexton]

Claim (work lane, work-2 / T5) — dispatching an implementation subagent this cycle. 2026-07-22T18:50:13Z

[kyle-sexton]