hrt-dark-mode-decision-gate

Resolve one escalation-worthy finding without asking the user, and log the reasoning to dark.md so a human can audit it later.

When this runs

• hrt-align-consistency-review / hrt-apply-code-review: DECISION and HIGH findings only. MECHANICAL is fixed directly by the caller, never reaches this skill.
• hrt-adversarial-authoring: a single-reviewer UNANCHORED finding only. ANCHORED, or both reviewers agree, auto-resolves in the caller and never reaches this skill.
• MUST NOT run on a finding that is a security requirement or mitigation already decided by stride-analysis-patterns / threat-mitigation-mapping / security-requirement-extraction — that decision is a hard constraint, not an opposing case to weigh. Escalate those to the user even in dark mode.

Fresh subagent, always

Caller MUST spawn a fresh subagent with no memory of the session that produced the finding. Give it only: the finding, and the grounding artifacts needed to evaluate it (discovery.md, proposal.md, specs/**/*.md, design.md, tasks.md, align.md as relevant; code and test output if called from apply). MUST NOT run inline in the calling session's own context — a same-context steel-man reuses the reasoning that produced the original recommendation instead of opposing it.