ADOPTERS: add multi (cross-model port of the ADHD loop)

[szarkans]

multi is a Claude Code plugin that runs the same job through several model families and treats disagreement as signal. Its /multi:adhd skill is a port of ADHD's loop with one change: each frame goes to a different model family — a Claude sub-agent, OpenAI Codex, and a cheap third through OpenCode/OpenRouter — so a run diverges by frame and by model priors at the same time, for roughly the price of one.

What was kept from ADHD:

• diverge → converge, with the critic off in Phase 1 and on in Phase 2
• the frames table as the vantage mechanism (3am on-call, inversion, biology, speedrunner, 10-year-old, …)
• traps get named, not silently dropped

What was added:

• backend rotation across frames, external models started first because they are slow
• an explicit confound note: since frame and model vary together, a run must never be reported as a model comparison. Comparing families is the same frame to every backend, and it costs 3x — the skill offers that instead of faking it.

Skill: skills/adhd/SKILL.md. MIT, upstream credited in the README (en/ru/zh) and in the skill's frontmatter.