ADR-0111 — Kit skills are Suspec concepts; style / stance / depth lives in the catalog

Context

The starter kit ships its .agents/skills/ as "essential" (ADR-0075 D2 / ADR-0064), and the catalog (suspec-skills) holds the optional layer. The boundary was stated as "what every workspace needs" — true but imprecise, and one skill sat on the wrong side of it: adversarial-review. A review is a Suspec concept (the review artifact, the no-self-verdict gate, coverage); but adversarial / three-lens / refute-by-default is a review style — Suspec mandates the artifact, never the style. A user reaching for the adversarial-review discipline in a repo with no Suspec install found it trapped in the kit.

Decision

A skill stays in the kit iff it authors or operates a Suspec concept. Concretely:

• Kit (required): a skill that builds a Suspec artifact type — spec, task, review, finding, intake, inventory, change-plan, audit, research, RFC, PRD, bug-report — or runs a core-loop method (Spec → Task → Run → Review → Close, incl. split-work, spec-check). These are the concepts the method is made of; a workspace can't follow Suspec without them.
• Catalog (suspec-skills): a skill that encodes a style, stance, or depth that is not itself a Suspec concept — a way of doing the work, not a Suspec artifact. Suspec requires neither it nor any particular version of it.

The test that decides the boundary case: does Suspec mandate this thing, or a style of doing a mandated thing? A review is mandated (kit: review-output builds the packet); adversarial review is a style (catalog). Relocate adversarial-review from the kit to suspec-skills. This also explains, post-hoc, why the catalog's existing residents are correctly placed — persona-skeptic/-challenger/ -surveyor (stances), empirical-proof (a discipline), the write-fix/-refactor/… depth guides: none is a Suspec artifact type.

Consequences

• The kit holds only artifact-authoring + loop guides; the catalog holds style/stance/depth. The split is now stated by a single test, not a per-skill audit.
• adversarial-review is installable into any repo (npx skills add), Suspec or not — its natural home. The skill moves intact; its best-in-class consolidation is a separate change (the Improvement Program WS-C, Wave 4).
• review-output (build the review packet — a Suspec concept) stays in the kit; the review style leaves. No core-loop guide moves.
• Accepted ADRs 0064/0075 are not edited (Nygard immutability); this ADR refines them and the README ledger carries the disposition.

Affected obligations / constraints

• Refines: ADR-0064 (the kit/advanced tiering criterion), ADR-0075 (the "essential guides" list — adversarial-review leaves it). Reaffirms: ADR-0093 (stances live in the catalog, not the kit), single-sourcing.
• Does NOT change: the core-loop guides, the artifact formats, the verdict model, or the checks contract.