harness-slimlisted

# Harness Slim This is the **post-adoption harness evaluator**. `gatekeeper` decides whether an external artifact is safe to adopt; `review` checks a code diff; `harness-slim` compares the live source, installation, runtime, context, and attention surfaces after adoption. ## Contract - Audit read-only. Propose before any move; execution routes through `careful` and the host's normal issue/worktree flow. - Treat responsibility overlap as the prune signal. Counts and usage are supporting evidence, never deletion authority. - Keep `actual_invocation`, `dispatch_selection`, and `load_proxy` separate. A skill-file read is load evidence, not invocation. - Preserve trust, permission, data-loss, and external-write gates unless a replacement proves the same behavior. ## Audit 1. Read the host contract and machine-state documentation. Resolve the authoritative source manifest, installed registrations/caches, cold-start context, available doctor commands, and usage evidence. Missing paths remain named evidence gaps; do not invent defaults. 2. Run read-only doctors first. Report source truth, installed truth, and live runtime truth independently; one green surface cannot mask another red one. 3. Inventory what hosts actually register. Marketplace/catalog availability, unregistered files, remote tools, and optional overlays are separate surfaces. 4. Normalize usage by event semantics and coverage. Zero-use pruning requires at least 30 days, 20 eligible op