[carlrannaberg]

Visualize installations currently include the complete 103-system catalog, generated previews, and preview authoring sources: about 31.7 MB even when an artifact uses one reference.

Make compact installation the default. Keep the index, Clean fallback, shared guidance, templates, runtime helpers, dependencies, and helper fixtures installed. Load other selected packages through scripts/catalog.mjs from an already-published immutable commit and cache locally; download previews only on request. Failed resolution reports the selected package as unavailable without silently substituting Clean.

• Move the full catalog to catalog/design-systems/ and preview kit to dev-scripts/preview-kit/. Package DESIGNs, token values, and generated previews remain byte-identical.
• Generate the bundled index and fallback from the published pin, while validating the working authoring catalog separately. Publish additions before a follow-up updates the distribution pin.
• Keep shared animation-patterns.md runtime-owned. Catalog preview authors link to it; editing motion guidance needs no catalog publication or repin.
• Update affected authoring tools/docs and regenerate the three mirrors. No plugin repin, release, or merge.

Installed payload falls from 801 files / 31,733,780 bytes to 180 files / about 11.4 MB (64% fewer bytes). This measures installed files, not Git download size or model context.

Validation: 11 catalog tests; 73 runtime tests; palette smoke; 72 fixture checks; authoring/schema; 35-template inventory, preview matrix and strict detector; all 103 catalog previews fresh; overflow, diagram and Mermaid self-tests; shellcheck and mirror checks. Actual skills CLI 1.7.4 installs, live pinned Editorial guidance/previews byte parity, denied-network cache reuse and offline Clean passed. For the motion follow-up, targeted checks passed and an isolated edit survived generator write plus freshness check with an unchanged catalog pin. Both fresh correctness and ownership reviewers passed.

Behavioral evidence:

• Fresh Editorial and offline Clean agents created artifacts; inspected raw calls/results and desktop/mobile light/dark renders. The original trial installs preceded the pin-isolation fix and differed only in two Clean README links at that point. Their saved spawn messages are encrypted; the raw files independently establish actions, not the prompt text. Clean's known three-full-template over-reading remains outside scope.
• An additional fresh agent used an exact 311746b skill export with a cold cache and OS-enforced network denial. The resolver failed, and the agent reported Editorial unavailable without creating HTML, inventing guidance or substituting Clean. Exact plaintext task/environment/spawn inputs, hashes and actual raw tool results are retained locally. The motion follow-up changes no resolver behavior.
• Direct invocation was tested; automatic discovery was not. CI39466758238, attempt1 passed on updated commit 9e91878: main validation and Windows Mermaid both succeeded.