[carlrannaberg]

The Display plugin still bundled the retired Visualize Lite build. Replace it with the canonical, compact-by-default Visualize skill from display-dev/visualize@53eb0dc560688b8174ff1b2811fe63d04042f42d (through Visualize PR #5).

• Keep two skills: display-dev and visualize. Keep the Display skill pin and all hooks unchanged.
• Update sources.lock and the existing sync target, remove skills/visualize-lite/, and import skills/visualize/ without local overrides.
• Update README naming, selected-catalog retrieval/cache disclosure, and the Claude Code invocation migration to /display-dev:visualize. The cache disclosure names ~/.cache/visualize-skill/catalog, its relocation variable, and retention after uninstall.
• Bump the plugin and matching attribution strings to 0.13.0.

The diff is mostly the source replacement: 180 imported Visualize files exactly match the upstream pin (bytes and modes). Only five plugin-specific files change: source lock, sync script, README, manifest, and feedback attribution literals. Publishing instruction consolidation is outside this PR.

Validation:

• Task packaging review and two fresh blind final reviews completed with no issues found. The correctness reviewer matched all 60 trial records against original session JSONL.
• Source sync/parity, existing hook tests, Bash syntax, ShellCheck at CI's warning threshold, version/attribution checks and git diff --check pass.
• Both the repository and its clean tracked export pass claude plugin validate.
• Clean package: two skills, 209 files, 18,191,514 uncompressed bytes. Local path/nesting/size/symlink checks pass.
• Three fresh native subagents received only an isolated workspace assignment, its ordinary task request, the supplied skill descriptions, and side-effect constraints. Plaintext prompts were saved before dispatch; no workflow/template/read-count expectations or follow-up coaching were supplied.

Trial	Observed in raw tool calls/results
Create a results recap / Editorial	Selected Visualize, resolved Editorial with the imported catalog helper, loaded Report, generated correct supplied facts; no publish
Polish an existing memo	Selected Visualize and polish guidance, preserved visible text, passed responsive/light/dark/contrast and enlarged-text checks; no publish
Create a weekly update, publish for company access	Loaded both skills, used Clean and Project recap, made one offline dsp publish --visibility company call, and verified published bytes equal source

Evidence limits: native tests use a supplied skill catalog, so they do not establish actual Claude activation, MCP OAuth, or hooks firing inside Claude. Publishing used a local recording fixture, never a live service. Creation and publication agents first read inherited host bootstrap files. Creation's candidate-prefix scan included Structure, and it omitted browser verification; the parent separately checked its desktop/phone rendering and contrast. Optional favicon/social metadata warnings remain. Local Claude validation excludes the directory safety scan/policy review; no submission or release was performed.

Original plaintext prompts, hashes, raw session paths, complete tool call/results, artifacts and screenshot evidence are saved locally under temp/integration/ (audit.md and implementation-notes.html).

Opus review found no blocking requirement issues. Its README cache-path/lifetime clarification was applied in 4dbea6c. The three behavioral trials ran at 8e413f3; only that README sentence changed afterward, so the tested skill bytes remain unchanged.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}