fix(taste-skill): narrow routing to supported web pages

[kbegiedza]

Summary

• Narrow taste-skill discovery to expressive web pages and eligible visual redesigns.
• Exclude decks, generic prototypes, dashboards, data tables, and multi-step product UI from the full landing-page workflow.
• Keep an explicit findings-only path for requested dense-product-UI visual audits using the playbook accessibility and anti-tell sections.

Why

The skill advertised itself as the default design driver for dashboards, prototypes, and decks even though its vendored playbook explicitly excludes dashboards, data tables, and multi-step product UI, and its runtime defaults to an HTML artifact. That mismatch could route unsupported work into the landing-page workflow.

Implementation

The change is limited to skills-seed/taste-skill/SKILL.md. It aligns the discovery description and opening scope with the playbook, and makes the dense-product-UI branch explicit: visual audit findings only, with no artifact build, HTML default, or publication. The vendored playbook and package topology are unchanged.

Verification

• node --test test/frontmatter.test.ts test/skill-conformance.test.ts - passed, 10/10, on host Node 24.11.1.
• git diff --check - passed.
• Three paired fresh-context Codex CLI / gpt-5.4-mini cases - deck overreach fixed; landing-page full workflow preserved; dense-table visual audit preserved as findings-only.
• Fresh independent package review - PASS with no actionable findings.
• Not run on the pinned Node 24.18.0 / npm >=11.10 toolchain - the host has Node 24.11.1 / npm 11.6.2.
• Not run: test/skills-seed.test.ts - pg-boss is absent because dependencies are not installed.
• Not run: Prettier - node_modules is absent.
• Not run: live QM discovery/activation - no live instance was used for this narrow skill-instruction change.

Review focus

• Confirm that the discovery boundary excludes unsupported artifact types without weakening landing-page, portfolio, marketing/editorial, or eligible visual-redesign routing.
• Confirm that explicit dense-product-UI visual audits remain useful without claiming build or publication authority.

Risks / limitations

Live QM routing and harness/model combinations other than Codex CLI with gpt-5.4-mini remain unvalidated. The PR is draft until CI or a pinned-toolchain run supplies the missing repository validation.

---

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