Security coverage to 95%+ (real tests) + embedded frontend-design skill

[tzone85]

Summary

Two requested improvements, both TDD'd and independently reviewed.

1. Security coverage: 73.0% → 98.6% / 95.2% — real behavior pins, not checkbox tests

• internal/security: 73.0% → 98.6%. Centerpiece: a fake-scanner harness (scanners_exec_test.go) — executable shell scripts named gosec/gitleaks/semgrep/govulncheck/npm on a controlled PATH emit canned real-format tool output, driving RunScanners end-to-end with zero network: language applicability, PATH availability, per-tool parsing, dedup, and the three-way ran/skipped/failed classification (a tool emitting garbage lands in failed and never masquerades as a clean run; missing tools are skipped visibly; inapplicable tools appear nowhere).
• engine/security_gate.go → 95.2% (statements). All LLM review paths driven via llm.ReplayClient: LLM findings merge with scanner findings, a critical diff finding blocks a story, the diff is asserted to travel inside <diff> data tags (injection mitigation), an LLM failure degrades to scanners-only, garbage responses pass cleanly.
• Previously-uncovered load-bearing logic now pinned: Covers was at 0% — the self-upskilling dedup that decides whether the agent re-learns a vulnerability class (rule-ID match, CWE alias, learning trigger, receiver immutability). Plus corrupt-KB-is-a-loud-error at both gate entry points, read-only-KB persist failure (best-effort, never aborts), vulnClassID fallback chain, cweOf extraction bounds, language-manifest branches, and parser error paths for all 5 tools.
• Remaining uncovered lines are error-log statements requiring failing event/projection stores — acknowledged, not gamed.

2. Embedded frontend-design skill — vxd-built UIs stop looking AI-generated

Research: Anthropic's frontend-design skill (read in full) + a web-research sweep (Anthropic's design-skills blog, anti-"AI slop" guides, WCAG 2026 checklists, LLM UI stack analyses).

• agent.FrontendDesignBrief — single-source design-standards block injected into every UI-facing story's goal prompt: two-pass token-first process (palette/type/layout/signature plan → self-critique for genericness → code derived from the plan), named banned defaults (Inter/Roboto, purple-gradient-on-white, cream #F4F1EA+serif+terracotta, acid-green-on-black, the template feature-card page, scattered animation), a non-negotiable accessibility floor (responsive to 360px, visible keyboard focus, WCAG AA contrast, prefers-reduced-motion, 44px touch targets, designed empty/loading/error states, CSS specificity discipline), and copy-as-design-material rules ("Save changes", never "Submit"). Size budget pinned at ≤6 KB.
• detectFrontend — owned-file extensions (strongest signal) + whole-word UI vocabulary regex. Word boundaries pinned by table tests: "pagination"≠page, "performance"≠form, "review"≠view; per review, "html" and "responsive" excluded (server-side HTML emails and "responsive API gateway" are backend work).
• Threaded into both dispatch paths (first attempt + retry) with a dangling-wire source-scan guard (TestExecutor_WiresFrontendDetection).
• Planner: ENGINEERING STANDARDS now require the first UI story to establish a design-token foundation (CSS custom properties / framework theme) consumed by later UI stories, and UI acceptance criteria to carry the accessibility floor.

Review

everything-claude-code:go-reviewer: approve, no CRITICAL/HIGH; 2 MEDIUM + 2 LOW all applied in the final commit (keyword false-positives removed with new negative cases; fake-tool harness exit-code fidelity + collision-resistant sentinel + honest Bin values). The suggested style keyword was deliberately not added — it would trade a minor miss for "code style" false positives.

Test plan

• go test ./... -count=1 — all packages pass
• go vet ./... clean; golangci-lint run — 0 issues
• internal/security 98.6% coverage; security_gate.go 95.2%
• Doc-coverage wiring tests pass (CLAUDE.md section + README bullet added; stale "12 checks" corrected to 16)
• NXD port noted as pending in CLAUDE.md