feat: add v0.3.0 design quality workflow

[sleegme]

Problem

Implements #8 for the planned v0.3.0 feature. The skill already protected routing and mutation boundaries, but its positive generation guidance was too abstract to reliably prevent generic AI/SaaS styling, superficial reference copying, competing visual ideas, or immediate generation without a bounded critique.

Behavioral constraints added

• Requires a compact pre-generation direction for create, authorized redesign, and design-bearing translate.
• Resolves subject/message, audience/context, focal point, hierarchy, density/rhythm, one composition principle, functional color and type roles, one subject-specific signature, and one flourish to omit.
• Interprets references as structural, compositional, photographic, typographic, brand-specific, and incidental traits, subordinate to evidence priority and the fixed/changeable contract.
• Treats concrete AI-design anti-patterns as discouraged defaults with subject-, evidence-, or user-supported exceptions.
• Requires exactly one same-model pre-finalization critique and one repair batch; no iterative refinement loop.
• Reports a concrete before/after repair or honestly marks the pass pending when the artifact cannot be inspected.

Consolidation and runtime loading

The former medium-neutral references/frontend-design.md guidance was consolidated into references/visual-direction.md, which now owns shared direction, reference-system interpretation, anti-default guidance, bounded critique, and four weak-to-strong examples.

Only three selectively loaded runtime references were added:

• references/generated-image-design.md
• references/editorial-social-design.md
• references/web-interface-design.md

Shared rules are not duplicated across those files. SKILL.md loads only the matching medium file; other media continue through the shared direction plus their existing specialist workflow.

Evaluation cases added

A separate evals/design-quality-cases.json layer covers:

1. image generation without a reference;
2. image generation with an original visual reference fixture;
3. five-frame mobile card news;
4. authorized visual redesign with fixed 1080 × 1350 dimensions and information architecture;
5. responsive web/interface composition;
6. realistic, non-idealized image generation.

The result schema and grade_design_quality_results.py require baseline/candidate artifact identifiers, client/model/suite provenance, exact observable coverage, pass/fail decisions, and concrete evidence. Graded output is explicitly a structural_record_check, requires independent visual review, and never supports a universal aesthetic claim.

Compatibility

• Preserves all eight routes.
• Preserves meaningful_design_exists semantics.
• Preserves evidence priority, fixed/changeable dimensions, explicit redesign authorization, analysis-only mutation boundaries, and unauthorized-drift protection.
• Leaves the route fixtures, semantic smoke suite, route-contract validator, and route regression tests unchanged.
• Keeps the deterministic runtime package project-neutral and tool-neutral.
• Adds no renderer, editor, multi-agent harness, model router, permanent model roles, or project-specific design system.

Relationship to #7

Issue #7 remains the separate v0.2.2 onboarding and release-rerun hardening patch. This branch started from the clean v0.2.1-based main; no v0.2.2 work was present to preserve. This PR does not modify the release workflow, implement #7 onboarding material, close #7, or expand its scope.

Validation and actual results

Automatically verified:

• python scripts/validate_project.py — passed.
• .tooling/bin/agentskills validate design-workflow using pinned skills-ref==0.1.1 — Valid skill: design-workflow.
• python -m compileall -q design-workflow/scripts scripts tests — passed.
• python -m unittest discover -s tests -v — 26/26 passed, including all existing route/mutation regressions, package allowlist tests, and new design-quality record tests.
• python design-workflow/scripts/validate_route_contract.py design-workflow/assets/route-contract.example.json — valid.
• git diff --check — passed.
• GitHub Actions CI — Python 3.11, 3.12, and 3.13 validation plus deterministic release-archive verification all passed on both push and pull-request triggers.
• Deterministic skill ZIP built twice and compared byte-for-byte — passed; SHA-256 ccfb14698f6d18917d679a927748d74a7707789fd0c0eac88e2e52d927d189c6.
• Deterministic project ZIP built twice and compared byte-for-byte — passed; SHA-256 5ad4d7867db3f4f136be96e1cfeaae76c09bfa38ba3d3aaec5f50fb366b3c353.

Structurally evaluated:

• Six required design-quality scenarios and all observable properties are validator-enforced.
• The design-quality grader validates provenance, exact case/property coverage, differing baseline/candidate artifact identifiers, and non-empty evidence.
• Structural checks do not inspect pixels or prove that critique repairs occurred.

Manually inspected:

• Rendered the original SVG reference fixture and confirmed its intended asymmetric focal cluster, repeated spacing/rule rhythm, and deliberately quiet field.
• Inspected the final runtime ZIP: it contains only intended skill legal files, metadata, assets, validator, existing route references, visual-direction.md, and the three medium references. It excludes repository docs, evals, tests, the removed frontend-design.md, and unrelated files.
• Reviewed the diff to confirm the release workflow and existing routing/boundary fixtures remain unchanged.

Pending real-client visual evaluation:

• ChatGPT-class baseline/candidate visual run: pending because no real ChatGPT-class image/design client is available in this environment.
• Gemini-class baseline/candidate visual run: pending because no real Gemini-class image/design client is available in this environment.
• evals/README.md provides the reproducible paired-run procedure, original fixture, provenance requirements, independent review process, and grading command. No result was fabricated.

Unsupported claims and limitations

This PR does not claim universal aesthetic improvement, objective beauty, or identical behavior across clients, models, and generators. Structural or textual graders cannot establish visual quality, verify pixels, or substitute for real-client artifact review. The skill remains a model-agnostic judgment and boundary layer; the connected client/tool still performs generation.

Closes #8