Add evaluated visual verification routing and model harness

[theclaymethod]

What changed

• preserve the committed Artifacture, Impeccable, and Unslop routing boundaries
• add provider-independent rendering, adapter, runner, aggregation, merge, telemetry, and policy-selection contracts
• replace the repetitive 112-state fixture set with a curated 36-pair / 72-state seed across five Artifacture-owned families
• add responsive viewports, visually distinct operating-model cases, and source-conditioned controls for claims whose validity depends on supplied evidence
• bind evidence identity to rendered pixels plus visible/source evidence while independently gating distinct-image coverage
• enforce smallest-first model ladders and corpus-declared batch sizes 1 and 2
• fail closed unless live graduation has human review, representative Artifacture capture provenance, and 10 distinct images and source artifacts per label/family

Why

Artifacture needs empirical evidence before assigning visual verification to a smaller model or increasing screenshot batch size. The corpus is now useful for seed development and contract testing without pretending synthetic fixtures are enough to qualify production routing.

Explicit limitations

• no provider measurements or model winner are claimed
• this seed can compare batch 1 with batch 2 only; it does not establish that 2 is globally maximal
• graduation still requires representative product captures, human adjudication, exact provider candidates, and complete cost/latency/cache telemetry

Validation

• npm test — 89/89 tests passed
• visual corpus render — 72 states
• dry-run request matrix — 324 requests across batch 1 and 2
• npm run ve:eval — 180 seeded violations, 7 clean fixtures, and 9 design-system cases passed
• npm run ve:check
• npm run check:manifests
• follow-up corpus, specification, and standards nemesis reviews found no remaining material defects in this overhaul