SceneProof v0.3.0

SceneProof 0.3 turns supplied visual targets into measurable evidence. Agents
can now compare a render with a reference directly instead of tuning geometry,
materials, and motion from memory.

Highlights

• Compares a render with a supplied reference using confidence-gated or exact
  masks, aligned silhouette evidence, paired luminance histograms, normalized
  pixel probes, and reference/current/difference artifacts.
• Reports unaligned viewport composition separately, so subject normalization
  cannot hide incorrect placement or delivery size.
• Extracts target-only silhouette profiles and warns when contrast or dynamic
  range makes a visual claim unjudgeable.
• Runs one-variable fixture-prop sweeps in one browser and bundle, with explicit
  geometry, appearance, composition, or balanced ranking objectives.
• Accepts labeled multi-view reference manifests and keeps every camera, mask,
  artifact, and score attributable to its own perspective.
• Adds adjacent frame-motion summaries and previous-render comparison maps so a
  successful command cannot silently certify a null visual result.

Reliability and ergonomics

• Isolation preserves Three.js lights.
• Camera presets respect the fixture camera's up axis.
• Scout accepts --isolate symmetrically, bare semantic IDs resolve across
  commands, and missing React props receive actionable render guidance.
• Transient esbuild service deaths receive a bounded retry, and every render
  reports whether SwiftShader CPU rasterization was used.

See the changelog
and Three.js fixture protocol
for the complete evidence contracts.

Verification

• Lint and strict TypeScript typecheck passed.
• 49 tests passed, including browser-backed React, Three.js, reference,
  multi-view, sweep, motion, and transient-service behavior.
• Compiled binary build and git diff --check passed.
• The packed tarball installed cleanly, reported version 0.3.0, exposed the
  new reference and sweep CLI surface, and imported sceneproof/three.