saccade 0.2.0

Added

• Add opt-in rendering evidence: required effect occupancy, intended experiment
  variables, spatial and layer scope, automatic evidence regions, fixed-camera
  temporal tiles, noisy offline references, preregistered blind trials and warmup
  convergence qualification. Capability discovery lists the new workflows;
  media records can expose descriptive spatial tiles through quality_tile_size.
  Structural classes and human preferences do not qualify performance or approve baselines.
• Add versioned media records with per-section status/provenance, reusable optional
  CPU sessions, deterministic saliency and declared crop fitness.
• Add the abi3 Python package, exact image/text index APIs, external ffmpeg
  keyframe sampling and saved-fingerprint usage matching.
• Add a bounded local HTTP API and unprivileged container recipe; configured
  compatible/Azure provider mappings are verified with fixtures only.
• Add pinned official SigLIP 2 checkpoint/tokenizer provenance and a reproducible
  CPU export script; text retrieval requires that joint model and remains uncalibrated.
• Add experimental review explain, review audit-mask, review check-ui and
  review assist batch submit|status|collect; optional Jev routing remains unqualified.
• Add the local Playwright matcher, sweep plan|compare, imgtune audit|search,
  design pull|compare, notify and verified --baseline last-good lookup.
• Add explicit registration/resampling, hash, dedupe, similar,
  index build|query|export-inputs|calibrate, text, assess, inspect-image,
  capabilities and comparison question routing.
• Add optional SVG/PDF rendering, offline C2PA validation and pinned Rust OCR.
  Model, OCR readability, forensic specificity and broad renderer qualification
  keep the limits recorded in their evidence and documentation.
• Add vision commands and explicit face/crop/watermark observations in image
  reports, fixture-only hosted mappings and advisory check-ui localization.
  Shared model contracts retain legacy readers. Browser/sweep masks neutralize
  excluded pixels before filtering; core score-exclusion defaults remain.
  Model export/parity and generated OCR review keep separate qualification gates.

Changed

• Raise the MSRV to Rust 1.89 and upgrade Butteraugli to 0.9.3; remove the
  vendored compatibility patch.
• Replace the local ocRs adapter with pinned PP-OCRv5 Latin on CPU ONNX Runtime
  1.22. Add explicitly selected Mistral document OCR with fixture-tested spend
  and egress controls. Strict scores, expectations and thresholds stay unchanged;
  a separate declared typographic-equivalence view handles curly apostrophes,
  narrow/nonbreaking/thin spaces and en/em dashes.
• Update Cargo, Python, plugin and MCP Registry release metadata to 0.2.0.

Known limitations

• PP-OCRv5 can misread œ in large serif text and omit dashes with some sans
  fonts. Missing spaces before € remain errors. See the
  strict and typographic scores.
  Live Mistral API and billing behavior remains unqualified.
• Constructed AI assist qualification could not start: the release has no frozen
  corpus with an observed immutable Gemini revision or exact-source gate receipt.
  All assist features remain experimental and unqualified; no provider spend occurred.
• Image/text retrieval calibration remains unqualified. GI effect occupancy is a
  supplied proxy, not proof of physical illumination or causality.
• LPIPS, DISTS and MUSIQ exports remain deferred. TrustMark neural inference
  does not establish payload decoding.