@onlyworlds/sdk 4.11.0

Added

• World lookup. buildWorldIndex(elements), lookupWorld(index, question, { limit }) and
  renderElement(element, index) find the elements a plain-language question is about, in a
  world held in memory: no model, no network, a few milliseconds. Ranking: names and short
  forms first, then name words, supertype and subtype, one link step in both directions, then
  other text and numbers (always below any name hit). Each hit says why it matched.
  renderElement gives a compact text card with links by name, which is what a model reads
  when one phrases the answer. Measured on two real worlds (Assembly's Lantern Ask trial):
  the answer was in the first three results for 60 of 60 questions that name their subject
  and 17 of 24 held-out questions that only describe it; this code returns the same top 8
  as the trial's on all 100 questions. Names in any script are matched (Chinese and Japanese
  only where spaces or punctuation separate the words: there is no word splitting). The question words
  that steer the link step ("where", "rules", "allies") are English.
• Photo preparation. preparePhoto(blob, { maxSide, quality, maxBytes }) shrinks and
  converts a picture in the browser before uploadImage: longest side 2048 by default, webp at
  quality 0.82, from the upload page's converter. A browser that cannot encode webp returns a
  PNG when asked; the result's real type is checked, and a photo then falls back to JPEG while a
  PNG stays PNG. When re-encoding would only make a file bigger, the original is kept.
  planPhoto is the pure part (sizes and type order). Walked in Chromium, WebKit and Firefox
  (a 4032×3024 photo became a 2048-wide webp of about 0.5 MB) and with webp encoding forced off.
  Needs a browser; in Node it throws a TypeError that says so. PHOTO_TYPES,
  PHOTO_MAX_BYTES are keel's accepted types and its 15 MB limit.