PhraseLens v0.6.0 — A vocabulary that sorts itself

Added

• Saved words are filed for you, and the collection can be cut apart. A vocabulary
  list past a screenful was a wall: a headword, a fixed expression and a whole sentence
  kept for its shape tiled together with nothing to tell them apart, and the only way
  back to one of them was remembering how it was spelled. Every word is now filed along
  five dimensions as it is saved — what kind of thing it is, its part of speech, the
  subject it belongs to, its register, and how hard it is, with the native rung ("N2",
  "B1") kept alongside for the languages that have one. Four of the five are closed
  sets, which is what lets the filter rail draw stable rows; topic is open, because what
  a reader's material is about is decided by the material, and a request offers the
  topics already in use back to the model so a collection converges on a handful of
  buckets instead of coining a synonym every twenty words.
• A filter rail with counts that stay honest. Rows inside a dimension widen the
  result; rows across dimensions narrow it. Each dimension's counts are taken with its
  own choices lifted, so arming Word leaves Phrase and Sentence standing with
  their real totals rather than collapsing every sibling to zero — no click lands
  somewhere empty. The rail folds into a pull-down on a narrow window.
• Grouping. Break the collection into pinned sections by any of the same dimensions,
  or by the month you collected it. Closed dimensions keep their declared order rather
  than sorting by size, and a word filed under two topics is browsable under both.
• Organize. Words collected before this existed are picked up by one pass, twenty per
  request through the provider you already configured, saving each batch as it lands so
  a cancelled or dropped run keeps everything it finished.

Every tag is optional at every layer. Words saved before this existed decode untouched,
a model that cannot judge a dimension is told to leave it out rather than guess, and
anything unfiled stays reachable through a Not tagged row instead of dropping out of
the collection. Tags are read on every launch from the file that holds the words
themselves, so an unrecognized value costs that value and not the collection.

Fixed

• The answer comes back in the language you picked, even when the source is a wall of
  another one. Summarize, Analyze and Explain in Context all ended their command
  prompt with the material itself, and a model that reads several hundred words of
  Japanese last answers in Japanese. Each now closes on the line naming the target
  language, and the shared answer-language rule states outright that the material never
  overrides it. Words, examples and quotes stay in the source language; everything
  written about them is in yours.

---

The app is signed but not notarized, so the first launch after downloading needs a right-click → Open.

Full changelog: v0.5.3...v0.6.0