Release v0.6.35: reduce resident memory and compact shared dictionaries

2026-11-04 14:08:05 +02:00
parent 004fee4976
commit d13e9849a7
21 changed files with 469 additions and 100 deletions
@@ -627,7 +627,12 @@ support automatic correction. Dictionary recognition is heuristic, not an unders
 subtitle vocabulary is incomplete, and uncommon names or jargon can need personal exceptions.

 No network service or language model is used. The three dictionary files occupy about 2.6 MB on disk;
-the English/Ukrainian lookup data uses less than 3 MiB of RAM and is loaded on first automatic check.
+the bundled EN/UK/RU lookup payload uses about 0.90 MiB of RAM. When automatic correction is enabled,
+selected dictionaries are loaded before typing and shared by the input predictor and worker.
+Words are stored in compact indexed blocks; all words and frequency ranks are retained, and editable
+UTF-8 dictionary files keep the same format. Custom dictionaries can use more memory.
+`build.ps1 -Test` checks every bundled word/rank and measures lookup speed; CMake exposes the same
+check as `dictionary-regressions`. The payload figure excludes the rest of the process and allocator overhead.
 Keep `dictionaries/` beside the executable when copying the app. These word lists derive from
 [FrequencyWords](https://github.com/hermitdave/FrequencyWords) and are separately licensed under
 CC BY-SA 4.0; attribution, pinned source revision and license are in `dictionaries/`.