Release Scan Tools 0.2.0

public/docs/ARCHITECTURE.md

@@ -5,13 +5,17 @@ serializable page settings. No page content enters local storage. Each source is
 decoded with `createImageBitmap`; its dimensions and byte size are checked before
 it joins the workspace.
 
-The renderer downscales exceptionally large source rasters to a 24-megapixel
-working ceiling. A projective map sends every destination pixel through the four
-selected source corners and samples it bilinearly. A bounded pixel pass applies
-brightness, contrast, perceptual grayscale and optional Otsu/manual threshold.
-Canvas rotation handles quarter turns and the final ±5 degree deskew. Preview,
-OCR and export call the same renderer with different disclosed pixel budgets, so
-the preview semantics match the downloaded result.
+The renderer uses a disposable module worker and `OffscreenCanvas` when the
+browser supports them. It downscales exceptionally large source rasters to a
+24-megapixel working ceiling, then a projective map sends every destination
+pixel through the four selected source corners and samples it bilinearly. A
+bounded pixel pass applies brightness, contrast, perceptual grayscale and
+optional Otsu/manual threshold. Canvas rotation handles quarter turns and the
+final ±5 degree deskew. Worker jobs report stages, are terminated by abort and
+cannot replace a newer preview. A bounded HTML-canvas fallback preserves the
+same result where Worker/OffscreenCanvas are unavailable. Preview, OCR and
+export call this renderer with different disclosed pixel budgets, so the preview
+semantics match the downloaded result.
 
 Deskew estimation downsamples to at most 420 pixels on the long edge, tests
 half-degree candidates from −5 to +5, and scores the horizontal projection of
@@ -24,3 +28,10 @@ runtime variants and English integer trained data are copied from pinned npm
 packages to same-origin release assets. Tesseract's persistent model cache is
 disabled, one in-memory worker is reused and can be terminated, and no fallback
 CDN path is used.
+
+OCR requests bounded block geometry, hOCR and TSV alongside plain text. Word
+boxes are normalized to the corrected OCR raster for an inert SVG overlay and
+are capped at 100,000 entries. hOCR is never injected into the application DOM;
+it and TSV are download-only evidence. Any subsequent page adjustment clears
+positioned OCR artifacts while retaining the user's editable plain text. PDFs
+remain image-only rather than silently embedding a potentially misaligned layer.