docs(release): notes for v0.1.25

Author: Schaerer Damien

app/RELEASE-NOTES-v0.1.25.md

@@ -0,0 +1,59 @@
+# Galactus Desktop v0.1.25
+
+A native macOS app for the Galactus MoE engine: run certified open-weight
+Mixture-of-Experts models fully on-device, including models several times
+larger than your RAM.
+
+Designed and developed by Noxalis Lab.
+
+This one gives the agent your documents: a spreadsheet of instructions on
+one side, Word and PDF files to change on the other, all of it on your Mac
+and none of it leaving it.
+
+## A spreadsheet finally reads as a table
+
+An .xlsx used to arrive as its list of words followed by a grid of numbers,
+because a sheet stores indexes into a shared string table and dates as the
+day counts they are. Anything driven by that table was one bad join away
+from the wrong result. Workbooks now read back as CSV, one block per sheet,
+carrying the Excel row number so a report can name a row you can find. Empty
+cells keep their column, dates read as dates, and a formula gives the value
+you see in the cell.
+
+## Word documents can be edited, and keep their layout
+
+Replace a sentence, add a paragraph after another, add one at the end. Only
+the parts that hold the text are rewritten; styles, images, numbering, page
+setup and everything else are copied through untouched, which a test checks
+entry by entry.
+
+It searches the joined text of each paragraph rather than each run, so a
+sentence with a bold word in the middle is still found, and it looks in
+headers, footers, footnotes, table cells and text boxes, not only the body.
+
+## PDFs too, with the trade a PDF imposes
+
+The same four operations. A replacement covers the old sentence and draws
+the new one in its place, and the page that changed is re-drawn as an image
+so the old words really are gone rather than merely hidden under a white
+box. That page loses its selectable text; every other page keeps it.
+
+## What a Word document cannot do, said out loud
+
+There are no pages in a .docx. Word computes them when it lays the file out,
+from the paper size, the fonts, the images and even the printer driver, so
+the same document paginates differently on two machines. Asking for a page
+is refused with the reason and with what to use instead: a range between two
+headings ("Article 4" to "Article 5"), a paragraph number, or an occurrence
+number. Searching also reports how many explicit page breaks a document
+carries, so the point can be shown rather than argued. On a PDF, pages exist
+and page numbers work.
+
+## Housekeeping
+
+A new skill, `documents-depuis-tableau`, teaches the order that keeps a
+batch honest: restate the table before touching anything, dry run every row
+with the operation that writes nothing, show it for approval, then edit and
+re-read. Nothing above adds a dependency: PDFKit and Vision ship with macOS,
+zipfile and ElementTree ship with Python, and both scripts refuse a document
+that declares XML entities before parsing it.