docs: record why the translation engine lives in-tree

A from-scratch CPU inference engine is the largest, riskiest part of
the codebase and the least connected to screenwriting, so it deserves a
written decision. It stays: the alternative (a ctranslate2/argos
dependency, a plugin or a separate binary) buys a single menu item at
the cost of cgo and per-distribution packaging, while the in-tree
engine keeps the one-binary, offline, GPL-clean build. The package doc
now says so, with the condition that should reopen the question
(more language pairs).

AGENTS.md

@@ -62,7 +62,8 @@ internal/layout         line wrapping + pagination (paper sizes, locked pages);
 internal/pdf            hand-written PDF 1.4 writer, standard Courier fonts, zero dependencies
 internal/spell          Hunspell .aff/.dic reader, checker (compounds too), suggestions, word scan
 internal/marian         MarianMT (OPUS-MT) inference in pure Go: safetensors, SentencePiece unigram,
-                        transformer encoder-decoder, beam search, verified download
+                        transformer encoder-decoder, beam search, verified download. In-tree on
+                        purpose (one binary, no cgo, offline by design): see the package doc.
 internal/translate      French → English script translation: language detection, screenplay term
                         tables, per-element rules, name protection, parallel sentence translation
 frontend/index.html     app shell: header bar, navigator, paper, HUD, <dialog>s