nx_manga_ocr.nx

nx_manga_ocr.nx -- MG22: FIND THE TEXT REGIONS ON A PAGE (lib; lib+gate split, no main). The inverse of the lettering compositor: mb_bubble_draw DRAWS text into a box, this FINDS the box again. That makes the referee free -- our own emitter can produce any fixture this detector must recover -- which is why this rung needs no corpus and no model, unlike MG21 CJK recognition. COMPOSES THE INCUMBENT AND RE-IMPLEMENTS NOTHING: nx_reader_panels already ships the ink primitives this needs -- rp_ink_map (integer Otsu binarize, ink vs paper), rp_rowink and rp_colink (ink counts over a row or column segment). Panel detection and text detection are the SAME projection machinery at different scales: a panel is separated by a wide gutter, a text line by a narrow leading. Writing a second ink scanner beside that one would be the duplicate-ruler defect. METHOD, and its honest scope: horizontal projection groups inked rows into BANDS separated by more than MO_LINE_GAP rows of paper, then a vertical projection inside each band gives its extent. That finds INK CLUSTERS -- text lines and blocks. It does NOT yet distinguish a speech bubble from a caption or from dense hatching, and it does not read the bubble OUTLINE; that is the next rung and this header says so rather than letting a reader assume otherwise. POLICY-AS-CONSTS, DEBT DECLARED: the four thresholds below are named rather than buried, but named is not data -- they belong in a conf row beside knowledge/manga/layouts.conf, exactly as nx_manga_bubble's lettering style does. Filed as MG22-conf rather than left silent.