nx_img_harvest.nx

nx_img_harvest.nx -- S-CLASS-EXCEED image-URL harvester (operator 2026-07-03: "no we want to get s class exceed" -- the generic <img src>-only extractor missed LAZY-LOADED + JS/JSON-embedded pages, which is how modern manga readers actually ship their page lists; gallery-dl's edge is exactly this). MULTI-STRATEGY, raw scan of the page bytes so it is agnostic to WHERE the URL lives: * <img src=..> (the old floor) * lazy attrs data-src / data-original / data-lazy-src / data-url (quoted values are just strings) * srcset="a 1x, b 2x" (each comma/space token checked -> highest set of candidates) * JS / JSON page arrays "pages":["https:\/\/cdn\/01.jpg", ...] window.__data / __NEXT_DATA__ / chapter.images Mechanism: scan for every QUOTED region (" or '), tokenize its content on space/comma, UNESCAPE JSON "\/"->"/", and if a token LOOKS like an image URL (path ends .jpg/.jpeg/.png/.webp/.gif/.avif before ?/#; NOT data:/.svg/ sprite/favicon UI-junk) resolve it (absolute passes through; relative via the proven nx_url_resolve) + DEDUP. Order = appearance order (matches reading order for HTML galleries + JSON arrays). Returns the count; fills the same offs/lens/url_buf interface nx_html_extract_imgs uses, so nx_manga_get swaps it in with no other change. The completeness magic-guard in nx_manga_get still fetch-verifies each (valid FULL file), so a stray non-page URL that slips the heuristic simply fails the fetch -- never a blank/false page. license_tier: ORIGINAL