nx_srcset_lib.nx

nx_srcset_lib.nx -- BR14: CHOOSE THE ONE IMAGE URL A RENDER SHOULD FETCH FOR THIS ELEMENT AT THIS VIEWPORT. The compare board's browser rung BR14 (symbol bi_srcset) and the named debt row `images-lazy`: "Lazy-load / srcset / data-src images are not fetched ... so media-heavy pages render most figures blank even though the decoders are byte-exact." The decoders were never the problem -- br_extract only ever looked at the `src` attribute, so a lazy-loaded <img> handed the renderer a `data:` placeholder (or nothing at all) and the real URL sat unread in srcset / data-src. WHAT THIS IS NOT: nx_img_harvest / nx_media_harvest COLLECT every image URL on a page (scraping -- they want ALL srcset candidates). This SELECTS exactly one, for one element, at one viewport. Same attribute, opposite question, so it is a second ruler for a second job, not a duplicate of the first. nx_cms_image's img_srcset_widths GENERATES a srcset ladder for publishing -- the inverse direction. NO ALLOCATION. The selection needs no candidate list: "smallest candidate that still covers the target, else the largest available" is two running extrema, computable in ONE pass over the attribute bytes. Scratch is PASSED IN (BI_SCRATCH_SLOTS) so nothing here allocates per <img>. NO NEW BYTES. Every URL this returns is a byte range INTO THE CALLER'S OWN BUFFER (off,len), which is exactly the contract br_extract/page.bsrc_off already speaks. A srcset candidate is a contiguous substring of the attribute value, so selection is a pair of integers -- no copy, no arena, no cap. INTEGER ONLY. A pixel-density descriptor is written as a decimal ("1.5x"), so every density here is PERMIL (1x = 1000, 1.5x = 1500, 2x = 2000). Permil is the estate's standing fixed-point unit and three fractional digits covers every density any shipping site writes. DELIBERATE DIVERGENCES FROM THE HTML SPEC LETTER, each named so the next reader does not have to rediscover which are bugs and which are choices: * descriptor suffix accepted in EITHER case (`2X` as well as `2x`). The spec's parser is lowercase-only. There is no other meaning for `2X`, and the divergence can only ever ADD a picture, never remove one -- it fails in the direction of rendering more, not less. * `sizes` is evaluated only in the two forms decidable WITHOUT a media-query engine (a bare `<n>px` and a bare `<n>vw`). Anything carrying a media condition or a list REFUSES BY NAME (BI_SIZES_UNPARSED) and falls back to HTML's own default source size of 100vw, so a caller can report that it did not honour the author's sizes instead of silently pretending it did. * fractional digits past permil resolution are TRUNCATED toward zero, never rounded, so a density can only ever select a LARGER candidate than the exact value would -- again the direction that renders more.