The strip stopped building every tile two commits ago, but the wall was left holding
every frame on purpose: the grid view was the one list that still made a card per
frame. A card is an `<article>`, an `<img>`, an object URL and a date the browser
formats, and on a roll of twenty thousand the first one did not paint for 7410ms —
19998 cards and 19998 pictures, 7007ms of long tasks, the largest of them 4882ms. The
object URL is only about a tenth of it (83µs each); the rest is the DOM and the i18n
date formatting behind every card.
The wall now draws the rows under its viewport and `WALL_ROWS` either side, measured
from `scrollTop` and the client height on scroll and on resize, and two spacers stand
in for the rows that are not drawn, each as tall as the rows it replaces so the
scrollbar still spans the whole shelf. The step from one row to the next is the
average of the rows in hand rather than the smallest of them: the cards do not all
stand the same height — a caption that wraps makes its row taller (274.34 / 259.36 /
274.36 measured) — and `offsetTop` is rounded to whole pixels, so the least step was
a pixel short on every one of thousands of rows and the scrollbar came up 1436px shy
of the end. The average puts `scrollHeight` back on the number the fully drawn wall
had, to the pixel (1085426).
Measured against a seeded roll of 20000: the first card 7410ms → 212ms, cards drawn
19998 → 25, long tasks 7007ms → none, and the scrollbar unchanged. A check on a roll
of 4000 walks the wall end to end — the last frame drawn at the bottom, the first
drawn again at the top, spacer 215719px either side, `scrollHeight` the same 217076
at both ends, no drift — and coming back to the strip still leaves eighteen tiles.
A card that has a thumbnail and has not been handed its URL yet now keeps its box in
silence instead of saying RAW, since the wall hands pictures out only around the eye;
the word is left for a frame that has no thumbnail to give.
3dtours
e53d6264ca