关于「首屏:像素分配、三项网页指标的算术、骨架屏的证据与首图字节预算」,《拇指热区与首屏留存工程》给出的定量结论是什么?

结论:首屏是一块定量的地。按工具栏展开的可见高度算,夜珀 B 版连 1:1 的首图加流内价格都放不下,价格是靠常驻操作栏才留在第一屏里的。LCP 的四段里,A 版最贵的是「发现首图之前」那一段:只压缩图片,第 75 百分位从 4.72 秒降到 3.65 秒,改了发现路径才到 1.71 秒。首图的预算按 p61 的模拟是 185 KB 留余量,两倍密度的 AVIF 只要 96.0 KB,三倍密度要 216.0 KB。CLS 的大头是一张没有尺寸的首图,一次就是 0.5971 分;INP 的大头是首屏期间的第三方长任务。骨架屏能否缩短主观等待,现有证据不足以下结论,夜珀把它当作一个待测的假说,而不是默认做法。

数据口径:全部品牌与数字都是虚构样本,由书中程序算出;涉及隐私同意与无障碍的法规只写一般框架,不是法律意见。 书中以标注框区分证据、计算、原典与假说,每条假说都写明什么结果会推翻它;原典考辨取《考工记》「天有时,地有气,材有美,工有巧」的良工四条件。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 1.71, 185, 3.65, 4.72, 75

segments in seconds: TTFB + load delay + load duration + render delay = LCP
slow ( 1.6 Mbit/s, 0.150 s) A0.80 + 2.25 + 2.75 + 0.15 = 5.95 s poor
slow ( 1.6 Mbit/s, 0.150 s) B0.80 + 0.00 + 1.13 + 0.02 = 1.95 s good
typical ( 9.0 Mbit/s, 0.060 s) A0.44 + 0.66 + 0.52 + 0.15 = 1.77 s good
typical ( 9.0 Mbit/s, 0.060 s) B0.44 + 0.00 + 0.23 + 0.02 = 0.69 s good
simulated layout A: p75 LCP 4.72 s; share <2.5 s 0.223; share > 4.0 s 0.373
simulated layout B: p75 LCP 1.71 s; share <2.5 s 0.943; share > 4.0 s 0.007
layout B pipeline: largest hero with simulated p75 LCP <2.5 s: 335 KB
layout B pipeline: largest hero keeping 0.5 s spare (p75 <2.0 s): 185 KB
layout A pipeline carrying B's 96 KB file: p75 LCP 3.65 s (needs work)
查看完整验算程序
# Chapter 6. p61: LCP of the Nightamber hero in four segments, a simulated p75, and a hero byte budget.
# LCP = time to first byte + resource load delay + resource load duration + element render delay.
# All page and network numbers are assumptions for the book's sample, not measurements (KB = 1000 bytes):
#   TTFB = 3 round trips to open the connection + 1 for the HTML + 0.20 s of server time.
#   A: the hero is a slide of a carousel script (300 KB) on a separate image host; the browser learns of it
#      only after the script arrives and runs (0.15 s) and a new connection opens (3 round trips);
#      1200 x 1500 JPEG of 420 KB; 0.15 s from arrival to paint while the carousel lays itself out.
#   B: the hero is an <img> in the HTML with width, height and fetchpriority="high", same host, 1:1,
#      780 x 780 AVIF of 96 KB; 0.02 s from arrival to paint.
#   Both: 100 KB of CSS and fonts download alongside the hero and share the line.
# Population: 20,000 simulated visits; round trip log-normal around 0.12 s (sigma 0.5), bandwidth log-normal
# around 4 Mbit/s (sigma 0.8); seed 6161. A synthetic population, not field data.
# Thresholds (good / poor) as published by Google at the time of writing: 2.5 s and 4.0 s at the 75th
# percentile; always check the current official Web Vitals documentation.
import math
import random

GOOD, POOR = 2.5, 4.0
SERVER, SHARED = 0.20, 100
PAGES = {"A": dict(hero=420, script=300, run=0.15, new_host=3, render=0.15),
         "B": dict(hero=96, script=0, run=0.0, new_host=0, render=0.02)}
PROFILES = [("slow", 1.6, 0.150), ("typical", 9.0, 0.060)]          # name, Mbit/s, round trip s


def secs(kb, mbps):
    return kb * 8 / 1000 / mbps


def lcp(page, mbps, rtt, hero=None):
    p = PAGES[page]
    ttfb = 4 * rtt + SERVER
    delay = (rtt + secs(p["script"], mbps) + p["run"] if p["script"] else 0.0) + p["new_host"] * rtt
    load = rtt + secs((p["hero"] if hero is None else hero) + SHARED, mbps)
    return ttfb, delay, load, p["render"]


def grade(t):
    return "good" if t <= GOOD else ("poor" if t > POOR else "needs work")


def p75(values):
    v = sorted(values)
    return v[math.ceil(0.75 * len(v)) - 1]


print("segments in seconds: TTFB + load delay + load duration + render delay = LCP")
for name, mbps, rtt in PROFILES:
    for page in PAGES:
        parts = lcp(page, mbps, rtt)
        print("   {:<7} ({:4.1f} Mbit/s, {:.3f} s) {}: {:.2f} + {:.2f} + {:.2f} + {:.2f} = {:5.2f} s  {}"
              .format(name, mbps, rtt, page, *parts, math.fsum(parts), grade(math.fsum(parts))))

rng = random.Random(6161)
visits = [(4.0 * math.exp(rng.gauss(0, 0.8)), 0.120 * math.exp(rng.gauss(0, 0.5))) for _ in range(20000)]
for page in PAGES:
    t = [math.fsum(lcp(page, m, r)) for m, r in visits]
    print("simulated layout {}: p75 LCP {:.2f} s; share <= {} s {:.3f}; share > {} s {:.3f}".format(
        page, p75(t), GOOD, sum(1 for x in t if x <= GOOD) / len(t), POOR,
        sum(1 for x in t if x > POOR) / len(t)))


def budget(limit):
    lo, hi = 0, 3000
    while lo < hi:
        mid = (lo + hi + 1) // 2
        if p75([math.fsum(lcp("B", m, r, mid)) for m, r in visits]) <= limit:
            lo = mid
        else:
            hi = mid - 1
    return lo


print("layout B pipeline: largest hero with simulated p75 LCP <= {} s: {} KB".format(GOOD, budget(GOOD)))
print("layout B pipeline: largest hero keeping 0.5 s spare (p75 <= {:.1f} s): {} KB".format(
    GOOD - 0.5, budget(GOOD - 0.5)))
t = [math.fsum(lcp("A", m, r, 96)) for m, r in visits]
print("layout A pipeline carrying B's 96 KB file: p75 LCP {:.2f} s ({})".format(p75(t), grade(p75(t))))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《移动端拇指热区动力学、暗黑奢品微交互与首屏视觉留存工程》第 6 章「首屏:像素分配、三项网页指标的算术、骨架屏的证据与首图字节预算」(版本 0.1.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-137 · v0.1.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026137,
  author    = {Fan, Yafu},
  title     = {Mobile Thumb-Zone Kinetics, Dark-Luxury Micro-Interactions and Above-the-Fold Visual Retention Engineering},
  year      = {2026},
  publisher = {5D Library},
  edition   = {0.1.0},
  note      = {DOI pending}
}

本书其他章节结论