SitUStatis Multibahasa: Jebakan Sumber dan Output di Direktori yang Sama

Situs ini punya 180 halaman dalam 5 bahasa. Suatu hari halaman bahasa Inggris memiliki nav bahasa Inggris tetapi seluruh badan teksnya berbahasa Mandarin — dan sudah daring selama beberapa hari. Penyebabnya bukan terjemahan hilang, melainkan urutan build.

Symptom: only the nav was translated

The dropdown correctly showed EN and the nav was in English, but the hero, cards and lists were all Chinese. This half-right, half-wrong shape is easy to misread as missing translation data.

The real diagnosis is not the page but the source: count how many language spans remain in each page's body.

# 统计 </nav> 之后的多语 span 数量
import re
s = open("src5/tools/chatgpt.html", encoding="utf-8").read()
body = s[s.find("</nav>"):]
print(len(re.findall(r'class="(?:zh|en|id|sl|it)"', body)))

The answer was 0. All 36 source pages had zero body spans; only the 25 in the header nav survived, because a separate script had re-added those later.

Root cause: source and output in one directory

The pipeline worked like this: generators wrote pages with five-language spans to the root directory, then the build step folded those spans into per-language pages and wrote the Chinese version back to the same path, since deployment needs Chinese at the root.

That step overwrites the source with an already-folded single-language page. The next snapshot then captures pages that have no spans left, so folding has nothing to choose from and all five languages come out Chinese.

Fix: make the order executable

Memorising the order does not work: conversations forget it, people get it wrong. Put the order in a script so it becomes part of the code, and make a missing step fail the build.

# build_all.py —— 唯一构建入口
STEPS = [
    ("i18n_lint.py",      "翻译数据自检"),
    ("gen_content.py",   "生成内容页（带 5 语 span）"),
    # ...
    ("inject_seo.py",    "注入 SEO 头（仅根页）"),
    ("snapshot_src5.py", "快照多语源 ← 必须在折叠之前"),
    ("build_i18n.py",    "折叠成 180 个单语页"),
    ("audit_site.py",    "全站自检，不过则不部署"),
]

• The snapshot step self-audits: any page with zero body spans or a duplicated footer exits with code 1
• The injection step is limited to root pages, skipping src5 and the language subdirectories

Verify: don't eyeball the pages

Eyeballing only covers the page you happened to open, and it cannot see time-order problems such as "source already degraded while output still looks fine". Let the script report the numbers.

# 构建末尾断言
assert span_count(body) > 0, "正文无多语 span（会退化成单语）"
assert footer_count(page) == 1, "页脚重复"
assert len(re.findall(r'rel="alternate" hreflang', head)) == 6

The last line came from a real misdiagnosis: counting hreflang across the whole page also counted the language menu's <a hreflang> attributes, wrongly reporting duplicates. The <head> actually had exactly six. When a number looks wrong, verify the measurement scope before changing product code.

Alat terkait