Three lectures render Chinese plot labels as tofu — no matplotlib font configuration

[mmcky]

Three lectures set Chinese text on matplotlib figures but carry no font configuration, so matplotlib falls back to DejaVu Sans, which has no CJK coverage. The labels render as tofu — blank rectangles — in a valid PNG, and nothing fails.

Lecture	Chinese plot text	How it broke
divergence_measures.md	7 sites, e.g. plt.ylabel('密度'), axes[0].set_xlabel('KL散度 KL(f, g)')	never had a block; labels hand-added 2025-09-12, after the blanket font pass
kesten_processes.md	4 sites, e.g. ax.set_ylabel("收益率"), ax.set_xlabel("日期")	regression — block deleted by resync da7b176 (#190), labels kept
jv.md	via titles = [r"$s$策略", r"$\phi$策略", "价值函数"]	block stripped by resync

Fix

Add the edition's standard block to each, above first plotting use:

FONTPATH = "fonts/SourceHanSerifSC-SemiBold.otf"
mpl.font_manager.fontManager.addfont(FONTPATH)
plt.rcParams['font.family'] = ['Source Han Serif SC']

Note fonts/, not _fonts/ — this edition keeps the asset at lectures/fonts/SourceHanSerifSC-SemiBold.otf and the other 103 configured lectures all use that path.

Context

Not yet visible to readers: this edition last published 2025-11-06 and the breakage is post-dated, so the live site still serves pre-resync content. That is timing, not safety — it ships on the next publish.

Root cause and the recommendation against reverting to usetex are in the program review: project-translation reports/2026-07-24-cjk-font-rendering-review.md. Engine side is action-translation#107 — the resync drops i18n adaptations. A further ~12 lectures here use matplotlib with English labels and no config, which becomes tofu the moment anyone localises a label.

CI cannot currently catch any of this: the missing-glyph warning is raised in the kernel subprocess, so -W never sees it.

[mmcky]

A runtime scan of the published site adds two pages, one of which is already fixed in main.

bin/check-tofu in project-translation reads the missing-glyph warnings out of the published HTML, where myst-nb renders the kernel's UserWarning as stderr output. Against https://quantecon.github.io/lecture-python.zh-cn (87 pages — the deployment is behind main, which has 121):

Page	Characters lost	Status
divergence_measures	与关切叠夫密尔度散熵的系诺重	this issue, fixed in #206
stats_examples	分布本特福	already fixed in main — its block was added by resync aa9ccb7 on 2026-07-19, after publish-2025nov06, so this resolves on the next publish
ols	一与之亡人入关图均对平年征收数死殖民率的系者间阶险风（）：	not yet explained

ols.md has carried the font block since c4c4b86 (2025-03-18), well before publish-2025nov06, so the block was live at publish and the labels still lost their glyphs. I have not diagnosed why. Worth checking for a later rcParams reset in the same document, a figure drawn before the block's cell, or an execution failure that skipped it — matplotlib.md in the programming edition turned out to be the first of those (plt.style.use() resets font.family), so that is the first thing to rule out.

Method note

Source-level analysis found divergence_measures, kesten_processes and jv here; the runtime scan found divergence_measures, ols and stats_examples. The overlap is one. Static analysis catches source defects including unpublished ones; runtime catches what readers actually see, plus labels built dynamically. Keep both.