Sync applies no localisation rules, so every organically-synced lecture lands with English figure labels and no font config

[mmcky]

translate init applies the localisation rules — code comments, figure labels, i18n font config. The sync path applies none of them, so every new lecture arriving through the automated production path lands unlocalised, and the divergence from the seeded corpus grows with each upstream merge.

Evidence

applyLocalizationRules is referenced in exactly two places, both CLI:

src/cli/index.ts
src/cli/commands/init.ts

Neither sync-orchestrator.ts nor file-processor.ts calls it. What the sync path does have is a set of translator-prompt rules telling the model to preserve localisation that is already present (translator.ts:276, :369, :636 — the #107 mitigation). For a new file there is nothing to preserve, so it arrives with English code comments, English figure labels, and no font block.

What it looks like in production

QuantEcon/lecture-python.zh-cn#202 is a clean natural experiment: ten new lectures for the same edition in the same PR, nine delivered by sync and one by init -f (retranslated by hand after the sync run failed the parity check on it, #156).

	# i18n lines	Plot labels
Nine sync-delivered lectures	0 each	English
One init-delivered lecture	4	14 translated to Chinese

Edition-wide the init-shaped result is the convention, not the exception: 103 of 122 lectures carry font config, because the corpus was seeded with init.

So a Chinese edition is now accumulating lectures whose figures are labelled in English, silently, through the path that is supposed to be the automated one.

Why it is easy to miss

Nothing fails. The build is green either way — an unlocalised lecture is valid MyST that executes fine, and its figures simply come out in English. There is no warning on the sync run and nothing in the review verdict covers it (the reviewer sees a faithful translation of the prose, which it is). It only becomes visible by diffing a synced lecture against a seeded one, which is not part of any workflow.

It also produces a false positive in review when the two paths meet. On #202 the review flagged the init-delivered lecture's font block as a blocker — "an apparent unauthorized insertion of font-loading code lines into a Python code cell ... out of scope for a translation task" — and dropped the verdict from PASS to WARN. That block is correct and necessary: without it the lecture's 14 Chinese plot labels render as tofu. The reviewer reached the wrong conclusion because the nine siblings in the same PR set the local expectation, and they are the ones that are wrong.

Suggested fix

1. Apply the localisation rules on the sync path for newly-created files, using the same applyLocalizationRules the CLI uses, with the rule set resolved from .translate/config.yml so an edition can opt out.
2. Failing that, report it — a sync that creates a file without localising it should say so on the PR, so the gap is visible rather than silent.

Related but distinct: #107 is resync removing existing localisation; this is sync never adding it. #141 is the injection being nondeterministic and the font path hard-coded — which also bit on #202, where init wrote _fonts/SourceHanSerifSC-SemiBold.otf into an edition that keeps the font at lectures/fonts/, so addfont() would have raised FileNotFoundError on the one lecture that needed it.

[mmcky]

Hold this — it may be unnecessary work, and the deciding question is not settled.

A review of CJK font handling across the estate (project-translation reports/2026-07-24-cjk-font-rendering-review.md) reached a conclusion that changes this issue's scope.

The problem this issue describes is real and was measured: nine lectures across three editions now render Chinese plot labels as blank boxes, one of them live to readers for three months. But the recommendation is not to wire injection into the sync path. It is to stop injecting per-lecture at all, and to apply font configuration once at build level — the current mechanism is 152 hand-maintained copies of the same four lines, in four textual dialects across two different font paths, which is what keeps drifting.

If that lands, the engine's job for zh-cn shrinks from inject correctly on every path to do not break what is already there:

Issue	Fate under a global mechanism
#178 (this one — wire injection into sync)	unnecessary
#141 (nondeterministic injection, hard-coded font path)	collapses to "emit nothing"
#182 (make the #107 preservation guardrail deterministic)	remains, and becomes the whole job
#181 (neutralise usetex on the sync path)	remains

So building #178 now is plausibly work we then unpick. The decision it waits on is which global mechanism the editions adopt — and that has its own open question, since the option the investigation preferred (an IPython startup hook) had its verification blocked by a safety classifier, leaving the more conventional route (install the font into the environment plus a matplotlibrc) as the better-verified one.

Leaving this open rather than closing it: if the editions decide against a global mechanism and stay with per-lecture blocks, this issue becomes necessary again exactly as written.

One correction to this issue's framing while it is open. It claims sync applies no localisation and cites English figure labels as evidence. That half stands, but the sharper statement is that the rule is unreachable from the shipped Action bundle at all — grep -c "SourceHanSerifSC-SemiBold" dist-action/index.js returns 0, and applyLocalizationRules has exactly one call site, src/cli/commands/init.ts. It is not that sync declines to localise; it is that the Action has never contained the capability.

[mmcky]

Two 2026-08-07 data points sharpening the hand fixes are bridge repairs consequence of this gap. (1) phillips_credibility — a sync-path lecture with no font config — shipped live reader-facing tofu the day lecture-python.zh-cn published for the first time in nine months; hand-fixed in zh-cn#244. (2) The hand font-fix on ar1_turningpts survived zh-cn#230's re-delivery only because the three-way merge happened to apply cleanly — a regenerated file with different surrounding context would have dropped it silently, and the rebase automation demonstrated that exact overwrite pattern on _toc.yml the same day (four reintroductions of a trimmed entry in one day; trail on zh-cn#210). Hand localisation on sync-delivered lectures is structurally temporary until this lands.

[mmcky]

PR #225 closed 2026-09-01 without merging — not on the code, which was sound, but on three stacked holds: the localisation-rule prompt measurably induces the #118 head-of-document corruption (33% vs 93% structural pass, p = 0.002) and the rule is unchanged; decision D2 removes sync's first-translation path, which is the only path that PR wired; and this issue's fate is gated on D1 in W2 (#260). Full rationale and the reusable pieces (the localize input pattern and the test file, with one assertion change) are in the closing comment on #225. When D1 is decided against a global mechanism, the work restarts as a fresh PR against init's gap-fill, not sync.