ci: stop refetching translations and rebuilding JS in downstream jobs

[jkmassel]

Summary

• Cut PR build wall-time from 13:15 → 9:00 (~32% faster) and per-build agent-time by ~10 minutes.
• Single source of truth for translation fetching and JS compilation: only Build React App does either, and every other job consumes the dist.tar.gz artifact.
• Drop node-fetch in favour of Node 20's built-in fetch so the translation fetch can run in parallel with npm ci.

Speedup

Compared against trunk #2251 (last clean run before this PR):

Step	Trunk #2251	This PR (#2307)	Δ
Build React App	2:43	1:50	−53s
Test Android Library	4:19	2:43	−1:36
iOS Simulator Tests	4:46	2:57	−1:49
Library Tests	3:22	1:18	−2:04
Build XCFramework	4:55	1:29	−3:26
Test Android Library Instrumented	2:11	0:57	−1:14
Test iOS E2E	5:16	5:22	similar
Test Web E2E	6:00	6:07	similar
Other steps	(no change)	(no change)	noise
Total wall-time	13:15	9:00	−4:15

Build XCFramework was on the wall-time critical path; trimming npm ci + JS build + translation fetch from it accounts for most of the headline win.

Root Cause

Three compounding inefficiencies:

1. build-react re-fetched translations every build

make build REFRESH_L10N=1 STRICT_L10N=1 defeated the Makefile's existence-based cache (Makefile:48), forcing all ~50 locales to be re-fetched from translate.wordpress.org on every CI build. Buildkite agents don't share state, so every build paid the full GlotPress round-trip cost.

2. Every downstream `: build` consumer re-fetched translations and re-ran `npm ci`

test-swift-library, test-swift-package, test-android, test-android-library-e2e, and build-resources-xcframework all declare build as a Make prereq. build: npm-dependencies prep-translations, so both ran first. On agents that just extracted an upstream dist.tar.gz:

• dist/ is populated → build's recipe short-circuited the JS build
• src/translations/ is empty (not in dist.tar.gz) → prep-translations ran the full fetcher
• node_modules is empty → npm-dependencies ran npm ci
• All of that work was then thrown away because dist/ already had everything baked in

3. Two Android jobs rebuilt dist/ from scratch

Test Android Library and Test Android Library Instrumented had no upstream dependency, so they walked the full : build chain — fetch all locales, npm ci, npm run build, copy assets — before finally running gradle.

Fix

.buildkite/pipeline.yml

• Build React App: backgrounds make prep-translations STRICT_L10N=1 while running make npm-dependencies in the foreground, then waits on the fetch before make build REFRESH_JS_BUILD=1. Shell vars escaped with $$ for Buildkite interpolation.
• Test Android Library and Test Android Library Instrumented: gain depends_on: build-react and download dist.tar.gz.

Makefile

• prep-translations: gains a "skip if dist/ exists" clause. Translations are baked into the bundle at JS build time, so downstream consumers don't need a fresh fetch. REFRESH_L10N=1 and direct invocation still force the fetch.
• build: drops npm-dependencies from its prereq list and invokes it via $(MAKE) inside the rebuild branch. Downstream : build consumers (which run gradle / xcodebuild / swift, not anything from node_modules/.bin) no longer pay for an npm ci they don't need.
• npm-dependencies: drops the "direct invocation forces install" clause — build's recipe now invokes it via $(MAKE) which would otherwise trigger the force. REFRESH_DEPS=1 remains as the explicit force.
• test-android and test-android-library-e2e: copy dist/ into Android assets unconditionally. build's recipe short-circuits copy-dist-android when dist/ already exists, so CI agents that download a dist artifact would otherwise have stale assets. Matches what test-android-e2e:329-331 and build-resources-xcframework:131 already do.

bin/prep-translations.js

Drop import fetch from 'node-fetch' in favour of Node 20's built-in global fetch. The script's surface (response.ok, response.status, response.headers.get, response.json()) is API-compatible with WHATWG Fetch in both implementations. Also drops node-fetch from package.json and package-lock.json (98 lines of lockfile).

Decision matrix for prep-translations

dist/	src/translations/	MAKECMDGOALS	REFRESH_L10N	Outcome
absent	empty	build (or anything)	unset	fetch
absent	populated	build (or anything)	unset	skip (bundles present)
present	either	not prep-translations	unset	skip (dist built) — new
either	either	prep-translations	unset	fetch (direct invoke)
either	either	anything	1	fetch

What we explored

1. CI-level cache with a weekly key ❌

Cache src/translations/ with a key that rolls weekly. Faster on most builds but accepts up to a week of staleness in shipped strings. Rejected — fresh-per-build is cheap once we stop fetching per-job.

2. Run prep-translations as a separate Buildkite step ❌

The first iteration of this PR did exactly that. It bought the agent-time win but added 2:39 of serial wall-time on the critical path because the fetch waited for its own agent startup + npm install. Replaced by the in-step parallel approach (see commit 9b4e35c1).

3. Swap hand-rolled fetcher for Fastlane::Helper::GlotpressDownloader

Separate concern — doesn't change CI speed. Tracked as a follow-up; see the release-toolkit helper.

Test plan

• CI green on this PR (build #2307, 9:00 total)
• Per-job wall-time savings on Library Tests, iOS Simulator Tests, Test Android Library, Build XCFramework, and Test Android Library Instrumented — confirms the dist/-based skip is firing
• Local: make prep-translations still fetches when invoked directly (explicit user intent)
• Local: make test-swift-library after a successful make build logs both "Skipping translations fetch (dist/ already built…)" and "Skipping NPM dependencies installation (node_modules already exists)" — and does not invoke npm ci
• Spot-check that STRICT_L10N=1 still fails the prep step on a real GlotPress error (not exercised by this PR's builds — verified previously)

[wpmobilebot]

XCFramework Build

This PR's XCFramework is available for testing. Add the following to your Package.swift:

.package(url: "https://github.com/wordpress-mobile/GutenbergKit", branch: "pr-build/513")

_{Built from 8b403ce}

[oguzkocer]

Looks good. I raised an issue, but that's not a regression, a possibly missed opportunity instead. So, I am approving it as is.

[oguzkocer]

I'm not familiar with this pipeline, but if I am understanding it correctly, the translations are baked into the bundle we get from dist/ which means if it exists, we don't need translations at all. However, the check [ -n "$$(find src/translations -maxdepth 1 -name '*.json' -print -quit 2>/dev/null)" ] makes it so that it'll still get downloaded - and not used?

If that understanding is correct, it'd mean that we are downloading translations in downstream CI builds where we have dist/ but not the translations folder, and then simply discard the newly downloaded translations because there is no build step.

I think we can skip that check and save some time in CI steps such as Test Android Library.