ci: Mitigate translation download 429 errors

[dcalhoun]

What?

Reduces translation download requests from ~192 to ~48 per CI run by building once and sharing artifacts, removing translations from postinstall, and adds robust 429 rate limit handling so the remaining 48 requests recover gracefully when throttled.

Why?

Fix CMM-1250.

Every Buildkite CI run makes ~192 HTTP requests to translate.wordpress.org because 4 of 6 pipeline steps independently download all 48 translation locales. This causes 429 (rate limiting) errors. The pipeline had no artifact sharing — the "Build React App" and "Publish Android Library" steps each ran a full make build REFRESH_L10N=1, and the lint/test steps triggered downloads via npm ci's postinstall hook.

Even after reducing to ~48 requests, builds were still failing because the download script's retry logic was too aggressive — batch size of 10 with 1-2s backoff meant retries also got 429'd immediately.

How?

1. Remove prep-translations from postinstall in package.json

Translation downloads are a build concern, not an install concern. The make targets (make build, make dev-server, etc.) already handle translations via the prep-translations dependency. This eliminates the translation downloads that were happening in lint/test CI steps (which run npm ci but don't need translations).

2. Restructure .buildkite/pipeline.yml

• Build React App gets a key: build-react and uploads dist/ as a Buildkite artifact
• Lint and Test JS are unchanged — npm ci no longer triggers translation downloads
• Publish Android Library gets depends_on: build-react, downloads the pre-built dist/ artifact instead of running its own make build

Step	Before	After
Build React App	48 requests	48 requests (unchanged, sole source)
Lint React App	48 requests	0 (postinstall no longer downloads)
Test JavaScript	48 requests	0 (postinstall no longer downloads)
Publish Android	48 requests	0 (uses artifact, no npm ci)
Total	~192	~48

3. Fail build on translation fetch errors in bin/prep-translations.js

Previously, translation fetch failures were silently swallowed. The script now properly propagates errors so CI fails visibly rather than producing a build with missing translations.

4. Improve 429 rate limit handling in bin/prep-translations.js

• Reduce batch size from 10 to 5 concurrent requests
• Add inter-batch delay of 2s between batches to spread load
• Detect 429 specifically via a RateLimitError class that carries the parsed Retry-After value
• Respect Retry-After header (supports both delta-seconds and HTTP-date formats)
• Exponential backoff for 429s without Retry-After: 5s, 10s, 20s, 40s + jitter
• Linear backoff for other errors: 1s, 2s, 3s (unchanged behavior)
• Increase max retries from 3 to 5
• Add 60s max backoff safety limit to avoid blocking CI indefinitely
• Log configuration at startup for CI visibility
• All parameters configurable via environment variables (L10N_BATCH_SIZE, L10N_BATCH_DELAY_MS, L10N_MAX_RETRIES, L10N_MAX_BACKOFF_MS)

Testing Instructions

1. Verify in Buildkite that:
   • Build React App step downloads translations and uploads dist.tar.gz
   • Lint and Test JS steps run npm ci without downloading translations
   • Publish Android Library step waits for build, downloads artifact, and publishes without running npm ci
   • Any 429 errors from translate.wordpress.org are recovered via backoff and retry
2. Local verification:
   • Run npm ci and confirm translations are not downloaded
   • Run npm run prep-translations -- --force and confirm translations download with the new config logged (batch=5, delay=2000ms, retries=5, maxBackoff=60000ms)
   • Run L10N_BATCH_SIZE=2 L10N_BATCH_DELAY_MS=5000 npm run prep-translations -- --force and confirm env var overrides are respected in the config log
   • Run make dev-server and confirm the editor works with translations

[dcalhoun]

Upload the build artifact for later tasks to avoid unnecessary duplicate builds.

[dcalhoun]

Rely upon the single build created by an earlier task.

[dcalhoun]

Expand the script robustness to include more sophisticated backoff logic when 429 errors occur.

[dcalhoun]

Remove the default translation downloads that occur in npm's postinstall script.

The downside is that someone running npm install rather than the project's documented recommendation of make dev-server or make build will result in absent translation strings for new clones.

The upside is that avoiding unnecessary or duplicative translation string fetches is far simpler.

[dcalhoun]

Add STRICT_L10N=1 to ensure the release build fails if downloading translation strings fails. Otherwise, the release may lack expected translations.

[nbradbury]

All good!