Summary

Production answered 40–60 % of requests with 500 Cannot redefine property: $i18n, on every route and in both locales, even for strictly sequential requests. Cloudflare error 1102 appeared in between.

Cause (reproduced against the Workers preview of main):

1. On Workers Free a request gets 10 ms CPU. A server render here needs more, so Cloudflare aborts renders midway with error 1102.
2. The abort leaves Vue's module-global currentInstance pointing at the dead request.
3. In the next request in that isolate, @nuxtjs/i18n resolves the Nuxt app via useNuxtApp(), which prefers getCurrentInstance()?.appContext.app.$nuxt. It gets the dead request's app, and Object.defineProperty(nuxt, "$i18n", …) throws.
4. The isolate stays broken until Cloudflare recycles it.

The Vue version is not the cause: 3.5.42 (production) and 3.5.43 (main) behave identically.

Changes

• Account upgraded to Workers Paid, done in the dashboard. It removes the trigger by raising the CPU limit to a 30 s default.
• limits.cpu_ms: 5000 in $production of nuxt.config.ts: a guard against runaway renders, far above a normal render. It will be re-tuned to 10× the measured p99 after 48 h. Free rejects limits, so this PR must only merge on Paid.
• server/plugins/error-response-headers.ts: 5xx responses get X-Robots-Tag: noindex. Before, @nuxtjs/robots had already stamped index, follow before the render failed.
• scripts/concurrency-check.mjs: load, locale-isolation and content-isolation checks. It runs in CI against pnpm preview:prod (new script, wrangler pinned as devDependency) and catches per-request state leaking into module scope. Run it against production after a deploy to catch CPU aborts, which wrangler dev cannot reproduce.
• AGENTS.md: new "Deploy" section.

⚠️ Action required in the Cloudflare dashboard

Production does not run main. Every main check run of Workers Builds reports a Version ID and a Preview Alias URL (main-launchpad…workers.dev). The branch is treated as a preview branch: versions are uploaded, nothing is deployed. Production still serves Vue 3.5.42, while main has had 3.5.43 since #345 (2026-09-19).

Fix: Workers → launchpad → Settings → Builds. Set the production branch to main and the deploy command to npx wrangler deploy. Until then this PR, like every PR since #341, will not reach visitors.

Verification

• pnpm test: 57 files, 186 tests green, including a new unit test for the robots plugin.
• ESLint 317/319 errors, 36/36 warnings, TypeScript 21/21. No rise against the baseline.
• .output/server/wrangler.json contains "limits": { "cpu_ms": 5000 }.
• curl -I on a failing route under preview:prod shows x-robots-tag: noindex.
• CI concurrency step green on this PR.
• After the dashboard fix and the first deploy: the entry bundle on onelitefeather.net carries the lockfile's Vue version, and the concurrency check against production is green.
• 48 h later: 5xx rate < 0.1 %, and no Exceeded CPU Time Limits / Exceeded Memory in the invocation statuses.

Follow-ups

• Upstream issue for nuxt-modules/i18n: the plugin should use the app it is handed rather than useNuxtApp().
• The planned edge-caching-and-seo change assumed the Free plan and gets revised for Paid.

🤖 Generated with Claude Code

https://claude.ai/code/session_012ZzvomAnwjAS3et96Avp7s