[jiro-by-nori]

Summary

🤖 Generated with Nori

Root cause (verified on the live site): the production service worker precached all 24,918 files in public/ (≈150 MB; ~24.6k are per-sprite atlas-packer inputs the game never requests) in one atomic cache.addAll with cache: 'reload'. Pages/Fastly answered the burst with 503 (server: Varnish, retry-after: 0, HTML body). One non-OK response rejected addAll, the worker was discarded with nothing cached, and every reload / reopened tab started the whole ~25k-request download again. In a clean profile on https://opencoaster.noriagentic.com/ the page never became controlled across first visit, 3 reloads, hard reload and tab reopen; each load issued 3k–24k worker requests, 21k of them aborted after the 503.

• tools/pwa/precache-urls.ts (used by vite.config.ts): precache only the bundle, shell and the public files the game requests (sprites/manifest.json, sprites/atlas/**, sprites/ui-icons/**; no CNAME). 24,918 → 278 URLs; worker script 1.68 MB → 12.6 KB. Offline play keeps every atlas page.
• src/pwa/service-worker.js: install downloads ≤6 files at a time, stops at the first failure, and resumes a retried install from what this build's cache already holds. Activation/fetch strategy unchanged; IndexedDB/localStorage never touched. No client reload logic (no reload loop risk).
• Docs: src/pwa/docs.md, docs/deploy/README.md, smoke/docs.md; corrected stale claims that UI icons are excluded from the atlas.

Test Plan

• vitest src/pwa tools/pwa tools/sprites/dist-icons.test.ts: resumable install, bounded concurrency + cache: 'reload', precache selection (each failed first on main).
• smoke/returning-visitor.spec.ts (production build under /opencoaster/, no COOP/COEP): returning visitor with an earlier build's cache, IndexedDB save and settings → worker controls with <1000 downloads, 3 reloads fetch only the navigation, old cache removed, save + settings intact, Load game lists the save. Fails on main's build (worker never controls within 60 s even on localhost).
• Old-worker → new-deploy upgrade (local, real main-build worker installed first): new build installs with 280 downloads, then each reload = 1 download, 54/54 assets from worker, old cache gone, save intact.
• lint, format, typecheck.
• Note: returning-visitor.spec.ts only runs under pnpm smoke:subpath, which CI does not run; adding it to ci.yml needs the workflow scope.
• Live verification after deploy (proof comment to follow).

Requested in Nori: https://demo.norisessions.com/chat/01a16c13-c814-7295-9955-b11ff875ed14

Share Nori with your team: https://www.npmjs.com/package/nori-skillsets

[jiro-by-nori]

Live proof (deploy run https://github.com/tilework-tech/opencoaster/actions/runs/41946045318, main @5b47730c)

	Before (main @152523b0)	After
Precache URLs / worker size	24,918 / 1.68 MB	278 / 12.6 KB
Install	never finished; a Fastly 503 aborted ~21k requests, then it restarted on every load	controlled after 7.1 s, 278 × 200
Each reload (×3), hard reload, reopened tab	3k–24k worker requests, never controlled	1 network request (shell), 56/56 assets from the worker, 0 console errors
Saved park	n/a	saved via the UI before the reloads; still listed afterwards and loads

The after run used the same session-Chrome profile the old worker had left behind (an ordinary returning visitor). It kept storage, and the stale cache was removed automatically.

Traces, scripts and README: https://github.com/tilework-tech/opencoaster/tree/7d261aef1cc3d0a4079195798260efe8723b35fb/pr-197