Research 0003 — WKWebView offline behavior and service-worker constraints

Status: Pending spike. Tracked centrally as sovereign epic task 20.10 and shared with sovereign workstream 0001; run once, both workstreams consume the finding. This doc is the place that finding gets recorded from this repo's side once the spike lands.

Question

Once ADR 0005 commits this shell to loading server.url over https (required for a service worker to exist at all), how much of Sovereign's PWA offline behavior actually survives inside a Capacitor WKWebView in practice? Assumption is not evidence here — this has to be measured against a real Capacitor build, not inferred from how service workers behave in Safari or Chrome directly.

Why this matters here specifically

• Determines whether offline is a feature of the first native release or a documented follow-up — a scope decision, not a go/no-go one. This leg always produces a usable answer either way.
• iOS is known to evict WKWebsiteDataStore under storage pressure or prolonged non-use. For this whole-instance shell the visible effect of eviction is a forced re-login and a cold cache — annoying, not silent data loss, since the instance (not the shell) is the source of truth. Still needs to be measured, not assumed.
• A negative result narrows this repo's Phase 1 scope; it does not block the release. See epics/shell.md kill criteria.

What the spike needs to answer

1. Does a service worker registered by the loaded instance actually install and activate inside a Capacitor WKWebView pointed at server.url?
2. What survives a forced offline toggle mid-session — cached shell chrome, cached data, neither?
3. How does WKWebsiteDataStore eviction actually behave under realistic storage pressure and idle duration on a physical device (not a simulator)?
4. Does anything differ meaningfully between iOS and Android here, given Android's WebView has different service-worker and storage-eviction behavior?

Recommendation

Not yet decided — pending the spike. Do not build Phase 1 offline-dependent UX ahead of this finding; see epics/shell.md leg 1 for sequencing.