Add relay disconnect UX: friendly errors, reconnect, cached identity

[wpfleger96]

This PR improves the desktop app's behavior when the relay is unreachable — replacing raw network errors with friendly disconnected states, adding reconnect affordances, and caching the user's identity locally so the sidebar doesn't revert to their npub key.

When WARP VPN needs its daily reauth, Cloudflare Access intercepts relay HTTPS requests and returns an HTML login page with HTTP 200. Previously the Rust backend passed this through as failed to parse query response: error decoding response body for url (https://sqprod.cloudflareaccess.com/...), rendering verbatim in the sidebar — confusing and non-actionable.

• Classify relay HTTP errors with a stable relay unreachable: prefix — detect Cloudflare Access intercepts (by final-response host), HTML bodies, and connection/timeout/DNS failures. This covers query_relay_at() and submit_event() as well as the media commands (upload_blob, fetch_blob_bytes), which previously surfaced raw reqwest errors (embedding full request URLs) and raw response bodies in upload/download toasts. Raw URLs and HTML bodies are never included in error strings.
• Add ConnectionBanner — a thin warning strip above the content pane that auto-appears after 2s when degraded, with a Reconnect button; also add a "Reconnect to relay" item to the profile popover (always visible); reconnect calls relayClient.preconnect() + queryClient.invalidateQueries() without unmounting the workspace or clearing drafts
• Auto-heal via useRelayAutoHeal: when the WebSocket recovers from a degraded state, invalidate all queries so errored queries (messages don't poll) refetch automatically without any user action — rate-limited to once per 15s so a flappy connection (e.g. VPN toggling) can't repeatedly fire an unfiltered invalidation at a relay that just recovered
• Single-source the disconnected-state copy as RELAY_UNREACHABLE_SHORT / RELAY_UNREACHABLE_MESSAGE in relayError.ts, shared by the sidebar banner, home screen, and channel canvas; relayErrorDetail() falls back to the full message instead of an empty string when a classified error carries no detail
• Cache the last-fetched self-profile (display name + avatar as a ≤256 KB data URL) in localStorage keyed by relay+pubkey via selfProfileStorage; ProfileAvatar falls back to the cached data URL when the live proxied URL fails to load, and renders initials immediately (instead of after the 200ms fallback delay) when there is no image source at all
• Harden the self-profile cache: writes are skipped entirely when the serialized payload is unchanged (the 30s profile poll otherwise rewrote ~341KB and re-notified listeners every cycle); avatarDataUrl is only accepted from storage when it starts with data:image/ since it flows into an <img src> sink; removing a workspace garbage-collects all cached profiles for that relay; the avatar fetch/reuse policy is extracted into pure shouldFetchAvatar / resolveAvatarDataUrl helpers with unit tests
• Add a relay-connectivity-screenshots Playwright spec (smoke project) covering the disconnected-state surfaces, backed by new mock-bridge error knobs (channelsReadError, feedReadError, canvasReadError), a get_canvas mock (previously an unsupported-command throw), and a __BUZZ_E2E_SET_RELAY_CONNECTION_STATE__ seam for driving the ConnectionBanner deterministically

[wpfleger96]

Screenshots of the disconnected-state UX, captured by the new relay-connectivity-screenshots.spec.ts (runs in the Playwright smoke project). To re-run locally: cd desktop && pnpm run build && pnpm exec playwright test tests/e2e/relay-connectivity-screenshots.spec.ts — PNGs land in desktop/test-results/relay-connectivity-screenshots/.

Sidebar — relay unreachable

The channels query fails with a classified relay unreachable: error, so the sidebar shows the friendly copy and a Reconnect affordance instead of a raw reqwest error string.

Connection banner while reconnecting

The thin strip appears above the content pane once the WebSocket has been degraded for 2s, with its own Reconnect button.

Home feed unreachable

The home feed error card uses the shared RELAY_UNREACHABLE_MESSAGE copy with a retry affordance.

Channel canvas unreachable

The same single-sourced copy inside the channel management sheet's canvas section.

Cached identity while offline

With a seeded self-profile cache, the profile card keeps the display name and the cached avatar data URL even though get_profile fails.

No cache — npub fallback

Without the cache, the same offline state falls back to the npub-derived name and initials. This is the contrast case for the shot above.

Profile popover — reconnect to relay

The always-available "Reconnect to relay" action in the profile popover, shown while the connection is degraded.