[baxen]

Summary

When the relay advertises buzz_v1 with the ping capability, web and desktop take unread state from the relay instead of NIP-RS. That keeps every device in agreement. Relays without it keep the existing NIP-RS path unchanged.

Depends on the relay PR block/buzz#8639. Replaces #427.

Flow

1. On connect, read-api.ts loads the sidebar from GET /buzz/v1/me/sidebar and applies the relay's unread, mention, latest and followed_threads values. server-unread.ts turns those rows into the unread decisions the sidebar, badge and Inbox already use. unread.ts and read-state.ts hand those surfaces over when v1 is active and stop publishing NIP-RS.
2. Reading a channel or thread queues a mark-through intent. The queue is durable, so it survives reloads and offline periods, and it posts to POST /me/read-state.
3. The relay pings the account's other devices with the channels and threads that changed. Those devices refetch only those.
4. The Inbox asks POST /me/coverage about the specific messages it shows. The relay's follow list decides which thread replies count.

Ordering and recovery

• Every answer carries read_revision. An older answer than one already applied is dropped.
• A revision gap waits 1.5 s for late pings, then does one full load.
• The socket's once-a-minute check probes a single channel and loads fully only if the revision moved. Reconnects always do a full load.
• A message read once stays read as its scope's anchor moves on.

Transport

• Web goes through the browser host's relay broker, which adds the read-v1-* routes.
• Desktop uses one Tauri command, relay_read_v1. It signs NIP-98 in Rust, bounds channel IDs to 20 UUIDs, and caps bodies at 64 KiB and responses at 1 MiB.

Testing

• At 5618e1ee5: pnpm check passes, and so do all 12,449 vitest tests, Tauri clippy and the pre-push hooks. In the Tauri cargo test run, 568 of 571 passed. The three host_command timing tests failed while the machine was loaded and passed when rerun alone; that file isn't touched here.
• Live acceptance on Oct 27 against a local relay running the relay PR's code: three browser profiles (Chromium and WebKit), this native app and the mobile app, all on one identity. It covered cross-device read clearing in 0.6–1.0 s, a 10-minute fault soak (63/63 checkpoints), thread follow, unfollow and refollow in the sidebar, menu and Inbox, offline replay, a dropped ping with reconnect, and an ordinary-use window with zero full reloads after startup.
• Not proven: the native window's "mark read while viewing" (the tests can't focus a window), and fan-out across several relay pods.

Draft until baxen tests the whole stack after the relay PR merges.

🤖 Generated with Claude Code