writer: NextBackend on the mdbase-next SDK (opt-in)

[callumalpass]

Part of the mdbase-next ship (sdk workstream). Adds NextBackend, a WriterBackend on the new TS SDK @mdbase-dev/sdk. It is opt-in: Connect stays the default backend, and the default production bundle is byte-identical (the next modes are compiled out unless DEV, VITE_WRITER_DEMO=1 or VITE_WRITER_NEXT=1).

SDK dependency: the SDK is not published yet. A packed build (c573c96) is vendored at apps/writer/vendor/mdbase-dev-sdk-c573c96.tgz and used through a file: spec. This switches to a published version once one exists.

What's ported

Writer need	mdbase-next
Discovery (manuscripts, index, library, annotations, comments, people)	One live({}, {effective: true}) query without bodies. Its snapshot and diffs feed the existing CollectionStore. This replaces watch + readMany refetch. A later snapshot (reset or reconnect) reconciles.
Bodies	Loaded on demand with get(id, {body: true}), only for open records, comment threads in scope and a source's annotations (8 reads at a time). readBody is cached until a live diff touches the path.
Record sessions (autosave)	New NextRecordSession behind the workspace's session interface. It saves with update(view, {patch, body}), so the SDK sends base fields and body edits and the replica merges concurrent edits. It uses the optimistic Write.records as the next base; saving lasts until Write.confirmed.
Conflicts	A rejected conflict receipt, a confirmed receipt with status: conflicted, or a hold on the record (onHolds) becomes the workspace's existing conflict state, with remote set. The existing compare/resolve UI works. Resolve dismisses receipt conflicts or calls resolveHold.
outcome_unknown	recovery state: the receipt is looked up by mutation ID before anything new is sent.
Comments and settings	create, and update(view, {patch}).
Create retry	create at a numbered path on conflict/path_taken, replacing the /exist/ regex. This also applies when the log rejects the path after the replica accepted it: a fresh create goes to the next numbered path, carrying any text written meanwhile, and onRecordMoved moves the open manuscript or chapter embed to the new path.
Files	files.list() once, files.download() on demand. Kept current from watchChanges (with reset), which is the fallback feed.
Errors	All 15 codes map to Writer text plus a recovery (backend/next-errors.ts). unavailable/no_device_online is a waiting for one of your devices state on the gate and in a sync status line, not an error.

Small refactor first (c09c322): WriterBackend.records and the workspace now use a small structural RecordSession interface instead of Connect's MdbaseRecordLease class types. Connect's sessions still satisfy it unchanged.

Stubbed / no equivalent yet

• People directory (account → person record): there isn't one. Person names still come from discovery, but new comments are unsigned (signing: unavailable), and reviewIdentityAccess is not offered.
• Metadata-only queries (query-metadata-v1 select projections): discovery carries whole frontmatter (without bodies), not a projection.
• Contract bindings in describe(): when a replica reports contract implementations they are used. Otherwise (MemoryReplica, and replicas until describe carries contracts) Writer falls back to the starter type names (writer-manuscript, reader-source, …).
• File mtime: RecordView has none, so manuscript "last modified" is not shown.

Blocked on the control workstream

How a web app gets a grant (registering client_pk at consent) and a route does not exist yet. Both sit behind one interface, NextControlPlane (src/connect/next-control.ts):

• route() is implemented against a proposed endpoint: GET {serverUrl}/v1/next/collections/:id/route → {targets: [{url, device, noise_pk}]}. An empty list means no device is online.
• grant() is a stub that returns ?grant= from the URL. Consent does not register the browser's key yet; loadOrCreateClientKey("mdbase-writer") creates the key that consent would register.

Try it

• pnpm dev, then open http://127.0.0.1:5320/?demo=next. This is the bundled demo collection on the SDK's MemoryReplica, through NextBackend. Writes stay pending for 400 ms, then confirm.
  • For a concurrent edit, run this in the console: const o = await writer.backend.otherClient(); const v = await o.get({path: "chapters/…"}, {body: true}); await o.update(v, {body: "…"}). The open chapter updates live. If you have unsaved text instead, the rejected merge shows the conflict dialog (MemoryReplica doesn't merge; a real replica does).
• To use a real replica: ?next&collection=<id>&grant=<id>[&server=…] (development builds, or VITE_WRITER_NEXT=1). This needs the control-plane endpoint above.

Tests

src/backend/next.test.ts and src/connect/next-control.test.ts run against MemoryReplica. They cover:

• discovery without bodies (no get calls, no bodies in the live set);
• live diffs;
• lazy and cached bodies, and scoped comment bodies;
• files list and download;
• autosave edit → pending → confirmed;
• frontmatter patches sent with their base view;
• a second client's edit arriving live;
• a concurrent edit over unsaved text becoming a conflict and being resolved;
• a rejected receipt mapped to an error;
• deletion, and shared sessions;
• a full ManuscriptWorkspace on NextBackend;
• create path-collision retry, at submit time and after a late rejection at the log, and no retry for other codes;
• comments create/patch, and people;
• all 15 error codes, waiting-for-device, and a refused connection;
• route parsing and HTTP mapping for the proposed endpoint;
• the status line;
• the next-demo seed.

pnpm typecheck and pnpm test (app) pass: 34 files, 208 tests. Existing tests are unchanged. vite build passes with and without the next modes enabled.

No deploys.

[callumalpass]

D3 (mdbase-next#167), late path collision on create: 04c4e94

• A create that the replica accepted, but the log later rejected with conflict/path_taken, is retried at the next numbered path. It uses the same policy and the same 20-attempt limit as a submit-time collision. The retry is a fresh create, with a new record ID and a new mutation ID. Any text written into the optimistic record meanwhile is carried over.
• WriterBackend.onRecordMoved reports the move. The workspace drops the rolled-back record's session and local draft, points the chapter embed at the new path, and opens it there. The app reopens a moved manuscript at its new path.
• A late rejection that isn't a path collision now shows on the status line. Before, it was sent as a collection problem, which marked the index as failed.
• Tests: three new tests use MemoryReplica.reject(). They cover a manuscript with edits made meanwhile, a new chapter's embed, and a non-collision rejection. With the vendored SDK (c573c96), reject() does roll back the optimistic create, and the tests check that. Typecheck passes, and the app tests pass: 34 files, 208 tests.

I've removed this item from the "Stubbed" list.