feat(sync): web cursor-pull gap-filler (PR B3)

[Rasbandit]

Summary

PR B3 (web) of the ordered cursor-sync feature — migrates the React SPA onto the durable ordered change feed as a gap-filler beside the live Phoenix socket. Builds on B1 (backend /sync/changes + cursor) and B2 (plugin).

The web has no local mirror — it renders from the server on demand and keeps an in-memory React Query cache invalidated live by the socket. The socket drops events with no replay while disconnected (tab backgrounded, laptop sleep, network blip). This PR adds a cursor pull that replays exactly the missed seq>cursor deltas and feeds each note row into the existing handleNoteChanged invalidation pipeline — a change row is treated as a signal, not data.

• Backend: ?fields=meta passthrough on GET /sync/changes (signal-only, no note content → no wasted server-side decrypt). Exposes the existing list_changes_by_seq(fields: :meta) capability; attachments/watermark/pagination untouched. Lenient parse_fields (unknown → full, forward-compatible).
• device-id.ts — mint + persist a per-install UUID, sent as X-Device-Id.
• cursor.ts — opaque cursor codec byte-matched to Engram.Sync.encode_cursor/2 + per-vault localStorage.
• cursor-sync.ts — bootstrap a fresh device's cursor from the manifest change_seq; pull fields=meta deltas; single-flight per vault; reseed on a stale-cursor 400/410 (HISTORY_EXPIRED dormant until PR D); console.warn instead of silently truncating at the page cap.
• Triggers — pull on mount, window focus, and socket reconnect (socket.onOpen). Socket stays the latency accelerator; the cursor pull is the durable convergence path. Single-flight dedupes overlapping triggers.

No migration (no phase/* label). Lays the foundation for a future PWA — the apply step swaps invalidate→write-to-IndexedDB in one place.

Spec: docs/superpowers/specs/2026-06-16-sync-cursor-pull-design.md (PR B3 row). Plan: docs/superpowers/plans/2026-06-17-sync-cursor-pull-b3-web.md.

Test Plan

• Backend: mix test test/engram_web/controllers/sync_changes_test.exs (6 pass — meta-strips-content, default-keeps-content, lenient-unknown, + existing pagination/watermark/change_seq)
• Frontend: bun run test src/api/ (97 pass — device_id, codec backend-parity, bootstrap seed, pull advance/multi-page/null-tail/empty/attachment-noop, 410 reseed, single-flight, focus/reconnect/cleanup wiring)
• bun run build (tsc + vite) green
• Manual real-browser check over the laptop CDP tunnel: confirm X-Device-Id header + engram.syncCursor.<vault> seeded, an edit from a second client propagates to the web on refocus without reload, and /sync/changes rows carry content: null

Note: full frontend suite has 2 pre-existing AbortError teardown errors in viewer/pdf tests, unrelated to this PR (missing local dev deps; CI installs fresh).

🤖 Generated with Claude Code

[Rasbandit]

Heads-up for reviewers: this branch also includes a pre-existing, unrelated bug fix (commit `fda77032`), bundled in at the maintainer's request after it surfaced during live selfhost browser testing of B3.

`fix(web): treat identical 3-way merge sides as agreement` — the web editor showed a merge conflict on every edit. Root cause (verified live): a REST save broadcasts its own `note_changed` to every socket on the vault topic including the saving tab (no per-device origin filter); that echo races the save's HTTP response so `baseRef` is momentarily stale when `merge3(staleBase, local, remote)` runs — with `remote` byte-identical to the local draft. `merge3` produced identical hunks over the same base lines, which the collision gate (a range overlaps itself) wrongly flagged. Fix: short-circuit equal sides (local === remote always agrees). Pure-frontend, covered by a new merge.test.ts case; not related to the cursor-sync change.