feat(v3): DOM-virtualize the Songs grid (#636 item 3 stage 2) (#643)

The v3 Songs grid appended every scrolled page and never released nodes,
so card-node count grew unbounded with scroll depth (24 → 624 → 2001 for a
2000-song library). Replace it with a windowed/recycled render: only the
visible window (± overscan) is in the DOM while a #v3-songs-gridsizer
element sized to ceil(total/cols)*rowH gives the scrollbar full-library
geometry; #v3-songs-grid is absolutely positioned to the first visible row.

- state.songs is a sparse, absolutely-indexed store filled a page at a time
  by ensureWindow(): the stage-1 keyset cursor for contiguous forward scroll
  (O(page)), OFFSET page= for jumps/restore/non-keyset providers. _loadPage
  shares an in-flight promise per page and an epoch guard discards a stale
  fetch that lands after a reset.
- A–Z rail seeks directly via sort_letters cumulative counts (O(1), no
  page-through); bounded scan fallback for legacy providers without it.
- Snapshot/restore is now scrollTop-based (geometry is stable). Select mode,
  accuracy badges, ⋮ menu, plugin card actions, and tree/folder coexistence
  survive cards recycling; renderWindow re-renders when select mode toggles.
- Plugins get window.v3Songs.visibleCards() + a v3:library-window-rendered
  event instead of assuming all cards are present (highway-stutter lesson).

Verified in a browser against a seeded 2001-song library: DOM bounded to
~60 nodes while the count reads "2001 songs", rail jump lands on the target
row, selection survives recycling, scroll-restore exact. Codex-reviewed
(3 findings fixed: stale-fetch epoch guard, await-in-flight page promise,
select-mode resync on cached re-entry).

Frontend-only. Tests: tests/browser/v3-grid-virtualization.spec.ts pins the
bounded-DOM invariant + direct rail jump; tests/js/v3_az_rail.test.js and
v3_songs_scroll.test.js updated to the new wiring.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -8,6 +8,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Added
+- **The v3 Songs grid is now DOM-virtualized — card-node count stays bounded no matter how big the library is or how far you scroll.** The grid used to append every scrolled page and never let go, so a 2000-song library grew the DOM from 24 → 624 → 2001 card nodes as you scrolled (layout/memory cost scaling with depth). It now renders only the **visible window** of cards (± a small overscan); a sizer element sized to the whole library (`ceil(total/cols) × rowH`) gives the scrollbar its full geometry while the grid is absolutely positioned to the first visible row. `state.songs` is a sparse, absolutely-indexed store fetched a page at a time on demand — using the stage-1 **keyset cursor** for contiguous forward scroll (O(page)) and falling back to `OFFSET page=` for jumps/restore/non-keyset providers (collections, remote). Verified bounded (~60 nodes for a 2001-song library while the count still reads "2001 songs"). The **A–Z rail now seeks directly**: `sort_letters` gives a letter's first-row index (cumulative of prior buckets), converted to a scrollTop in O(1) — no more paging through every intervening row (a bounded forward scan covers the rare legacy provider without `sort_letters`). Select-mode selections, accuracy badges, the ⋮ card menu, plugin card actions, scroll-restore (now scrollTop-based, since geometry is stable), and the tree/folder views all survive cards leaving and re-entering the DOM. Plugins that decorate cards get a stable `window.v3Songs.visibleCards()` accessor + a `v3:library-window-rendered` event instead of assuming every card is present (the highway-stutter lesson). Stage 2 of the virtualized-grid project (got-feedback/feedBack#636 item 3), building on the stage-1 keyset data layer below. Frontend-only: `static/v3/songs.js`, `static/v3/v3.css`. Tests: `tests/browser/v3-grid-virtualization.spec.ts` (bounded-DOM invariant across a 2001-song scroll + direct rail jump), updated `tests/js/v3_az_rail.test.js` + `tests/js/v3_songs_scroll.test.js`.
 - **Keyset (cursor) pagination for the library grid — the data layer for an upcoming virtualized grid, and a latent paging bug fixed along the way.** Every library sort now carries a unique `filename` tiebreak, making the order **total** — which fixes a latent bug where rows sharing a sort key (e.g. two songs by the same artist) could be skipped or duplicated across `OFFSET` pages. `GET /api/library` gains an opaque `after` cursor + a `next_cursor` in the response: passing the cursor back fetches the next page with a **WHERE-seek** instead of `OFFSET`, so deep paging is O(page) regardless of depth. The seek is NULL-aware and exactly `OFFSET`-equivalent (verified across artist/title/recent, ascending + descending, including the legacy `dir=desc` shape and NULL sort keys); unknown/compound sorts and bad cursors fall back to `OFFSET`, and only the local provider is handed a cursor (collections/remote page by `OFFSET`). New composite `(artist NOCASE, filename)` / `(title NOCASE, filename)` / `(mtime, filename)` indexes cover the order. This is stage 1 of the virtualized-grid project (got-feedback/feedBack#636 item 3); the DOM-recycling render window builds on it next. Tests: `tests/test_library_keyset.py` (keyset==OFFSET parity, stable tiebreak, dir=desc, NULL keys, cursor fallback).
 - **Smart collections — save a set of library filters as a live, auto-updating source.** A collection is a saved `/api/library` query (e.g. "Drop-D tunings", "sloppak only", "recently added") that stays live: it's registered as a **library provider**, so it shows up in the v3 Songs source picker and inherits the whole grid UI — paging, stats, the A–Z rail, art — for free, with **no new screen**. Storage reuses the playlist subsystem (a `playlists.rules` JSON blob = a smart collection; membership is the live filter result, not stored songs, and collections are excluded from the manual-playlist list + read-only to playlist mutations). New `GET`/`POST`/`PUT`/`DELETE /api/collections`; a per-collection `SmartCollectionProvider` delegates `query_page`/`query_stats`/`query_artists` to the local DB with the stored rules applied; providers are re-registered from a boot scan so collections survive a restart. Rules mirror the raw `/api/library` query params (unknown keys dropped, never 500). Frontend: a "＋ Save as collection" action in the v3 filter drawer (shown when filters are active) names the current filter set and switches to it. The charrette's "the homelab primitive FeedBack was missing" pick (got-feedback/feedBack#636 item 2); richer rule fields (accuracy, genre, difficulty) follow as the metadata work lands. Tests: `tests/test_collections_api.py`, `tests/js/v3_collections.test.js`.
 - **The settings backup now includes your library database + custom art — your scores, favorites, playlists, and play history are no longer the one thing a backup can't save.** `GET /api/settings/export` gains an additive `core_server_files` section carrying a **consistent snapshot of `web_library.db`** (taken via the SQLite online-backup API, so it's a complete single file even while the server is running) plus any custom **playlist covers** and **avatar** (`CONFIG_DIR/playlist_covers/`, `CONFIG_DIR/avatars/`). On `POST /api/settings/import` the database is **staged** to `web_library.db.restore` rather than written over the live, open DB; it's swapped in at the next startup (`_apply_pending_db_restore`, before the connection opens), which also clears the old WAL sidecars so a stale `-wal` can't be replayed onto the restored file — the import response sets `restart_required: true` and warns accordingly. Custom art is written immediately. The bundle stays backward-compatible (older servers ignore the new section). Came out of the library design charrette (dev-ops lens's top "protect irreplaceable data" pick, got-feedback/feedBack#636). _Known gap:_ custom uploaded **song** art is still commingled with the rebuildable thumbnail cache in `art_cache/`, so it isn't bundled yet (a tracked follow-up). Tests: `tests/test_settings_export_library_db.py` (snapshot consistency, staged-not-live restore, sidecar clearing, traversal rejection, full round-trip).