quak files COLLECTION fetches that collection's files alone (closes #217)

`quak files COLLECTION` used to wait for a full refresh, which on an
empty cache fetched the files of every collection first. It now reads
the collection list, then calls the new `Library.refreshCollection()`.
That brings one collection up to date in the store as a refresh would.
Once the stored cursor is past 0, it only updates a collection the store
holds. For a new one, `quak files COLLECTION` waits for a refresh. A
refresh from a cursor of 0 replaces all the store holds, so collections
then sit in the server's order. `open()` awaits the first refresh while
the cursor is 0. The store lists what did not decrypt in a fixed order.

Model: opus-5-5

README.md

@@ -2,20 +2,31 @@
 
 quak is a WTFPL-licensed TypeScript client library and CLI by
 [@sneak](https://sneak.berlin) for the [Ente](https://ente.io) end-to-end
-encrypted photo hosting service. It logs in, enumerates collections and files,
-and downloads individual images while decrypting them on the way to disk.
+encrypted photo hosting service. It logs in, lists collections and their files,
+and downloads and decrypts single files.
 
-quak also includes a resilient backup command that downloads every file in the
-account into a deduplicated local directory tree, skipping files that already
-exist on disk and continuing past individual download failures instead of
-crashing. For each file it persists the basic metadata fields quak keeps (title,
-file type, creation and modification time, latitude, longitude, content hash),
-its update time, the private and public magic metadata in full, Ente's ML data
-for it when there is any, and, for an image (for a live photo, its image), its
-original's EXIF and XMP, with its dimensions for a JPEG only; for a video it
-keeps none of these three. It runs unattended from cron (see "Running the backup
-from cron"). A helper subcommand can detect and regenerate missing thumbnails,
-encrypting and uploading them back to the server.
+`quak backup` downloads every file in the account into a local directory tree.
+It downloads and stores each file once, however many collections hold it. It
+skips files already on disk. It logs a failed download and moves on to the next
+file. Beside each file it writes a JSON file that holds:
+
+- the file's title and file type
+- its creation and modification times
+- its latitude and longitude
+- its content hash
+- its update time
+- its private and public magic metadata, in full
+- the ML data that Ente has for it, without the CLIP embedding
+- its keywords in place of that embedding (see "Backup layout")
+- for an image, the EXIF and XMP of its original, and its dimensions if it is a
+  JPEG
+- for a live photo, the same for its image
+
+The backup runs unattended from cron (see "Running the backup from cron").
+
+`quak helper list-missing-thumbnails` finds files whose thumbnail is missing.
+`quak helper fix-missing-thumbnails` makes a new thumbnail for such a file, then
+encrypts it and uploads it to the server.
 
 ## Getting Started
 
@@ -59,10 +70,11 @@ const client = await Client.login({
     password: "your-password",
 });
 
-// Open a cache-backed library. On an empty cache this awaits one server
-// refresh; on an existing cache it returns immediately and refreshes in the
-// background. Later refreshes start `refreshIntervalSeconds` (default 3) after
-// the previous one ends.
+// Open a cache-backed library. On an empty cache, or one that only
+// `lib.refreshCollection()` wrote, this awaits one server refresh. Otherwise
+// it returns immediately and refreshes in the background. A background timer
+// then refreshes again `refreshIntervalSeconds` (default 3) after its previous
+// refresh ends.
 const lib = await Library.open({ client });
 
 // Default reads answer synchronously from the local cache — no network.
@@ -371,6 +383,11 @@ quak/
                         backup-metadata: the metadata quak keeps, as JSON
         exif.ts         EXIF read from an image's bytes with exifreader
         mldata-fetch.ts fetch + decrypt per-file ML data
+        clip-keywords.ts
+                        a file's keywords from its CLIP embedding
+        clip-keyword-vectors.ts
+                        Ente's Discover categories with their text
+                        embeddings, written by bin/clip-keyword-vectors
         filename.ts     safe file names from server metadata
         errors.ts       error types shared across layers, and how an error
                         is described
@@ -387,6 +404,9 @@ quak/
         index.ts        public library exports
     bin/
         quak.ts         CLI entrypoint (commander.js)
+        clip-keyword-vectors, clip-keyword-vectors.py
+                        write src/clip-keyword-vectors.ts and a test
+                        fixture (see Source attribution)
     examples/
         download-albums.ts
                         download every album's photos and metadata
@@ -631,17 +651,17 @@ The CLI stores the snapshot at the platform-appropriate data directory via
 `0600`. The key material is stored in cleartext in the JSON; treat this file as
 you would treat the password itself. A missing file is reported as "not logged
 in"; a file that exists but is corrupt is reported as such, naming the bad
-field. When the refresh a command starts with gets HTTP 401 from the server,
-because it no longer accepts the saved session's token, the command prints one
-line, `quak: the saved session is no longer valid; run "quak login"`, with no
-stack trace. All three exit with status 3, which means the user must run
-`quak login` again. `quak logout` is the exception: with no file it says there
-is no session and exits 0, and it handles a corrupt file or a failed server call
-as described below. `quak backup` meets an expired session on the refresh that
-starts every run, before it touches any file but its lock (see "Backup layout").
-A session that stops working partway through a backup instead fails each
-remaining file into `failures.json`, so that run exits 1 and the next one stops
-at its refresh with status 3. No command but `quak login` ever prompts.
+field. The server answers HTTP 401 when it no longer accepts the saved session's
+token. A command that gets that answer prints one line,
+`quak: the saved session is no longer valid; run "quak login"`, with no stack
+trace. All three exit with status 3, which means the user must run `quak login`
+again. `quak logout` is the exception: with no file it says there is no session
+and exits 0, and it handles a corrupt file or a failed server call as described
+below. `quak backup` meets an expired session on the refresh that starts every
+run, before it touches any file but its lock (see "Backup layout"). A session
+that stops working partway through a backup instead fails each remaining file
+into `failures.json`, so that run exits 1 and the next one stops at its refresh
+with status 3. No command but `quak login` ever prompts.
 
 `quak logout` ends the session on the server, so the token in `session.json`
 stops working even in a copy of the file, and then deletes the file. If the
@@ -671,22 +691,40 @@ quak helper fix-missing-thumbnails [--file ids] [--json]  generate + upload miss
 ```
 
 Every command except `login`, `whoami` and `logout` runs on the cache-backed
-library. The read commands — `collections`, `files`, `get`, `get-thumb`,
-`backup-metadata`, `helper list-missing-thumbnails` and
-`helper fix-missing-thumbnails` — force a fresh server round-trip before they
-answer, so they report current account state rather than whatever the cache last
-held. `collections` asks the server for the collection list alone, in one
-request, and neither fetches any collection's files nor changes the cache. The
-others wait for a refresh. With an empty cache, as on a first run, that refresh
-fetches the files of every collection, one collection after another, and stores
-nothing until it has them all, which on a large account takes a while. Meanwhile
-the command logs a line on stderr for each collection whose files it has
-fetched, such as `Fetched the files of 3 of 120 collections`, with `--json` too;
-`backup` does the same while the refresh it starts with runs. stdout holds only
-the command's output. If that round-trip fails, the command prints the error on
-one line and exits 1, or 3 when the server no longer accepts the saved session
-(see "Session handling"). An album or file quak cannot decrypt does not fail it;
-the command answers from what did decrypt (see "Default reads vs. fresh reads").
+library. Each read command makes a fresh server round-trip before it answers. It
+then reports the account's current state, not what the cache last held. The read
+commands are:
+
+- `collections`
+- `files`
+- `get`
+- `get-thumb`
+- `backup-metadata`
+- `helper list-missing-thumbnails`
+- `helper fix-missing-thumbnails`
+
+`collections` asks the server for the collection list alone, in one request. It
+fetches no collection's files and leaves the cache as it is. `files` with a
+collection ID asks for the collection list the same way. It then asks for the
+files of that collection alone, and stores them in the cache. It asks for no
+files when the cache's copy of that collection is current. The other read
+commands, and `files` without a collection ID, wait for a refresh. So does
+`files` with the ID of a collection that is new since the last refresh.
+
+The first refresh fetches the files of every collection, one collection after
+another. That includes any collection that `files` with a collection ID stored
+before it. The refresh stores nothing until it has them all, and on a large
+account that takes a while. Meanwhile the command logs a line on stderr for each
+collection whose files it has fetched, such as
+`Fetched the files of 3 of 120 collections`. It logs these lines with `--json`
+too. `backup` logs them during the refresh it starts with. stdout holds only the
+command's output.
+
+If the server round-trip fails, the command prints the error on one line. It
+exits 1, or 3 when the server no longer accepts the saved session (see "Session
+handling"). An album or file that quak cannot decrypt does not fail the command.
+The command answers from what did decrypt (see "Default reads vs. fresh reads").
+
 `--cache-dir` overrides where the cache lives; without it each account gets its
 own directory under the per-user cache path.
 
@@ -741,9 +779,10 @@ its `Failed files:` list. Every other line is logged as it is without `--json`.
 exit status 1.
 
 `files <collectionID>` lists the files in that collection. `files` with no
-argument lists every file in the account, collection by collection, each file
-once: a file in several collections is listed with the first of them, which is
-its `collectionID` in `--json`.
+argument lists every file in the account once, collection by collection. The
+collections come in the order the server first sent them in a refresh. A file in
+several collections is listed with the first of them, which is its
+`collectionID` in `--json`.
 
 An ID given to `files`, `get`, `get-thumb` or
 `helper fix-missing-thumbnails --file` is written in decimal digits, no larger
@@ -783,9 +822,11 @@ as it was and goes into `failures.json` (see "Backup layout").
 
 `backup-metadata <dir>` writes `account.json` and, for each collection, a
 directory `collections/<id>-<name>/` holding `_collection.json` and a
-`<fileID>.json` for each of its files. The collection's name has unsafe
-characters replaced and is cut short between two characters so that the
-directory's name fits in 255 bytes; the ID keeps the directories apart.
+`<fileID>.json` for each of its files. Each `<fileID>.json` holds the file's ML
+data and keywords as `quak backup` writes them (see "Backup layout"). The
+collection's name has unsafe characters replaced and is cut short between two
+characters so that the directory's name fits in 255 bytes; the ID keeps the
+directories apart.
 
 `backup-metadata` fetches ML data in requests of up to 200 files. When a request
 fails, the error is logged, each of its files is written with the reason in an
@@ -818,11 +859,11 @@ the smallest does not.
                                     below)
         YYYY-MM-DD.<fileID>.json    the file's basic metadata fields quak
                                     keeps, its update time, its private and
-                                    public magic metadata, its ML data, and,
-                                    for an image (for a live photo, its
-                                    image), its original's EXIF and XMP, with
-                                    dimensions for a JPEG only; none of these
-                                    three for a video
+                                    public magic metadata, its ML data and
+                                    keywords, and, for an image (for a live
+                                    photo, its image), its original's EXIF
+                                    and XMP, with dimensions for a JPEG
+                                    only; none of these three for a video
         YYYY-MM-DD.<fileID>.livephoto.json
                                     which of a live photo's two files is which
     collections/
@@ -853,20 +894,51 @@ JSON holds its `id`, `collectionID`, `ownerID`, `metadata` and `updationTime`,
 and its `magicMetadata` and `pubMagicMetadata` when it has them. Update times
 are in microseconds, as Ente records them.
 
-A file's JSON holds Ente's ML data for it (its faces and its CLIP embedding) as
-`mlData`, the same payload `backup-metadata` writes; a file Ente has no ML data
-for has no `mlData`. The backup waits for the library's ML data fetch to finish
-before it writes the JSON files. If that fetch fails, each file whose ML data is
-not in the cache gets the reason in `mlDataError` instead and counts as failed,
-and the next run fetches it again. The same goes for a file with no ML data in
-the cache whose ML data the server sent but quak could not decrypt, gunzip or
-parse. Its `mlDataError` says why, and `failures.json` records it with an error
-starting `ML data: `. quak keeps that reason in memory only and caches nothing
-for the file, so the next run asks for its ML data again. An ML data fetch that
-fails, such as one the library's refresh starts while the backup writes the JSON
-files, leaves the reasons as the last fetch that succeeded left them. The JSON
-files are rewritten on every run, so ML data that arrived since the last run
-appears.
+A file's JSON holds Ente's ML data for it as `mlData`, such as the faces Ente
+found, without its CLIP embedding. It holds `keywords` in place of the
+embedding.
+
+The Discover section of Ente's photos app has 16 categories:
+
+- Identity (no videos)
+- Screenshots (no videos)
+- QR Codes (no videos)
+- Receipts
+- Notes
+- Memes
+- Visiting Cards (no videos)
+- Babies
+- Pets
+- Selfies
+- Wallpapers (no videos)
+- Food
+- Celebrations
+- Sunset
+- Hills
+- Greenery
+
+A file's keywords are the titles of the categories it falls in. A file's score
+for a category is the cosine similarity of its CLIP embedding and the text
+embedding of the category's query. As in Ente's app, the file falls in the
+category when that score is at least the category's minimum score. A video never
+falls in a category marked "no videos". The keywords are listed highest score
+first.
+
+A file with no ML data has neither `mlData` nor `keywords`. A file whose ML data
+holds no usable CLIP embedding has no `keywords`. `backup-metadata` writes both
+fields the same way.
+
+The backup waits for the library's ML data fetch to finish before it writes the
+JSON files. If that fetch fails, each file whose ML data is not in the cache
+gets the reason in `mlDataError` instead and counts as failed, and the next run
+fetches it again. The same goes for a file with no ML data in the cache whose ML
+data the server sent but quak could not decrypt, gunzip or parse. Its
+`mlDataError` says why, and `failures.json` records it with an error starting
+`ML data: `. quak keeps that reason in memory only and caches nothing for the
+file, so the next run asks for its ML data again. An ML data fetch that fails,
+such as one the library's refresh starts while the backup writes the JSON files,
+leaves the reasons as the last fetch that succeeded left them. The JSON files
+are rewritten on every run, so ML data that arrived since the last run appears.
 
 A file's JSON holds, as `imageMetadata`, what `backup-metadata --exif` records
 from its original, or from a live photo's image: `format`, `width` and `height`
@@ -1050,15 +1122,17 @@ against a fake of the Ente HTTP API, and `yarn test` verifies them.
 ### Opening a library
 
 `Library.open(options)` loads the on-disk cache, starts the background refresh
-loop, and resolves to a `Library`. On an empty cache it awaits the first
-refresh, so it opens onto the account's data whenever the server is reachable;
-if that refresh fails, it opens with no data and records the error in
-`lib.status()`. On an existing cache it returns immediately and refreshes in the
-background, so an unreachable server does not block opening. With
-`refreshOnOpen: false` it starts no refresh and returns at once onto what the
-cache holds, empty or not; a refresh then starts only when `fresh()` or
-`lib.backup()` asks, or the background timer fires. The CLI's read commands open
-it so (see "CLI surface").
+loop, and resolves to a `Library`. On an empty cache, or one that only
+`lib.refreshCollection()` wrote, it awaits one refresh. It then opens onto the
+account's data whenever the server is reachable. If that refresh fails, it opens
+onto what the cache holds, and records the error in `lib.status()`. On any other
+cache it returns immediately and refreshes in the background. An unreachable
+server then does not block opening.
+
+With `refreshOnOpen: false`, `open()` starts no refresh. It returns at once onto
+what the cache holds, even when the cache is empty. A refresh then starts only
+when `fresh()` or `lib.backup()` asks, or the background timer fires. The CLI's
+read commands open the library this way (see "CLI surface").
 
 `LibraryOptions`:
 
@@ -1085,22 +1159,35 @@ when the library opens; `lib.downloadDirectory` holds the result.
 `onProgress` receives a `RefreshEvent` with status `"started"` as a refresh, an
 ML data fetch or a precache sweep starts, and `"done"` or `"failed"` as it ends.
 In between, a refresh sends `"progress"` after each collection whose files it
-fetched, with `fetched`, how many collections it has fetched the files of so
-far, and `total`, how many it will. A refresh fetches the files of each
-collection that is new or changed since the cache's copy, so on an empty cache
-it fetches those of every collection.
+fetched. That event's `fetched` is the number of collections whose files the
+refresh has fetched so far. Its `total` is the number of collections whose files
+it will fetch in all. A refresh fetches the files of each collection that is new
+or changed since the cache's copy. On the first refresh, that is every
+collection.
 
 Concurrency is set through `pools`: construct
 `new RequestPools({ metadataConcurrency, contentConcurrency, thumbnailConcurrency })`
 and pass it. The three pools default to 10 / 5 / 25 (see Request pools below).
 
-`lib.status()` returns a `LibraryStatus` (collection/file counts, last
-refresh/ML times and errors, originals usage and effective limit, precache
-progress, and `closed`). `lib.close()` stops the background timer; it is
-idempotent, and an in-flight refresh is left to finish. The promise it returns
-resolves once that refresh (including its cache write), the ML data fetch and
-the precache fetches already running have all finished, so the cache directory
-can then be removed.
+`lib.status()` returns a `LibraryStatus`, which holds:
+
+- the collection and file counts
+- the time of the last refresh that succeeded, and the error of the latest
+  refresh if it failed
+- the same two for the ML data fetch
+- the originals cache's usage and its effective limit
+- the precache progress
+- `closed`
+
+`lib.close()` stops the background timer, and is idempotent. A refresh or
+`lib.refreshCollection()` in flight is left to finish. `close()` returns a
+promise that resolves once all of these have finished:
+
+- that refresh or `lib.refreshCollection()`, with its cache write
+- the ML data fetch
+- the precache fetches already running
+
+The cache directory can then be removed.
 
 `lib.status().mlDataErrors` lists, as `{ fileID, error }`, each file whose ML
 data the last ML data fetch that succeeded got but could not decrypt, gunzip or
@@ -1114,39 +1201,72 @@ cache.
 
 Default reads — `lib.albums`, `lib.photos`, `lib.timeline` — answer
 synchronously from the copy held in RAM and never touch the network. A refresh
-changes that copy only once all its server requests have succeeded, and the
-background timer starts the next refresh `refreshIntervalSeconds` after the
-previous one ends. So a default read is immediate, but only as current as the
-last refresh whose requests all succeeded; right after opening an existing
-cache, it is the copy on disk.
+changes that copy only once all its server requests have succeeded.
+`lib.refreshCollection()` changes it the same way for one collection. So a
+default read is immediate, but only as current as the last refresh or
+`lib.refreshCollection()` whose requests all succeeded. Right after `open()`
+returns without awaiting a refresh, that copy is the one loaded from disk.
 
-`await lib.fresh()` waits for a refresh to complete and persist, and returns the
-same `{ albums, photos, timeline }` namespaces, which then reflect a completed
-server round-trip. When a refresh is already running, background or not,
-`fresh()` waits for that one, so its answer can come from requests made before
-the call; only when none is running does it start one. A refresh that fails
+The background timer takes a turn `refreshIntervalSeconds` after its last turn
+ends. Each turn starts a refresh, unless a refresh or `lib.refreshCollection()`
+is already running.
+
+`await lib.fresh()` waits for a refresh to complete and persist. It then returns
+the same `{ albums, photos, timeline }` namespaces, which reflect a completed
+server round-trip. If a `lib.refreshCollection()` is running, `fresh()` first
+waits for it to end. When a refresh is already running, background or not,
+`fresh()` waits for that one. Its answer can then come from requests made before
+the call. Only when none is running does it start one. A refresh that fails
 rejects the caller (default reads stay silent and keep serving the last good
-copy). The CLI's read commands use fresh reads (issue
-https://git.eeqj.de/sneak/quak/issues/75).
+copy). Every CLI read command calls `fresh()`, except `quak collections`.
+`quak files` with a collection ID calls it only when `lib.refreshCollection()`
+resolves to `false`.
 
-`await lib.freshCollections()` is the fresh read of the collection list alone:
-one collections request, answered with every live collection that decrypts, as
-the `Collection` objects `listCollections()` returns, in the server's order. It
-fetches no collection's files and stores nothing, not even the list: a refresh
-fetches a collection's files only when the collection is new or has changed
-since the stored copy, so a collection stored without its files would never get
-them. `quak collections` uses it (issue
-https://git.eeqj.de/sneak/quak/issues/210).
+`await lib.freshCollections()` is the fresh read of the collection list alone.
+It sends one collections request. It answers with every live collection that
+decrypts, in the server's order, as the `Collection` objects that
+`listCollections()` returns. It fetches no collection's files and stores
+nothing, not even the list. A refresh from a cursor past 0 fetches a
+collection's files only when the collection is new or has changed since the
+stored copy. So does `lib.refreshCollection()`. Neither would fetch the files of
+a collection stored without them. `quak collections` uses `freshCollections()`.
 
-An album or file in the server's answer that quak cannot decrypt, or whose
-metadata is not JSON, does not fail the refresh. The refresh leaves it out, an
-album with all its files, and drops any older copy the cache held, so reads
-answer from what did decrypt. The library keeps a list of what it left out, with
-the reason, in `metadata.json`: the next refresh asks the server only for what
-changed, so it would not see the record again. A record stays on that list until
-a later refresh gets a copy that decrypts or learns that the server deleted it.
-`lib.backup()` reports each such album, and each such file in its scope that has
-no copy that decrypted (see "Backup layout").
+`await lib.refreshCollection(collection)` brings one collection up to date in
+the cache, given its `Collection` from `freshCollections()`. It does for that
+collection what a refresh does for a changed one, and fetches no other
+collection's files. It fetches the collection's files changed since the cache's
+copy, or all of them when the cache has no copy. It then stores and saves them.
+It sends no request when the cache's copy has the same `updationTime` or a later
+one. It resolves to `true` when the cache then holds the collection.
+
+`lib.refreshCollection()` adds a collection the cache does not hold only while
+the refresh cursor is 0. The cursor moves past 0 once a refresh receives the
+account's collections. After that, for a collection the cache does not hold, it
+sends no request and resolves to `false`. Such a collection is new since that
+refresh, or the refresh dropped it. Only a refresh may then add it, after the
+collections the cache holds.
+
+A refresh from a cursor of 0 replaces everything the cache holds. It fetches the
+files of every collection, even those `lib.refreshCollection()` stored. It
+stores the collections in the server's order. The cache then holds what it would
+had `lib.refreshCollection()` never run.
+
+`lib.refreshCollection()` leaves the refresh cursor where it was, so the next
+refresh still fetches the files of every other collection that changed. It never
+runs at the same time as a refresh. It first waits for a running refresh or
+`lib.refreshCollection()` to end. While it runs, a `fresh()` or `lib.backup()`
+waits for it to end. `quak files` with a collection ID uses it.
+
+A refresh does not fail on an album or file that quak cannot decrypt, or whose
+metadata is not JSON. The refresh leaves out such a file, and such an album
+together with all its files. It also drops any older copy that the cache held,
+so reads answer from what did decrypt. The library keeps a list of what it left
+out, with the reason, in `metadata.json`. The next refresh asks the server only
+for what changed, so it would not see them again. An entry stays on that list
+until a later refresh or `lib.refreshCollection()` stores a copy that decrypts,
+or learns that the server deleted it. `lib.backup()` reports each such album,
+and each such file in its scope that has no copy that decrypted (see "Backup
+layout").
 
 ### Read surface
 
@@ -1240,10 +1360,17 @@ The GUI-facing records hold no key material and no binary, so they survive
 - `LibrarySnapshot`: `{ albums, photos, takenAt }`.
 
 `lib.snapshot()` returns a `LibrarySnapshot` (albums newest-updated first,
-photos newest first). `lib.subscribe({ onChange })` delivers a `LibraryChange`
-(`albumsChanged`, `photosChanged`, `fileIDsRemoved`, `albumIDsRemoved`,
-`refreshedAt`) whenever a refresh alters the projection, and returns
-`{ unsubscribe }`; a refresh that changes nothing delivers nothing.
+photos newest first). `lib.subscribe({ onChange })` returns `{ unsubscribe }`.
+It calls `onChange` with a `LibraryChange` whenever a refresh or
+`lib.refreshCollection()` changes these records. A refresh or
+`lib.refreshCollection()` that changes nothing does not call it. A
+`LibraryChange` holds:
+
+- `albumsChanged`
+- `photosChanged`
+- `fileIDsRemoved`
+- `albumIDsRemoved`
+- `refreshedAt`
 
 ### Thumbnails, ML search, and backup
 
@@ -1259,29 +1386,44 @@ photos newest first). `lib.subscribe({ onChange })` delivers a `LibraryChange`
   `SimilarResult[]` (`{ fileID, score }`, cosine similarity, most similar first,
   default limit 20). quak bundles no text encoder, so `searchByEmbedding` takes
   a query vector the caller produced elsewhere.
-- `await lib.backup(opts?)` → `BackupResult`. It takes the lock in the download
-  directory, and fails at once with an error whose `code` is `ELOCKED` while
-  another backup of it runs. A caller can instead take the lock itself, before
-  it opens its library, as `quak backup` does: `await lockBackupDirectory(dir)`
-  takes it, failing the same way, and returns a `BackupLock`, whose `release()`
-  removes the lock unless it was lost and whose `lost` holds why it was lost.
-  The caller passes it to the backup as `lock`, and releases it once the backup
-  ends. It waits for a refresh as `fresh()` does, puts every in-scope original
-  not already at its save path there as `photo.download()` does (and, with
-  `includeThumbnails`, fetches thumbnails) through the content cache, waits for
-  an ML data fetch, and rebuilds the on-disk backup tree with a durable failure
-  ledger. Each file's JSON holds its ML data and, for an image (for a live
-  photo, its image), its original's EXIF and XMP, with its dimensions for a JPEG
-  only; a video's JSON holds none of these three. A fetched original is written
-  straight to its save path and not into the cache, which then counts it as
-  present; one the cache already held is copied from there. `BackupOptions`:
-  `downloadDirectory` (falls back to the library's), `includeOriginals` (default
-  `true`), `includeThumbnails` (default `false`), `onlyAlbumNames`, `verify`
-  (default `false`), `onProgress`, and `lock`. See Backup layout above for the
-  tree it writes, and for the albums `BackupResult` lists in
-  `undecryptableAlbums` and `failedAlbums`. Whether the backup took the lock or
-  its caller did, once the lock is lost the backup rejects with
-  `lost the lock on <dir>: <reason>` before it starts another file.
+- `await lib.backup(opts?)` → `BackupResult`. It takes the lock, `backup.lock`
+  in the download directory. It fails at once with an error whose `code` is
+  `ELOCKED` while another backup of that directory runs. A caller can instead
+  take the lock itself before it opens its library, as `quak backup` does.
+  `await lockBackupDirectory(dir)` takes it, failing the same way, and returns a
+  `BackupLock`. Its `release()` removes the lock unless it was lost, and its
+  `lost` holds why it was lost. The caller passes it to the backup as `lock`,
+  and releases it once the backup ends.
+
+    The backup waits for a refresh as `fresh()` does. It then fetches each
+    in-scope original that is not yet at its save path, as `photo.download()`
+    does, through the content cache. With `includeThumbnails` it also fetches
+    thumbnails. It waits for an ML data fetch, then rebuilds the backup tree. It
+    records each file that failed in `failures.json`.
+
+    Each file's JSON holds its ML data and keywords. An image's JSON also holds
+    the EXIF and XMP of its original, and its dimensions if it is a JPEG. A live
+    photo's JSON holds these for its image. A video's JSON holds no EXIF, XMP or
+    dimensions.
+
+    A fetched original is written straight to its save path, not into the cache.
+    The cache then counts it as present. An original that the cache already held
+    is copied from there.
+
+    `BackupOptions` holds:
+    - `downloadDirectory`, which falls back to the library's
+    - `includeOriginals`, `true` by default
+    - `includeThumbnails`, `false` by default
+    - `onlyAlbumNames`
+    - `verify`, `false` by default
+    - `onProgress`
+    - `lock`
+
+    "Backup layout" above describes the tree that the backup writes. It also
+    describes the albums that `BackupResult` lists in `undecryptableAlbums` and
+    `failedAlbums`. Once the lock is lost, the backup rejects with
+    `lost the lock on <dir>: <reason>` before it starts another file. This holds
+    whether the backup or its caller took the lock.
 
 ### Request pools
 
@@ -1364,6 +1506,14 @@ is rewritten in TypeScript in this repository. Protocol fidelity is verified
 against the upstream implementations in `web/packages/base/`,
 `mobile/apps/photos/lib/`, and `cli/`.
 
+The keywords come from Ente too. `src/clip-keyword-vectors.ts` holds the
+Discover categories of Ente's photos app, from
+`mobile/apps/photos/assets/discover.json`, each with its query's text embedding
+from Ente's CLIP text model. `bin/clip-keyword-vectors` writes that file and the
+CLIP embedding of a public-domain photo that the tests use. It runs Ente's
+models in Docker, with every download pinned by sha256. It is run by hand only,
+followed by `make fmt`.
+
 ## For LLMs
 
 If you are an LLM agent working on this repository, read and follow these

TODO.md

@@ -26,6 +26,33 @@ declares one.
 
 # Completed Steps
 
+- 2026-10-25: `quak files COLLECTION` fetches the files of that collection alone
+  (https://git.eeqj.de/sneak/quak/issues/217). It used to wait for a refresh,
+  which on an empty cache fetched the files of every collection first. It now
+  reads the collection list through `lib.freshCollections()`, then brings that
+  one collection up to date through the new `lib.refreshCollection()`. That
+  stores and saves its files as a refresh would, and leaves the refresh cursor
+  where it was. When the stored copy is current, it asks for none of the files.
+  It never runs at the same time as a refresh. Once a refresh has received the
+  account's collections, it stores no collection the cache does not hold, and
+  `quak files COLLECTION` waits for a refresh instead. A refresh from a cursor
+  of 0 replaces all the cache holds, so collections keep the server's order. A
+  library opened on a cache that only `lib.refreshCollection()` wrote awaits its
+  first refresh. `quak files` with no argument still waits for a refresh. The
+  store lists what did not decrypt by collection ID and file ID.
+
+- 2026-10-22: The JSON `quak backup` and `quak backup-metadata` write for a file
+  holds `keywords` in place of the CLIP embedding in its ML data
+  (https://git.eeqj.de/sneak/quak/issues/218). The keywords are the titles of
+  the Discover categories of Ente's photos app that the file falls in, by the
+  app's rule: the cosine similarity of the file's CLIP embedding and the text
+  embedding of a category's query is at least the category's minimum score, and
+  videos stay out of the categories that leave them out. `src/clip-keywords.ts`
+  applies the rule. `src/clip-keyword-vectors.ts` holds the 16 categories with
+  their text embeddings. `bin/clip-keyword-vectors` writes it with Ente's CLIP
+  text model, and writes the CLIP embedding of a public-domain photo of a sunset
+  that the tests use.
+
 - 2026-10-18: `-v`/`--verbose` raises the CLI's log level once for each repeat,
   and `-q`/`--quiet` prints only fatal errors
   (https://git.eeqj.de/sneak/quak/issues/212). The levels are fatal, error,

bin/clip-keyword-vectors

@@ -0,0 +1,31 @@
+#!/bin/sh
+# Writes src/clip-keyword-vectors.ts and test/clip-keywords-sunset.json by
+# running bin/clip-keyword-vectors.py in a Python image pinned by digest, with
+# its packages pinned by hash. Run it by hand from a checkout, then `make fmt`;
+# no gate runs it. Besides the Python image, it downloads about 265 MB, most of
+# it Ente's two CLIP models.
+set -eu
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd -P)"
+
+# python:3.13-slim (Python 3.13.17 on Debian 13.7), 2026-10-21. The wheels
+# below are the ones for this image on linux/amd64.
+IMAGE=python@sha256:57b29a7c050c4e679e4bea572e14b3965cf4b5839de2da49d55e49ffc4024f39
+
+main() {
+    cd "$ROOT"
+    docker run --rm -i --platform linux/amd64 \
+        --user "$(id -u):$(id -g)" -e HOME=/tmp \
+        -v "$ROOT:/repo" -w /repo "$IMAGE" sh -c '
+            python -m venv /tmp/venv &&
+            /tmp/venv/bin/pip install --quiet --disable-pip-version-check \
+                --no-cache-dir --no-deps --require-hashes -r /dev/stdin &&
+            /tmp/venv/bin/python bin/clip-keyword-vectors.py' <<'EOF'
+numpy==2.5.3 --hash=sha256:a5fa86b80fd24bcd1aff83ad23be44ea323de3f787be8f8b15d4a65621e25321
+onnxruntime==1.30.0 --hash=sha256:86f940afc801ea9681a4da8af84fbe95e1d9ea7d80903952cc1bfad54faad38f
+pillow==12.3.0 --hash=sha256:0847a763afefb695bc912d7c131e7e0632d4edc1d8698f58ddabec8e46b8b6d3
+tokenizers==0.23.2 --hash=sha256:41c2f84d172449b4dadb9cdc508e3e364076613c35b16e76ecfe47a60d1e3305
+EOF
+}
+
+main "$@"

bin/clip-keyword-vectors.py

@@ -0,0 +1,158 @@
+"""Write src/clip-keyword-vectors.ts and test/clip-keywords-sunset.json.
+
+bin/clip-keyword-vectors runs this in a pinned Python image. It downloads
+Ente's Discover list, Ente's CLIP text and image models, the CLIP tokenizer
+and a public-domain photo of a sunset, checks each against its sha256, and
+writes:
+
+- src/clip-keyword-vectors.ts: each Discover category of Ente's photos app,
+  with the text embedding Ente's CLIP text model gives its query;
+- test/clip-keywords-sunset.json: the CLIP embedding Ente's image model gives
+  the photo, prepared as Ente's apps prepare a photo.
+
+Both embeddings are scaled to length 1, as Ente's apps scale them.
+"""
+
+import hashlib
+import io
+import json
+import urllib.request
+
+import numpy as np
+import onnxruntime
+from PIL import Image, ImageOps
+from tokenizers import Tokenizer
+
+ENTE_COMMIT = "f5ad70a860ed25d0c8057b0efb4bd6eaefe4b0c5"
+DISCOVER_PATH = "mobile/apps/photos/assets/discover.json"
+DISCOVER_URL = (
+    f"https://raw.githubusercontent.com/ente-io/ente/{ENTE_COMMIT}/{DISCOVER_PATH}"
+)
+DISCOVER_SHA256 = "cf70ac7116b592f7921ca4cf67cdbb421a8e6b287fbe44b2be678b8d198e5c24"
+
+TEXT_MODEL = "mobileclip_s2_text_opset18_quant.onnx"
+TEXT_MODEL_SHA256 = "d92f33dfcff83077fc2e0d3414250710efbb51795dfd89767bdbefb5fdc47322"
+IMAGE_MODEL = "mobileclip_s2_image_gelu_opset20.onnx"
+IMAGE_MODEL_SHA256 = "205a430af825e501c5138e5bb9abea942482a7a4fd4a680e98e47cf0830dce7e"
+
+# The CLIP tokenizer Hugging Face publishes for OpenAI's CLIP. Its vocabulary
+# and merges are those of bpe_simple_vocab_16e6.txt, the vocabulary Ente's
+# apps tokenize with.
+TOKENIZER_URL = (
+    "https://huggingface.co/openai/clip-vit-base-patch32/resolve/"
+    "3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268/tokenizer.json"
+)
+TOKENIZER_SHA256 = "b556ac8c99757ffb677208af34bc8c6721572114111a6e0aaf5fa69ff0b8d842"
+
+# "Sunset off the Massachusetts coast", a NOAA photo in the public domain:
+# https://commons.wikimedia.org/wiki/File:Sunset_off_the_Massachusetts_coast_-_NOAA.jpg
+PHOTO_URL = (
+    "https://upload.wikimedia.org/wikipedia/commons/d/db/"
+    "Sunset_off_the_Massachusetts_coast_-_NOAA.jpg"
+)
+PHOTO_SHA256 = "be62f3289c9bf7b75af68cd6978021d54e8a531d5d4cc493f6e9fbf5e1ab74d3"
+
+# Ente's text model takes 77 token ids, padded with 0; its image model takes
+# a 256x256 RGB image.
+TOKEN_COUNT = 77
+IMAGE_SIZE = 256
+
+
+def download(url, sha256):
+    # Wikimedia refuses requests without a User-Agent.
+    request = urllib.request.Request(url, headers={"User-Agent": "quak"})
+    with urllib.request.urlopen(request) as response:
+        data = response.read()
+    digest = hashlib.sha256(data).hexdigest()
+    if digest != sha256:
+        raise SystemExit(f"{url} has sha256 {digest}, not {sha256}")
+    return data
+
+
+def ente_model(name, sha256):
+    return onnxruntime.InferenceSession(
+        download(f"https://models.ente.com/{name}", sha256),
+        providers=["CPUExecutionProvider"],
+    )
+
+
+def run(session, tensor):
+    (output,) = session.run(None, {session.get_inputs()[0].name: tensor})
+    embedding = output[0]
+    return embedding / np.linalg.norm(embedding)
+
+
+def text_embedding(session, tokenizer, query):
+    ids = tokenizer.encode(query).ids
+    ids += [0] * (TOKEN_COUNT - len(ids))
+    return run(session, np.array([ids], dtype=np.int32))
+
+
+def image_embedding(session, photo):
+    image = ImageOps.exif_transpose(Image.open(io.BytesIO(photo))).convert("RGB")
+    # Scale to cover IMAGE_SIZE x IMAGE_SIZE and crop the centre.
+    square = ImageOps.fit(image, (IMAGE_SIZE, IMAGE_SIZE), Image.Resampling.BILINEAR)
+    pixels = np.asarray(square, dtype=np.float32) / 255
+    return run(session, pixels.transpose(2, 0, 1)[np.newaxis])
+
+
+def numbers(embedding):
+    return ", ".join(str(value) for value in embedding)
+
+
+def write_categories(prompts, embeddings):
+    lines = [
+        "// Written by bin/clip-keyword-vectors; do not edit by hand.",
+        "//",
+        "// The Discover categories of Ente's photos app, in their order, from",
+        f"// {DISCOVER_PATH} in https://github.com/ente-io/ente",
+        f"// at commit {ENTE_COMMIT}. Each has the text",
+        f"// embedding Ente's CLIP text model ({TEXT_MODEL})",
+        "// gives its query, scaled to length 1. `showVideo` is false for a category",
+        "// that leaves out videos.",
+        "",
+        "export const DISCOVER_CATEGORIES = [",
+    ]
+    for prompt, embedding in zip(prompts, embeddings):
+        lines += [
+            "    {",
+            f"        title: {json.dumps(prompt['title'])},",
+            f"        query: {json.dumps(prompt['query'])},",
+            f"        minScore: {json.dumps(prompt['minScore'])},",
+            f"        showVideo: {json.dumps(prompt.get('showVideo', True))},",
+            f"        textEmbedding: [{numbers(embedding)}],",
+            "    },",
+        ]
+    lines.append("];")
+    with open("src/clip-keyword-vectors.ts", "w", encoding="utf-8") as file:
+        file.write("\n".join(lines) + "\n")
+
+
+def write_fixture(embedding):
+    with open("test/clip-keywords-sunset.json", "w", encoding="utf-8") as file:
+        file.write(
+            "{\n"
+            f'    "photo": {json.dumps(PHOTO_URL)},\n'
+            f'    "sha256": {json.dumps(PHOTO_SHA256)},\n'
+            f'    "embedding": [{numbers(embedding)}]\n'
+            "}\n"
+        )
+
+
+def main():
+    prompts = json.loads(download(DISCOVER_URL, DISCOVER_SHA256))["prompts"]
+    tokenizer = Tokenizer.from_str(
+        download(TOKENIZER_URL, TOKENIZER_SHA256).decode("utf-8")
+    )
+    text_model = ente_model(TEXT_MODEL, TEXT_MODEL_SHA256)
+    embeddings = [
+        text_embedding(text_model, tokenizer, prompt["query"]) for prompt in prompts
+    ]
+    write_categories(prompts, embeddings)
+
+    image_model = ente_model(IMAGE_MODEL, IMAGE_MODEL_SHA256)
+    write_fixture(image_embedding(image_model, download(PHOTO_URL, PHOTO_SHA256)))
+
+
+if __name__ == "__main__":
+    main()

src/backup.ts

@@ -14,9 +14,9 @@
 //         YYYY/YYYY-MM/YYYY-MM-DD/
 //             YYYY-MM-DD.<fileID>.<ext>    the decrypted bytes (the save path)
 //             YYYY-MM-DD.<fileID>.json     per-file metadata sidecar, with
-//                                          the file's ML data and its
-//                                          original's EXIF, XMP and
-//                                          dimensions
+//                                          the file's ML data, its
+//                                          keywords, and its original's
+//                                          EXIF, XMP and dimensions
 //         collections/<name>/<title>       symlink to the original
 //         collections/<name>.json          per-collection metadata
 //         account.json                     the account's email and user ID
@@ -77,6 +77,7 @@ import { readFile } from "node:fs/promises";
 import { dirname, extname, join, relative, resolve } from "node:path";
 import lockfile from "proper-lockfile";
 
+import { clipKeywords } from "./clip-keywords.js";
 import { removeLeftoverTempFiles } from "./download/index.js";
 import { describeError } from "./errors.js";
 import {
@@ -497,8 +498,9 @@ const imageMetadataFor = async (
     }
 };
 
-// The file's JSON: its basic fields, its magic metadata, its ML data or the
-// reason the ML data is missing, and its image metadata.
+// The file's JSON: its basic fields, its magic metadata, its ML data without
+// the CLIP embedding and the keywords that embedding gives, or the reason the
+// ML data is missing, and its image metadata.
 const writeSidecar = (
     path: string,
     file: EnteFile,
@@ -514,7 +516,13 @@ const writeSidecar = (
     };
     if (file.magicMetadata) meta.magicMetadata = file.magicMetadata;
     if (file.pubMagicMetadata) meta.pubMagicMetadata = file.pubMagicMetadata;
-    if (ml.mlData) meta.mlData = ml.mlData;
+    if (ml.mlData) {
+        const { clip: _clip, ...mlDataWithoutClip } = ml.mlData;
+        meta.mlData = mlDataWithoutClip;
+        const isVideo = file.metadata.fileType === "video";
+        const keywords = clipKeywords(ml.mlData, isVideo);
+        if (keywords) meta.keywords = keywords;
+    }
     if (ml.mlDataError) meta.mlDataError = ml.mlDataError;
     if (image.imageMetadata) meta.imageMetadata = image.imageMetadata;
     if (image.imageMetadataError) {

src/cli-commands.ts

@@ -128,12 +128,14 @@ const readLibraryClient = (client: Client): LibraryClient => ({
     contentSource: () => client.contentSource(),
 });
 
-// Open a library for a single point command: the aggressive background precache
-// (issue #48) is off — a one-shot `collections` or `get` must not start
-// downloading the whole account — and no refresh starts as it opens or on a
-// timer: each command asks for the one it needs, and `collections` needs none
-// (issue #210). While a refresh runs, each collection whose files it has
-// fetched is logged, so a large account shows progress rather than nothing.
+// Open a library for a single point command. The aggressive background
+// precache (issue #48) is off, because a one-shot `collections` or `get` must
+// not start downloading the whole account. No refresh starts as it opens or on
+// a timer. Each command asks for the refresh it needs. `collections` needs none
+// (issue #210). `files COLLECTION` needs one only when `refreshCollection()`
+// resolves to `false` (issue #217). While a refresh runs,
+// each collection whose files it has fetched is logged, so a large account
+// shows progress rather than nothing.
 const openReadLibrary = (ctx: CliContext, client: Client): Promise<Library> =>
     Library.open({
         client: readLibraryClient(client),

src/cli-read.ts

@@ -2,33 +2,42 @@
 //
 // `collections`, `files`, `get`, and `get-thumb` must answer for
 // the account's state at the moment the command runs, not for whatever the
-// local cache last happened to hold (owner amendment, issue #36). `files`,
-// `get` and `get-thumb` therefore force a server round-trip through
-// `Library.fresh()` and only then read — so a collection, file, or metadata
-// change made elsewhere is visible. `collections` asks the library for the
-// collection list alone through `Library.freshCollections()` (issue #210): a
-// refresh would also fetch the files of every collection, which on a large
-// account takes long and which `collections` does not print.
+// local cache last happened to hold (owner amendment, issue #36). `files`
+// without a collection ID, `get` and `get-thumb` therefore force a server
+// round-trip through `Library.fresh()`, and only then read. A collection, file,
+// or metadata change made elsewhere is then visible. `collections` asks the
+// library for the collection list alone through `Library.freshCollections()`
+// (issue #210). A refresh would also fetch the files of every collection, which
+// `collections` does not print, and on a large account that takes long.
+// `files COLLECTION` reads the collection list the same way, for the same
+// reason. It then brings that one collection up to date through
+// `Library.refreshCollection()` (issue #217), and fetches no other collection's
+// files. Once a refresh has received the account's collections,
+// `refreshCollection()` stores no collection the library does not hold.
+// `files COLLECTION` then forces a refresh instead.
 //
-// `collections` lists in the server's order and `files` in the library's own
-// enumeration order (`listFiles()`), the orders the pre-library CLI printed,
-// rather than the `albums`/`photos` projection's newest-first order, which
-// re-sorts the rows. The field values still come from each record's raw
-// metadata via `cli-output.ts`. `files` without a collection ID lists every
-// file in the account once (issue #211): the library keeps a record of a file
-// for each collection that holds it, and only the first collection's record is
-// listed.
+// `collections` lists in the server's order. `files` lists in the library's
+// own enumeration order: `listCollections()`, then `listFiles()` for each.
+// Those are the orders the pre-library CLI printed. The `albums`/`photos`
+// projection would re-sort the rows newest first. After a refresh, the
+// library holds collections in the order the server first sent them. The
+// field values still come from each record's raw metadata via
+// `cli-output.ts`. `files` without a collection ID lists every file in the
+// account once (issue #211).
+// The library keeps a record of a file for each collection that holds it, and
+// only the first collection's record is listed.
 
 import type { Collection, EnteFile } from "./model/types.js";
 import type { Photo, PhotosAPI } from "./library/index.js";
 
 // The slice of `Library` these helpers read. `Library` satisfies it
-// structurally; a test can drive them with a stand-in that records the
-// `fresh()` and `freshCollections()` calls and serves records in a known
-// enumeration order.
+// structurally. A test can drive them with a stand-in that records the
+// `fresh()`, `freshCollections()` and `refreshCollection()` calls. The
+// stand-in serves records in a known enumeration order.
 export interface FreshReadLibrary {
     fresh(): Promise<unknown>;
     freshCollections(): Promise<Collection[]>;
+    refreshCollection(collection: Collection): Promise<boolean>;
     getCollection(id: number): Collection | undefined;
     listCollections(): Collection[];
     listFiles(collectionID: number): EnteFile[];
@@ -42,21 +51,29 @@ export const freshCollections = (
     lib: FreshReadLibrary,
 ): Promise<Collection[]> => lib.freshCollections();
 
-// The files of one collection, current as of a forced refresh, in enumeration
-// order. `undefined` (not an empty list) when the collection does not exist, so
-// the caller can tell "no such collection" from "an empty collection". Without
-// a collection ID, every file in the account, each once, as held by the first
-// collection that holds it.
+// The files of one collection, in enumeration order. `undefined` (not an empty
+// list) when the collection does not exist, so the caller can tell "no such
+// collection" from "an empty collection". They are current as of one
+// collections request and the requests that bring that collection alone up to
+// date. When `refreshCollection()` resolves to `false`, they are current as of
+// a forced refresh instead. Without a collection ID, every file in the account,
+// current as of a forced refresh. Each file is listed once, as held by the
+// first collection that holds it.
 export const freshFiles = async (
     lib: FreshReadLibrary,
     collectionID?: number,
 ): Promise<EnteFile[] | undefined> => {
-    await lib.fresh();
     if (collectionID !== undefined) {
-        return lib.getCollection(collectionID)
-            ? lib.listFiles(collectionID)
-            : undefined;
+        const collections = await lib.freshCollections();
+        const collection = collections.find((c) => c.id === collectionID);
+        if (!collection) return undefined;
+        if (!(await lib.refreshCollection(collection))) {
+            await lib.fresh();
+            if (!lib.getCollection(collectionID)) return undefined;
         }
+        return lib.listFiles(collectionID);
+    }
+    await lib.fresh();
     const listed = new Set<number>();
     const files: EnteFile[] = [];
     for (const collection of lib.listCollections()) {

src/clip-keywords.ts

@@ -0,0 +1,57 @@
+// Keywords for a file from its CLIP embedding.
+//
+// Ente stores no keywords. A file's ML data holds a CLIP embedding, the 512
+// numbers Ente's image model gave the photo, and Ente's photos app sorts files
+// into the categories of its Discover section with it: a file falls in a
+// category when the cosine similarity of its CLIP embedding and the text
+// embedding of the category's query is at least the category's `minScore`,
+// unless the file is a video and the category leaves out videos. A file's
+// keywords are the titles of the categories it falls in. The categories and
+// their text embeddings are in `clip-keyword-vectors.ts`.
+
+import { DISCOVER_CATEGORIES } from "./clip-keyword-vectors.js";
+import { clipEmbedding } from "./library/mldata.js";
+import type { MLData } from "./mldata-fetch.js";
+
+// How many numbers a CLIP embedding from Ente's image model has.
+const EMBEDDING_LENGTH = 512;
+
+const norm = (vector: number[]): number => Math.hypot(...vector);
+
+const cosineSimilarity = (a: number[], b: number[]): number => {
+    let dot = 0;
+    a.forEach((value, i) => {
+        dot += value * (b[i] ?? 0);
+    });
+    return dot / (norm(a) * norm(b));
+};
+
+// The titles of the Discover categories a file falls in, highest score first,
+// or undefined when its ML data has no CLIP embedding of 512 finite numbers
+// that are not all zero. Tests pass their own `categories`.
+export const clipKeywords = (
+    mlData: MLData,
+    isVideo: boolean,
+    categories = DISCOVER_CATEGORIES,
+): string[] | undefined => {
+    const embedding = clipEmbedding(mlData);
+    if (
+        !embedding ||
+        embedding.length !== EMBEDDING_LENGTH ||
+        norm(embedding) === 0
+    ) {
+        return undefined;
+    }
+
+    const matches: { title: string; score: number }[] = [];
+    for (const category of categories) {
+        if (isVideo && !category.showVideo) continue;
+        const score = cosineSimilarity(embedding, category.textEmbedding);
+        if (score >= category.minScore) {
+            matches.push({ title: category.title, score });
+        }
+    }
+    return matches
+        .sort((a, b) => b.score - a.score)
+        .map((match) => match.title);
+};

src/library/index.ts

@@ -1,28 +1,37 @@
 // The library surface over the local cache.
 //
 // `Library.open()` loads the on-disk metadata store (issue #41), then starts
-// the refresh loop. When the cache loaded empty it awaits the first refresh,
-// so the library never opens onto an empty store it could have filled; when an
-// existing copy loaded, that first refresh runs in the background and `open()`
-// returns as soon as the cached data is ready to serve — a slow or unreachable
-// server no longer stalls opening. A background timer then refreshes every
-// `refreshIntervalSeconds`. Every default read is answered from RAM — no
-// default read touches the network. There is deliberately no `sync()`, no
-// `refresh()`, no `serverReachable` flag, and no "before each read" mode
-// (design #36).
+// the refresh loop. When the stored cursor is 0, `open()` awaits the first
+// refresh. No refresh has then stored any collection: the cache is empty, or
+// holds only what `refreshCollection()` stored. The library must not serve
+// that as the whole account. Otherwise that first refresh runs in the
+// background. `open()` returns as soon as the cached data is ready to serve,
+// so a slow or unreachable server does not stall opening. A background timer
+// then takes a turn `refreshIntervalSeconds` after its last turn ends. Each
+// turn starts a refresh, unless a refresh or `refreshCollection()` is already
+// running. Every default read is answered from RAM — no default read touches
+// the network. There is deliberately no `sync()`, no `refresh()`, no
+// `serverReachable` flag, and no "before each read" mode (design #36).
 //
-// `fresh()` is the one exception (issue #75, an owner amendment to #36): it
-// forces a refresh, awaits it, and only then hands back the read namespaces, so
-// a caller that needs server-current data can ask for it. Concurrent `fresh()`
-// calls coalesce onto one in-flight refresh, and a refresh that fails rejects
-// the caller (the default reads stay silent and serve the last good copy). The
-// default methods and the background loop are unchanged.
+// `fresh()` is an exception (issue #75, an owner amendment to #36). It awaits
+// a refresh, and only then hands back the read namespaces. A caller that needs
+// server-current data uses it. Concurrent `fresh()` calls coalesce onto one
+// in-flight refresh. A refresh that fails rejects the caller. The default
+// reads stay silent and serve the last good copy. The default methods and the
+// background loop are unchanged.
 //
 // `freshCollections()` is the fresh read of the collection list alone (issue
-// #210): one collections request, no collection's files, and nothing stored,
-// for a caller such as `quak collections` that needs no more. A caller that
-// asks for each refresh it needs can also open the library with
-// `refreshOnOpen: false`, so that `open()` starts none.
+// #210). It sends one collections request and stores nothing.
+// `quak collections` uses it. `refreshCollection()` brings one collection
+// from that list up to date in the store (issue #217). It fetches that
+// collection's files alone, as a refresh would. `quak files COLLECTION` uses
+// it. It never runs at the same time as a refresh. Once the stored cursor is
+// past 0, it only updates a collection the store holds, and only a refresh
+// adds one. A refresh from a stored cursor of 0 replaces all the store holds.
+// So after a refresh, the store holds collections in the order the server
+// first sent them. A caller that asks for
+// each refresh it needs can also open the library with `refreshOnOpen: false`,
+// so that `open()` starts none.
 //
 // A refresh reports through `onProgress` each collection whose files it has
 // fetched, so a large account, whose first refresh fetches the files of every
@@ -31,10 +40,12 @@
 // A refresh stages all of its network work first and only mutates the store
 // once every fetch has succeeded. A refresh that fails partway therefore never
 // becomes visible to reads: the last good snapshot stays in place, and the
-// failure surfaces through `onProgress` and `status()` instead. A commit that
-// mutates RAM but then fails to persist keeps `status().lastError` set and the
-// store marked unsaved until a later save actually lands, so a stuck disk is
-// never masked by a subsequent empty refresh.
+// failure surfaces through `onProgress` and `status()` instead. A save can
+// still fail after the store has changed in RAM. The store then stays marked
+// unsaved until a later save lands. Until then, every `commit()` retries the
+// save. A refresh whose save fails sets `status().lastError`, which stays set
+// until a later refresh succeeds. `refreshCollection()` never sets or clears
+// it. So a stuck disk is never hidden by a later refresh that changes nothing.
 
 import { rm } from "node:fs/promises";
 import { join, resolve } from "node:path";
@@ -159,21 +170,23 @@ export interface LibraryClient {
         payloads: Map<number, MLData>;
         errors: Map<number, unknown>;
     }>;
-    // The byte source for the on-disk content cache. Optional so a mock client
-    // that only serves metadata still satisfies the interface; when absent (and
-    // no explicit `contentSource` is passed to `open`) the content cache is
-    // disabled and `Photo.original`/`thumbnail` and `thumbnails.ensure` throw.
+    // The byte source for the on-disk content cache. Optional, so a mock
+    // client that only serves metadata still satisfies the interface. Without
+    // it, and with no `contentSource` passed to `open`, the content cache is
+    // disabled. `Photo.original`/`thumbnail` and `thumbnails.ensure` then
+    // throw.
     contentSource?(): ContentSource;
 }
 
-// A progress event for one unit of background work. A metadata "refresh" or an
-// ML "fetchMLData" pass each fire "started" before their network work and then
-// exactly one of "done" or "failed"; "failed" carries the error message and
-// an ML "done" reports how many payloads it stored and, in `mlDataErrors`,
-// each file whose ML data did not decrypt, gunzip or parse, with the reason.
-// In between, a refresh fires "progress" after each collection whose files it
-// fetched: `fetched` is how many collections it has fetched the files of so
-// far, `total` how many it will. The precache fills
+// A progress event for one unit of background work. A metadata "refresh" and
+// an ML "fetchMLData" pass each fire "started" before their network work. Each
+// then fires exactly one of "done" or "failed". A "failed" event carries the
+// error message. An ML "done" reports in `fetched` how many payloads it
+// stored. Its `mlDataErrors` lists each file whose ML data did not decrypt,
+// gunzip or parse, with the reason. In between, a refresh fires "progress"
+// after each collection whose files it fetched. Its `fetched` is how many
+// collections it has fetched the files of so far, and `total` how many it
+// will. The precache fills
 // ("precacheThumbnails"/"precacheOriginals", #48) fire "started"/"done" around
 // each sweep that has work, "done" reporting the count newly cached.
 export interface RefreshEvent {
@@ -201,9 +214,10 @@ export interface LibraryOptions {
     // directory at open. The content cache treats an original already stored
     // at its save path as present.
     downloadDirectory?: string;
-    // How long after one refresh ends the background timer starts the next.
-    // `Infinity` starts none: the library then refreshes only as it opens and
-    // when `fresh()` or `backup()` asks.
+    // How long after its last turn ends the background timer takes the next.
+    // A turn starts a refresh unless a refresh or `refreshCollection()` is
+    // running. With `Infinity` the timer takes no turn: the library then
+    // refreshes only as it opens and when `fresh()` or `backup()` asks.
     refreshIntervalSeconds?: number;
     // Whether `open()` starts a refresh. Defaults to true. With false, `open()`
     // returns at once onto what the cache holds, empty or not, and a refresh
@@ -242,8 +256,9 @@ export interface LibraryStatus {
     // Wall-clock ms of the last refresh that succeeded, or undefined if none
     // has yet.
     lastRefreshAt?: number;
-    // The message from the most recent refresh, set only while that refresh
-    // failed; cleared by the next success.
+    // The error of the most recent refresh, set only when it failed. The next
+    // refresh that succeeds clears it. `refreshCollection()` never sets or
+    // clears it.
     lastError?: string;
     // Wall-clock ms of the last ML fetch pass that succeeded, or undefined if
     // none has yet (or ML fetching is disabled).
@@ -306,12 +321,17 @@ export class Library {
     private readonly precache?: Precache;
 
     private timer?: ReturnType<typeof setTimeout>;
-    // The in-flight refresh cycle, or undefined when none runs. One slot serves
-    // both paths: the background loop skips when it is set, and a fresh read
-    // (issue #75) coalesces onto it or starts one. The promise carries the
-    // cycle's real outcome (it rejects on failure); the background loop ignores
-    // that, a fresh read propagates it.
+    // The in-flight refresh cycle or `refreshCollection()`, or undefined when
+    // neither runs. One slot holds both, so they never run at the same time.
+    // The background loop skips while it is set. A fresh read (issue #75)
+    // coalesces onto a running refresh. It waits for a running
+    // `refreshCollection()` to end, and then joins or starts a refresh. The
+    // promise carries the real outcome: it rejects on failure. The background
+    // loop ignores a failure. A fresh read passes on the failure of the
+    // refresh it joins or starts, and `refreshCollection()` passes on its own.
     private cycle?: Promise<void>;
+    // Whether `cycle` is a refresh cycle rather than a `refreshCollection()`.
+    private cycleIsRefresh = false;
     // Guards the ML fetch pass so a slow backfill never runs twice at once; a
     // refresh whose pass is still running kicks nothing new. Holds the running
     // pass, so `close()` and `backup()` can wait for it. It rejects when the
@@ -324,19 +344,20 @@ export class Library {
     private lastMLError?: string;
     // For each file whose ML data the last ML fetch pass that succeeded got
     // but could not decrypt, gunzip or parse, why not, by fileID. A pass
-    // gathers its own and puts them here when it ends without failing, so
-    // `status()` and `backup()` never see the list of a pass that is still
-    // running or that failed. Kept in RAM only, never in the cache.
+    // gathers its own and puts them here only when it ends without failing.
+    // So `status()` and `backup()` never see the list of a pass still running
+    // or one that failed. Kept in RAM only, never in the cache.
     private mlDataErrors = new Map<number, unknown>();
-    // The plain-record projection as of the last refresh, and the GUI change
-    // subscribers. A refresh that alters the projection notifies each with the
-    // delta; `lastRecords` is kept current every refresh so a subscriber that
-    // joins later diffs against the state its own `snapshot()` already returned.
+    // The plain-record projection as of the last store change, and the GUI
+    // change subscribers. A refresh or `refreshCollection()` that alters the
+    // projection notifies each subscriber with the delta. `lastRecords`
+    // advances on each `commit()` that changes the store. A subscriber that
+    // joins later then diffs against the state its own `snapshot()` returned.
     private readonly subscribers = new Set<(change: LibraryChange) => void>();
     private lastRecords: DerivedRecords;
     // RAM holds changes disk has not yet accepted (an earlier save failed).
-    // Cleared only when a save actually succeeds; keeps the store trying to
-    // persist and the failure visible in `status()` until then.
+    // Cleared only when a save succeeds. Until then, every `commit()` retries
+    // the save.
     private unsaved = false;
 
     private constructor(args: {
@@ -366,7 +387,7 @@ export class Library {
         this.lastRecords = this.deriveNow();
 
         // The read namespaces derive fresh from the store on each call, so they
-        // always reflect the latest refresh.
+        // always reflect what it holds now.
         const derive = (): DerivedRecords => this.deriveNow();
         const root = this.downloadDirectory;
         const saves: SavePathLookup = {
@@ -393,14 +414,16 @@ export class Library {
         this.mldata = makeMLDataAPI(() => this.mlStore);
     }
 
-    // Load the cache and start the refresh loop. With an empty cache the first
-    // refresh is awaited, so `open()` resolves onto populated data whenever the
-    // server is reachable; that awaited refresh may still fail, and the library
-    // then opens empty with the failure recorded in `status()`. With an
-    // existing cache the first refresh runs in the background and `open()`
-    // returns as soon as the cached data is ready — an unreachable server does
-    // not block opening. With `refreshOnOpen: false` there is no first
-    // refresh: `open()` returns onto the cache as it is.
+    // Load the cache and start the refresh loop. When the stored cursor is 0,
+    // the first refresh is awaited. Such a cache is empty, or holds only what
+    // `refreshCollection()` stored. `open()` then resolves onto the whole
+    // account whenever the server is reachable. If that refresh fails, the
+    // library opens onto what the cache holds, with the failure recorded in
+    // `status()`. With any other cache the first refresh runs in the
+    // background, and `open()` returns as soon as the cached data is ready.
+    // An unreachable server then does not block opening. With
+    // `refreshOnOpen: false` there is no first refresh: `open()` returns onto
+    // the cache as it is.
     static async open(opts: LibraryOptions): Promise<Library> {
         const { userID } = opts.client.whoami();
         const cacheDirectory =
@@ -416,8 +439,8 @@ export class Library {
         let store = await MetadataStore.load(metadataPath);
         // A cache directory given explicitly can hold another account's cache.
         // Its records and cursor are not this account's, so delete it and the
-        // ML data beside it and start empty. A user ID of 0 means the cache
-        // was never refreshed and so holds nothing to discard.
+        // ML data beside it and start empty. A user ID of 0 means nothing was
+        // ever stored there, so there is nothing to discard.
         if (store.userID !== 0 && store.userID !== userID) {
             await rm(metadataPath, { force: true });
             await rm(join(cacheDirectory, "mldata"), {
@@ -440,10 +463,11 @@ export class Library {
             ? await MLDataStore.open(join(cacheDirectory, "mldata"))
             : undefined;
 
-        // Build the content cache from an explicit source or the client's own,
-        // and take its record of what is already cached (and reap orphan temp
-        // files) before the first projection, so cached paths are present from
-        // the start and the first refresh raises no spurious path-change diff.
+        // Build the content cache from an explicit source or the client's own.
+        // Take its record of what is already cached, and reap orphan temp
+        // files, before the first projection. Cached paths are then present
+        // from the start, and the first `commit()` raises no spurious
+        // path-change diff.
         const source = opts.contentSource ?? opts.client.contentSource?.();
         let cache: ContentCache | undefined;
         let precache: Precache | undefined;
@@ -497,22 +521,26 @@ export class Library {
             precache,
         });
 
-        // Start filling from whatever the loaded store already holds; each
-        // refresh below re-kicks with the new files (and retries any that
-        // failed). An empty store starts empty here and fills after its first
-        // refresh.
+        // Start filling from whatever the loaded store already holds. Each
+        // later `commit()` re-kicks the fills with its new files, and retries
+        // any that failed. An empty store starts empty here, and fills once a
+        // `commit()` stores files.
         precache?.start();
 
         if (opts.refreshOnOpen === false) {
             // The caller asks for each refresh it needs.
             lib.scheduleNext();
-        } else if (store.loadedFromDisk) {
-            // An existing copy already answers reads; refresh in the background
-            // and start the interval once that first cycle settles.
+        } else if (store.collectionsSinceTime > 0) {
+            // The stored cursor has moved, so a refresh has stored the
+            // account's collections and the cache's copy answers reads.
+            // Refresh in the background, and start the interval once that
+            // first cycle settles.
             void lib.runRefresh().then(() => lib.scheduleNext());
         } else {
-            // Nothing was cached: wait for the first refresh to fill the store
-            // (or fail) rather than resolve onto an empty library.
+            // The stored cursor is still 0, so no refresh has stored any
+            // collection: the cache is empty, or holds only what
+            // `refreshCollection()` stored. Wait for the first refresh to fill
+            // the store (or fail) rather than resolve onto part of the account.
             await lib.runRefresh();
             lib.scheduleNext();
         }
@@ -550,9 +578,9 @@ export class Library {
         return snapshotFrom(this.deriveNow(), Date.now());
     }
 
-    // Deliver a `LibraryChange` whenever a refresh alters the projection. A
-    // refresh that changes nothing delivers nothing. The returned handle's
-    // `unsubscribe` stops delivery.
+    // Deliver a `LibraryChange` whenever a refresh or `refreshCollection()`
+    // alters the projection. One that changes nothing delivers nothing. The
+    // returned handle's `unsubscribe` stops delivery.
     subscribe(args: { onChange: (change: LibraryChange) => void }): {
         unsubscribe: () => void;
     } {
@@ -595,12 +623,13 @@ export class Library {
         };
     }
 
-    // Fresh reads (issue #75, owner amendment to design #36). Force a refresh,
-    // wait for it to complete and persist, then hand back the same
-    // `albums`/`photos`/`timeline` namespaces — now guaranteed to reflect a
-    // completed server round-trip. Concurrent calls coalesce onto one refresh;
-    // a refresh that fails rejects here, where the default namespaces would
-    // instead stay silent and serve the last good copy.
+    // Fresh reads (issue #75, owner amendment to design #36). Wait for a
+    // refresh to complete and persist, then hand back the same
+    // `albums`/`photos`/`timeline` namespaces. They then reflect a completed
+    // server round-trip. A running `refreshCollection()` is waited for first.
+    // Concurrent calls coalesce onto one refresh. A refresh that fails rejects
+    // here, where the default namespaces would stay silent and serve the last
+    // good copy.
     async fresh(): Promise<FreshReads> {
         await this.refreshNow();
         return {
@@ -613,10 +642,10 @@ export class Library {
     // The fresh read of the collection list alone (issue #210): every live
     // collection that decrypts, in the server's order, from one collections
     // request. Unlike a refresh, it fetches no collection's files, which on a
-    // large account take long. Nothing is stored: a refresh fetches a
-    // collection's files only when the collection is new or has changed since
-    // the stored copy, so a collection stored here without its files would
-    // never have them fetched.
+    // large account take long. It stores nothing. A refresh from a stored
+    // cursor past 0 fetches a collection's files only when the collection is
+    // new or has changed since the stored copy. So does `refreshCollection()`.
+    // Neither would fetch the files of a collection stored here without them.
     async freshCollections(): Promise<Collection[]> {
         if (this.closed) throw new Error("the library is closed");
         const { collections } = await this.client.collectionsSince({
@@ -625,19 +654,44 @@ export class Library {
         return collections;
     }
 
+    // Bring one collection up to date in the store (issue #217), given its
+    // record from `freshCollections()`. It does for that collection what a
+    // refresh does for a changed one. It fetches no other collection's files.
+    // It resolves to whether the store then holds the collection.
+    //
+    // It adds a collection the store does not hold only while the stored
+    // cursor is 0, before any refresh has received a collection. After that,
+    // it stores nothing for such a collection and resolves to false. Only a
+    // refresh may then add it, after the collections the store holds.
+    //
+    // It runs in the same slot as a refresh, so the two never run at the same
+    // time. It first waits for a running refresh or `refreshCollection()` to
+    // end. While its own work runs, a fresh read or `backup()` waits for it,
+    // and the background loop skips its turn.
+    async refreshCollection(collection: Collection): Promise<boolean> {
+        while (this.cycle) await this.cycle.catch(() => {});
+        if (this.closed) throw new Error("the library is closed");
+        let held = false;
+        await this.startCycle(false, async () => {
+            held = await this.refreshCollectionOnce(collection);
+        });
+        return held;
+    }
+
     // Back up every in-scope file to `opts.downloadDirectory`, or else the
     // library's, each original at its save path, with a durable failure
-    // ledger (issue #51). Takes the lock in that directory first, unless the
-    // caller passes the lock it took as `opts.lock`, failing at once while
-    // another backup of it runs, and before it starts another file once the
-    // lock is lost (see `lockBackupDirectory`). Waits for a
-    // completed refresh, as `fresh()` does, joining one already running, and
-    // rejects before touching any file but the lock when it fails. Then puts
-    // pending originals at their save paths as `Photo.download()` does (and
-    // optional thumbnails) through the content cache and pools, waits for an
-    // ML data fetch, and rebuilds the derived symlink/JSON views from the
-    // model, each file's JSON with its ML data, beside an `account.json` with
-    // the account's email and user ID.
+    // ledger (issue #51). It first takes the lock in that directory, unless
+    // the caller passes the lock it took as `opts.lock`. Taking it fails at
+    // once while another backup of that directory runs. Once the lock is lost,
+    // the backup fails before it starts another file (see
+    // `lockBackupDirectory`). It waits for a completed refresh as `fresh()`
+    // does, joining one already running. When that refresh fails, it rejects
+    // before touching any file but the lock. It then puts pending originals
+    // at their save paths as `Photo.download()` does, through the content
+    // cache and pools. With `includeThumbnails` it also fetches thumbnails. It
+    // waits for an ML data fetch. It rebuilds the derived symlink/JSON views
+    // from the model, each file's JSON with its ML data. An `account.json`
+    // beside them holds the account's email and user ID.
     // Throws before any network work when no content cache backs the
     // originals it must fetch.
     backup(opts?: BackupOptions): Promise<BackupResult> {
@@ -674,12 +728,15 @@ export class Library {
         );
     }
 
-    // Stop the background timer. Idempotent. An in-flight refresh is left to
-    // finish; it will not schedule another cycle once closed. The returned
-    // promise resolves once that refresh (including its cache write), the ML
-    // fetch pass and the precache fetches already running have all finished,
-    // so a caller can then remove the cache directory. A refresh failure is
-    // reported through `status()`, not thrown here.
+    // Stop the background timer. Idempotent. A refresh or `refreshCollection()`
+    // in flight is left to finish, and nothing starts after it. The returned
+    // promise resolves once all of these have finished:
+    //  - that refresh or `refreshCollection()`, with its cache write;
+    //  - the ML fetch pass;
+    //  - the precache fetches already running.
+    // A caller can then remove the cache directory. A refresh failure is
+    // reported through `status()`, and a `refreshCollection()` failure to its
+    // own caller. Neither is thrown here.
     async close(): Promise<void> {
         this.closed = true;
         const precacheClosed = this.precache?.close();
@@ -702,33 +759,41 @@ export class Library {
         this.timer.unref?.();
     }
 
-    // The background loop's refresh: run a cycle unless one is already in flight
-    // (or the library is closed), and never let a failure escape — the
-    // background path reports errors through `status()`/`onProgress`, it does
-    // not throw. Resolves once the cycle it started (or skipped past) settles.
+    // The background loop's refresh: run a cycle unless a refresh or
+    // `refreshCollection()` is in flight (or the library is closed), and never
+    // let a failure escape. The background path reports errors through
+    // `status()`/`onProgress`; it does not throw. It resolves once the cycle
+    // it started settles, or at once when it starts none.
     private runRefresh(): Promise<void> {
         if (this.closed || this.cycle) return Promise.resolve();
-        return this.startCycle().catch(() => {});
+        return this.startCycle(true, () => this.refreshCycle()).catch(() => {});
     }
 
-    // A fresh read's refresh (issue #75): force a cycle and await it, rejecting
-    // if it fails. Concurrent fresh reads coalesce onto the one in-flight cycle
-    // — the background loop's included — so they never fan out into redundant
-    // server round-trips.
-    private refreshNow(): Promise<void> {
-        if (this.closed) {
-            return Promise.reject(new Error("the library is closed"));
+    // A fresh read's refresh (issue #75): join the running refresh or start
+    // one, and await it, rejecting if it fails. Concurrent fresh reads
+    // coalesce onto the one in-flight cycle — the background loop's included —
+    // so they never fan out into redundant server round-trips. A running
+    // `refreshCollection()` is waited for first.
+    private async refreshNow(): Promise<void> {
+        while (this.cycle && !this.cycleIsRefresh) {
+            await this.cycle.catch(() => {});
         }
-        return this.cycle ?? this.startCycle();
+        if (this.closed) throw new Error("the library is closed");
+        return this.cycle ?? this.startCycle(true, () => this.refreshCycle());
     }
 
-    // Start one refresh cycle and record it as the in-flight cycle so every
-    // caller coalesces onto it. The returned promise carries the cycle's real
-    // outcome; each caller attaches the handling its own path needs, and the
-    // slot is cleared once the cycle settles.
-    private startCycle(): Promise<void> {
-        const cycle = this.refreshCycle();
+    // Start `work` and record it in the in-flight slot, so every other caller
+    // sees it. `isRefresh` says whether it is a refresh cycle, which a fresh
+    // read can coalesce onto. The returned promise carries the real outcome.
+    // Each caller attaches the handling its own path needs, and the slot is
+    // cleared once the work settles.
+    private startCycle(
+        isRefresh: boolean,
+        work: () => Promise<void>,
+    ): Promise<void> {
+        const cycle = work();
         this.cycle = cycle;
+        this.cycleIsRefresh = isRefresh;
         void cycle.then(
             () => {
                 if (this.cycle === cycle) this.cycle = undefined;
@@ -751,9 +816,9 @@ export class Library {
             this.lastError = undefined;
             this.emit({ operation: "refresh", status: "done" });
             // Backfill ML data for the files this refresh knows about. It runs
-            // outside the refresh's success/failure so a fetch or disk problem
-            // there never marks the metadata refresh failed, and it is not
-            // awaited so it never stalls the refresh interval. Its failure is
+            // outside the refresh's success or failure, so a fetch or disk
+            // problem there never marks the metadata refresh failed. It is not
+            // awaited, so it never stalls the refresh interval. Its failure is
             // reported through `status()` and `onProgress`.
             void this.fetchMLDataNow().catch(() => {});
         } catch (err) {
@@ -766,11 +831,14 @@ export class Library {
 
     // Fetch every change since the stored cursor, then commit. All network
     // reads happen before any store mutation, so a fetch that throws leaves the
-    // store untouched and the previous snapshot intact. A collection or file
-    // that does not decrypt is left out instead, as if deleted, with the reason
-    // kept in the store: the cursor moves past it, so no later refresh sees it
-    // again until the server's copy changes.
+    // store untouched and the previous snapshot intact.
+    //
+    // From a stored cursor of 0, the server lists every live collection. The
+    // refresh then replaces all the store holds, and counts every collection
+    // as new. Only `refreshCollection()` can have stored anything before it.
+    // So the store ends as it would had `refreshCollection()` never run.
     private async refreshOnce(): Promise<void> {
+        const replace = this.store.collectionsSinceTime === 0;
         const page = await this.client.collectionsSince({
             sinceTime: this.store.collectionsSinceTime,
         });
@@ -778,16 +846,19 @@ export class Library {
         // Stage per-collection file diffs. A collection's files are
         // re-enumerated only when its updationTime has advanced past the cached
         // copy; an unchanged album's file list cannot have changed. New
-        // collections enumerate from the beginning of time. Each collection
+        // collections enumerate from the beginning of time. When the refresh
+        // replaces the store, every collection counts as new. Each collection
         // whose files are fetched is reported, with how many there are to
         // fetch.
+        const storedCopy = (id: number): Collection | undefined =>
+            replace ? undefined : this.store.getCollection(id);
         const changedCollections = page.collections.filter((collection) => {
-            const known = this.store.getCollection(collection.id);
+            const known = storedCopy(collection.id);
             return !known || collection.updationTime > known.updationTime;
         });
         const filePages: { collectionID: number; page: FilesPage }[] = [];
         for (const collection of changedCollections) {
-            const known = this.store.getCollection(collection.id);
+            const known = storedCopy(collection.id);
             const filePage = await this.client.filesSince({
                 collectionID: collection.id,
                 collectionKey: collection.key,
@@ -802,8 +873,56 @@ export class Library {
             });
         }
 
-        // Network work done; commit to the store and persist only if something
-        // actually changed.
+        await this.commit(page, filePages, replace);
+    }
+
+    // `refreshCollection()`'s work, run in the slot: fetch the files of one
+    // collection changed since the stored copy, then commit. It resolves to
+    // whether the store then holds the collection.
+    private async refreshCollectionOnce(
+        collection: Collection,
+    ): Promise<boolean> {
+        const known = this.store.getCollection(collection.id);
+        // A refresh has received the account's collections, and this one is
+        // not stored. It is new since then, or a refresh dropped it. Leave it
+        // to the next refresh, which adds new collections after the stored
+        // ones, in the server's order.
+        if (!known && this.store.collectionsSinceTime > 0) return false;
+        // The stored copy is as new or newer, so there is nothing to fetch.
+        if (known && collection.updationTime <= known.updationTime) return true;
+        const filePage = await this.client.filesSince({
+            collectionID: collection.id,
+            collectionKey: collection.key,
+            sinceTime: known ? known.updationTime : 0,
+        });
+        // The stored cursor stays where it is, so the next refresh still
+        // fetches the files of every other collection that changed.
+        await this.commit(
+            {
+                collections: [collection],
+                deleted: [],
+                cursor: this.store.collectionsSinceTime,
+            },
+            [{ collectionID: collection.id, page: filePage }],
+            false,
+        );
+        return true;
+    }
+
+    // Commit a collections page and the file pages fetched for its collections
+    // to the store. With `replace`, the page lists every live collection, so
+    // the commit first deletes every stored collection. A collection or file
+    // that does not decrypt is left out, as if deleted. The reason is kept in
+    // the store. Until the server's copy changes, only a refresh from a stored
+    // cursor of 0 fetches it again. Last, save the store when it holds changes
+    // not yet saved.
+    private async commit(
+        page: CollectionsPage,
+        filePages: { collectionID: number; page: FilesPage }[],
+        replace: boolean,
+    ): Promise<void> {
+        // The network work is done, so change the store now. `changed` says
+        // whether this commit changed anything.
         let changed = false;
 
         if (this.store.userID !== this.userID) {
@@ -811,6 +930,13 @@ export class Library {
             changed = true;
         }
 
+        if (replace) {
+            for (const { id } of this.store.listCollections()) {
+                this.store.deleteCollection(id);
+                changed = true;
+            }
+        }
+
         for (const id of page.deleted) {
             if (this.store.getCollection(id)) {
                 this.store.deleteCollection(id);
@@ -864,10 +990,10 @@ export class Library {
         if (changed) this.unsaved = true;
 
         // Reproject and notify subscribers of the delta. This tracks RAM (what
-        // reads see), so it fires whether or not the save below succeeds; a
-        // save failure surfaces separately through `status().lastError`.
-        // `lastRecords` advances every changed refresh so the next diff is
-        // against current state.
+        // reads see), so it fires whether or not the save below succeeds. A
+        // failed save rejects the caller, and a refresh also records it in
+        // `status().lastError`. `lastRecords` advances on every commit that
+        // changes something, so the next diff is against current state.
         if (changed) {
             const next = this.deriveNow();
             if (this.subscribers.size > 0) {
@@ -878,16 +1004,16 @@ export class Library {
             // Recompute the fill orders and pinned set against the new library.
             this.precache?.update(next);
         }
-        // Re-kick the fills every cycle: a finished sweep starts afresh to pick
-        // up new files and retry any that failed, and a running one is left be.
+        // Re-kick the fills on every commit. A finished sweep starts afresh to
+        // pick up new files and retry any that failed. A running one is left
+        // be.
         this.precache?.start();
 
         // Persist whenever RAM holds changes disk has not accepted — including
         // changes an earlier cycle staged whose save failed. `unsaved` clears
-        // only once a save lands, so a save failure both stays visible through
-        // `status().lastError` (the throw below records it) and keeps being
-        // retried, instead of a later empty refresh silently clearing it while
-        // the on-disk cache is still behind RAM.
+        // only once a save lands, so a later commit retries a failed save even
+        // when it changes nothing. A failed save rejects the caller, and a
+        // refresh also records it in `status().lastError`.
         if (this.unsaved) {
             await this.store.save();
             this.unsaved = false;
@@ -903,13 +1029,14 @@ export class Library {
         return this.mlFetch;
     }
 
-    // One ML fetch pass: fetch, decrypt and store the ML data for every file
-    // the store knows about that is not cached (or whose `updationTime` has
-    // advanced), through the metadata pool, and update the CLIP index. For a
-    // file whose ML data does not decrypt, gunzip or parse, it gathers the
-    // reason instead, and puts the reasons it gathered in `mlDataErrors` when
-    // it ends without failing. A failure leaves `mlDataErrors` as it was and
-    // is reported through `status()` and `onProgress`, then thrown.
+    // One ML fetch pass. It fetches, decrypts and stores the ML data of every
+    // file the store knows about that is not cached, or whose `updationTime`
+    // has advanced. It fetches through the metadata pool, and updates the CLIP
+    // index. For a file whose ML data does not decrypt, gunzip or parse, it
+    // gathers the reason instead. It puts the reasons it gathered in
+    // `mlDataErrors` when it ends without failing. A failure leaves
+    // `mlDataErrors` as it was and is reported through `status()` and
+    // `onProgress`, then thrown.
     private async runMLFetch(): Promise<void> {
         const mldata = this.mlStore;
         // Bind so the call keeps the client as its receiver when invoked
@@ -977,9 +1104,9 @@ export class Library {
         }));
     }
 
-    // The distinct files the store holds, one entry per fileID (a file in
-    // several collections shares its ML data), each carrying the key and the
-    // newest `updationTime` seen across its memberships.
+    // The distinct files the store holds, one entry per fileID, since a file
+    // in several collections shares its ML data. Each entry carries the key
+    // and the newest `updationTime` seen across the file's memberships.
     private uniqueFiles(): {
         id: number;
         key: Uint8Array;

src/library/mldata.ts

@@ -61,7 +61,7 @@ export interface MLIndex {
 // Pull the CLIP embedding out of a payload, or undefined when it is absent or
 // misshapen. Kept strict so a bad payload is skipped rather than corrupting the
 // packed index.
-const clipEmbedding = (payload: MLData): number[] | undefined => {
+export const clipEmbedding = (payload: MLData): number[] | undefined => {
     const clip = payload.clip;
     if (typeof clip !== "object" || clip === null) return undefined;
     const embedding = (clip as { embedding?: unknown }).embedding;

src/library/precache.ts

@@ -122,7 +122,7 @@ export class Precache {
     }
 
     // Recompute the fill orders and the pinned set from the current projection.
-    // Called at open and after every refresh that changes the library.
+    // Called at open and after every change to the store.
     update(records: DerivedRecords): void {
         const photos = [...records.photos.values()].sort(
             (a, b) => b.takenAt - a.takenAt || b.fileID - a.fileID,

src/library/records.ts

@@ -71,13 +71,14 @@ export interface LibrarySnapshot {
     takenAt: number;
 }
 
+// What one refresh or `refreshCollection()` changed in the records.
 export interface LibraryChange {
-    // Full records for albums/photos added or changed by the refresh.
+    // Full records for albums/photos added or changed.
     albumsChanged: AlbumRecord[];
     photosChanged: PhotoRecord[];
     fileIDsRemoved: number[];
     albumIDsRemoved: number[];
-    // Wall-clock milliseconds of the refresh that produced this change.
+    // Wall-clock milliseconds when the store changed.
     refreshedAt: number;
 }
 

src/library/store.ts

@@ -3,7 +3,7 @@
 // The store keeps one `metadata.json` file holding the account's server
 // state: the user id, a schema version, the cursor for the incremental
 // collections listing, the decrypted collection and file records, and the
-// collections and files the refresh could not decrypt. The whole file is read
+// collections and files that did not decrypt. The whole file is read
 // into RAM on load and rewritten as a whole on save; there is no partial
 // update and no lock file. A separate refresh unit populates the store from
 // the server — this module only stores what it is given.
@@ -34,9 +34,11 @@ const FILE_MODE = 0o600;
 type StoredCollection = Omit<Collection, "key"> & { key: string };
 type StoredFile = Omit<EnteFile, "key"> & { key: string };
 
-// A collection, or with `fileID` a file in it, that the last refresh to fetch
-// it could not decrypt or parse, and why. It is kept until a later refresh
-// fetches it again, because a refresh fetches only what changed.
+// A collection, or with `fileID` a file in it, that did not decrypt or parse,
+// and why. A refresh or `refreshCollection()` records it on fetching such a
+// copy. A later fetch drops it on getting a copy that decrypts, or on learning
+// that the server deleted it. Each fetch asks only for what changed, so the
+// store keeps the record between fetches.
 export interface Undecryptable {
     collectionID: number;
     fileID?: number;
@@ -70,12 +72,6 @@ export class MetadataStore {
     userID = 0;
     collectionsSinceTime: Microseconds = 0;
 
-    // True when `load` populated this store from a valid existing file; false
-    // on a first run or a missing/corrupt/wrong-version file that loaded empty.
-    // `Library.open` reads it to decide whether the first refresh may run in
-    // the background (an existing copy already serves reads) or must be awaited.
-    loadedFromDisk = false;
-
     private readonly collections = new Map<number, Collection>();
     private readonly files = new Map<string, EnteFile>();
     private undecryptable: Undecryptable[] = [];
@@ -102,7 +98,6 @@ export class MetadataStore {
             if (parsed.schemaVersion !== METADATA_SCHEMA_VERSION) {
                 return store;
             }
-            store.loadedFromDisk = true;
             store.userID = parsed.userID ?? 0;
             store.collectionsSinceTime = parsed.collectionsSinceTime ?? 0;
             for (const stored of parsed.collections ?? []) {
@@ -229,8 +224,14 @@ export class MetadataStore {
         this.files.delete(fileKey(collectionID, fileID));
     }
 
+    // Sorted by collection ID, then file ID, so the order does not depend on
+    // when each record was fetched.
     listUndecryptable(): Undecryptable[] {
-        return [...this.undecryptable];
+        return [...this.undecryptable].sort(
+            (a, b) =>
+                a.collectionID - b.collectionID ||
+                (a.fileID ?? 0) - (b.fileID ?? 0),
+        );
     }
 
     // Replaces any earlier record of the same collection or file.

src/metadata-backup.ts

@@ -1,6 +1,7 @@
 import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import * as jpeg from "jpeg-js";
+import { clipKeywords } from "./clip-keywords.js";
 import type { Client } from "./client.js";
 import { describeError } from "./errors.js";
 import { readExifTags } from "./exif.js";
@@ -97,7 +98,8 @@ const extractExif = async (
 
 // Dump every decrypted metadata layer the account holds into a directory tree
 // of plain JSON: account, per-collection, and per-file records including the
-// private and public magic metadata and (by default) the ML data. Collections
+// private and public magic metadata and the ML data, with the keywords of its
+// CLIP embedding in place of the embedding (see clip-keywords.ts). Collections
 // and files are enumerated from the library's cache, which the caller refreshes
 // first. Returns how many files have no ML data because their ML data request
 // failed or their ML data did not decrypt, gunzip or parse; they are still
@@ -230,8 +232,16 @@ export const runMetadataBackup = async (
         if (file.pubMagicMetadata)
             fileMeta.pubMagicMetadata = file.pubMagicMetadata;
 
+        // The ML data without the CLIP embedding, and the keywords that
+        // embedding gives.
         const ml = mlDataMap.get(file.id);
-        if (ml) fileMeta.mlData = ml;
+        if (ml) {
+            const { clip: _clip, ...mlDataWithoutClip } = ml;
+            fileMeta.mlData = mlDataWithoutClip;
+            const isVideo = file.metadata.fileType === "video";
+            const keywords = clipKeywords(ml, isVideo);
+            if (keywords) fileMeta.keywords = keywords;
+        }
         const mlError = mlDataErrors.get(file.id);
         if (mlError) fileMeta.mlDataError = mlError;
 

test/cli/backup-end-to-end.test.ts

@@ -86,10 +86,17 @@ const DAY = "2024/2024-03/2024-03-01";
 const TOAST = new TextEncoder().encode("a few seconds of the wedding toast");
 
 // Ente's ML data for beach.jpg: a face it found, and the photo's CLIP
-// embedding.
+// embedding, here that of a photo of a sunset (see clip-keywords.test.ts).
 const ML_DATA = {
     face: { faces: [{ faceID: "100_1", score: 0.98 }] },
-    clip: { embedding: [0.5, 0.6, 0.7] },
+    clip: {
+        embedding: JSON.parse(
+            readFileSync(
+                new URL("../clip-keywords-sunset.json", import.meta.url),
+                "utf-8",
+            ),
+        ).embedding,
+    },
 };
 
 // Each file's ID, its owner, its original as uploaded, and the metadata its
@@ -486,7 +493,10 @@ describe("the first backup of the account", () => {
             },
             updationTime: TAKEN,
         });
-        expect(beach.mlData).toEqual(ML_DATA);
+        // The JSON holds the ML data without its CLIP embedding, and the
+        // keywords the embedding gives in its place.
+        expect(beach.mlData).toEqual({ face: ML_DATA.face });
+        expect(beach.keywords).toEqual(["Sunset"]);
         expect(beach.imageMetadata.exif.Make.description).toBe("Canon");
         expect(beach.imageMetadata.exif.Model.description).toBe("EOS R5");
 
@@ -497,6 +507,7 @@ describe("the first backup of the account", () => {
             metadata: { title: "toast.mp4", fileType: "video" },
         });
         expect(toast).not.toHaveProperty("mlData");
+        expect(toast).not.toHaveProperty("keywords");
         expect(toast).not.toHaveProperty("imageMetadata");
         expect(toast).not.toHaveProperty("imageMetadataError");
 

test/cli/backup.test.ts

@@ -52,6 +52,7 @@ import { tmpdir } from "node:os";
 import { describe, it, expect, beforeEach, afterEach, vi } from "vitest";
 
 import { runBackup, type BackupLibrary } from "../../src/backup.js";
+import { DISCOVER_CATEGORIES } from "../../src/clip-keyword-vectors.js";
 import { Library } from "../../src/library/index.js";
 import type { ContentSource } from "../../src/library/content.js";
 import type { CollectionsPage, FilesPage } from "../../src/client.js";
@@ -946,9 +947,19 @@ describe("the backup lock", () => {
     });
 });
 
-// Also serves ML data: file 100 has `ML_PAYLOAD`, the other two have none.
-// While `mlError` is set, every ML data request fails with it.
-const ML_PAYLOAD = { face: { faces: [] }, clip: { embedding: [0.5, 0.25] } };
+// Also serves ML data: file 100 has `ML_PAYLOAD`, whose CLIP embedding is that
+// of a photo of a sunset (see clip-keywords.test.ts), and the other two have
+// none. While `mlError` is set, every ML data request fails with it.
+const SUNSET_EMBEDDING: number[] = JSON.parse(
+    readFileSync(
+        new URL("../clip-keywords-sunset.json", import.meta.url),
+        "utf-8",
+    ),
+).embedding;
+const ML_PAYLOAD = {
+    face: { faces: [] },
+    clip: { embedding: SUNSET_EMBEDDING },
+};
 
 class MLClient extends MockClient {
     mlError?: string;
@@ -968,16 +979,99 @@ const fileJSON = (outDir: string, fileID: number): Record<string, unknown> =>
     JSON.parse(readFileSync(saved(outDir, `${fileID}.json`), "utf-8"));
 
 describe("ML data in each file's JSON", () => {
-    it("writes a file's ML data, and no ML field for a file that has none", async () => {
+    // The JSON holds the ML data without its CLIP embedding, and in its place
+    // the keywords the embedding gives.
+    it("writes a file's ML data and keywords, and no ML field for a file that has none", async () => {
         const lib = await openLibrary(stubSource(), new MLClient());
         const outDir = join(root, "backup");
 
         const result = await lib.backup({ downloadDirectory: outDir });
 
         expect(result.failed).toBe(0);
-        expect(fileJSON(outDir, 100).mlData).toEqual(ML_PAYLOAD);
+        expect(fileJSON(outDir, 100).mlData).toEqual({ face: { faces: [] } });
+        expect(fileJSON(outDir, 100).keywords).toEqual(["Sunset"]);
         expect(fileJSON(outDir, 101)).not.toHaveProperty("mlData");
         expect(fileJSON(outDir, 101)).not.toHaveProperty("mlDataError");
+        expect(fileJSON(outDir, 101)).not.toHaveProperty("keywords");
+        await lib.close();
+    });
+
+    // A file's ML data can hold faces and no CLIP embedding. Its JSON then has
+    // the ML data and no keywords.
+    it("writes no keywords for ML data without a CLIP embedding", async () => {
+        const faces = { face: { faces: [{ faceID: "101_1", score: 0.9 }] } };
+
+        // Gives file 101 ML data holding only faces.
+        class FacesClient extends MockClient {
+            async fetchMLData(args: { fileIDs: number[] }): Promise<{
+                payloads: Map<number, MLData>;
+                errors: Map<number, unknown>;
+            }> {
+                const payloads = new Map<number, MLData>();
+                if (args.fileIDs.includes(101)) payloads.set(101, faces);
+                return { payloads, errors: new Map() };
+            }
+        }
+        const lib = await openLibrary(stubSource(), new FacesClient());
+        const outDir = join(root, "backup");
+
+        const result = await lib.backup({ downloadDirectory: outDir });
+
+        expect(result.failed).toBe(0);
+        expect(fileJSON(outDir, 101).mlData).toEqual(faces);
+        expect(fileJSON(outDir, 101)).not.toHaveProperty("keywords");
+        await lib.close();
+    });
+
+    // Ente's photos app keeps videos out of some Discover categories, Identity
+    // among them. Here a photo and a video share one CLIP embedding, Identity's
+    // own text embedding, so the photo gets the keyword Identity and the video
+    // does not.
+    it("keeps a video out of a category that leaves out videos", async () => {
+        const identity = DISCOVER_CATEGORIES.find(
+            (c) => c.title === "Identity",
+        )!;
+        expect(identity.showVideo).toBe(false);
+        const toast = file(102, 1, "toast.mp4");
+        const video = {
+            ...toast,
+            metadata: { ...toast.metadata, fileType: "video" as const },
+        };
+
+        // Serves beach.jpg and the video in the Vacation album, and gives
+        // each ML data holding only that embedding.
+        class VideoClient extends MockClient {
+            override async filesSince(args: {
+                collectionID: number;
+            }): Promise<FilesPage> {
+                const files =
+                    args.collectionID === 1
+                        ? [file(100, 1, "beach.jpg"), video]
+                        : [];
+                return { files, deleted: [], cursor: 1 };
+            }
+            async fetchMLData(args: { fileIDs: number[] }): Promise<{
+                payloads: Map<number, MLData>;
+                errors: Map<number, unknown>;
+            }> {
+                const payloads = new Map<number, MLData>();
+                for (const id of args.fileIDs) {
+                    payloads.set(id, {
+                        clip: { embedding: identity.textEmbedding },
+                    });
+                }
+                return { payloads, errors: new Map() };
+            }
+        }
+        const lib = await openLibrary(stubSource(), new VideoClient());
+        const outDir = join(root, "backup");
+
+        const result = await lib.backup({ downloadDirectory: outDir });
+
+        expect(result.failed).toBe(0);
+        expect(fileJSON(outDir, 100).keywords).toContain("Identity");
+        expect(fileJSON(outDir, 102)).toHaveProperty("keywords");
+        expect(fileJSON(outDir, 102).keywords).not.toContain("Identity");
         await lib.close();
     });
 
@@ -993,7 +1087,7 @@ describe("ML data in each file's JSON", () => {
         const result = await lib.backup({ downloadDirectory: outDir });
 
         // File 100's ML data was cached by the first run and is kept.
-        expect(fileJSON(outDir, 100).mlData).toEqual(ML_PAYLOAD);
+        expect(fileJSON(outDir, 100).mlData).toEqual({ face: { faces: [] } });
         expect(fileJSON(outDir, 100)).not.toHaveProperty("mlDataError");
         for (const fileID of [101, 200]) {
             expect(fileJSON(outDir, fileID).mlDataError).toBe(
@@ -1026,7 +1120,7 @@ describe("ML data in each file's JSON", () => {
 
         expect(result.failed).toBe(0);
         expect(result.skipped).toBe(3);
-        expect(fileJSON(outDir, 100).mlData).toEqual(ML_PAYLOAD);
+        expect(fileJSON(outDir, 100).mlData).toEqual({ face: { faces: [] } });
         for (const fileID of [100, 101, 200]) {
             expect(fileJSON(outDir, fileID)).not.toHaveProperty("mlDataError");
         }

test/cli/commands.test.ts

@@ -124,20 +124,35 @@ const FILES: Record<number, EnteFile[]> = {
 // thumbnail as empty. `withNewFile` adds new.jpg (102) to Vacation, advancing
 // the collection's updationTime as the server does. `beachInWork` adds
 // beach.jpg (100) to Work as well, so that one file is in both albums.
-// `refreshError` makes listing collections fail with that message.
+// `newAlbums` adds the albums A (3) and B (4), in that order, which both hold
+// shared.jpg (300). `refreshError` makes listing collections fail with that
+// message.
 const fakeClient = (
     opts: {
         failID?: number;
         emptyThumbID?: number;
         withNewFile?: boolean;
         beachInWork?: boolean;
+        newAlbums?: boolean;
         refreshError?: string;
     } = {},
 ) => {
-    const collections = opts.withNewFile
+    let collections = opts.withNewFile
         ? [{ ...COLLECTIONS[0], updationTime: 2 }, COLLECTIONS[1]]
         : COLLECTIONS;
     let files = FILES;
+    if (opts.newAlbums) {
+        collections = [
+            ...collections,
+            { ...collection(3, "A"), updationTime: 2 },
+            { ...collection(4, "B"), updationTime: 2 },
+        ];
+        files = {
+            ...files,
+            3: [file(300, 3, "shared.jpg")],
+            4: [file(300, 4, "shared.jpg")],
+        };
+    }
     if (opts.withNewFile) {
         files = {
             ...files,
@@ -535,9 +550,7 @@ describe("files", () => {
 
     it("exits 1 for an unknown collection", async () => {
         expect(await filesCommand(context(), "9", {})).toBe(1);
-        expect(stderr.text).toBe(
-            FETCHED_BOTH_ALBUMS + "Collection 9 not found\n",
-        );
+        expect(stderr.text).toBe("Collection 9 not found\n");
     });
 
     // Without a collection ID it lists every file in the account once (issue
@@ -571,6 +584,57 @@ describe("files", () => {
         ]);
         expect(stderr.text).toBe(FETCHED_BOTH_ALBUMS);
     });
+
+    // `files 2` on an empty cache stores Work alone. The refresh that `files`
+    // then waits for must still store Vacation first, as the server lists it.
+    it("without a collection ID lists in the server's order after files 2 on an empty cache", async () => {
+        const ctx = context(fakeClient({ beachInWork: true }));
+        expect(await filesCommand(ctx, "2", {})).toBe(0);
+        stdout.text = "";
+        stderr.text = "";
+
+        expect(await filesCommand(ctx, undefined, { json: true })).toBe(0);
+        expect(
+            JSON.parse(stdout.text).map(
+                (row: { id: number; collectionID: number }) => [
+                    row.id,
+                    row.collectionID,
+                ],
+            ),
+        ).toEqual([
+            [100, 1],
+            [101, 1],
+            [200, 2],
+        ]);
+        expect(stderr.text).toBe(FETCHED_BOTH_ALBUMS);
+    });
+
+    // After a refresh, the server adds A and then B. `files 4` must not store B
+    // ahead of A, so it waits for a refresh, which stores both in that order.
+    it("without a collection ID lists new albums in the server's order after files with the second one's ID", async () => {
+        expect(await filesCommand(context(), undefined, {})).toBe(0);
+        const ctx = context(fakeClient({ newAlbums: true }));
+        stdout.text = "";
+
+        expect(await filesCommand(ctx, "4", {})).toBe(0);
+        expect(stdout.text).toBe("300\timage\tshared.jpg\n");
+        stdout.text = "";
+
+        expect(await filesCommand(ctx, undefined, { json: true })).toBe(0);
+        expect(
+            JSON.parse(stdout.text).map(
+                (row: { id: number; collectionID: number }) => [
+                    row.id,
+                    row.collectionID,
+                ],
+            ),
+        ).toEqual([
+            [100, 1],
+            [101, 1],
+            [200, 2],
+            [300, 3],
+        ]);
+    });
 });
 
 describe("get and get-thumb", () => {
@@ -1394,7 +1458,7 @@ describe("backup-metadata --exif", () => {
 // collection list alone and stores nothing.)
 describe("backup-metadata and the thumbnail helpers refresh first", () => {
     beforeEach(async () => {
-        expect(await filesCommand(context(), "1", {})).toBe(0);
+        expect(await filesCommand(context(), undefined, {})).toBe(0);
         stdout.text = "";
         stderr.text = "";
     });
@@ -1579,11 +1643,13 @@ describe("an ID given on the command line", () => {
 
 // An account with many albums (issue #210). On an empty cache `collections`
 // used to wait for a refresh that fetched and decrypted the files of every
-// album, one album after another, before it printed anything, so on such an
-// account it seemed to hang. The account here has 40 albums of one photo each,
-// encrypted as Ente's clients encrypt them, on the fake Ente server of
-// `test/ente-server.ts`, which also records the path of each request quak
-// sends. The session is the one logging in to that server saves.
+// album, one album after another. It printed nothing until then, so on such an
+// account it seemed to hang. `files` with an album's ID did the same before it
+// listed that album (issue #217). The account here has 40 albums of one photo
+// each, encrypted as Ente's clients encrypt them. The fake Ente server of
+// `test/ente-server.ts` serves them. A wrapper around it records the path and
+// query of each request quak sends. The session is the one logging in to that
+// server saves.
 describe("an account with many albums", () => {
     const ALBUMS = 40;
     const PASSWORD = "many-albums";
@@ -1664,7 +1730,8 @@ describe("an account with many albums", () => {
                     : input instanceof URL
                       ? input.href
                       : input.url;
-            paths.push(new URL(url).pathname);
+            const { pathname, search } = new URL(url);
+            paths.push(pathname + search);
             return ente(input, init);
         }) as typeof globalThis.fetch;
         const client = await Client.login({
@@ -1698,42 +1765,68 @@ describe("an account with many albums", () => {
             albums += `${id}\talbum\tAlbum ${id}\n`;
         expect(stdout.text).toBe(albums);
         expect(stderr.text).toBe("");
-        expect(paths).toEqual(["/collections/v2"]);
+        expect(paths).toEqual(["/collections/v2?sinceTime=0"]);
 
-        // It stored nothing that would make a refresh skip an album's files:
-        // `files` after it still finds the album's photo.
+        // It stored nothing that would make `refreshCollection()` skip an
+        // album's files: `files` after it still finds the album's photo.
         stdout.text = "";
         expect(await filesCommand(ctx, "7", {})).toBe(0);
         expect(stdout.text).toBe("700\timage\tphoto-7.jpg\n");
     });
 
-    it("files counts on stderr each album whose files its refresh fetched", async () => {
+    it("files fetches the files of the album it lists and of no other", async () => {
         expect(await filesCommand(accountContext(), "7", {})).toBe(0);
 
         expect(stdout.text).toBe("700\timage\tphoto-7.jpg\n");
+        expect(stderr.text).toBe("");
+        expect(paths).toEqual([
+            "/collections/v2?sinceTime=0",
+            "/collections/v2/diff?collectionID=7&sinceTime=0",
+        ]);
+    });
+
+    // The first listing stored the album, so the second finds its stored copy
+    // current and asks for no album's files.
+    it("files run again for the same album sends no request for its files", async () => {
+        const ctx = accountContext();
+        expect(await filesCommand(ctx, "7", {})).toBe(0);
+        stdout.text = "";
+        paths.length = 0;
+
+        expect(await filesCommand(ctx, "7", {})).toBe(0);
+        expect(stdout.text).toBe("700\timage\tphoto-7.jpg\n");
+        expect(paths).toEqual(["/collections/v2?sinceTime=0"]);
+    });
+
+    it("files with no album's ID counts on stderr each album whose files its refresh fetched", async () => {
+        expect(await filesCommand(accountContext(), undefined, {})).toBe(0);
+
+        let photos = "";
         let progress = "";
         for (let n = 1; n <= ALBUMS; n++) {
+            photos += `${n * 100}\timage\tphoto-${n}.jpg\n`;
             progress += `Fetched the files of ${n} of ${ALBUMS} collections\n`;
         }
+        expect(stdout.text).toBe(photos);
         expect(stderr.text).toBe(progress);
     });
 
     // `--json` shapes stdout only, so the refresh's progress is still counted
     // on stderr.
-    it("files --json counts on stderr each album whose files its refresh fetched", async () => {
-        expect(await filesCommand(accountContext(), "7", { json: true })).toBe(
-            0,
-        );
+    it("files --json with no album's ID counts on stderr each album whose files its refresh fetched", async () => {
+        expect(
+            await filesCommand(accountContext(), undefined, { json: true }),
+        ).toBe(0);
 
-        expect(JSON.parse(stdout.text)).toEqual([
-            {
+        const rows = JSON.parse(stdout.text);
+        expect(rows).toHaveLength(ALBUMS);
+        expect(rows[6]).toEqual({
             id: 700,
             title: "photo-7.jpg",
             fileType: "image",
             creationTime: 1000,
             collectionID: 7,
-            },
-        ]);
+        });
         let progress = "";
         for (let n = 1; n <= ALBUMS; n++) {
             progress += `Fetched the files of ${n} of ${ALBUMS} collections\n`;

test/cli/log.test.ts

@@ -234,8 +234,8 @@ describe("-q", () => {
         });
     });
 
-    // The refresh's progress is left out; the line saying why the command
-    // failed is the fatal one, and it is printed.
+    // The line saying why the command failed is the fatal one, and it is
+    // printed.
     it("prints only the line a failing command exits with", async () => {
         expect(await quak("-q", "files", "9")).toEqual({
             code: 1,

test/cli/metadata-backup.test.ts

@@ -41,6 +41,7 @@ import { beforeAll, afterAll, describe, expect, it, vi } from "vitest";
 import { init, toBase64, encryptBlob } from "../../src/crypto/index.js";
 import * as jpegJs from "jpeg-js";
 import { Client } from "../../src/client.js";
+import { DISCOVER_CATEGORIES } from "../../src/clip-keyword-vectors.js";
 import { Library } from "../../src/library/index.js";
 import {
     runMetadataBackup,
@@ -56,8 +57,8 @@ import {
     type MetaMockState,
 } from "../ente-server.js";
 
-// One file per ML data request, so the two files of the mock account are
-// fetched in two requests and one of them can fail on its own.
+// One file per ML data request, so each file of the mock account is fetched in
+// a request of its own and one of them can fail on its own.
 vi.mock("../../src/mldata-fetch.js", async (importOriginal) => ({
     ...(await importOriginal<typeof import("../../src/mldata-fetch.js")>()),
     MLDATA_BATCH_SIZE: 1,
@@ -66,6 +67,21 @@ vi.mock("../../src/mldata-fetch.js", async (importOriginal) => ({
 const TEST_EMAIL = "metabackup@example.com";
 const TEST_PASSWORD = "metapass";
 
+// The CLIP embedding of a photo of a sunset (see clip-keywords.test.ts), which
+// the ML data of beach.jpg holds.
+const SUNSET_EMBEDDING: number[] = JSON.parse(
+    readFileSync(
+        new URL("../clip-keywords-sunset.json", import.meta.url),
+        "utf-8",
+    ),
+).embedding;
+
+// Identity, a Discover category of Ente's photos app that leaves out videos.
+const IDENTITY = DISCOVER_CATEGORIES.find((c) => c.title === "Identity")!;
+
+// ML data holding faces and no CLIP embedding, which file 302 has.
+const FACES_ONLY = { face: { faces: [{ faceID: "face-def", score: 0.9 }] } };
+
 let mock: MetaMockState;
 let testDir: string;
 
@@ -204,6 +220,65 @@ const buildMetaMock = async (): Promise<MetaMockState> => {
         updationTime: 1710000000000000,
     };
 
+    // A file in coll 10 with the ML data given. `fileType` is Ente's number: 0
+    // an image, 1 a video.
+    const vacationFile = (
+        id: number,
+        title: string,
+        fileType: number,
+        mlData: object,
+    ) => {
+        const fileKey = sodium.crypto_secretstream_xchacha20poly1305_keygen();
+        const { ciphertext: encKey, nonce } = encryptSecretbox(fileKey, ck1);
+        const meta = encryptStreamBlob(
+            new TextEncoder().encode(
+                JSON.stringify({
+                    title,
+                    fileType,
+                    creationTime: 1700000000000000,
+                    modificationTime: 1700000000000000,
+                }),
+            ),
+            fileKey,
+        );
+        const ml = encryptBlob(
+            new Uint8Array(gzipSync(Buffer.from(JSON.stringify(mlData)))),
+            fileKey,
+        );
+        return {
+            raw: {
+                id,
+                collectionID: 10,
+                ownerID: 42,
+                encryptedKey: toBase64(encKey),
+                keyDecryptionNonce: toBase64(nonce),
+                metadata: {
+                    encryptedData: toBase64(meta.ciphertext),
+                    decryptionHeader: toBase64(meta.header),
+                },
+                file: {
+                    decryptionHeader: toBase64(sodium.randombytes_buf(24)),
+                },
+                thumbnail: {
+                    decryptionHeader: toBase64(sodium.randombytes_buf(24)),
+                },
+                updationTime: 1700000000000000,
+            },
+            mlData: {
+                encryptedData: toBase64(ml.ciphertext),
+                decryptionHeader: toBase64(ml.header),
+            },
+        };
+    };
+    // Files 300 and 301 in coll 10: a photo and a video whose ML data holds
+    // only a CLIP embedding, Identity's own text embedding.
+    const identityML = { clip: { embedding: IDENTITY.textEmbedding } };
+    const identityPhoto = vacationFile(300, "passport.jpg", 0, identityML);
+    const identityVideo = vacationFile(301, "passport.mp4", 1, identityML);
+    // File 302 in coll 10: a photo whose ML data holds faces and no CLIP
+    // embedding.
+    const facesPhoto = vacationFile(302, "party.jpg", 0, FACES_ONLY);
+
     // Encrypt ML data for file 100 (gzipped JSON, encrypted with file key)
     const mlPayload = JSON.stringify({
         face: {
@@ -230,7 +305,7 @@ const buildMetaMock = async (): Promise<MetaMockState> => {
         clip: {
             version: 1,
             client: "test",
-            embedding: [0.5, 0.6, 0.7],
+            embedding: SUNSET_EMBEDDING,
         },
     });
     const gzipped = gzipSync(Buffer.from(mlPayload));
@@ -270,12 +345,23 @@ const buildMetaMock = async (): Promise<MetaMockState> => {
         keyAttributes: account.keyAttributes,
         encryptedToken: account.encryptedToken,
         collections: [rawColl1, rawColl2],
-        filesByCollection: { 10: [rawFile1], 20: [rawFile2] },
+        filesByCollection: {
+            10: [
+                rawFile1,
+                identityPhoto.raw,
+                identityVideo.raw,
+                facesPhoto.raw,
+            ],
+            20: [rawFile2],
+        },
         encryptedMLData: {
             100: {
                 encryptedData: toBase64(mlCiphertext),
                 decryptionHeader: toBase64(mlHeader),
             },
+            300: identityPhoto.mlData,
+            301: identityVideo.mlData,
+            302: facesPhoto.mlData,
         },
         fileCiphertexts: { 100: encFileBody1 },
     };
@@ -432,7 +518,9 @@ describe("quak backup-metadata", () => {
         expect(account.email).toBe(TEST_EMAIL);
     });
 
-    it("fetches and decrypts ML data by default", async () => {
+    // The CLIP embedding is left out of the ML data, and the keywords it gives
+    // are written in its place.
+    it("fetches and decrypts ML data by default, with keywords for the CLIP embedding", async () => {
         const outDir = join(testDir, "ml-data");
         await runBackup(outDir);
 
@@ -452,8 +540,41 @@ describe("quak backup-metadata", () => {
         expect(fileMeta.mlData.face.faces[0].faceID).toBe("face-abc");
         expect(fileMeta.mlData.face.faces[0].score).toBeCloseTo(0.98);
         expect(fileMeta.mlData.face.faces[0].detection.box.x).toBeCloseTo(0.1);
-        expect(fileMeta.mlData.clip).toBeDefined();
-        expect(fileMeta.mlData.clip.embedding).toEqual([0.5, 0.6, 0.7]);
+        expect(fileMeta.mlData).not.toHaveProperty("clip");
+        expect(fileMeta.keywords).toEqual(["Sunset"]);
+    });
+
+    // Ente's photos app keeps videos out of some Discover categories, Identity
+    // among them. Files 300 and 301, a photo and a video, share one CLIP
+    // embedding, Identity's own text embedding, so the photo gets the keyword
+    // Identity and the video does not.
+    it("keeps a video out of a category that leaves out videos", async () => {
+        expect(IDENTITY.showVideo).toBe(false);
+        const outDir = join(testDir, "video-keywords");
+        await runBackup(outDir);
+
+        const vacation = join(outDir, "collections", "10-Vacation");
+        const fileJSON = (fileID: number) =>
+            JSON.parse(readFileSync(join(vacation, `${fileID}.json`), "utf-8"));
+        expect(fileJSON(300).keywords).toContain("Identity");
+        expect(fileJSON(301)).toHaveProperty("keywords");
+        expect(fileJSON(301).keywords).not.toContain("Identity");
+    });
+
+    // A file's ML data can hold faces and no CLIP embedding. File 302's JSON
+    // then has the ML data and no keywords.
+    it("writes no keywords for ML data without a CLIP embedding", async () => {
+        const outDir = join(testDir, "no-clip");
+        await runBackup(outDir);
+
+        const fileMeta = JSON.parse(
+            readFileSync(
+                join(outDir, "collections", "10-Vacation", "302.json"),
+                "utf-8",
+            ),
+        );
+        expect(fileMeta.mlData).toEqual(FACES_ONLY);
+        expect(fileMeta).not.toHaveProperty("keywords");
     });
 
     it("extracts EXIF from downloaded files when --exif is set", async () => {
@@ -541,7 +662,7 @@ describe("quak backup-metadata when a file's ML data cannot be read", () => {
                 "utf-8",
             ),
         );
-        expect(ok.mlData.clip.embedding).toEqual([0.5, 0.6, 0.7]);
+        expect(ok.keywords).toEqual(["Sunset"]);
         expect(ok.mlDataError).toBeUndefined();
 
         const failed = JSON.parse(
@@ -599,7 +720,7 @@ describe("quak backup-metadata when a file's ML data cannot be read", () => {
                 "utf-8",
             ),
         );
-        expect(ok.mlData.clip.embedding).toEqual([0.5, 0.6, 0.7]);
+        expect(ok.keywords).toEqual(["Sunset"]);
     });
 
     it("exits 0 when every ML data request succeeds", async () => {

test/cli/read.test.ts

@@ -3,13 +3,18 @@
  * issue #36, issue #52).
  *
  * The `collections`, `files`, `get`, and `get-thumb` commands must answer for
- * current server state, not the local cache. `files`, `get` and `get-thumb`
- * force a `Library.fresh()` round-trip before they read; the stand-in library
- * below serves nothing until `fresh()` has been awaited, so a helper that read
- * without refreshing would come back empty and fail here. `collections` asks
- * the library for the collection list alone (`freshCollections()`, issue
- * #210), and must not wait for a refresh, which would fetch the files of every
- * collection.
+ * current server state, not the local cache. `files` without a collection ID,
+ * `get` and `get-thumb` force a `Library.fresh()` round-trip before they read.
+ * The stand-in library below serves nothing until `fresh()` has been awaited.
+ * A helper that read without refreshing would come back empty and fail here.
+ * `collections` asks the library for the collection list alone
+ * (`freshCollections()`, issue #210). It must not wait for a refresh, which
+ * would fetch the files of every collection. `files COLLECTION` reads that list
+ * too, then brings that one collection up to date (`refreshCollection()`,
+ * issue #217). The stand-in serves a collection's files once `fresh()` has run,
+ * or once `refreshCollection()` has resolved to `true` for that collection.
+ * When `refreshCollection()` resolves to `false`, `files COLLECTION` forces a
+ * refresh instead.
  *
  * `collections` and `files` also list in the library's enumeration order
  * (`listCollections`/`listFiles`) — the order the pre-library CLI printed — not
@@ -64,12 +69,18 @@ const file = (
 
 // A library that reveals its records only after `fresh()` has been awaited, and
 // serves them in the enumeration order it was given. `freshCollections()`
-// answers the collection list in that order without a refresh. `photos.byID`
-// returns a stand-in `Photo` carrying just the fileID the helper passes
-// through.
+// answers the collection list in that order without a refresh. When
+// `refreshCollection()` resolves to true, it reveals the files of that one
+// collection.
+// `photos.byID` returns a stand-in `Photo` carrying just the fileID the helper
+// passes through.
 class FakeLibrary implements FreshReadLibrary {
     freshCalls = 0;
     freshCollectionsCalls = 0;
+    refreshCollectionIDs: number[] = [];
+    // What `refreshCollection()` resolves to. With false it reveals nothing,
+    // as a library that has had a refresh does for a collection it lacks.
+    refreshCollectionHolds = true;
     private refreshed = false;
 
     constructor(
@@ -88,6 +99,11 @@ class FakeLibrary implements FreshReadLibrary {
         return this.collections;
     }
 
+    async refreshCollection(collection: Collection): Promise<boolean> {
+        this.refreshCollectionIDs.push(collection.id);
+        return this.refreshCollectionHolds;
+    }
+
     getCollection(id: number): Collection | undefined {
         if (!this.refreshed) return undefined;
         return this.collections.find((c) => c.id === id);
@@ -98,7 +114,11 @@ class FakeLibrary implements FreshReadLibrary {
     }
 
     listFiles(collectionID: number): EnteFile[] {
-        return this.refreshed
+        const current =
+            this.refreshed ||
+            (this.refreshCollectionHolds &&
+                this.refreshCollectionIDs.includes(collectionID));
+        return current
             ? this.files.filter((f) => f.collectionID === collectionID)
             : [];
     }
@@ -138,18 +158,25 @@ describe("CLI read helpers (issue #36 amendment, issue #52)", () => {
         expect(rows.map((c) => c.id)).not.toEqual(newestFirst);
     });
 
-    it("freshFiles refreshes first, lists in enumeration order, keeps raw fields", async () => {
+    it("freshFiles brings the collection alone up to date, lists in enumeration order, keeps raw fields", async () => {
         // Enumeration order by id 10, 11, 12; creationTimes ascending, so a
-        // newest-first sort would reverse them.
+        // newest-first sort would reverse them. Collection 2's file is not
+        // listed.
         const files = [
             file(10, 1, 1_700_000_000_000_000),
             file(11, 1, 1_700_000_000_000_001),
             file(12, 1, 1_700_000_000_000_002),
+            file(20, 2, 1_700_000_000_000_003),
         ];
-        const lib = new FakeLibrary([collection(1, 100)], files);
+        const lib = new FakeLibrary(
+            [collection(1, 100), collection(2, 100)],
+            files,
+        );
 
         const rows = await freshFiles(lib, 1);
-        expect(lib.freshCalls).toBe(1);
+        expect(lib.freshCollectionsCalls).toBe(1);
+        expect(lib.refreshCollectionIDs).toEqual([1]);
+        expect(lib.freshCalls).toBe(0);
         expect(rows?.map((f) => f.id)).toEqual([10, 11, 12]);
 
         // Field values come from raw metadata: microsecond creationTime and the
@@ -203,10 +230,25 @@ describe("CLI read helpers (issue #36 amendment, issue #52)", () => {
         ]);
     });
 
+    it("freshFiles forces a refresh when refreshCollection resolves to false", async () => {
+        const lib = new FakeLibrary(
+            [collection(1, 100)],
+            [file(10, 1, 1_700_000_000_000_000)],
+        );
+        lib.refreshCollectionHolds = false;
+
+        const rows = await freshFiles(lib, 1);
+        expect(lib.refreshCollectionIDs).toEqual([1]);
+        expect(lib.freshCalls).toBe(1);
+        expect(rows?.map((f) => f.id)).toEqual([10]);
+    });
+
     it("freshFiles returns undefined for an unknown collection", async () => {
         const lib = new FakeLibrary([collection(1, 100)], []);
         const rows = await freshFiles(lib, 999);
-        expect(lib.freshCalls).toBe(1);
+        expect(lib.freshCollectionsCalls).toBe(1);
+        expect(lib.refreshCollectionIDs).toEqual([]);
+        expect(lib.freshCalls).toBe(0);
         expect(rows).toBeUndefined();
     });
 

test/clip-keywords-sunset.json

@@ -0,0 +1,109 @@
+{
+    "photo": "https://upload.wikimedia.org/wikipedia/commons/d/db/Sunset_off_the_Massachusetts_coast_-_NOAA.jpg",
+    "sha256": "be62f3289c9bf7b75af68cd6978021d54e8a531d5d4cc493f6e9fbf5e1ab74d3",
+    "embedding": [
+        -0.0021123015, -0.04437615, -0.062354308, 0.006249992, 0.022141185,
+        0.011159019, 0.01227392, -0.00030532264, 0.03186437, 0.09196922,
+        0.039111234, -0.037084445, -0.032962628, -0.019820206, -0.013141825,
+        0.0047183544, 0.017137334, 0.046581306, 0.0047622235, 0.04982521,
+        0.023128154, 0.032080866, 0.037709303, 0.019241406, 0.0032100384,
+        -0.014275843, 0.049337693, 0.025953151, 0.014263995, -0.0062328726,
+        -0.040504098, -0.015083303, -0.008123875, -0.006233639, -0.05352683,
+        -0.0048083956, 0.0763486, -0.013155067, 0.025114346, 0.022016674,
+        0.0001588023, -0.010341572, -0.0037518279, -0.019181008, 0.04549337,
+        -0.0010725886, 0.004636866, 0.017177915, -0.0117317885, 0.0026097358,
+        0.030738998, 0.013158572, -0.08856403, -0.0121391015, 0.011477229,
+        -0.0033104012, 0.035524756, -0.018607879, -0.0071739843, -0.031906743,
+        0.0056713386, -0.0044195866, 0.00292471, 0.00061686034, -0.043075223,
+        0.09040427, 0.001522537, -0.011445652, 0.009609886, -0.0053201183,
+        -0.0057452377, 0.00972853, -0.0024397348, -0.021255083, 0.015893811,
+        -0.006020548, -0.09428023, 0.014277583, -0.017975299, -0.011878315,
+        0.011825521, 0.00014203975, -0.0020437292, -0.05305343, -0.024840152,
+        -0.011336872, -0.002796541, -0.00036171873, 0.004530499, 0.044486605,
+        0.021602718, -0.01292857, 0.024530916, 0.019559948, -0.039928272,
+        -0.0681174, 0.049101, 0.02461532, 0.0365531, -0.04070358, -0.011011939,
+        -0.0027188526, -0.020370876, 0.012904856, 0.043157287, -0.051658224,
+        -0.3726012, 0.019549074, -0.0499252, -0.04800559, -0.011493859,
+        0.02893256, 0.060341954, 0.0013929688, -0.051283866, -0.024286803,
+        0.044124287, -0.005995806, -0.062817656, 0.007369614, -0.021470634,
+        -0.028147545, 0.00411659, 0.04234773, 0.03787647, 0.008579682,
+        -0.038568925, -0.027498025, -0.028416378, 0.0022012168, 0.010906191,
+        -0.013745787, 0.018267838, 0.013531097, -0.061559845, 0.004682517,
+        -0.005993303, 0.0028113858, -0.0028286583, 0.0026029332, -0.06693403,
+        -0.0072381413, -0.02419584, 0.0037009937, -0.0011296495, -0.030310456,
+        0.0005999609, 0.009945354, 0.024915686, 0.038672045, 0.023009432,
+        -0.05093051, -0.09101409, -0.06686167, -0.06211481, -0.022600474,
+        0.0024902367, -0.0067171026, -0.016769778, -0.0032803628, 0.008623083,
+        -0.0016110836, 0.014925073, 0.033811048, 0.012829449, -0.052575592,
+        0.03728281, -0.0051201633, 0.021238064, -0.031086605, -0.0043402985,
+        0.25054494, -0.028396713, -0.010725523, -0.023445267, 0.021751683,
+        0.003130971, 0.012270548, -0.014836475, 0.014466885, 0.0021478906,
+        -0.088153616, 0.015356315, 0.039639276, -0.040283255, 0.014923924,
+        -0.029303031, -0.004475201, 0.047299225, -0.06726505, -0.012491676,
+        -0.002080582, 0.03506341, 0.04146718, 0.0074114986, -0.009582633,
+        0.006623749, -0.05454898, -0.0034387247, -0.014780821, 0.03698328,
+        -0.034441486, 0.22032379, -0.021792699, 0.008001907, 0.0341019,
+        0.0025784625, -0.0071689896, -0.014151225, -0.050547987, -0.030372301,
+        -0.03288388, -0.042987447, -0.0033656165, 0.021502906, -0.0030374376,
+        -0.024413, -0.03506603, 0.030891743, 0.0069994032, 0.004887366,
+        0.030707076, 0.0065348223, 0.0008240555, 0.016809713, -0.024151647,
+        -0.0040458064, 0.044567548, 0.03908212, -0.061782435, 0.0026002831,
+        0.02003242, -0.0115451, -0.045135964, -0.0663542, 0.005240975,
+        0.0010777109, -0.009276002, -0.011919081, 0.027654119, -0.013634551,
+        0.0129154315, -0.013195512, 0.040571272, -0.099290505, -0.04656693,
+        -0.20979854, 0.04181004, -0.06769166, -0.04121946, 0.018260444,
+        0.005617506, -0.15694222, 0.032472793, 0.014189598, 0.008345103,
+        -0.041118417, 0.0074527003, -0.011494104, 0.004906474, -0.0027100458,
+        0.0009994594, -0.041061398, -0.045274835, 0.015187793, 0.034275036,
+        -0.02813289, 0.02098019, -0.015043198, 0.03716667, 0.013542642,
+        -0.008104835, 0.014965486, -0.027978558, 0.021163663, -0.006942737,
+        -0.034969814, -0.038961593, 0.016802894, -0.012318175, -0.047821593,
+        -0.01933216, 0.0064432374, 0.0322358, -0.020437784, 0.02659447,
+        -0.03022022, -0.019443465, -0.094178624, -0.017009107, 0.03860671,
+        -0.00023695534, -0.033089854, -0.014677462, 0.012356399, 0.018040499,
+        -0.19360113, 0.011710817, -0.019871691, 0.020124774, 0.024948105,
+        -0.024573864, -0.028349072, -0.041731454, 0.032646254, 0.015098165,
+        0.0018447012, -0.01129152, -0.02481687, 0.036048, 0.20797811,
+        0.04317757, -0.017239755, 0.033431605, -0.011384422, 0.0099174455,
+        -0.19739223, -0.008901062, -0.056426458, 0.014591864, 0.0044474266,
+        0.008209326, -0.012241819, 0.042553645, -0.021552453, 0.008611734,
+        0.048944127, -0.013872049, 0.019179989, 0.008183264, -0.053223096,
+        -0.027327595, -0.0033542586, -0.012353858, -0.010941209, 0.010212546,
+        0.029714134, -0.011385593, -0.0013529479, 0.014505385, 0.01148663,
+        0.012613636, -0.0074430504, -0.014008101, -0.028146038, -0.042778917,
+        0.02143165, 0.016812252, -0.06286215, 0.02085532, -0.035128277,
+        0.0241282, 0.01397908, -0.034992814, 0.013153766, 0.0068236147,
+        0.013594184, -0.06859393, 0.051849406, 0.029675119, -0.0037699796,
+        -0.00013216575, 0.02916647, 0.031859048, -0.015006051, 0.0001439966,
+        0.0026584554, -0.0046948595, 0.039754897, 0.034873154, 0.0012870439,
+        -0.011516082, -0.074089095, 0.019241294, 0.00076195505, 0.0027334464,
+        -0.036206458, -0.0011109262, -0.048838835, -0.00204685, 0.01179991,
+        0.0068136845, 0.009852495, 0.017308988, 0.03159545, 0.058803353,
+        0.038116634, 0.008836143, -0.010460637, -0.010994099, -0.0012926942,
+        -0.048370842, 0.037696566, -0.08093239, -0.0041242186, -0.005866373,
+        0.02048607, -0.12549639, 0.011705739, -0.0021425558, 0.018078592,
+        -0.038228087, 0.004245971, -0.11560507, 0.01381203, -0.02734893,
+        0.029595908, 0.101082794, -0.042842254, -0.08397301, -0.1546237,
+        -0.037624564, 0.0069606556, -0.008020479, 0.0043268185, 0.016504597,
+        0.044554908, -0.05240209, 0.020519335, -0.050994243, -0.019861093,
+        0.022066997, -0.0188911, 0.012919696, -0.022872692, -0.012961374,
+        -0.020071112, 0.0074553066, 0.06402014, 0.04065211, 0.014139608,
+        0.06221546, -0.02574805, 0.010114545, 0.020619567, 0.032264773,
+        -0.009148828, 0.022849955, -0.016593235, -0.032456633, -0.006737303,
+        -0.0037493357, -0.065912865, -0.0024181998, -0.0037048983, -0.019304488,
+        0.009085037, 0.057529982, -0.0378338, 0.01792042, 0.01885012,
+        -0.0043600523, 0.017382368, -0.027792772, 0.006749118, -0.0053297766,
+        -0.011628587, 0.015736576, 0.024801968, 0.011543913, -0.01159873,
+        -0.000759976, 0.042183932, 0.05764386, -0.06422416, 0.010577374,
+        -0.02261249, 0.02282793, -0.05141075, -0.11287078, -0.020670144,
+        -0.020953743, 0.030885734, -0.012597732, -0.121472664, 0.028447151,
+        -0.0415664, -0.010061708, 0.06441338, 0.0271207, -0.0016693748,
+        -0.0045694243, 0.020342957, 0.030528478, 0.007129517, -0.0029600926,
+        -0.007061397, 0.06036486, -0.004736018, -0.0090786135, -0.07601116,
+        -0.039556272, -0.051269546, -0.019473685, 0.029858151, 0.028464556,
+        0.012424476, 0.013742254, 0.008742927, 0.006065751, -0.009892663,
+        0.07699999, -0.003443197, 0.0027863425, 0.029197995, 0.0211013,
+        0.01098379, 0.016684106, -0.028917195, 0.008122037, 0.0265241,
+        -0.04461223
+    ]
+}

test/clip-keywords.test.ts

@@ -0,0 +1,129 @@
+// `clipKeywords`: the keywords quak writes for a file in place of its CLIP
+// embedding.
+//
+// Ente's ML data for a file holds a CLIP embedding, 512 numbers its image
+// model gave the photo. Ente's photos app puts a file in a category of its
+// Discover section when the cosine similarity of that embedding and the text
+// embedding of the category's query is at least the category's minimum score,
+// and keeps videos out of some categories. `clipKeywords` applies that rule
+// and returns the titles of the categories a file falls in, highest score
+// first; `quak backup` and `quak backup-metadata` write them as `keywords`.
+
+import { readFileSync } from "node:fs";
+import { describe, expect, it } from "vitest";
+
+import { DISCOVER_CATEGORIES } from "../src/clip-keyword-vectors.js";
+import { clipKeywords } from "../src/clip-keywords.js";
+
+// 512 numbers: the ones given, then zeros.
+const vector = (...start: number[]): number[] => [
+    ...start,
+    ...new Array<number>(512 - start.length).fill(0),
+];
+
+const withEmbedding = (embedding: unknown) => ({
+    face: { faces: [] },
+    clip: { embedding },
+});
+
+// Two made-up categories, each with a text embedding along one axis. The
+// embedding [3, 4] has a cosine similarity of 3/5 with "Across" and 4/5 with
+// "Up". Videos are kept out of "Across".
+const ACROSS = {
+    title: "Across",
+    query: "across",
+    minScore: 0.6,
+    showVideo: false,
+    textEmbedding: vector(1),
+};
+const UP = {
+    title: "Up",
+    query: "up",
+    minScore: 0.5,
+    showVideo: true,
+    textEmbedding: vector(0, 1),
+};
+
+describe("clipKeywords", () => {
+    it("gives each category whose minimum score the similarity reaches, highest score first", () => {
+        expect(
+            clipKeywords(withEmbedding(vector(3, 4)), false, [ACROSS, UP]),
+        ).toEqual(["Up", "Across"]);
+    });
+
+    it("leaves out a category when the similarity is just below its minimum score", () => {
+        const stricter = { ...ACROSS, minScore: 0.6 + 1e-9 };
+
+        expect(
+            clipKeywords(withEmbedding(vector(3, 4)), false, [stricter, UP]),
+        ).toEqual(["Up"]);
+    });
+
+    it("keeps a video out of a category that leaves out videos", () => {
+        expect(
+            clipKeywords(withEmbedding(vector(3, 4)), true, [ACROSS, UP]),
+        ).toEqual(["Up"]);
+    });
+
+    it("gives an empty list for an embedding that falls in no category", () => {
+        expect(
+            clipKeywords(withEmbedding(vector(0, 0, 1)), false, [ACROSS, UP]),
+        ).toEqual([]);
+    });
+
+    it.each([
+        ["no CLIP embedding", { face: { faces: [] } }],
+        ["a CLIP part that is not an object", { clip: null }],
+        ["a CLIP part without an embedding", { clip: {} }],
+        ["an embedding of 511 numbers", withEmbedding(vector(3, 4).slice(1))],
+        ["an embedding with NaN in it", withEmbedding(vector(3, NaN))],
+        ["an embedding with Infinity in it", withEmbedding(vector(Infinity))],
+        ["an embedding of strings", withEmbedding(vector(3, 4).map(String))],
+        ["an embedding of zeros", withEmbedding(vector())],
+    ])("gives no keywords for ML data with %s", (_, mlData) => {
+        expect(clipKeywords(mlData, false, [ACROSS, UP])).toBeUndefined();
+    });
+});
+
+// The CLIP embedding Ente's image model gives a public-domain NOAA photo of a
+// sunset, made by bin/clip-keyword-vectors from the photo at `photo`.
+const SUNSET: { photo: string; sha256: string; embedding: number[] } =
+    JSON.parse(
+        readFileSync(
+            new URL("clip-keywords-sunset.json", import.meta.url),
+            "utf-8",
+        ),
+    );
+
+describe("Ente's Discover categories", () => {
+    it("are the 16 of Ente's photos app, each with a text embedding of norm 1", () => {
+        expect(DISCOVER_CATEGORIES.map((c) => c.title)).toEqual([
+            "Identity",
+            "Screenshots",
+            "QR Codes",
+            "Receipts",
+            "Notes",
+            "Memes",
+            "Visiting Cards",
+            "Babies",
+            "Pets",
+            "Selfies",
+            "Wallpapers",
+            "Food",
+            "Celebrations",
+            "Sunset",
+            "Hills",
+            "Greenery",
+        ]);
+        for (const category of DISCOVER_CATEGORIES) {
+            expect(category.textEmbedding).toHaveLength(512);
+            expect(Math.hypot(...category.textEmbedding)).toBeCloseTo(1, 5);
+        }
+    });
+
+    it("give a photo of a sunset the keyword Sunset and no other", () => {
+        expect(
+            clipKeywords({ clip: { embedding: SUNSET.embedding } }, false),
+        ).toEqual(["Sunset"]);
+    });
+});

test/library/fresh.test.ts

@@ -17,6 +17,14 @@
  *  4. `freshCollections()` reads the collection list alone, with one request
  *     and no collection's files, and stores nothing, so the next refresh still
  *     fetches the files of every collection (issue #210).
+ *  5. `refreshCollection()` brings one collection up to date in the store
+ *     (issue #217). It leaves the stored cursor where it was, so the next
+ *     refresh still fetches the files of every other collection that changed.
+ *     It never runs at the same time as a refresh. After a refresh, it stores
+ *     no collection the store does not hold: not a new one, and not one a
+ *     refresh deleted. `close()` waits for it. A library opened on a cache that
+ *     only it wrote awaits its first refresh. That refresh replaces what it
+ *     stored, in the server's order.
  *
  * The client is the same metadata-only mock the background-refresh tests use:
  * no crypto, no network, scripted pages, and a record of each call. A long
@@ -30,6 +38,7 @@ import { tmpdir } from "node:os";
 import { join } from "node:path";
 
 import { Library } from "../../src/library/index.js";
+import { MetadataStore } from "../../src/library/store.js";
 import type { CollectionsPage, FilesPage } from "../../src/client.js";
 import type { Collection, EnteFile } from "../../src/model/types.js";
 
@@ -144,6 +153,26 @@ const openSeeded = async (
     return { client, lib };
 };
 
+// Hold the client's next files request until `release()` is called. `started`
+// resolves once that request has been sent.
+const holdNextFilesRequest = (
+    client: MockClient,
+): { started: Promise<void>; release: () => void } => {
+    let release!: () => void;
+    const held = new Promise<void>((r) => (release = r));
+    let sent!: () => void;
+    const started = new Promise<void>((r) => (sent = r));
+    const realFilesSince = client.filesSince.bind(client);
+    client.filesSince = async (args) => {
+        client.filesSince = realFilesSince;
+        const page = await realFilesSince(args);
+        sent();
+        await held;
+        return page;
+    };
+    return { started, release };
+};
+
 describe("Library.fresh", () => {
     let dir: string;
     let cacheDirectory: string;
@@ -287,10 +316,10 @@ describe("Library.freshCollections", () => {
         rmSync(dir, { recursive: true, force: true });
     });
 
-    // A refresh fetches a collection's files only when the collection is new
-    // or its updationTime has advanced past the stored copy's. Had the
-    // collection list been stored, the next refresh would take every
-    // collection for one it already holds the files of, and fetch none.
+    // A refresh from a stored cursor past 0 fetches a collection's files only
+    // when the collection is new or its updationTime has advanced past the
+    // stored copy's. So does `refreshCollection()`. Neither would fetch the
+    // files of a collection stored here without them.
     it("asks for the collection list alone and stores nothing", async () => {
         const client = new MockClient();
         const page = {
@@ -340,3 +369,243 @@ describe("Library.freshCollections", () => {
         }
     });
 });
+
+describe("Library.refreshCollection", () => {
+    let dir: string;
+    let cacheDirectory: string;
+
+    beforeEach(() => {
+        dir = mkdtempSync(join(tmpdir(), "quak-refresh-collection-"));
+        cacheDirectory = join(dir, "cache");
+    });
+
+    afterEach(() => {
+        rmSync(dir, { recursive: true, force: true });
+    });
+
+    // Collections 1 and 2 are stored at updationTime 100, with a stored cursor
+    // of 100, and then both change on the server.
+    it("fetches the files of one changed collection and leaves the others to the next refresh", async () => {
+        const client = new MockClient();
+        client.collectionsQueue.push({
+            collections: [collection(1, 100), collection(2, 100)],
+            deleted: [],
+            cursor: 100,
+        });
+        client.filesFor(1, {
+            files: [file(1001, 1, 90)],
+            deleted: [],
+            cursor: 90,
+        });
+        client.filesFor(2, {
+            files: [file(2001, 2, 90)],
+            deleted: [],
+            cursor: 90,
+        });
+        const lib = await Library.open({
+            client,
+            cacheDirectory,
+            refreshIntervalSeconds: SLOW_INTERVAL,
+        });
+        try {
+            client.filesCalls.length = 0;
+            client.filesFor(1, {
+                files: [file(1002, 1, 110)],
+                deleted: [],
+                cursor: 110,
+            });
+            client.filesFor(2, {
+                files: [file(2002, 2, 120)],
+                deleted: [],
+                cursor: 120,
+            });
+
+            // Collection 1's files alone, since the stored copy.
+            expect(await lib.refreshCollection(collection(1, 110))).toBe(true);
+            expect(client.filesCalls).toEqual([
+                { collectionID: 1, sinceTime: 100 },
+            ]);
+            expect(lib.listFiles(1).map((f) => f.id)).toEqual([1001, 1002]);
+            expect(lib.listFiles(2).map((f) => f.id)).toEqual([2001]);
+
+            // The stored copy is now current, so no request.
+            await lib.refreshCollection(collection(1, 110));
+            expect(client.filesCalls).toHaveLength(1);
+
+            // The next refresh asks from the stored cursor, 100, so it still
+            // sees collection 2 changed and fetches its files.
+            client.collectionsQueue.push({
+                collections: [collection(1, 110), collection(2, 120)],
+                deleted: [],
+                cursor: 120,
+            });
+            await lib.fresh();
+            expect(client.collectionsSinceTimes).toEqual([0, 100]);
+            expect(client.filesCalls).toEqual([
+                { collectionID: 1, sinceTime: 100 },
+                { collectionID: 2, sinceTime: 100 },
+            ]);
+            expect(lib.listFiles(2).map((f) => f.id)).toEqual([2001, 2002]);
+        } finally {
+            await lib.close();
+        }
+    });
+
+    // Collection 1 is stored at 100. `refreshCollection()` asks for it at 110,
+    // which adds file 1002. While that files request waits, a fresh read
+    // starts a refresh. The server tells that refresh that collection 1 was
+    // deleted at 130.
+    it("holds off a refresh that starts during the files request, which then deletes the collection", async () => {
+        const { client, lib } = await openSeeded(cacheDirectory);
+        const files = holdNextFilesRequest(client);
+        try {
+            client.filesFor(1, {
+                files: [file(1002, 1, 110)],
+                deleted: [],
+                cursor: 110,
+            });
+            client.collectionsQueue.push({
+                collections: [],
+                deleted: [1],
+                cursor: 130,
+            });
+
+            const refreshing = lib.refreshCollection(collection(1, 110));
+            await files.started;
+            const reading = lib.fresh();
+            // The refresh has not sent its collections request yet.
+            await new Promise((r) => setImmediate(r));
+            expect(client.collectionsSinceTimes).toEqual([0]);
+
+            files.release();
+            await refreshing;
+            await reading;
+            expect(client.collectionsSinceTimes).toEqual([0, 100]);
+            expect(lib.getCollection(1)).toBeUndefined();
+            expect(lib.listFiles(1)).toEqual([]);
+        } finally {
+            files.release();
+            await lib.close();
+        }
+    });
+
+    // `freshCollections()` lists collection 1 at 110. Before
+    // `refreshCollection()` is called, a refresh learns that it was deleted at
+    // 130.
+    it("stores nothing for a collection that a refresh deleted before the call", async () => {
+        const { client, lib } = await openSeeded(cacheDirectory);
+        try {
+            client.collectionsQueue.push(
+                { collections: [collection(1, 110)], deleted: [], cursor: 110 },
+                { collections: [], deleted: [1], cursor: 130 },
+            );
+            const [listed] = await lib.freshCollections();
+            await lib.fresh();
+            client.filesCalls.length = 0;
+
+            expect(await lib.refreshCollection(listed)).toBe(false);
+            expect(client.filesCalls).toEqual([]);
+            expect(lib.getCollection(1)).toBeUndefined();
+        } finally {
+            await lib.close();
+        }
+    });
+
+    // A refresh has stored collection 1. Collection 2 is new on the server
+    // since then. Only the next refresh may add it, after collection 1.
+    it("stores nothing for a collection that is new since the last refresh", async () => {
+        const { client, lib } = await openSeeded(cacheDirectory);
+        try {
+            client.filesCalls.length = 0;
+
+            expect(await lib.refreshCollection(collection(2, 110))).toBe(false);
+            expect(client.filesCalls).toEqual([]);
+            expect(lib.getCollection(2)).toBeUndefined();
+        } finally {
+            await lib.close();
+        }
+    });
+
+    // The caller may remove the cache directory once `close()` resolves, so
+    // the cache write must have landed by then.
+    it("close() waits for a refreshCollection() in flight", async () => {
+        const { client, lib } = await openSeeded(cacheDirectory);
+        const files = holdNextFilesRequest(client);
+        client.filesFor(1, {
+            files: [file(1002, 1, 110)],
+            deleted: [],
+            cursor: 110,
+        });
+
+        const refreshing = lib.refreshCollection(collection(1, 110));
+        await files.started;
+        let closed = false;
+        const closing = lib.close().then(() => {
+            closed = true;
+        });
+        await new Promise((r) => setImmediate(r));
+        expect(closed).toBe(false);
+
+        files.release();
+        await closing;
+        const saved = await MetadataStore.load(
+            join(cacheDirectory, "metadata.json"),
+        );
+        expect(saved.listFiles(1).map((f) => f.id)).toEqual([1001, 1002]);
+        await refreshing;
+    });
+
+    // `quak files 2` on an empty cache stores collection 2 alone. A library
+    // opened later on that cache must not serve it as the whole account. Its
+    // first refresh replaces what the cache holds, in the server's order.
+    it("open() awaits the first refresh on a cache that only refreshCollection() wrote", async () => {
+        const client = new MockClient();
+        client.filesFor(2, {
+            files: [file(2001, 2, 90)],
+            deleted: [],
+            cursor: 90,
+        });
+        const first = await Library.open({
+            client,
+            cacheDirectory,
+            refreshOnOpen: false,
+            refreshIntervalSeconds: SLOW_INTERVAL,
+        });
+        expect(await first.refreshCollection(collection(2, 100))).toBe(true);
+        await first.close();
+
+        client.collectionsQueue.push({
+            collections: [collection(1, 100), collection(2, 100)],
+            deleted: [],
+            cursor: 100,
+        });
+        client.filesFor(1, {
+            files: [file(1001, 1, 90)],
+            deleted: [],
+            cursor: 90,
+        });
+        client.filesFor(2, {
+            files: [file(2001, 2, 90)],
+            deleted: [],
+            cursor: 90,
+        });
+        client.filesCalls.length = 0;
+        const lib = await Library.open({
+            client,
+            cacheDirectory,
+            refreshIntervalSeconds: SLOW_INTERVAL,
+        });
+        try {
+            expect(lib.status().lastRefreshAt).toBeDefined();
+            expect(lib.listCollections().map((c) => c.id)).toEqual([1, 2]);
+            // It fetches collection 2's files again, from the beginning.
+            expect(client.filesCalls).toEqual([
+                { collectionID: 1, sinceTime: 0 },
+                { collectionID: 2, sinceTime: 0 },
+            ]);
+            expect(lib.listFiles(2).map((f) => f.id)).toEqual([2001]);
+        } finally {
+            await lib.close();
+        }
+    });
+});

test/library/library.test.ts

@@ -20,9 +20,9 @@
  * 5. `close()` stops the timer and is idempotent, and its promise resolves
  *    only once an in-flight refresh has written the cache file.
  * 6. `cacheDirectory` defaults to the env-paths cache dir plus the user id.
- * 7. `open()` branches on the cache: an empty cache awaits the first refresh
- *    (it has nothing to serve yet); an existing cache serves its copy at once
- *    and refreshes in the background, so a slow or dead server never stalls
+ * 7. `open()` branches on the stored cursor. At 0, as on an empty cache, it
+ *    awaits the first refresh. Past 0 it serves the cache's copy at once and
+ *    refreshes in the background, so a slow or dead server never stalls
  *    opening.
  * 8. A save failure that leaves RAM ahead of disk keeps `status().lastError`
  *    set and keeps retrying the write; a later empty refresh does not clear it.

test/library/snapshot.test.ts

@@ -4,7 +4,8 @@
  * These are the surface the GUI consumes across Electron IPC. `snapshot()` is
  * synchronous — it reads the in-RAM store and projects it into plain records
  * (no keys) — and `subscribe({ onChange })` delivers a `LibraryChange` whenever
- * a background refresh actually changes the derived records. The contracts:
+ * a refresh or `refreshCollection()` changes the derived records. The
+ * contracts:
  *
  *  1. `snapshot()` deduplicates a file across memberships into one record with
  *     every album id, orders photos newest first, and carries no key material.

test/library/store.test.ts

@@ -172,7 +172,7 @@ describe("MetadataStore", () => {
         expect(store.listCollections()).toEqual([]);
     });
 
-    it("loads an empty store when the list of what the refresh could not decrypt is damaged", async () => {
+    it("loads an empty store when the list of what did not decrypt is damaged", async () => {
         // A backup reads that list on every run, so a damaged one that loaded
         // would fail every backup until the cache was deleted. A string is
         // damaged too, though JavaScript can loop over its characters.
@@ -266,4 +266,23 @@ describe("MetadataStore", () => {
         expect(store.getFile(12345, 67890)).toBeUndefined();
         expect(store.getFile(55555, 67890)?.key).toEqual(inB.key);
     });
+
+    // A backup reports these in this order. It must not depend on whether a
+    // refresh or `Library.refreshCollection()` fetched each one first.
+    it("lists what did not decrypt by collection ID, then file ID", async () => {
+        const store = await MetadataStore.load(path);
+        store.putUndecryptable({ collectionID: 2, fileID: 7, error: "a" });
+        store.putUndecryptable({ collectionID: 3, error: "b" });
+        store.putUndecryptable({ collectionID: 1, fileID: 9, error: "c" });
+        store.putUndecryptable({ collectionID: 2, fileID: 5, error: "d" });
+
+        expect(
+            store.listUndecryptable().map((r) => [r.collectionID, r.fileID]),
+        ).toEqual([
+            [1, 9],
+            [2, 5],
+            [2, 7],
+            [3, undefined],
+        ]);
+    });
 });