ML data that does not decrypt is recorded with its reason (closes #198)

`fetchMLDataBatch` dropped an entry that did not decrypt, gunzip or parse,
so the file looked like one with no ML data and the run exited 0. It now
returns the reason beside the payloads, also for ML data that is not a
JSON object. The library gives the reasons from its last ML data fetch
that succeeded in `status().mlDataErrors` and the `fetchMLData` "done"
event; a fetch that fails leaves them as they were. `quak backup` writes
the reason as `mlDataError` and records the file in `failures.json`;
`backup-metadata` writes `mlDataError` and exits 1.

Judgement call: a file whose ML data is cached keeps it in its JSON when a
newer copy does not decode, as when the fetch fails.

Model: opus-5-5

README.md

@@ -159,10 +159,11 @@ After a `1`, the next run fetches each missing original again, fetches the ML
 data that is not cached, and rebuilds the album links. An album quak could not
 decrypt, or a file in `failures.json` because quak could not decrypt it, keeps
 the exit code at `1` until the server sends a copy that decrypts, or deletes it
-(see "Backup layout"). An original that a `--verify` run could not read, or put
-back with bytes that still do not match, stays at its save path. Only a later
-`--verify` run checks it again: a run without `--verify` leaves it as it is and
-takes the file out of `failures.json`, so that run can exit 0.
+(see "Backup layout"). An original that a `--verify` run could not read, or
+could not replace with a download that matches, stays at its save path as it
+was. Only a later `--verify` run checks it again: a run without `--verify`
+leaves it as it is and takes the file out of `failures.json`, so that run can
+exit 0.
 
 ## Examples
 
@@ -692,13 +693,15 @@ reads no tag from is recorded, base64, as `exifRaw`, with the reason in
 and `helper fix-missing-thumbnails` take `--json` for machine-readable output.
 
 `backup --verify` also hashes the originals already in the backup and replaces
-any that do not match the content hash Ente records; one it cannot replace goes
-into `failures.json` (see "Backup layout").
+any that do not match the content hash Ente records; one it cannot replace stays
+as it was and goes into `failures.json` (see "Backup layout").
 
 `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
-`mlDataError` field instead of `mlData`, and the dump goes on. The exit code is
-non-zero if any ML data request failed.
+`mlDataError` field instead of `mlData`, and the dump goes on. A file whose ML
+data does not decrypt, gunzip or parse is logged and written the same way, with
+why in `mlDataError`. The exit code is non-zero if any file was written with
+`mlDataError`.
 
 `helper fix-missing-thumbnails` regenerates thumbnails for JPEG images only,
 because the bundled decoder (`jpeg-js`) decodes only JPEG. A non-JPEG image
@@ -763,8 +766,15 @@ A file's JSON holds Ente's ML data for it (its faces and its CLIP embedding) as
 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 JSON files are rewritten on every run, so
-ML data that arrived since the last run appears.
+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`
@@ -878,33 +888,41 @@ original already at its save path the way a download is checked (see "On-disk
 cache layout" below): its bytes, read in chunks, or a live photo's image and
 video, joined as `<imageHash>:<videoHash>`. An original that matches the content
 hash its metadata records is left as it is. One that does not is logged on one
-line naming the file, deleted (a live photo's image and video both), and put
-back in the same run like a missing one: downloaded, or copied from the cache if
-the cache holds it. What is put back is hashed too, because a copy from the
-cache is not checked as a download is. If it still does not match, it stays at
-its save path and the file goes into `failures.json`, as it does when the
-download fails. A file whose metadata records no hash is left as it is and
-counted as unchecked. A stored original that cannot be read is left as it is and
-counts as failed. The summary and `--json` add the counts `verified`,
-`mismatched` and `unchecked`, all of originals that were already stored; one
-first downloaded in this run is in none of them. A mismatch that was put back
-with matching bytes does not make the exit code non-zero. Without `--verify`
-nothing is hashed, the summary is unchanged, and the three counts are 0 in
-`--json`.
+line naming the file, and replaced once a matching copy is downloaded, in the
+same run. The copy is downloaded even when the cache holds the original, because
+a copy from the cache is not checked as a download is. It is written to a
+temporary file beside the original, a live photo's image and video each to its
+own, hashed the same way, and renamed over the original only when it matches; a
+live photo's image and video are both hashed before either is renamed. Until
+then the original stays as it was: when the download fails, or what it brings
+does not match either, the temporary files are deleted, the original keeps its
+bytes and its album links, and the file goes into `failures.json`. A run that
+stops between renaming a live photo's new image and its new video leaves the new
+image beside the old video, which a later `--verify` run replaces if the video
+was the part that did not match. A file whose metadata records no hash is left
+as it is and counted as unchecked. A stored original that cannot be read is left
+as it is and counts as failed. The summary and `--json` add the counts
+`verified`, `mismatched` and `unchecked`, all of originals that were already
+stored; one first downloaded in this run is in none of them. A mismatch that was
+replaced with matching bytes does not make the exit code non-zero. Without
+`--verify` nothing is hashed, the summary is unchanged, and the three counts are
+0 in `--json`.
 
 Each original is written to a temporary file in the same directory, synced to
 disk, and renamed into place, so an original is either complete or absent, even
 after a power cut. A downloaded original's temporary file is named
 `.quak-<pid>-<random>.tmp`, one copied from the cache
-`.quak-backup-YYYY-MM-DD.<fileID>.<ext>-<pid>-<random>.tmp`. A run that is
-killed can leave one of these temporary files behind; the next backup deletes
-those whose process is no longer running. The content cache uses the same
-scheme, and opening a library deletes the temporary files in the cache whose
-process is no longer running, so a download another process has in progress in
-the same cache is left alone. The rename replaces whatever was at the
-destination rather than writing through it: a symlink there is replaced, not
-followed, and the new file has the temporary file's permissions, not those of
-the file it replaced.
+`.quak-backup-YYYY-MM-DD.<fileID>.<ext>-<pid>-<random>.tmp`. A copy that
+`--verify` downloads to replace an original is written to such a
+`.quak-<pid>-<random>.tmp` too, a live photo's image and video to
+`.quak-<pid>-<random>.<ext>`, each with its own extension. A run that is killed
+can leave one of these temporary files behind; the next backup deletes those
+whose process is no longer running. The content cache uses the same scheme, and
+opening a library deletes the temporary files in the cache whose process is no
+longer running, so a download another process has in progress in the same cache
+is left alone. The rename replaces whatever was at the destination rather than
+writing through it: a symlink there is replaced, not followed, and the new file
+has the temporary file's permissions, not those of the file it replaced.
 
 ## TODO
 
@@ -978,6 +996,14 @@ 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().mlDataErrors` lists, as `{ fileID, error }`, each file whose ML
+data the last ML data fetch that succeeded got but could not decrypt, gunzip or
+parse; the `fetchMLData` event with status `"done"` carries the same list. A
+fetch that is still running leaves the list as it is until it ends, and a fetch
+that fails leaves it as it was. Such ML data is not cached, so the next ML data
+fetch asks for it again. The reasons are kept in memory only, never in the
+cache.
+
 ### Default reads vs. fresh reads
 
 Default reads — `lib.albums`, `lib.photos`, `lib.timeline` — answer

TODO.md

@@ -17,17 +17,38 @@ pre-1.0
 
 # Next Step
 
-Three bugs are open: https://git.eeqj.de/sneak/quak/issues/198,
-https://git.eeqj.de/sneak/quak/issues/199 and
-https://git.eeqj.de/sneak/quak/issues/205. The check of quak against a live Ente
-account, https://git.eeqj.de/sneak/quak/issues/187, needs an account only sneak
-has, so it waits on him.
+One bug is open: https://git.eeqj.de/sneak/quak/issues/205. The check of quak
+against a live Ente account, https://git.eeqj.de/sneak/quak/issues/187, needs an
+account only sneak has, so it waits on him.
 
 Tagging and releases are decided by sneak alone, and happen only when he
 declares one.
 
 # Completed Steps
 
+- 2026-10-16: ML data the server sends that does not decrypt, gunzip or parse is
+  no longer taken for no ML data (issue 198). `fetchMLDataBatch` returns why for
+  each such file beside the payloads; ML data that parses to something other
+  than a JSON object, such as `null`, is one of them. The library keeps the
+  reasons from its last ML data fetch that succeeded in memory, not in the
+  cache, and gives them in `status().mlDataErrors` and in the `fetchMLData`
+  event with status `"done"`; a fetch still running or one that fails does not
+  change them. `quak backup` writes the reason as `mlDataError` in the file's
+  JSON when no ML data is cached for it, and records the file in `failures.json`
+  with an error starting `ML data: `; `backup-metadata` writes `mlDataError`
+  too. Both exit 1. Tests serve ML data encrypted under a key that is not the
+  file's.
+
+- 2026-10-16: With `--verify`, an original whose hash does not match is no
+  longer removed before its replacement is downloaded (issue 199).
+  `replaceOriginal` in `src/library/content.ts` downloads the replacement to a
+  temporary file beside it, a live photo's image and video each to its own,
+  hashes it and renames it over the original only when it matches. When the
+  download fails or does not match either, the original keeps its bytes and
+  links, and the file goes into `failures.json`. The replacement is always
+  downloaded, through the new `fetchOriginal` of the content cache, never copied
+  from the cache.
+
 - 2026-10-16: The `~/bin/quak` that `make install` builds retries a refused
   connection and a name that does not resolve, and the backup records either as
   transient (issue 206). Bun's `fetch` reports them with codes of its own,

src/backup.ts

@@ -41,8 +41,9 @@
 //
 // With `verify`, each original already at its save path is hashed as the
 // download check hashes it, and one that does not match the content hash its
-// metadata records is removed and fetched again in the same run. What is put
-// back is hashed too, and recorded as failed if it still does not match.
+// metadata records is replaced once a matching copy is downloaded, in the same
+// run. Until then it stays as it is: a download that fails, or that does not
+// match either, leaves it in place and is recorded as failed.
 //
 // Resilience (issue #8): no per-file condition aborts the run. A failed
 // download, a failed symlink, or ML data missing because the ML data fetch
@@ -61,7 +62,6 @@
 // code.
 
 import {
-    createReadStream,
     lstatSync,
     mkdirSync,
     readdirSync,
@@ -77,12 +77,6 @@ import { readFile } from "node:fs/promises";
 import { dirname, extname, join, relative, resolve } from "node:path";
 import lockfile from "proper-lockfile";
 
-import {
-    chunkHashFinal,
-    chunkHashInit,
-    chunkHashUpdate,
-    init,
-} from "./crypto/index.js";
 import { removeLeftoverTempFiles } from "./download/index.js";
 import { describeError } from "./errors.js";
 import {
@@ -94,8 +88,10 @@ import {
 import {
     copyAtomic,
     placeOriginal,
+    replaceOriginal,
     savePath,
     storedAtSavePath,
+    storedHash,
 } from "./library/content.js";
 import { representative } from "./library/records.js";
 import type { Undecryptable } from "./library/store.js";
@@ -152,9 +148,9 @@ export interface BackupResult {
     // Originals already at their save path and left untouched.
     skipped: number;
     // With `verify`, the originals already at their save path whose hash
-    // matched, those whose hash did not (each removed and fetched again), and
-    // those whose metadata records no hash (left as they are). All three are
-    // zero without `verify`.
+    // matched, those whose hash did not (each replaced once a matching copy is
+    // downloaded), and those whose metadata records no hash (left as they
+    // are). All three are zero without `verify`.
     verified: number;
     mismatched: number;
     unchecked: number;
@@ -192,6 +188,15 @@ export interface BackupLibrary {
         fileID: number,
         destination: string,
     ): Promise<{ path: string; videoPath?: string }>;
+    // Fetch an original's bytes to `destination` through the content pool,
+    // a live photo as its image and its video beside it, even when the cache
+    // or the library's save path already holds it, and return where they
+    // landed. With `verify`, an original whose hash does not match is replaced
+    // with what this fetches.
+    fetchOriginal(
+        fileID: number,
+        destination: string,
+    ): Promise<{ path: string; videoPath?: string }>;
     thumbnail(fileID: number): Promise<{ path: string }>;
     // How many originals the backup works on at once: the limit of the
     // library's content pool.
@@ -203,6 +208,11 @@ export interface BackupLibrary {
     // A file's cached ML data, as `lib.mldata.forFile()` returns it, or
     // undefined when none is cached.
     mlData(fileID: number): Promise<MLData | undefined>;
+    // Why the ML data the last ML data fetch that succeeded got for a file did
+    // not decrypt, gunzip or parse, or undefined when it did or the fetch got
+    // none. A fetch still running does not change it. A fetch that fails
+    // leaves it as it was.
+    mlDataError(fileID: number): unknown;
 }
 
 type FailureClass = "transient" | "permanent" | "unknown";
@@ -444,26 +454,6 @@ const saveLedger = (path: string, ledger: Map<number, FailureEntry>): void => {
     );
 };
 
-// The content hash of the original stored at `stored`, computed as the download
-// check computes it: over each file's bytes, read in chunks, and for a live
-// photo `<imageHash>:<videoHash>`.
-const storedHash = async (stored: {
-    path: string;
-    videoPath?: string;
-}): Promise<string> => {
-    await init();
-    const hashFile = async (path: string): Promise<string> => {
-        const state = chunkHashInit();
-        for await (const chunk of createReadStream(path)) {
-            chunkHashUpdate(state, chunk as Buffer);
-        }
-        return chunkHashFinal(state);
-    };
-    const hash = await hashFile(stored.path);
-    if (stored.videoPath === undefined) return hash;
-    return `${hash}:${await hashFile(stored.videoPath)}`;
-};
-
 // A file's EXIF, XMP and dimensions as its JSON holds them: what
 // `extractImageMetadata` found in its original, or why the original could not
 // be read.
@@ -685,11 +675,13 @@ const runLockedBackup = async (
     // through the content cache/pools, as `Photo.download()` does, and fetch
     // the optional thumbnails; a present file is left as is. With `verify`, a
     // present original is hashed first, and one that does not match the hash
-    // its metadata records is removed and fetched like a missing one. One
-    // that cannot be read is recorded as failed and left where it is.
-    // `storeOriginal` does this for one original, and records its failure.
+    // its metadata records is downloaded again and replaced by
+    // `replaceOriginal`, which leaves it as it was when that download fails
+    // or does not match either. One that cannot be read is recorded as failed
+    // and left where it is. `storeOriginal` does this for one original, and
+    // records its failure.
     const storeOriginal = async (file: EnteFile): Promise<void> => {
-        let stored = storedAtSavePath(downloadDirectory, file);
+        const stored = storedAtSavePath(downloadDirectory, file);
         let mismatch = false;
         if (stored !== undefined && verify) {
             try {
@@ -703,11 +695,6 @@ const runLockedBackup = async (
                     );
                     mismatched++;
                     mismatch = true;
-                    rmSync(stored.path);
-                    if (stored.videoPath !== undefined) {
-                        rmSync(stored.videoPath);
-                    }
-                    stored = undefined;
                 }
             } catch (err) {
                 log(
@@ -722,29 +709,25 @@ const runLockedBackup = async (
                 return;
             }
         }
-        if (stored !== undefined) {
+        if (stored !== undefined && !mismatch) {
             skipped++;
             return;
         }
         try {
             log(`Fetching original ${file.metadata.title} (${file.id})...`);
-            // A fetched original is written straight to its save path (a
-            // live photo beside it); only one that was already cached
-            // elsewhere is copied.
-            const placed = await placeOriginal(
-                downloadDirectory,
-                file,
-                (dest) => lib.original(file.id, dest),
-            );
-            storedThisRun.add(file.id);
-            // A copy from the cache is not checked as a download is and can
-            // hold the same bad bytes, so what is put back after a mismatch
-            // is hashed too. A bad copy stays where it is and the file fails.
-            if (mismatch && (await storedHash(placed)) !== file.metadata.hash) {
-                throw new Error(
-                    "the original put back does not match its content hash either",
+            if (mismatch) {
+                await replaceOriginal(downloadDirectory, file, (dest) =>
+                    lib.fetchOriginal(file.id, dest),
+                );
+            } else {
+                // A fetched original is written straight to its save path (a
+                // live photo beside it); only one that was already cached
+                // elsewhere is copied.
+                await placeOriginal(downloadDirectory, file, (dest) =>
+                    lib.original(file.id, dest),
                 );
             }
+            storedThisRun.add(file.id);
             downloaded++;
         } catch (err) {
             log(
@@ -813,19 +796,18 @@ const runLockedBackup = async (
     // Phase 2: rebuild the derived views from the model. Sidecars first, for
     // every present original (this repairs stale ones), each with the file's
     // ML data once an ML data fetch has completed, and its image metadata.
-    // When the fetch fails, a file with no cached ML data gets the reason
-    // instead and is recorded as failed. The next run fetches its ML data
-    // again because none is cached.
+    // When the fetch got a file's ML data but could not decrypt, gunzip or
+    // parse it, or when the fetch failed, a file with no cached ML data gets
+    // the reason instead and is recorded as failed. The next run fetches its
+    // ML data again because none is cached.
     if (includeOriginals) {
         let mlDataFailure: unknown;
-        let mlDataError: string | undefined;
         try {
             log("Fetching ML data...");
             await lib.fetchMLData();
         } catch (err) {
             mlDataFailure = err;
-            mlDataError = describeError(err);
-            log(`FAILED ML data: ${mlDataError}`);
+            log(`FAILED ML data: ${describeError(err)}`);
         }
         for (const file of distinct.values()) {
             if (lock.lost) throw lock.lost;
@@ -842,13 +824,19 @@ const runLockedBackup = async (
                 storedThisRun.has(file.id),
             );
             const mlData = await lib.mlData(file.id);
-            if (mlData === undefined && mlDataError !== undefined) {
-                writeSidecar(path, file, { mlDataError }, image);
+            const failure = lib.mlDataError(file.id) ?? mlDataFailure;
+            if (mlData === undefined && failure !== undefined) {
+                writeSidecar(
+                    path,
+                    file,
+                    { mlDataError: describeError(failure) },
+                    image,
+                );
                 recordFailure(
                     file.id,
                     file.metadata.title,
                     collectionName.get(file.collectionID) ?? "",
-                    new Error("ML data", { cause: mlDataFailure }),
+                    new Error("ML data", { cause: failure }),
                 );
             } else {
                 writeSidecar(path, file, { mlData }, image);

src/cli-commands.ts

@@ -394,11 +394,11 @@ export const backupMetadataCommand = async (
         // Refresh first so the dump holds current account state, not what the
         // cache last held; a failed refresh throws.
         await lib.fresh();
-        const { failedMLBatches } = await runMetadataBackup(lib, client, dir, {
+        const result = await runMetadataBackup(lib, client, dir, {
             exif: opts.exif || opts.all,
             onProgress: (msg) => ctx.stderr.write(msg + "\n"),
         });
-        return failedMLBatches > 0 ? 1 : 0;
+        return result.filesWithMLDataError > 0 ? 1 : 0;
     } finally {
         await lib.close();
     }

src/client.ts

@@ -373,13 +373,17 @@ export class Client {
     }
 
     // Fetch machine-learning data (face detections + CLIP embeddings) for up
-    // to a batch of files, each decrypted with its own key. One request; the
-    // library batches at `MLDATA_BATCH_SIZE` and schedules each batch through
-    // its metadata request pool.
+    // to a batch of files, each decrypted with its own key, and for each file
+    // whose ML data did not decrypt, gunzip or parse, why not. One request;
+    // the library batches at `MLDATA_BATCH_SIZE` and schedules each batch
+    // through its metadata request pool.
     async fetchMLData(args: {
         fileIDs: number[];
         fileKeys: Map<number, Uint8Array>;
-    }): Promise<Map<number, MLData>> {
+    }): Promise<{
+        payloads: Map<number, MLData>;
+        errors: Map<number, unknown>;
+    }> {
         this.assertLoggedIn();
         return fetchMLDataBatch(this.api, args.fileIDs, args.fileKeys);
     }

src/download/index.ts

@@ -199,11 +199,13 @@ const isRunning = (pid: number): boolean => {
 };
 
 // Delete the temp files a killed process left in `dir`: the writer's
-// `.quak-<pid>-<random>.tmp` and the backup copy's
-// `.quak-backup-<name>-<pid>-<random>.tmp`. Only files whose process is no
-// longer running are removed, so another process writing into the same
-// directory keeps its own. A reused process ID can only keep a leftover a while
-// longer, never remove a live one.
+// `.quak-<pid>-<random>.tmp`, the backup copy's
+// `.quak-backup-<name>-<pid>-<random>.tmp`, and the image and video of a live
+// photo downloaded to replace a stored one, `.quak-<pid>-<random>.<ext>` (see
+// `replaceOriginal`). Only files whose process is no longer running are
+// removed, so another process writing into the same directory keeps its own. A
+// reused process ID can only keep a leftover a while longer, never remove a
+// live one.
 export const removeLeftoverTempFiles = (dir: string): void => {
     let names: string[];
     try {
@@ -212,7 +214,9 @@ export const removeLeftoverTempFiles = (dir: string): void => {
         return;
     }
     for (const name of names) {
-        const match = /^\.quak-(?:.*-)?(\d+)-[0-9a-z]*\.tmp$/.exec(name);
+        const match = /^\.quak-(?:.*-)?(\d+)-[0-9a-z]*\.[0-9A-Za-z]+$/.exec(
+            name,
+        );
         if (match && !isRunning(Number(match[1]))) {
             rmSync(join(dir, name), { force: true });
         }

src/library/content.ts

@@ -13,6 +13,8 @@
 // originals, the thumbnail pool for thumbnails. The pool limits concurrency,
 // orders on-demand work ahead of background, and dedups by key so a fileID
 // requested twice while the first is still in flight downloads once.
+// `fetchOriginal` passes no key, so it is not shared with a fetch of the same
+// file already in flight.
 //
 // Integrity. The reused streaming decrypt is the enforced guarantee: every
 // chunk is authenticated and the writer renames the file into place only once
@@ -27,9 +29,10 @@
 // It also names each original's save path under the download directory
 // (`savePath`), where `Photo.download()` and `lib.backup()` put it. A copy in
 // the cache does not count as saved there, but is copied there rather than
-// fetched again.
+// fetched again. An original stored there whose bytes do not match its content
+// hash is replaced only by a download that does (`replaceOriginal`).
 
-import { existsSync, readFileSync, statSync } from "node:fs";
+import { createReadStream, existsSync, readFileSync, statSync } from "node:fs";
 import {
     chmod,
     copyFile,
@@ -44,6 +47,12 @@ import {
 import { basename, dirname, extname, join } from "node:path";
 
 import type { ApiClient } from "../api/client.js";
+import {
+    chunkHashFinal,
+    chunkHashInit,
+    chunkHashUpdate,
+    init,
+} from "../crypto/index.js";
 import {
     downloadFile,
     downloadThumbnail,
@@ -349,6 +358,26 @@ export const storedAtSavePath = (
     return storedOriginal(dirname(path), basename(path, extname(path)), file);
 };
 
+// The content hash of the original stored at `stored`, computed as the download
+// check computes it: over each file's bytes, read in chunks, and for a live
+// photo `<imageHash>:<videoHash>`.
+export const storedHash = async (stored: {
+    path: string;
+    videoPath?: string;
+}): Promise<string> => {
+    await init();
+    const hashFile = async (path: string): Promise<string> => {
+        const state = chunkHashInit();
+        for await (const chunk of createReadStream(path)) {
+            chunkHashUpdate(state, chunk as Buffer);
+        }
+        return chunkHashFinal(state);
+    };
+    const hash = await hashFile(stored.path);
+    if (stored.videoPath === undefined) return hash;
+    return `${hash}:${await hashFile(stored.videoPath)}`;
+};
+
 // Copy bytes into `dest` via a temp file in the same directory plus rename, so
 // `dest` appears only once it is whole ("present means complete"). As in the
 // download writer, the temp file is fsynced before the rename and the directory
@@ -405,6 +434,62 @@ export const placeOriginal = async (
     return { path, videoPath };
 };
 
+// Replace the original of `file` stored at its save path under `root` with a
+// new copy, keeping the stored one until the copy is complete and its hash,
+// computed as `storedHash` computes it, is the content hash the file's metadata
+// records. `get` is given a temporary file beside the save path and writes the
+// copy there, a live photo as its image and its video beside that file, each
+// with its own extension. Only a copy that matches is renamed over the stored
+// one, and a live photo's image and video are both hashed before either is
+// renamed. The image is renamed first: should the run stop, or the video's
+// rename fail, before the video is renamed, the new image is left beside the
+// old video, and a later backup with `verify` replaces the video if the video
+// was the part that did not match. When `get` fails or the copy does not match,
+// the temporary files are removed and the stored original stays as it was.
+export const replaceOriginal = async (
+    root: string,
+    file: EnteFile,
+    get: (dest: string) => Promise<{ path: string; videoPath?: string }>,
+): Promise<{ path: string; videoPath?: string }> => {
+    const dest = savePath(root, file);
+    const dir = dirname(dest);
+    // Named as the download writer names its temporary files; a live photo's
+    // image and video get this name with their own extensions. The next
+    // backup removes any that a killed run leaves (`removeLeftoverTempFiles`).
+    const tmp = join(
+        dir,
+        `.quak-${process.pid}-${Math.random().toString(36).slice(2)}.tmp`,
+    );
+    let got: { path: string; videoPath?: string } | undefined;
+    try {
+        got = await get(tmp);
+        if ((await storedHash(got)) !== file.metadata.hash) {
+            throw new Error(
+                "the downloaded original does not match its content hash either",
+            );
+        }
+        if (got.videoPath === undefined) {
+            await rename(got.path, dest);
+            await fsyncPath(dir);
+            return { path: dest };
+        }
+        const path = withExtension(dest, extname(got.path));
+        const videoPath = withExtension(dest, extname(got.videoPath));
+        await rename(got.path, path);
+        await rename(got.videoPath, videoPath);
+        const name = basename(dest, extname(dest));
+        await writeLivePhotoJSON(dir, name, { path, videoPath });
+        return { path, videoPath };
+    } finally {
+        // Whatever was not renamed into place. A failure to remove it must
+        // not replace the error that explains what went wrong.
+        for (const path of [tmp, got?.path, got?.videoPath]) {
+            if (path === undefined) continue;
+            await rm(path, { force: true }).catch(() => undefined);
+        }
+    }
+};
+
 export class ContentCache implements PhotoContent, ThumbnailsAPI {
     private readonly pools: RequestPools;
     private readonly source: ContentSource;
@@ -543,6 +628,35 @@ export class ContentCache implements PhotoContent, ThumbnailsAPI {
         };
     }
 
+    // Fetch an original through the content pool to `destination`, a live
+    // photo as its image and its video beside it, and return where it landed.
+    // Unlike `backupOriginal`, it fetches even when the cache or the save path
+    // already holds the original, and it records nothing: what it writes is
+    // the caller's. It is not shared with a fetch of the same file already
+    // under way, which writes somewhere else.
+    async fetchOriginal(
+        fileID: number,
+        destination: string,
+    ): Promise<{ path: string; videoPath?: string }> {
+        const file = this.getFile(fileID);
+        if (!file) throw new Error(`content cache: unknown file ${fileID}`);
+        return this.pools.content.run(
+            async () => {
+                const fetched = await this.fetchInto(
+                    file,
+                    destination,
+                    "original",
+                    undefined,
+                );
+                for (const path of [fetched.path, fetched.videoPath]) {
+                    if (path !== undefined) await chmod(path, FILE_MODE);
+                }
+                return fetched;
+            },
+            { priority: "on-demand" },
+        );
+    }
+
     async ensure(args: EnsureOptions): Promise<EnsureResult[]> {
         return this.ensureThumbnails(args);
     }

src/library/index.ts

@@ -89,6 +89,7 @@ export {
 } from "./content.js";
 export { type MLDataAPI, type SimilarResult } from "./mlsearch.js";
 import type { CollectionsPage, FilesPage } from "../client.js";
+import { describeError } from "../errors.js";
 import { MLDATA_BATCH_SIZE, type MLData } from "../mldata-fetch.js";
 import type { Collection, EnteFile } from "../model/types.js";
 import { runBackup, type BackupOptions, type BackupResult } from "../backup.js";
@@ -138,12 +139,16 @@ export interface LibraryClient {
         sinceTime: number;
     }): Promise<FilesPage>;
     // Fetch ML data (face detections + CLIP embeddings) for up to a batch of
-    // files. Optional: a client without it simply disables ML fetching, leaving
-    // the metadata refresh untouched.
+    // files, and for each file whose ML data did not decrypt, gunzip or parse,
+    // why not. Optional: a client without it simply disables ML fetching,
+    // leaving the metadata refresh untouched.
     fetchMLData?(args: {
         fileIDs: number[];
         fileKeys: Map<number, Uint8Array>;
-    }): Promise<Map<number, MLData>>;
+    }): Promise<{
+        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
@@ -154,9 +159,11 @@ export interface LibraryClient {
 // 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. The precache fills
-// ("precacheThumbnails"/"precacheOriginals", #48) fire "started"/"done" around
-// each sweep that has work, "done" reporting the count newly cached.
+// 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.
+// The precache fills ("precacheThumbnails"/"precacheOriginals", #48) fire
+// "started"/"done" around each sweep that has work, "done" reporting the count
+// newly cached.
 export interface RefreshEvent {
     operation:
         | "refresh"
@@ -166,6 +173,7 @@ export interface RefreshEvent {
     status: "started" | "done" | "failed";
     error?: string;
     fetched?: number;
+    mlDataErrors?: { fileID: number; error: string }[];
 }
 
 export type RefreshProgressCallback = (event: RefreshEvent) => void;
@@ -225,6 +233,11 @@ export interface LibraryStatus {
     lastMLFetchAt?: number;
     // The most recent ML fetch pass's error, set only while it failed.
     lastMLError?: string;
+    // Each file whose ML data the last ML fetch pass that succeeded got but
+    // could not decrypt, gunzip or parse, with the reason. A pass still
+    // running leaves it as it is. A pass that fails leaves it as it was. Such
+    // ML data is not cached, so the next pass asks for it again.
+    mlDataErrors: { fileID: number; error: string }[];
     // ML payloads stored on disk and CLIP embeddings in the index; undefined
     // when ML fetching is disabled.
     mlStored?: number;
@@ -292,6 +305,12 @@ export class Library {
     private lastError?: string;
     private lastMLFetchAt?: number;
     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.
+    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
@@ -542,6 +561,7 @@ export class Library {
             lastError: this.lastError,
             lastMLFetchAt: this.lastMLFetchAt,
             lastMLError: this.lastMLError,
+            mlDataErrors: this.describeMLDataErrors(),
             mlStored: ml?.stored,
             mlIndexed: ml?.indexed,
             originalsUsedBytes: originals?.usedBytes,
@@ -606,10 +626,13 @@ export class Library {
                 undecryptable: () => this.store.listUndecryptable(),
                 original: (fileID, destination) =>
                     cache!.backupOriginal(fileID, destination),
+                fetchOriginal: (fileID, destination) =>
+                    cache!.fetchOriginal(fileID, destination),
                 thumbnail: (fileID) => cache!.thumbnail(fileID),
                 contentConcurrency: this.pools.content.concurrency,
                 fetchMLData: () => this.fetchMLDataNow(),
                 mlData: (fileID) => this.mldata.forFile({ fileID }),
+                mlDataError: (fileID) => this.mlDataErrors.get(fileID),
             },
             { ...opts, downloadDirectory },
         );
@@ -836,8 +859,11 @@ export class Library {
 
     // 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. A
-    // failure is reported through `status()` and `onProgress`, then thrown.
+    // 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.
     private async runMLFetch(): Promise<void> {
         const mldata = this.mlStore;
         // Bind so the call keeps the client as its receiver when invoked
@@ -847,7 +873,14 @@ export class Library {
 
         const files = this.uniqueFiles();
         const needed = mldata.neededFor(files);
-        if (needed.length === 0) return;
+        // This pass's reasons, put in `this.mlDataErrors` when it ends without
+        // failing. A file whose ML data did not decode in the pass before is
+        // still among `needed`, so this pass gives its reason anew.
+        const mlDataErrors = new Map<number, unknown>();
+        if (needed.length === 0) {
+            this.mlDataErrors = mlDataErrors;
+            return;
+        }
 
         this.emit({ operation: "fetchMLData", status: "started" });
         try {
@@ -862,20 +895,25 @@ export class Library {
             for (let i = 0; i < needed.length; i += MLDATA_BATCH_SIZE) {
                 if (this.closed) break;
                 const batch = needed.slice(i, i + MLDATA_BATCH_SIZE);
-                const payloads = await this.pools.metadata.run(
+                const { payloads, errors } = await this.pools.metadata.run(
                     () => fetchMLData({ fileIDs: batch, fileKeys }),
                     { priority: "background" },
                 );
+                for (const [id, error] of errors) {
+                    mlDataErrors.set(id, error);
+                }
                 stored += (await mldata.storeFetched(payloads, updation))
                     .stored;
             }
 
+            this.mlDataErrors = mlDataErrors;
             this.lastMLFetchAt = Date.now();
             this.lastMLError = undefined;
             this.emit({
                 operation: "fetchMLData",
                 status: "done",
                 fetched: stored,
+                mlDataErrors: this.describeMLDataErrors(),
             });
         } catch (err) {
             const error = err instanceof Error ? err.message : String(err);
@@ -885,6 +923,14 @@ export class Library {
         }
     }
 
+    // `mlDataErrors` as `status()` and the ML "done" event give it.
+    private describeMLDataErrors(): { fileID: number; error: string }[] {
+        return [...this.mlDataErrors].map(([fileID, err]) => ({
+            fileID,
+            error: describeError(err),
+        }));
+    }
+
     // 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.

src/metadata-backup.ts

@@ -2,6 +2,7 @@ import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import * as jpeg from "jpeg-js";
 import type { Client } from "./client.js";
+import { describeError } from "./errors.js";
 import { readExifTags } from "./exif.js";
 import type { Library, Photo } from "./library/index.js";
 import { sanitizeFileName } from "./filename.js";
@@ -91,14 +92,15 @@ const extractExif = async (
 // of plain JSON: account, per-collection, and per-file records including the
 // private and public magic metadata and (by default) the ML data. Collections
 // and files are enumerated from the library's cache, which the caller refreshes
-// first. Returns how many ML data requests failed; their files are still
+// 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
 // written, with `mlDataError` in place of `mlData`.
 export const runMetadataBackup = async (
     lib: Library,
     client: Client,
     outDir: string,
     opts?: MetadataBackupOptions,
-): Promise<{ failedMLBatches: number }> => {
+): Promise<{ filesWithMLDataError: number }> => {
     const log = opts?.onProgress ?? (() => {});
     const wantExif = opts?.exif ?? false;
 
@@ -164,24 +166,28 @@ export const runMetadataBackup = async (
     }
 
     // One failed request (retries exhausted) must not end the dump: its files
-    // get the reason in `mlDataError` and the other batches go on.
+    // get the reason in `mlDataError` and the other batches go on. So does a
+    // file whose ML data does not decrypt, gunzip or parse.
     log("Fetching ML data (face detections, CLIP embeddings)...");
     const mlDataMap = new Map<number, MLData>();
     const mlDataErrors = new Map<number, string>();
-    let failedMLBatches = 0;
     const fileIDs = [...fileKeys.keys()];
     for (let i = 0; i < fileIDs.length; i += MLDATA_BATCH_SIZE) {
         const batch = fileIDs.slice(i, i + MLDATA_BATCH_SIZE);
         try {
-            const result = await fetchMLDataBatch(
+            const { payloads, errors } = await fetchMLDataBatch(
                 client.getApiClient(),
                 batch,
                 fileKeys,
             );
-            for (const [id, payload] of result) mlDataMap.set(id, payload);
+            for (const [id, payload] of payloads) mlDataMap.set(id, payload);
+            for (const [id, err] of errors) {
+                const reason = describeError(err);
+                log(`ML data of file ${id} could not be read: ${reason}`);
+                mlDataErrors.set(id, reason);
+            }
         } catch (err) {
             const reason = err instanceof Error ? err.message : String(err);
-            failedMLBatches++;
             log(
                 `ML data request for ${batch.length} file(s) failed: ${reason}`,
             );
@@ -229,5 +235,5 @@ export const runMetadataBackup = async (
     }
 
     log("Metadata backup complete.");
-    return { failedMLBatches };
+    return { filesWithMLDataError: mlDataErrors.size };
 };

src/mldata-fetch.ts

@@ -28,43 +28,52 @@ interface RawRemoteFileData {
     updatedAt?: number;
 }
 
-// Decrypt one entry with its file key and gunzip the JSON payload. Returns
-// undefined when the key is unknown or the entry does not decrypt/parse, so one
-// corrupt file never fails a whole batch.
-const decodeEntry = (
-    entry: RawRemoteFileData,
-    key: Uint8Array | undefined,
-): MLData | undefined => {
-    if (!key) return undefined;
-    try {
-        const decrypted = decryptBlob(
-            fromBase64(entry.encryptedData),
-            fromBase64(entry.decryptionHeader),
-            key,
-        );
-        const json = gunzipSync(Buffer.from(decrypted)).toString("utf-8");
-        return JSON.parse(json) as MLData;
-    } catch {
-        return undefined;
+// Decrypt one entry with its file key and gunzip the JSON payload. Throws why
+// when the entry does not decrypt, gunzip or parse, or parses to something
+// other than a JSON object, such as `null`, a number or an array.
+const decodeEntry = (entry: RawRemoteFileData, key: Uint8Array): MLData => {
+    const decrypted = decryptBlob(
+        fromBase64(entry.encryptedData),
+        fromBase64(entry.decryptionHeader),
+        key,
+    );
+    const json = gunzipSync(Buffer.from(decrypted)).toString("utf-8");
+    const payload: unknown = JSON.parse(json);
+    if (
+        typeof payload !== "object" ||
+        payload === null ||
+        Array.isArray(payload)
+    ) {
+        throw new Error("not a JSON object");
     }
+    return payload as MLData;
 };
 
 // Fetch ML data for up to `MLDATA_BATCH_SIZE` ids in a single request. This is
 // the unit the request pools schedule; callers with more ids split them into
-// batches and submit each batch to the pool.
+// batches and submit each batch to the pool. Returns the payload of each file
+// whose entry decodes and, in `errors`, why not for each file whose entry does
+// not, so one corrupt file never fails a whole batch. An entry for a file
+// whose key is not known, which quak did not ask for, is skipped.
 export const fetchMLDataBatch = async (
     api: ApiClient,
     fileIDs: number[],
     fileKeys: Map<number, Uint8Array>,
-): Promise<Map<number, MLData>> => {
+): Promise<{ payloads: Map<number, MLData>; errors: Map<number, unknown> }> => {
     const { data } = await api.postJSON<{ data: RawRemoteFileData[] }>(
         "/files/data/fetch",
         { type: "mldata", fileIDs },
     );
-    const result = new Map<number, MLData>();
+    const payloads = new Map<number, MLData>();
+    const errors = new Map<number, unknown>();
     for (const entry of data ?? []) {
-        const payload = decodeEntry(entry, fileKeys.get(entry.fileID));
-        if (payload) result.set(entry.fileID, payload);
+        const key = fileKeys.get(entry.fileID);
+        if (!key) continue;
+        try {
+            payloads.set(entry.fileID, decodeEntry(entry, key));
+        } catch (err) {
+            errors.set(entry.fileID, err);
+        }
     }
-    return result;
+    return { payloads, errors };
 };

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

@@ -9,7 +9,7 @@
  * data this file encrypts with libsodium, as Ente's clients encrypt it. So
  * quak's own code fetches each original, decrypts it, checks it against the
  * hash its uploader recorded and unpacks a live photo, and quak's own retries
- * repeat a failed request. Two tests run `lib.backup()` instead, as a program
+ * repeat a failed request. Three tests run `lib.backup()` instead, as a program
  * using quak as a library does.
  *
  * The account, user 42, has three albums:
@@ -877,6 +877,133 @@ describe("quak backup", () => {
         ]);
     });
 
+    // The server sends beach.jpg's ML data encrypted under a key that is not
+    // the file's, so quak cannot decrypt it. The file's JSON has the reason
+    // in place of the ML data, and the file goes into `failures.json` with
+    // it, so the exit code is 1. The other two files have no ML data, which
+    // is no failure.
+    it("records a file whose ML data does not decrypt in its JSON and in failures.json, and exits 1", async () => {
+        const otherKey = sodium.crypto_secretstream_xchacha20poly1305_keygen();
+        const mlData = encryptStreamBlob(
+            gzipSync(JSON.stringify(ML_DATA)),
+            otherKey,
+        );
+        const fetch: typeof globalThis.fetch = async (input, init) => {
+            if (new URL(String(input)).pathname !== "/files/data/fetch") {
+                return server(input, init);
+            }
+            const data = [
+                {
+                    fileID: BEACH.id,
+                    encryptedData: toBase64(mlData.ciphertext),
+                    decryptionHeader: toBase64(mlData.header),
+                },
+            ];
+            return new Response(JSON.stringify({ data }), {
+                headers: { "content-type": "application/json" },
+            });
+        };
+        const dir = mkdtempSync(join(root, "backup-"));
+
+        const { code, stderr } = await backup(dir, {}, fetch);
+
+        expect(code).toBe(1);
+        const reason = "secretstream chunk authentication failed";
+        const beachJSON = readJSON(join(dir, DAY, "2024-03-01.100.json"));
+        expect(beachJSON).not.toHaveProperty("mlData");
+        expect(beachJSON.mlDataError).toBe(reason);
+        const failures = readJSON(join(dir, "failures.json"));
+        expect(Object.keys(failures.files)).toEqual(["100"]);
+        expect(failures.files["100"]).toMatchObject({
+            fileID: 100,
+            title: "beach.jpg",
+            error: `ML data: ${reason}`,
+        });
+        expect(stderr).toContain(`beach.jpg (id 100): ML data: ${reason}\n`);
+        for (const fileID of [200, 300]) {
+            const json = readJSON(join(dir, DAY, `2024-03-01.${fileID}.json`));
+            expect(json).not.toHaveProperty("mlData");
+            expect(json).not.toHaveProperty("mlDataError");
+        }
+    });
+
+    // A program using quak as a library backs the account up, and another ML
+    // data fetch, as one the library's refresh starts, fails while the backup
+    // writes the JSON files. The backup's own fetch got beach.jpg's ML data
+    // under a key that is not the file's; the fetch that fails leaves that
+    // reason in place, so beach.jpg's JSON and `failures.json` still give it.
+    it("keeps the reason in beach.jpg's JSON and in failures.json when an ML data fetch fails while lib.backup() writes the JSON files", async () => {
+        const otherKey = sodium.crypto_secretstream_xchacha20poly1305_keygen();
+        const mlData = encryptStreamBlob(
+            gzipSync(JSON.stringify(ML_DATA)),
+            otherKey,
+        );
+        let mlDataFetchFails = false;
+        const fetch: typeof globalThis.fetch = async (input, init) => {
+            if (new URL(String(input)).pathname !== "/files/data/fetch") {
+                return server(input, init);
+            }
+            if (mlDataFetchFails) throw new Error("the connection was reset");
+            const data = [
+                {
+                    fileID: BEACH.id,
+                    encryptedData: toBase64(mlData.ciphertext),
+                    decryptionHeader: toBase64(mlData.header),
+                },
+            ];
+            return new Response(JSON.stringify({ data }), {
+                headers: { "content-type": "application/json" },
+            });
+        };
+        const dir = mkdtempSync(join(root, "backup-"));
+        const lib = await Library.open({
+            client: loadSession(join(sessionDir, "session.json"), {
+                fetch,
+                retry: { ...UNATTENDED_RETRY_OPTIONS, sleep: async () => {} },
+            })!,
+            cacheDirectory: `${dir}-cache`,
+            downloadDirectory: dir,
+            refreshIntervalSeconds: Infinity,
+            precacheThumbnails: false,
+            precacheOriginals: false,
+        });
+        try {
+            // The backup reads a file's cached ML data just before it writes
+            // that file's JSON. The first time, a refresh starts an ML data
+            // fetch whose request fails, and the backup goes on once that
+            // fetch has failed.
+            const forFile = lib.mldata.forFile.bind(lib.mldata);
+            vi.spyOn(lib.mldata, "forFile").mockImplementationOnce(
+                async (args) => {
+                    mlDataFetchFails = true;
+                    await lib.fresh();
+                    await vi.waitFor(
+                        () => expect(lib.status().lastMLError).toBeDefined(),
+                        { timeout: 2000, interval: 5 },
+                    );
+                    return forFile(args);
+                },
+            );
+
+            const result = await lib.backup();
+
+            expect(result.failed).toBe(1);
+            const reason = "secretstream chunk authentication failed";
+            const beachJSON = readJSON(join(dir, DAY, "2024-03-01.100.json"));
+            expect(beachJSON).not.toHaveProperty("mlData");
+            expect(beachJSON.mlDataError).toBe(reason);
+            const failures = readJSON(join(dir, "failures.json"));
+            expect(Object.keys(failures.files)).toEqual(["100"]);
+            expect(failures.files["100"]).toMatchObject({
+                fileID: 100,
+                title: "beach.jpg",
+                error: `ML data: ${reason}`,
+            });
+        } finally {
+            await lib.close();
+        }
+    });
+
     // Something changed beach.jpg on disk after it was backed up. A run with
     // --verify hashes every stored original, finds that one no longer matches
     // the hash Ente recorded, says so, and downloads it again; it put back the

test/cli/backup.test.ts

@@ -191,9 +191,12 @@ class HostileClient extends MockClient {
 }
 
 // A content source that writes byte buffers of the expected length and can be
-// told to fail one fileID's original, to exercise per-file resilience.
+// told to fail one fileID's original, as a server error or as a 404, or to
+// write other bytes for it, to exercise per-file resilience.
 interface StubSource extends ContentSource {
     failID?: number;
+    notFoundID?: number;
+    otherBytesID?: number;
     failThumbID?: number;
     originalCalls: number;
 }
@@ -204,8 +207,10 @@ const stubSource = (): StubSource => {
         original: async ({ file: f, destination }) => {
             s.originalCalls++;
             if (s.failID === f.id) throw new Error("HTTP 500 from server");
+            if (s.notFoundID === f.id) throw new Error("HTTP 404 from server");
             const size = SIZE_BY_ID[f.id] ?? 10;
-            writeFileSync(destination, Buffer.alloc(size));
+            const fill = s.otherBytesID === f.id ? 2 : 0;
+            writeFileSync(destination, Buffer.alloc(size, fill));
             return { bytesWritten: size };
         },
         thumbnail: async ({ file: f, destination }) => {
@@ -696,17 +701,26 @@ describe("lib.backup", () => {
         // A child that has already exited: its process ID is not running.
         const exitedPID = spawnSync(process.execPath, ["-e", ""]).pid;
         const leftover = `.quak-backup-2026-03-01.100.jpg-${exitedPID}-abc123.tmp`;
+        // A live photo's image and video downloaded to replace a stored one,
+        // before they were renamed into place.
+        const replacement = [
+            `.quak-${exitedPID}-abc789.heic`,
+            `.quak-${exitedPID}-abc789.mov`,
+        ];
         // This test's own process stands in for a backup running at the same
         // time.
         const inProgress = `.quak-backup-2026-03-01.101.jpg-${process.pid}-def456.tmp`;
-        writeFileSync(join(day, leftover), "partial");
-        writeFileSync(join(day, inProgress), "partial");
+        for (const name of [leftover, ...replacement, inProgress]) {
+            writeFileSync(join(day, name), "partial");
+        }
         const lib = await openLibrary(stubSource());
 
         await lib.backup({ downloadDirectory: outDir });
 
         const names = readdirSync(day);
         expect(names).not.toContain(leftover);
+        expect(names).not.toContain(replacement[0]);
+        expect(names).not.toContain(replacement[1]);
         expect(names).toContain(inProgress);
         lib.close();
     });
@@ -938,13 +952,14 @@ const ML_PAYLOAD = { face: { faces: [] }, clip: { embedding: [0.5, 0.25] } };
 
 class MLClient extends MockClient {
     mlError?: string;
-    async fetchMLData(args: {
-        fileIDs: number[];
-    }): Promise<Map<number, MLData>> {
+    async fetchMLData(args: { fileIDs: number[] }): Promise<{
+        payloads: Map<number, MLData>;
+        errors: Map<number, unknown>;
+    }> {
         if (this.mlError) throw new Error(this.mlError);
-        const result = new Map<number, MLData>();
-        if (args.fileIDs.includes(100)) result.set(100, ML_PAYLOAD);
-        return result;
+        const payloads = new Map<number, MLData>();
+        if (args.fileIDs.includes(100)) payloads.set(100, ML_PAYLOAD);
+        return { payloads, errors: new Map() };
     }
 }
 
@@ -1294,7 +1309,8 @@ describe("image metadata in each file's JSON", () => {
 });
 
 // With `verify`, each original already stored is hashed, and one that does not
-// match the content hash its metadata records is downloaded again.
+// match the content hash its metadata records is downloaded again, and replaced
+// only by a download that matches.
 describe("backup with verify", () => {
     // MockClient's files, each recording the hash of what `stubSource` writes
     // for it, except diagram.png (200), which records none.
@@ -1337,6 +1353,12 @@ describe("backup with verify", () => {
     const good = Buffer.alloc(SIZE_BY_ID[100]!);
     const corrupt = Buffer.alloc(SIZE_BY_ID[100]!, 1);
 
+    // The temporary files left in the date folder of the backup in `outDir`.
+    const tempFiles = (outDir: string): string[] =>
+        readdirSync(join(outDir, DAY)).filter((name) =>
+            name.startsWith(".quak-"),
+        );
+
     it("leaves an original that matches its hash, and one with no hash, as they are", async () => {
         const { lib, source, outDir } = await backedUp();
         const calls = source.originalCalls;
@@ -1381,13 +1403,16 @@ describe("backup with verify", () => {
         expect(log.filter((msg) => msg.startsWith("MISMATCH"))).toEqual([
             "MISMATCH original beach.jpg (100): its bytes do not match its content hash",
         ]);
+        expect(tempFiles(outDir)).toEqual([]);
         await lib.close();
     });
 
-    it("records a failed download in failures.json, with the original removed", async () => {
+    // A copy whose hash does not match is still mostly the photo, so the
+    // backup keeps it until it holds a better one.
+    it("keeps an original that does not match its hash when its download answers 404", async () => {
         const { lib, source, outDir } = await backedUp();
         writeFileSync(saved(outDir, "100.jpg"), corrupt);
-        source.failID = 100;
+        source.notFoundID = 100;
 
         const result = await lib.backup({
             downloadDirectory: outDir,
@@ -1400,18 +1425,18 @@ describe("backup with verify", () => {
             failed: 1,
         });
         expect(Object.keys(readLedger(outDir).files)).toEqual(["100"]);
-        expect(existsSync(saved(outDir, "100.jpg"))).toBe(false);
+        expect(readFileSync(saved(outDir, "100.jpg"))).toEqual(corrupt);
+        expect(
+            readlinkSync(join(outDir, "collections", "Vacation", "beach.jpg")),
+        ).toBe(linkTo("100.jpg"));
+        expect(tempFiles(outDir)).toEqual([]);
         await lib.close();
     });
 
-    it("records in failures.json an original the cache puts back with the same bad bytes", async () => {
-        const lib = await openLibrary(stubSource(), new HashedClient());
-        const outDir = join(root, "backup");
-        // The cache holds a bad copy, and a backup copies an original the
-        // cache holds to its save path.
-        const cached = await lib.photos.byID({ fileID: 100 })!.original();
-        writeFileSync(cached.path, corrupt);
-        await lib.backup({ downloadDirectory: outDir });
+    it("keeps an original that does not match its hash when its download does not match either", async () => {
+        const { lib, source, outDir } = await backedUp();
+        writeFileSync(saved(outDir, "100.jpg"), corrupt);
+        source.otherBytesID = 100;
 
         const result = await lib.backup({
             downloadDirectory: outDir,
@@ -1424,10 +1449,39 @@ describe("backup with verify", () => {
             failed: 1,
         });
         expect(result.errors.map((e) => e.error)).toEqual([
-            "the original put back does not match its content hash either",
+            "the downloaded original does not match its content hash either",
         ]);
         expect(Object.keys(readLedger(outDir).files)).toEqual(["100"]);
         expect(readFileSync(saved(outDir, "100.jpg"))).toEqual(corrupt);
+        expect(tempFiles(outDir)).toEqual([]);
+        await lib.close();
+    });
+
+    // The cache can hold the same bad bytes, as a copy from it is not checked
+    // as a download is, so the replacement is always downloaded.
+    it("downloads the replacement even when the cache holds the original", async () => {
+        const source = stubSource();
+        const lib = await openLibrary(source, new HashedClient());
+        const outDir = join(root, "backup");
+        // The cache holds a bad copy, and a backup copies an original the
+        // cache holds to its save path.
+        const cached = await lib.photos.byID({ fileID: 100 })!.original();
+        writeFileSync(cached.path, corrupt);
+        await lib.backup({ downloadDirectory: outDir });
+        const calls = source.originalCalls;
+
+        const result = await lib.backup({
+            downloadDirectory: outDir,
+            verify: true,
+        });
+
+        expect(result).toMatchObject({
+            downloaded: 1,
+            mismatched: 1,
+            failed: 0,
+        });
+        expect(source.originalCalls).toBe(calls + 1);
+        expect(readFileSync(saved(outDir, "100.jpg"))).toEqual(good);
         await lib.close();
     });
 
@@ -1511,12 +1565,16 @@ describe("backup album folders", () => {
             writeFileSync(path, `original ${fileID}`);
             return { path };
         },
+        fetchOriginal: async () => {
+            throw new Error("no verify in this stand-in");
+        },
         thumbnail: async () => {
             throw new Error("no thumbnails in this stand-in");
         },
         contentConcurrency: 5,
         fetchMLData: async () => {},
         mlData: async () => undefined,
+        mlDataError: () => undefined,
         whoami: () => ({ email: "backup@example.com", userID: USER_ID }),
     });
 
@@ -2244,11 +2302,12 @@ describe("backup of live photos", () => {
             Buffer.from(IMAGE),
         );
         expect(readFileSync(video)).toEqual(Buffer.from(VIDEO));
+        expect(readdirSync(join(outDir, DAY)).sort()).toEqual(stored);
         expect(tree(outDir)).toEqual(linked);
         await lib.close();
     });
 
-    it("removes both files of a live photo that does not match its hash when downloading it again fails", async () => {
+    it("keeps both files of a live photo that does not match its hash when downloading it again fails", async () => {
         const { file: live, body } = await asLivePhoto(
             file(500, 10, "IMG_0500.HEIC"),
         );
@@ -2266,8 +2325,14 @@ describe("backup of live photos", () => {
 
         expect(result).toMatchObject({ mismatched: 1, failed: 1 });
         expect(Object.keys(readLedger(outDir).files)).toEqual(["500"]);
-        expect(existsSync(saved(outDir, "500.heic"))).toBe(false);
-        expect(existsSync(saved(outDir, "500.mov"))).toBe(false);
+        expect(readFileSync(saved(outDir, "500.heic"))).toEqual(
+            Buffer.from(IMAGE),
+        );
+        expect(readFileSync(saved(outDir, "500.mov"), "utf-8")).toBe(
+            "another few seconds of video",
+        );
+        expect(readdirSync(join(outDir, DAY)).sort()).toEqual(stored);
+        expect(tree(outDir)).toEqual(linked);
         await lib.close();
     });
 

test/cli/metadata-backup.test.ts

@@ -489,15 +489,15 @@ describe("quak backup-metadata", () => {
     });
 });
 
-describe("quak backup-metadata when an ML data request fails", () => {
-    // Run the CLI command against the mock and return its exit code, stderr
-    // and output directory.
-    const runCommand = async (failMLDataFor?: number) => {
+describe("quak backup-metadata when a file's ML data cannot be read", () => {
+    // Run the CLI command against `fetch`, the mock unless given, and return
+    // its exit code, stderr and output directory.
+    const runCommand = async (fetch = buildMetaFetch(mock)) => {
         const client = await Client.login({
             email: TEST_EMAIL,
             password: TEST_PASSWORD,
             apiOptions: {
-                fetch: buildMetaFetch(mock, failMLDataFor),
+                fetch,
                 retry: { sleep: async () => {} },
             },
         });
@@ -518,7 +518,9 @@ describe("quak backup-metadata when an ML data request fails", () => {
     };
 
     it("writes every file, marks the failed batch's files, and exits 1", async () => {
-        const { code, stderr, outDir } = await runCommand(200);
+        const { code, stderr, outDir } = await runCommand(
+            buildMetaFetch(mock, 200),
+        );
 
         expect(code).toBe(1);
         expect(stderr).toContain("ML data request for 1 file(s) failed");
@@ -543,6 +545,49 @@ describe("quak backup-metadata when an ML data request fails", () => {
         expect(failed.mlDataError).toContain("500");
     });
 
+    // The server sends ML data for diagram.png encrypted under a key that is
+    // not the file's, so quak cannot decrypt it.
+    it("writes the reason for ML data that does not decrypt, and exits 1", async () => {
+        const otherKey = sodium.crypto_secretstream_xchacha20poly1305_keygen();
+        const { header, ciphertext } = encryptBlob(
+            new Uint8Array(gzipSync(Buffer.from("{}"))),
+            otherKey,
+        );
+        const { code, stderr, outDir } = await runCommand(
+            buildMetaFetch({
+                ...mock,
+                encryptedMLData: {
+                    ...mock.encryptedMLData,
+                    200: {
+                        encryptedData: toBase64(ciphertext),
+                        decryptionHeader: toBase64(header),
+                    },
+                },
+            }),
+        );
+
+        expect(code).toBe(1);
+        const reason = "secretstream chunk authentication failed";
+        expect(stderr).toContain(
+            `ML data of file 200 could not be read: ${reason}\n`,
+        );
+        const failed = JSON.parse(
+            readFileSync(
+                join(outDir, "collections", "20-__Work", "200.json"),
+                "utf-8",
+            ),
+        );
+        expect(failed.mlData).toBeUndefined();
+        expect(failed.mlDataError).toBe(reason);
+        const ok = JSON.parse(
+            readFileSync(
+                join(outDir, "collections", "10-Vacation", "100.json"),
+                "utf-8",
+            ),
+        );
+        expect(ok.mlData.clip.embedding).toEqual([0.5, 0.6, 0.7]);
+    });
+
     it("exits 0 when every ML data request succeeds", async () => {
         const { code } = await runCommand();
         expect(code).toBe(0);

test/library/mldata.test.ts

@@ -12,9 +12,11 @@
  *
  *   2. `Library` wiring: after each refresh the library fetches ML data through
  *      the metadata pool for every known file not yet cached, is incremental on
- *      later refreshes, and refetches a file whose `updationTime` advanced.
- *      Opening a cache directory written by another account starts empty,
- *      its ML data included (issue #104).
+ *      later refreshes, and refetches a file whose `updationTime` advanced. A
+ *      file whose ML data did not decode is reported with the reason, which
+ *      stays listed until the next pass ends, and asked for again. Opening a
+ *      cache directory written by another account starts empty, its ML data
+ *      included (issue #104).
  *
  * Embedding values are chosen to be exactly representable as float32 so the
  * round-trip through `clip.f32` compares equal.
@@ -26,7 +28,7 @@ import { tmpdir } from "node:os";
 import { join } from "node:path";
 
 import { MLDataStore } from "../../src/library/mldata.js";
-import { Library } from "../../src/library/index.js";
+import { Library, type RefreshEvent } from "../../src/library/index.js";
 import type { CollectionsPage, FilesPage } from "../../src/client.js";
 import type { MLData } from "../../src/mldata-fetch.js";
 import type { Collection, EnteFile } from "../../src/model/types.js";
@@ -256,12 +258,14 @@ const file = (
 
 // A mock client that serves scripted collection/file pages and per-file ML
 // payloads, recording every ML fetch request so incremental behaviour is
-// provable.
+// provable. A file in `mlErrorByFile` has ML data that did not decode, for
+// the reason given there.
 class MLMockClient {
     userID = USER_ID;
     collectionsQueue: CollectionsPage[] = [];
     filesByCollection = new Map<number, FilesPage[]>();
     mlByFile = new Map<number, MLData>();
+    mlErrorByFile = new Map<number, Error>();
     mlFetchCalls: number[][] = [];
 
     whoami(): { email: string; userID: number } {
@@ -298,14 +302,20 @@ class MLMockClient {
     async fetchMLData(args: {
         fileIDs: number[];
         fileKeys: Map<number, Uint8Array>;
-    }): Promise<Map<number, MLData>> {
+    }): Promise<{
+        payloads: Map<number, MLData>;
+        errors: Map<number, unknown>;
+    }> {
         this.mlFetchCalls.push([...args.fileIDs]);
-        const result = new Map<number, MLData>();
+        const payloads = new Map<number, MLData>();
+        const errors = new Map<number, unknown>();
         for (const id of args.fileIDs) {
             const p = this.mlByFile.get(id);
-            if (p) result.set(id, p);
+            if (p) payloads.set(id, p);
+            const error = this.mlErrorByFile.get(id);
+            if (error) errors.set(id, error);
         }
-        return result;
+        return { payloads, errors };
     }
 
     filesFor(collectionID: number, ...pages: FilesPage[]): void {
@@ -449,6 +459,137 @@ describe("Library ML-data fetch on refresh", () => {
         }
     });
 
+    it("reports a file whose ML data did not decode through status() and onProgress, and asks for it again", async () => {
+        const client = new MLMockClient();
+        client.collectionsQueue.push({
+            collections: [collection(1, 100)],
+            deleted: [],
+            cursor: 100,
+        });
+        client.filesFor(1, {
+            files: [file(1001, 1, 90), file(1002, 1, 95)],
+            deleted: [],
+            cursor: 95,
+        });
+        client.mlByFile.set(1001, payload([0.5, 0.25, 0.75]));
+        client.mlErrorByFile.set(
+            1002,
+            new Error("secretstream chunk authentication failed"),
+        );
+        const events: RefreshEvent[] = [];
+
+        const lib = await Library.open({
+            client,
+            cacheDirectory,
+            refreshIntervalSeconds: Infinity,
+            onProgress: (event) => events.push(event),
+        });
+        try {
+            await vi.waitFor(
+                () => expect(lib.status().lastMLFetchAt).toBeGreaterThan(0),
+                { timeout: 2000, interval: 5 },
+            );
+            const reasons = [
+                {
+                    fileID: 1002,
+                    error: "secretstream chunk authentication failed",
+                },
+            ];
+            expect(lib.status().mlDataErrors).toEqual(reasons);
+            expect(lib.status().mlStored).toBe(1);
+            expect(events.filter((e) => e.operation === "fetchMLData")).toEqual(
+                [
+                    { operation: "fetchMLData", status: "started" },
+                    {
+                        operation: "fetchMLData",
+                        status: "done",
+                        fetched: 1,
+                        mlDataErrors: reasons,
+                    },
+                ],
+            );
+
+            // Nothing was cached for 1002, so the pass a refresh starts asks
+            // for it again, and once its ML data decodes the reason is gone.
+            client.mlErrorByFile.clear();
+            client.mlByFile.set(1002, payload([1, -2, 0.5]));
+            await lib.fresh();
+            await vi.waitFor(
+                () => {
+                    expect(lib.status().mlStored).toBe(2);
+                    expect(lib.status().mlDataErrors).toEqual([]);
+                },
+                { timeout: 2000, interval: 5 },
+            );
+            expect(client.mlFetchCalls.at(-1)).toEqual([1002]);
+        } finally {
+            await lib.close();
+        }
+    });
+
+    it("still lists the last pass's reasons while the next pass runs", async () => {
+        const client = new MLMockClient();
+        client.collectionsQueue.push({
+            collections: [collection(1, 100)],
+            deleted: [],
+            cursor: 100,
+        });
+        client.filesFor(1, {
+            files: [file(1002, 1, 95)],
+            deleted: [],
+            cursor: 95,
+        });
+        client.mlErrorByFile.set(
+            1002,
+            new Error("secretstream chunk authentication failed"),
+        );
+        const reasons = [
+            { fileID: 1002, error: "secretstream chunk authentication failed" },
+        ];
+        let release!: () => void;
+        const held = new Promise<void>((r) => (release = r));
+
+        const lib = await Library.open({
+            client,
+            cacheDirectory,
+            refreshIntervalSeconds: Infinity,
+        });
+        try {
+            await vi.waitFor(
+                () => expect(lib.status().lastMLFetchAt).toBeGreaterThan(0),
+                { timeout: 2000, interval: 5 },
+            );
+            expect(lib.status().mlDataErrors).toEqual(reasons);
+
+            // Hold the next ML data fetch open until the test releases it.
+            let fetchStarted!: () => void;
+            const started = new Promise<void>((r) => (fetchStarted = r));
+            const realFetch = client.fetchMLData.bind(client);
+            client.fetchMLData = async (args) => {
+                fetchStarted();
+                await held;
+                return realFetch(args);
+            };
+            client.mlErrorByFile.clear();
+            client.mlByFile.set(1002, payload([1, -2, 0.5]));
+
+            // The pass a refresh starts asks for 1002 again. While it is
+            // held, the reason from the pass before is still listed; once it
+            // ends, its own empty list takes the place of that one.
+            await lib.fresh();
+            await started;
+            expect(lib.status().mlDataErrors).toEqual(reasons);
+            release();
+            await vi.waitFor(
+                () => expect(lib.status().mlDataErrors).toEqual([]),
+                { timeout: 2000, interval: 5 },
+            );
+        } finally {
+            release();
+            await lib.close();
+        }
+    });
+
     it("close() resolves only after a running ML data fetch has stored its payloads", async () => {
         const client = new MLMockClient();
         client.collectionsQueue.push({

test/mldata-fetch/mldata-fetch.test.ts

@@ -0,0 +1,54 @@
+// `fetchMLDataBatch`, against a server that sends one file's ML data.
+//
+// ML data that decrypts and gunzips but is not a JSON object is reported with
+// the reason, as ML data that does not parse is, and is never returned as a
+// payload: the library would fail to store it.
+
+import { gzipSync } from "node:zlib";
+import sodium from "libsodium-wrappers-sumo";
+import { beforeAll, describe, expect, it } from "vitest";
+
+import { ApiClient } from "../../src/api/client.js";
+import { encryptBlob, init, toBase64 } from "../../src/crypto/index.js";
+import { fetchMLDataBatch } from "../../src/mldata-fetch.js";
+
+beforeAll(async () => {
+    await init();
+    await sodium.ready;
+});
+
+// Fetch the ML data of file 1 from a server that sends `json`, gzipped and
+// encrypted with the file's key.
+const fetchServing = async (json: string) => {
+    const key = sodium.crypto_secretstream_xchacha20poly1305_keygen();
+    const { header, ciphertext } = encryptBlob(
+        new Uint8Array(gzipSync(json)),
+        key,
+    );
+    const data = [
+        {
+            fileID: 1,
+            encryptedData: toBase64(ciphertext),
+            decryptionHeader: toBase64(header),
+        },
+    ];
+    const api = new ApiClient({
+        fetch: async () =>
+            new Response(JSON.stringify({ data }), {
+                headers: { "content-type": "application/json" },
+            }),
+    });
+    return fetchMLDataBatch(api, [1], new Map([[1, key]]));
+};
+
+describe("fetchMLDataBatch", () => {
+    it.each(["null", "5", "[]"])(
+        "gives a reason for ML data that is %s",
+        async (json) => {
+            const { payloads, errors } = await fetchServing(json);
+
+            expect(payloads.size).toBe(0);
+            expect(errors.get(1)).toEqual(new Error("not a JSON object"));
+        },
+    );
+});