Wait for the interrupted operation to clean up before exit (closes #159)

On SIGINT/SIGTERM the process could exit before the interrupted command cleanup defers ran, leaving decrypted data in the temp directory (the blob cache and the decrypted snapshot database).

RunApp now mirrors fx run sequence: start, block on app.Wait(), then app.Stop(), returning only after Stop completes. fx delivers both an OS interrupt and the finished operation Shutdowner.Shutdown() on one channel. Stop runs the OnStop hooks; the operation hook cancels the command and waits for its goroutine to return (bounded by shutdownTimeout) before exit. The old code returned as soon as app.Done fired, without Stop, so a real interrupt unwound to os.Exit while cleanup still ran. Restore loops check the context between chunks and blobs so the wait ends promptly. A cli test drives RunApp through the OnStop hook.

Model: opus-4-8

AGENTS.md

@@ -104,7 +104,12 @@ Version: 2025-06-08
 
 13. Pre-1.0: NEVER write database migrations. There are no live databases
     anywhere — every user's local index can be rebuilt from a fresh full
-    backup. When the schema changes, just change `schema.sql` (and any code
+    backup. To change the schema, edit `internal/database/schema/001.sql`
-    that touches the affected tables). The local index is disposable until
+    (and any code that touches the affected tables) directly; do not add new
-    1.0 ships and is tagged.
+    numbered schema files. Those numbered files and the `schema_migrations`
+    table they populate only bootstrap a fresh database — they are not an
+    upgrade path. The local index is disposable until 1.0 ships and is
+    tagged; once 1.0 is tagged that clause expires and the question of
+    upgrading existing indexes returns. See [`docs/DATAMODEL.md`](docs/DATAMODEL.md)
+    for the full explanation.
 

ARCHITECTURE.md

@@ -63,7 +63,7 @@ A content-addressed unit of data. Files are split into variable-size chunks usin
 - `ChunkHash`: SHA256 hash of chunk content (primary key)
 - `Size`: Chunk size in bytes
 
-Chunk sizes vary between `avgChunkSize/4` and `avgChunkSize*4` (typically 16KB-256KB for 64KB average).
+Chunk sizes vary between `avgChunkSize/4` and `avgChunkSize*4` (2.5MB-40MB for the 10MB default average).
 
 #### FileChunk (`database.FileChunk`)
 Maps files to their constituent chunks:
@@ -120,7 +120,7 @@ The CLI uses fx for dependency injection. Here's the instantiation order:
 ```go
 // cli/app.go: NewApp()
 fx.New(
-    fx.Supply(config.ConfigPath(opts.ConfigPath)),  // 1. Config path
+    fx.Supply(config.Path(opts.ConfigPath)),         // 1. Config path
     fx.Supply(opts.LogOptions),                      // 2. Log options
     fx.Provide(globals.New),                         // 3. Globals
     fx.Provide(log.New),                             // 4. Logger config
@@ -193,7 +193,7 @@ scanner := v.ScannerFactory(snapshot.ScannerParams{
 - **Created by**: `chunker.NewChunker(avgChunkSize)`
 - **When**: Inside `snapshot.NewScanner()`
 - **Configuration**:
-  - `avgChunkSize`: From config (typically 64KB)
+  - `avgChunkSize`: From config (default 10MB)
   - `minChunkSize`: avgChunkSize / 4
   - `maxChunkSize`: avgChunkSize * 4
 
@@ -286,7 +286,6 @@ Key methods:
 - `CreateSnapshot(ctx, hostname, version, commit)` → Create snapshot record
 - `CompleteSnapshot(ctx, snapshotID)` → Mark snapshot complete
 - `ExportSnapshotMetadata(ctx, dbPath, snapshotID)` → Export to S3
-- `CleanupIncompleteSnapshots(ctx, hostname)` → Remove failed snapshots
 
 ### `internal/database`
 SQLite database for local index. Single-writer mode for thread safety.
@@ -307,7 +306,7 @@ Repository interfaces:
 ```
 CreateSnapshot(opts)
     │
-    ├─► CleanupIncompleteSnapshots()   // Critical: avoid dedup errors
+    ├─► PruneDatabase()   // Critical: avoid dedup errors
     │
     ├─► SnapshotManager.CreateSnapshot()   // Create DB record
     │

README.md

@@ -84,6 +84,57 @@ VAULTIK_AGE_SECRET_KEY='AGE-SECRET-KEY-...' vaultik snapshot restore <snapshot-i
 # 0 3 * * * vaultik snapshot create --cron --prune --keep-newer-than 4w
 ```
 
+## restoring on another machine
+
+Restoring on a host that never ran the backup — a replacement machine
+after the original is gone — is the case vaultik is built for. That host
+needs only three things: the `vaultik` binary, the age **private** key,
+and the storage credentials for the destination. It does **not** need the
+local index, the original config file, or the original hostname.
+
+```sh
+# install
+go install sneak.berlin/go/vaultik/cmd/vaultik@latest
+
+# create a config and point it at the ORIGINAL backup destination
+vaultik config init
+vaultik config set storage_url "s3://bucket/prefix?endpoint=https://s3.example.com"
+vaultik config set s3.access_key_id "..."
+vaultik config set s3.secret_access_key "..."
+
+# see what is on the destination store
+vaultik snapshot list
+```
+
+`snapshot list` reads the destination store without the private key. A
+snapshot that is not in this host's (empty) local index is shown as
+remote-only: its row is identified by `<remote only:...>` rather than by
+a `hostname_name_timestamp` name, because the name lives only in the
+local index and the encrypted database and cannot be recovered from the
+store. Its timestamp and compressed size are real. (See the `snapshot
+list` description under [command details](#command-details) for the full
+explanation.)
+
+Use that remote key — the hex printed inside `<remote only:...>`, or the
+full `remote_key` from `snapshot list --json` — to restore and verify:
+
+```sh
+# restore everything to /tmp/restored, then check every restored file's
+# chunk hashes
+VAULTIK_AGE_SECRET_KEY='AGE-SECRET-KEY-...' \
+  vaultik snapshot restore --verify <remote-key> /tmp/restored
+
+# optionally, deep-verify the snapshot against the store (downloads and
+# cryptographically checks every blob)
+VAULTIK_AGE_SECRET_KEY='AGE-SECRET-KEY-...' \
+  vaultik snapshot verify --deep <remote-key>
+```
+
+`age_recipients` (the public key) is not needed to restore — only the
+private key in `VAULTIK_AGE_SECRET_KEY`. Both the abbreviated key printed
+in the table and the full 64-character key from `--json` are accepted; a
+leading part of the key is enough as long as it is unambiguous.
+
 ---
 
 ## cli
@@ -96,10 +147,10 @@ vaultik [--config <path>] config edit
 vaultik [--config <path>] config get <key>
 vaultik [--config <path>] config set <key> <value>
 vaultik [--config <path>] snapshot create [snapshot-names...] [--cron] [--prune] [--keep-newer-than <duration>]
-vaultik [--config <path>] snapshot list [--json]
+vaultik [--config <path>] snapshot list [--json]                  # alias: ls
 vaultik [--config <path>] snapshot verify <snapshot-id> [--deep] [--json]
 vaultik [--config <path>] snapshot purge [--keep-latest | --older-than <duration>] [--snapshot <name>...] [--force]
-vaultik [--config <path>] snapshot remove <snapshot-id> [--dry-run] [--force] [--local-only] [--json]
+vaultik [--config <path>] snapshot remove <snapshot-id> [--dry-run] [--force] [--local-only] [--json]  # alias: rm
 vaultik [--config <path>] snapshot restore <snapshot-id> <target-dir> [paths...] [--verify]
 vaultik [--config <path>] prune [--force] [--json]
 vaultik [--config <path>] info
@@ -116,7 +167,24 @@ vaultik version
 * `--verbose`, `-v`: Enable verbose output (on stderr — see below)
 * `--debug`: Enable debug output (on stderr — see below)
 * `--quiet`, `-q`: Suppress non-error output (also suppresses startup banner)
-* `--skip-errors`: Continue past per-file errors instead of aborting (applies to `snapshot create` and `restore`)
+* `--skip-errors`: Skip files that cannot be read when creating a snapshot, or that cannot be restored when restoring, instead of aborting. Packing and storage errors (which would leave a chunk recorded but not stored) still abort the run.
+
+### locking
+
+Commands that write persistent state — `snapshot create`, `snapshot
+purge`, `snapshot remove`, `prune`, and `remote nuke` — take a
+process-wide lock at `$XDG_DATA_HOME/vaultik/vaultik.pid`
+(`~/.local/share/vaultik/vaultik.pid` on Linux) for the whole run. Only
+one of them runs at a time: a second one exits immediately with an
+"already running" error rather than waiting, so two writers can never
+corrupt the local index or the destination store.
+
+Read-only commands — `info`, `snapshot list`, `snapshot verify`, and
+`remote info` — do not take the lock and are never blocked, so they run
+even while a backup is in progress. `snapshot restore` does not take the
+lock either: it writes only to the target directory you name, not the
+local index or the destination store. `config`, `database delete`,
+`completion`, and `version` do not take the lock.
 
 ### stdout and stderr
 
@@ -152,6 +220,8 @@ and `vaultik prune --json | jq .` both work as written.
 * `VAULTIK_AGE_SECRET_KEY`: Age private key for decryption (required for `snapshot restore` and `snapshot verify --deep`)
 * `VAULTIK_CONFIG`: Path to config file (overridden by `--config`)
 * `VAULTIK_INDEX_PATH`: Override local SQLite index path
+* `VAULTIK_CPUPROFILE`: Write a CPU profile to this path for the duration of the run (development/debugging)
+* `VAULTIK_MEMPROFILE`: Write a heap profile to this path when the run exits (development/debugging)
 
 ### shell completion
 
@@ -245,6 +315,8 @@ local index alone, and still exits zero.
 * Default (shallow): checks that all blobs referenced in the manifest exist in storage
 * `--deep`: Downloads and decrypts each blob, verifies chunk hashes against the
   encrypted metadata database
+* Accepts the same identifiers as `snapshot restore`: a snapshot ID, or a
+  remote-only snapshot's remote key (or an unambiguous leading part of it)
 * `--json`: Output results as JSON
 
 **`snapshot purge`**: Remove old snapshots based on criteria. Retention is
@@ -275,6 +347,10 @@ on the destination in one go, use `vaultik remote nuke --force`.
 
 **`snapshot restore`**: Restore files from a backup snapshot.
 * Requires `VAULTIK_AGE_SECRET_KEY` environment variable
+* Accepts a snapshot ID, or — for a snapshot only on the destination
+  store — its remote key (or an unambiguous leading part of it) as shown
+  by `snapshot list`. See
+  [restoring on another machine](#restoring-on-another-machine).
 * Optional path arguments to restore specific files/directories (default: all)
 * Preserves file permissions, timestamps, ownership (ownership requires root),
   symlinks, and empty directories
@@ -338,6 +414,10 @@ both are set.
 
 ## architecture
 
+For an implementation-level view of the internals — the data model, the
+`fx` dependency-injection wiring, and the scanner — see
+[`ARCHITECTURE.md`](ARCHITECTURE.md).
+
 ### remote storage layout
 
 ```
@@ -415,25 +495,30 @@ derivation.
 
 ### compression
 
-* zstd compression at configurable level (1-19, default 3)
+* zstd compression at configurable level (1-19, default 3). The level is
+  accepted as 1-19 but maps onto zstd's four internal speed presets:
+  1-2 fastest, 3-5 default, 6-9 better, 10-19 best. Levels within the
+  same band compress identically.
 * Applied before encryption at the blob level
 
 ---
 
 ## configuration reference
 
-Run `vaultik config init` to generate a fully commented config file.
+Run `vaultik config init` to generate a fully commented config file; a
-Key fields:
+complete annotated example also lives in
+[`config.example.yml`](config.example.yml). Key fields:
 
 | Field | Default | Description |
 |-------|---------|-------------|
 | `age_recipients` | (required) | Age public keys for encryption |
+| `age_secret_key` | (unset) | Age private key for decryption (`snapshot restore`, `snapshot verify --deep`). Setting it in the config file places the private key on the backed-up host, defeating the public-key-only design (see "why" above). Prefer the `VAULTIK_AGE_SECRET_KEY` environment variable, supplied only on the machine you restore from. |
 | `snapshots` | (required) | Named snapshot definitions with paths and excludes |
 | `storage_url` | | Storage backend URL (`s3://`, `file://`, `rclone://`) |
 | `s3.*` | | Legacy S3 configuration (endpoint, bucket, credentials) |
 | `exclude` | | Global exclude patterns (applied to all snapshots) |
 | `chunk_size` | `10MB` | Average chunk size for content-defined chunking |
-| `blob_size_limit` | `10GB` | Maximum blob size before splitting |
+| `blob_size_limit` | `10GB` | Maximum blob size before splitting. Must be at least four times `chunk_size` (the largest chunk the chunker can emit), otherwise a single-chunk blob could exceed the limit |
 | `compression_level` | `3` | zstd compression level (1-19) |
 | `hostname` | system hostname | Hostname used in snapshot IDs |
 | `index_path` | platform data dir | Local SQLite index path |
@@ -457,9 +542,13 @@ Key fields:
   sequentially. Restore speed is bound by single-stream throughput.
 * **Device nodes, named pipes, and sockets are silently skipped.** Only
   regular files, directories, and symlinks are backed up.
-* **No database migrations.** If the local SQLite schema changes between
+* **No upgrade path between versions.** There is no supported way to carry
-  versions, delete the local database (`vaultik database delete`) and run
+  an existing local index across a schema change; if the local SQLite
-  a full backup. Remote storage is unaffected.
+  schema changes between versions, delete the local database (`vaultik
+  database delete`) and run a full backup. Remote storage is unaffected.
+  (The binary does embed numbered schema files and a `schema_migrations`
+  table to bootstrap a fresh database — see [`docs/DATAMODEL.md`](docs/DATAMODEL.md)
+  — but that is not an upgrade path.)
 * **Files that change during backup may be inconsistent.** There is no
   filesystem snapshot or freeze. If a file is modified between the scan
   and chunk phases, the backed-up copy may reflect a partial write.
@@ -525,14 +614,12 @@ priority.
 
 ### infrastructure
 
-* **Cross-machine restore documentation.** The "restore from
+* **Cross-version schema upgrades.** There is no upgrade path between
-  another host" workflow works but isn't documented as a
+  released versions — pre-1.0 schema changes are handled by `vaultik
-  first-class operation in this README. Worth a dedicated section
+  database delete` plus a full re-scan (see
-  once it's settled.
+  [`docs/DATAMODEL.md`](docs/DATAMODEL.md)). Post-1.0 we'll need a
-* **Schema migrations.** Currently nonexistent — pre-1.0 schema
+  migration story to keep existing index databases usable across
-  changes are handled by `vaultik database delete` plus a full
+  upgrades.
-  re-scan. Post-1.0 we'll need a migration story to keep existing
-  index databases usable across upgrades.
 * **Storage backend coverage tests.** S3, file://, and rclone://
   all share the Storer interface but the rclone path is the least
   exercised in CI.
@@ -541,9 +628,17 @@ priority.
 
 ## output style
 
-All user-facing output goes through helpers in `internal/ui` and conforms
+The operational narration of the long-running commands — the Begin,
-to a uniform style. Color is enabled when stdout is a TTY and the
+Complete, Progress, and status lines of `snapshot create`, `prune`,
-`NO_COLOR` environment variable is unset (https://no-color.org/).
+`snapshot restore`, and the like — goes through helpers in `internal/ui`
+and conforms to the uniform style below. Some commands instead write
+plain text straight to stdout (`version`, `info`, `config`, the
+`database delete` prompt, and the `snapshot list` table); that output is
+unstyled and does not honor `--quiet`. Routing it through `internal/ui`
+is tracked in
+[issue #149](https://git.eeqj.de/sneak/vaultik/issues/149). Color is
+enabled when stdout is a TTY and the `NO_COLOR` environment variable is
+unset (https://no-color.org/).
 
 `internal/ui` writes to stdout; it is the output the user asked for.
 Structured log records are a different thing and go through

TODO.md

@@ -25,6 +25,53 @@ release" is exactly the contradiction
 
 # Completed Steps
 
+- 2026-09-21: Stopped an interrupted blob upload from making a later
+  backup deduplicate against data that was never stored
+  ([issue #148](https://git.eeqj.de/sneak/vaultik/issues/148)). The
+  packer commits a blob's `chunks`, `blob_chunks`, and `blobs` rows
+  before the upload is attempted, so a failed upload left chunk rows
+  behind and the next run skipped re-uploading them, producing a
+  snapshot that reported success but could not be restored. A run now
+  deduplicates only against chunks held by a blob whose `uploaded_ts` is
+  set, and at startup drops any un-uploaded blob rows (and the chunks
+  they orphan) so the affected data is re-chunked and re-uploaded. Blobs
+  recorded with no remote backend are marked uploaded so this invariant
+  holds uniformly.
+
+- 2026-09-22: Made restore refuse any snapshot path that would write
+  outside the target directory
+  ([issue #154](https://git.eeqj.de/sneak/vaultik/issues/154)).
+  `restoreFile` and `verifyRestoredFiles` joined the stored path onto the
+  target with no containment check, so a `..` segment or an absolute path
+  escaped the target and a restored symlink could redirect a later child
+  write anywhere on disk. Every stored path is now rejected unless
+  `filepath.IsLocal` accepts it with the leading separator removed, and
+  each existing ancestor directory below the target is `Lstat`ed to refuse
+  descending through a symlink; honest symlinks pointing outside the tree
+  are still written verbatim. age decryption proves a snapshot is
+  readable, not honest, and restore usually runs as root.
+
+- 2026-09-21: Stopped `--json` from silencing stderr diagnostics
+  ([issue #112](https://git.eeqj.de/sneak/vaultik/issues/112)). `--json`
+  used to be folded into `Quiet`, which pinned the log level to `WARN`,
+  so `prune --json` gave a machine consumer no record of the local index
+  rows it deleted even under `--verbose`. `--json` now quiets only the
+  stdout UI (the JSON document must stay clean, per
+  [issue #108](https://git.eeqj.de/sneak/vaultik/issues/108)); the stderr
+  log level follows `--verbose`/`--debug` again. The coupling was
+  removed the same way for `snapshot verify`, `snapshot remove`, and
+  `remote info`, which carried it for the same outdated reason.
+
+- 2026-09-21: Stopped `prune` from reporting a failed row count as 0
+  ([issue #96](https://git.eeqj.de/sneak/vaultik/issues/96)). The seven
+  `getTableCount` reads in `PruneDatabase` discarded their error, so a
+  query that could not run became a plausible `0` and the before/after
+  delta computed from it looked like real work. Each read now logs at
+  warn on failure and renders as `unknown`, never `0`, so an empty table
+  is distinguishable from one that could not be queried. The counts have
+  no `--json` representation — under `--json` the summary is suppressed
+  entirely — so nothing there can show a false `0`.
+
 - 2026-09-21: Made the s3 storage backend report a missing object as
   `storage.ErrNotFound`, like the `file` and `rclone` backends and as the
   `Storer` interface documents. `S3Storer.Get` and `Stat` returned the raw
@@ -33,7 +80,6 @@ release" is exactly the contradiction
   helper (reused by `HeadObject`) and a test that a missing key maps to
   `ErrNotFound`
   ([issue #129](https://git.eeqj.de/sneak/vaultik/issues/129)).
-
 - 2026-09-21: Fixed `verify --deep` reporting healthy snapshots as
   corrupt. Its final blob-integrity check hashed the encrypted
   downloaded bytes with a single SHA256 and compared that to the blob

config.example.yml

@@ -257,16 +257,16 @@ exclude:
 
 # Storage URL - use either this OR the s3 section below
 # Supports: s3://bucket/prefix, file:///path, rclone://remote/path
-storage_url: "rclone://las1stor1//srv/pool.2024.04/backups/heraklion"
+storage_url: "rclone://myremote/path/to/backups"
 
 # S3-compatible storage configuration
 #s3:
 #  # S3-compatible endpoint URL
 #  # Examples: https://s3.amazonaws.com, https://storage.googleapis.com
-#  endpoint: http://10.100.205.122:8333
+#  endpoint: https://s3.example.com
 #
 #  # Bucket name where backups will be stored
-#  bucket: testbucket
+#  bucket: mybucket
 #
 #  # Prefix (folder) within the bucket for this host's backups
 #  # Useful for organizing backups from multiple hosts
@@ -274,8 +274,8 @@ storage_url: "rclone://las1stor1//srv/pool.2024.04/backups/heraklion"
 #  #prefix: "hosts/myserver/"
 #
 #  # S3 access credentials
-#  access_key_id: Z9GT22M9YFU08WRMC5D4
+#  access_key_id: YOUR_ACCESS_KEY
-#  secret_access_key: Pi0tPKjFbN4rZlRhcA4zBtEkib04yy2WcIzI+AXk
+#  secret_access_key: YOUR_SECRET_KEY
 #
 #  # S3 region
 #  # Default: us-east-1
@@ -304,6 +304,8 @@ storage_url: "rclone://las1stor1//srv/pool.2024.04/backups/heraklion"
 
 # Maximum blob size
 # Multiple chunks are packed into blobs up to this size
+# Must be at least four times chunk_size (the largest chunk the chunker can
+# emit); a smaller limit would let a single-chunk blob exceed it.
 # Supports: 1GB, 10G, 500MB, 1GiB, etc.
 # Default: 10GB
 #blob_size_limit: 10GB

docs/DATAMODEL.md

@@ -5,11 +5,30 @@
 Vaultik uses a local SQLite database to track file metadata, chunk mappings, and blob associations during the backup process. This database serves as an index for incremental backups and enables efficient deduplication.
 
 **Important Notes:**
-- **No Migration Support (pre-1.0)**: Vaultik does not support database schema
+
-  migrations. The local index is treated as disposable — if the schema changes,
+This section is the authoritative explanation of the schema/migration story;
-  delete the local SQLite database (`vaultik database delete`) and run a full
+other documents (the README and `AGENTS.md`) link here.
-  backup. The remote storage is unaffected; the new index will re-deduplicate
+
-  against existing remote blobs.
+- **No upgrade path between versions (pre-1.0)**: Vaultik has no supported way to
+  carry an existing local index across a schema change. The index is disposable
+  — if the on-disk schema changes between versions, delete the local SQLite
+  database (`vaultik database delete`) and run a full backup. Remote storage is
+  unaffected; the new index re-deduplicates against existing remote blobs. This
+  is the standing project policy, and it is separate from the schema bootstrap
+  described next.
+- **Schema bootstrap**: a fresh database is populated from numbered SQL files
+  embedded in the binary under `internal/database/schema/`. `000.sql` creates the
+  `schema_migrations` table; `001.sql` creates the application tables. On opening
+  a database the code applies each numbered file that has not yet run and records
+  its version in `schema_migrations`. This bootstraps a new database; it does not
+  upgrade an existing one between released versions.
+- **Changing the schema (pre-1.0)**: edit `internal/database/schema/001.sql` (and
+  the code that touches the affected tables) directly. Do not add new numbered
+  files — there is no installed base to migrate.
+- **Disposability expires at 1.0**: the index is treated as disposable only until
+  1.0 ships and is tagged. Once 1.0 is tagged that clause expires and the
+  question of upgrading existing indexes returns. It is deliberately left open
+  here.
 - **Version Compatibility**: In rare cases, you may need to use the same version
   of Vaultik to restore a backup as was used to create it. This ensures
   compatibility with the metadata format stored in S3.

docs/REPOSTRUCTURE.md

@@ -145,10 +145,10 @@ An observer cannot determine:
 ## Pruning Safety
 
 The prune operation is safe because:
-1. It only deletes blobs not referenced in any manifest
+1. It keeps every blob listed in any snapshot's manifest and deletes only blobs that no manifest references
 2. Manifests are unencrypted and can be read without keys
-3. The operation compares the latest local DB snapshot with the latest S3 snapshot to ensure consistency
+3. If any manifest cannot be downloaded or decoded, prune deletes nothing and exits with an error, rather than treating that snapshot's blobs as unreferenced
-4. Pruning will fail if these don't match, preventing accidental deletion of needed blobs
+4. Prune requires exclusive access to the destination: running it during a concurrent backup can race a snapshot whose manifest is not yet written, so do not prune while a backup is in progress
 
 ## Restoration Requirements
 

internal/blob/packer.go

@@ -487,7 +487,7 @@ func (p *Packer) closeBlobWriter() (string, int64, error) {
 		return "", 0, fmt.Errorf("seeking to start: %w", err)
 	}
 
-	finalHash := p.currentBlob.writer.Sum256()
+	finalHash := p.currentBlob.writer.ContentID()
 
 	return hex.EncodeToString(finalHash), finalSize, nil
 }

internal/blobgen/compress.go

@@ -1,89 +0,0 @@
-// Package blobgen implements the blob data pipeline: streaming zstd
-// compression, age encryption, and SHA256 content hashing for blob
-// creation, plus the matching decrypt/decompress/verify reader.
-package blobgen
-
-import (
-	"bytes"
-	"encoding/hex"
-	"fmt"
-	"io"
-)
-
-// CompressResult contains the results of compression
-type CompressResult struct {
-	Data             []byte
-	UncompressedSize int64
-	CompressedSize   int64
-	SHA256           string
-}
-
-// CompressData compresses and encrypts data, returning the result with hash
-func CompressData(
-	data []byte, compressionLevel int, recipients []string,
-) (*CompressResult, error) {
-	var buf bytes.Buffer
-
-	// Create writer
-	w, err := NewWriter(&buf, compressionLevel, recipients)
-	if err != nil {
-		return nil, fmt.Errorf("creating writer: %w", err)
-	}
-
-	// Write data
-	_, err = w.Write(data)
-	if err != nil {
-		_ = w.Close()
-
-		return nil, fmt.Errorf("writing data: %w", err)
-	}
-
-	// Close to flush
-	err = w.Close()
-	if err != nil {
-		return nil, fmt.Errorf("closing writer: %w", err)
-	}
-
-	return &CompressResult{
-		Data:             buf.Bytes(),
-		UncompressedSize: int64(len(data)),
-		CompressedSize:   int64(buf.Len()),
-		SHA256:           hex.EncodeToString(w.Sum256()),
-	}, nil
-}
-
-// CompressStream compresses and encrypts from reader to writer, returning
-// the number of uncompressed bytes written and the content hash.
-func CompressStream(
-	dst io.Writer, src io.Reader, compressionLevel int, recipients []string,
-) (int64, string, error) {
-	// Create writer
-	w, err := NewWriter(dst, compressionLevel, recipients)
-	if err != nil {
-		return 0, "", fmt.Errorf("creating writer: %w", err)
-	}
-
-	closed := false
-
-	defer func() {
-		if !closed {
-			_ = w.Close()
-		}
-	}()
-
-	// Copy data
-	_, err = io.Copy(w, src)
-	if err != nil {
-		return 0, "", fmt.Errorf("copying data: %w", err)
-	}
-
-	// Close to flush
-	err = w.Close()
-	if err != nil {
-		return 0, "", fmt.Errorf("closing writer: %w", err)
-	}
-
-	closed = true
-
-	return w.BytesWritten(), hex.EncodeToString(w.Sum256()), nil
-}

internal/blobgen/compress_test.go

@@ -1,80 +0,0 @@
-package blobgen_test
-
-import (
-	"bytes"
-	"crypto/rand"
-	"strings"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-	"sneak.berlin/go/vaultik/internal/blobgen"
-)
-
-// testRecipient is a static age recipient for tests.
-const testRecipient = "age1cplgrwj77ta54dnmydvvmzn64ltk83ankxl5sww04mrtmu62kv3s89gmvv"
-
-// TestCompressStreamNoDoubleClose is a regression test for issue #28.
-// It verifies that CompressStream does not panic or return an error due to
-// double-closing the underlying blobgen.Writer. Before the fix in PR #33,
-// the explicit Close() on the happy path combined with defer Close() would
-// cause a double close.
-func TestCompressStreamNoDoubleClose(t *testing.T) {
-	t.Parallel()
-
-	input := []byte("regression test data for issue #28 double-close fix")
-
-	var buf bytes.Buffer
-
-	written, hash, err := blobgen.CompressStream(
-		&buf, bytes.NewReader(input), 3, []string{testRecipient})
-	require.NoError(t, err, "CompressStream should not return an error")
-	assert.Positive(t, written, "expected bytes written > 0")
-	assert.NotEmpty(t, hash, "expected non-empty hash")
-	assert.Positive(t, buf.Len(), "expected non-empty output")
-}
-
-// TestCompressStreamLargeInput exercises CompressStream with a larger payload
-// to ensure no double-close issues surface under heavier I/O.
-func TestCompressStreamLargeInput(t *testing.T) {
-	t.Parallel()
-
-	data := make([]byte, 512*1024) // 512 KB
-	_, err := rand.Read(data)
-	require.NoError(t, err)
-
-	var buf bytes.Buffer
-
-	written, hash, err := blobgen.CompressStream(
-		&buf, bytes.NewReader(data), 3, []string{testRecipient})
-	require.NoError(t, err)
-	assert.Positive(t, written)
-	assert.NotEmpty(t, hash)
-}
-
-// TestCompressStreamEmptyInput verifies CompressStream handles empty input
-// without double-close issues.
-func TestCompressStreamEmptyInput(t *testing.T) {
-	t.Parallel()
-
-	var buf bytes.Buffer
-
-	_, hash, err := blobgen.CompressStream(
-		&buf, strings.NewReader(""), 3, []string{testRecipient})
-	require.NoError(t, err)
-	assert.NotEmpty(t, hash)
-}
-
-// TestCompressDataNoDoubleClose mirrors the stream test for CompressData,
-// ensuring the explicit Close + error-path Close pattern is also safe.
-func TestCompressDataNoDoubleClose(t *testing.T) {
-	t.Parallel()
-
-	input := []byte("CompressData regression test for double-close")
-
-	result, err := blobgen.CompressData(input, 3, []string{testRecipient})
-	require.NoError(t, err)
-	assert.Positive(t, result.CompressedSize)
-	assert.Equal(t, result.UncompressedSize, int64(len(input)))
-	assert.NotEmpty(t, result.SHA256)
-}

internal/blobgen/reader.go

@@ -64,7 +64,9 @@ func (r *Reader) Close() error {
 	return nil
 }
 
-// Sum256 returns the SHA256 hash of all data read
+// Sum256 returns the single SHA-256 of the plaintext read so far. This is the
+// first hash only; the stored object name is its double hash, which callers
+// obtain by passing this digest to DoubleSHA256.
 func (r *Reader) Sum256() []byte {
 	return r.hasher.Sum(nil)
 }

internal/blobgen/reader_test.go

@@ -0,0 +1,54 @@
+package blobgen_test
+
+import (
+	"bytes"
+	"io"
+	"testing"
+
+	"filippo.io/age"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/blobgen"
+)
+
+// TestMultipleRecipients verifies that data written for several recipients can
+// be read back by each recipient's identity. Moved from internal/crypto, which
+// held the only multi-recipient test; blobgen is now the sole encryption path.
+func TestMultipleRecipients(t *testing.T) {
+	t.Parallel()
+
+	identities := make([]*age.X25519Identity, 3)
+	recipients := make([]string, 3)
+
+	for i := range identities {
+		identity, err := age.GenerateX25519Identity()
+		require.NoError(t, err)
+
+		identities[i] = identity
+		recipients[i] = identity.Recipient().String()
+	}
+
+	plaintext := []byte("Secret message for multiple recipients")
+
+	var encrypted bytes.Buffer
+
+	writer, err := blobgen.NewWriter(&encrypted, 3, recipients)
+	require.NoError(t, err)
+	_, err = writer.Write(plaintext)
+	require.NoError(t, err)
+	require.NoError(t, writer.Close())
+
+	// Every recipient's identity must recover the original plaintext.
+	for i, identity := range identities {
+		reader, err := blobgen.NewReader(
+			bytes.NewReader(encrypted.Bytes()), identity)
+		require.NoError(t, err, "recipient %d should open the reader", i+1)
+
+		got, err := io.ReadAll(reader)
+		require.NoError(t, err, "recipient %d should read the plaintext", i+1)
+		require.NoError(t, reader.Close())
+
+		assert.Equal(t, plaintext, got,
+			"recipient %d should recover the original plaintext", i+1)
+	}
+}

internal/blobgen/writer.go

@@ -1,3 +1,6 @@
+// Package blobgen implements the blob data pipeline: streaming zstd
+// compression, age encryption, and SHA256 content hashing for blob
+// creation, plus the matching decrypt/decompress/verify reader.
 package blobgen
 
 import (
@@ -12,6 +15,18 @@ import (
 	"github.com/klauspost/compress/zstd"
 )
 
+// DoubleSHA256 returns the double SHA-256 of content whose single SHA-256
+// digest is sum: it hashes that digest once more. Stored objects are named by
+// this second hash so that a name never reveals whether known content is
+// present — an attacker who knows a plaintext, and thus its SHA-256, still
+// cannot derive the stored name without hashing the digest again. Both a blob
+// and the metadata database export are named this way.
+func DoubleSHA256(sum []byte) []byte {
+	h := sha256.Sum256(sum)
+
+	return h[:]
+}
+
 // Zstd compression level bounds accepted by NewWriter.
 const (
 	minCompressionLevel = 1
@@ -27,6 +42,11 @@ const reservedCompressionCPUs = 2
 var ErrInvalidCompressionLevel = errors.New(
 	"invalid compression level: must be between 1 and 19")
 
+// errInvalidRecipient is returned when a recipient string does not parse as
+// an X25519 age1... public key. It omits the value, which can be sensitive.
+var errInvalidRecipient = errors.New(
+	"not a valid X25519 age1... recipient")
+
 // Writer wraps compression and encryption with SHA256 hashing.
 // Data flows: input -> tee(hasher, compressor -> encryptor -> destination)
 // The hash is computed on the uncompressed input for deterministic content-addressing.
@@ -57,10 +77,12 @@ func NewWriter(
 	// Parse recipients
 	var ageRecipients []age.Recipient
 
-	for _, recipient := range recipients {
+	for i, recipient := range recipients {
+		// The recipient string can be sensitive (e.g. a secret key pasted by
+		// mistake), so the error names its position, never its value.
 		r, err := age.ParseX25519Recipient(recipient)
 		if err != nil {
-			return nil, fmt.Errorf("parsing recipient %s: %w", recipient, err)
+			return nil, fmt.Errorf("%w: recipient %d", errInvalidRecipient, i)
 		}
 
 		ageRecipients = append(ageRecipients, r)
@@ -123,17 +145,12 @@ func (w *Writer) Close() error {
 	return nil
 }
 
-// Sum256 returns the double SHA256 hash of the uncompressed input data.
+// ContentID returns the double SHA-256 of the uncompressed input data: the
-// Double hashing (SHA256(SHA256(data))) prevents information leakage about
+// name under which this content is stored. It is the second hash of the
-// the plaintext - an attacker cannot confirm existence of known content
+// running SHA-256, via DoubleSHA256; see that function for why content is
-// by computing its hash and checking for a matching blob filename.
+// named this way rather than by its plain SHA-256.
-func (w *Writer) Sum256() []byte {
+func (w *Writer) ContentID() []byte {
-	// First hash: SHA256(plaintext)
+	return DoubleSHA256(w.hasher.Sum(nil))
-	firstHash := w.hasher.Sum(nil)
-	// Second hash: SHA256(firstHash) - this is the blob ID
-	secondHash := sha256.Sum256(firstHash)
-
-	return secondHash[:]
 }
 
 // BytesWritten returns the number of uncompressed bytes written

internal/blobgen/writer_test.go

@@ -12,7 +12,7 @@ import (
 	"sneak.berlin/go/vaultik/internal/blobgen"
 )
 
-// TestWriterHashIsDoubleHash verifies that Writer.Sum256() returns
+// TestWriterHashIsDoubleHash verifies that Writer.ContentID() returns
 // the double hash SHA256(SHA256(plaintext)) for security.
 // Double hashing prevents attackers from confirming existence of known content.
 func TestWriterHashIsDoubleHash(t *testing.T) {
@@ -43,7 +43,7 @@ func TestWriterHashIsDoubleHash(t *testing.T) {
 	require.NoError(t, err)
 
 	// Get the hash from the writer
-	writerHash := hex.EncodeToString(writer.Sum256())
+	writerHash := hex.EncodeToString(writer.ContentID())
 
 	// Calculate the expected double hash: SHA256(SHA256(plaintext))
 	firstHash := sha256.Sum256(testData)
@@ -60,7 +60,7 @@ func TestWriterHashIsDoubleHash(t *testing.T) {
 
 	// The writer hash should match the double hash
 	assert.Equal(t, expectedDoubleHash, writerHash,
-		"Writer.Sum256() should return SHA256(SHA256(plaintext)) for security")
+		"Writer.ContentID() should return SHA256(SHA256(plaintext)) for security")
 
 	// Verify it's NOT the single hash (would leak information)
 	assert.NotEqual(t, singleHashStr, writerHash,
@@ -93,8 +93,8 @@ func TestWriterDeterministicHash(t *testing.T) {
 	require.NoError(t, err)
 	require.NoError(t, writer2.Close())
 
-	hash1 := hex.EncodeToString(writer1.Sum256())
+	hash1 := hex.EncodeToString(writer1.ContentID())
-	hash2 := hex.EncodeToString(writer2.Sum256())
+	hash2 := hex.EncodeToString(writer2.ContentID())
 
 	// Hashes should be identical (deterministic)
 	assert.Equal(t, hash1, hash2, "Same input should produce same hash")
@@ -108,3 +108,20 @@ func TestWriterDeterministicHash(t *testing.T) {
 	t.Logf("Encrypted size 1: %d bytes", buf1.Len())
 	t.Logf("Encrypted size 2: %d bytes", buf2.Len())
 }
+
+// TestNewWriterSecretKeyNotEchoed verifies that a secret key mistakenly passed
+// as a recipient does not appear in the returned error. A recipient string can
+// be sensitive, so the error must name only the position, not the value.
+func TestNewWriterSecretKeyNotEchoed(t *testing.T) {
+	t.Parallel()
+
+	secretKey := "AGE-SECRET-KEY-19CR5YSFW59HM4TLD6GX" +
+		"VEDMZFTVVF7PPHKUT68TXSFPK7APHXA2QS2NJA5"
+
+	var buf bytes.Buffer
+
+	_, err := blobgen.NewWriter(&buf, 3, []string{secretKey})
+	require.Error(t, err)
+	assert.NotContains(t, err.Error(), secretKey,
+		"error must not echo the recipient value")
+}

internal/chunker/chunker.go

@@ -33,9 +33,10 @@ type Chunker struct {
 	maxChunkSize int
 }
 
-// chunkSizeSpread is the FastCDC-recommended factor between the average
+// ChunkSizeSpread is the FastCDC-recommended factor between the average
 // chunk size and the minimum (avg/spread) and maximum (avg*spread) sizes.
-const chunkSizeSpread = 4
+// The largest chunk the chunker can emit is therefore avg*ChunkSizeSpread.
+const ChunkSizeSpread = 4
 
 // NewChunker creates a new chunker with the specified average chunk size.
 // The actual chunk sizes will vary between avgChunkSize/4 and avgChunkSize*4
@@ -45,8 +46,8 @@ func NewChunker(avgChunkSize int64) *Chunker {
 	// FastCDC recommends min = avg/4 and max = avg*4
 	return &Chunker{
 		avgChunkSize: int(avgChunkSize),
-		minChunkSize: int(avgChunkSize / chunkSizeSpread),
+		minChunkSize: int(avgChunkSize / ChunkSizeSpread),
-		maxChunkSize: int(avgChunkSize * chunkSizeSpread),
+		maxChunkSize: int(avgChunkSize * ChunkSizeSpread),
 	}
 }
 

internal/cli/app.go

@@ -7,12 +7,9 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"os"
-	"os/signal"
 	"path/filepath"
 	"strings"
 	"sync"
-	"syscall"
 	"time"
 
 	"github.com/adrg/xdg"
@@ -33,14 +30,33 @@ import (
 // may take before we give up.
 const shutdownTimeout = 30 * time.Second
 
-// AppOptions contains common options for creating the fx application.
+// lockMode says whether a command mutates persistent state — the local
-// It includes the configuration file path, logging options, and additional
+// index database or the remote store — and so must hold the process-wide
-// fx modules and invocations that should be included in the application.
+// PID lock, or only reads that state and may run alongside a mutator.
+type lockMode int
+
+const (
+	// mutating commands (snapshot create, snapshot purge, snapshot remove,
+	// prune, remote nuke) write the local index or the remote store. They
+	// hold the PID lock so that at most one runs at a time.
+	mutating lockMode = iota
+	// readOnly commands (info, snapshot list, snapshot verify, remote info,
+	// snapshot restore) do not write the local index or the remote store,
+	// so they run without the lock and are never blocked by a running
+	// mutator. restore writes only to the target directory it is given.
+	readOnly
+)
+
+// AppOptions contains common options for creating and running the fx
+// application: the configuration file path, logging options, additional fx
+// modules and invocations, and whether the command mutates persistent
+// state (which decides whether it takes the PID lock).
 type AppOptions struct {
 	ConfigPath string
 	LogOptions log.Options
 	Modules    []fx.Option
 	Invokes    []fx.Option
+	Mode       lockMode
 }
 
 // setupGlobals records the startup time and, when an output-suppression
@@ -49,6 +65,11 @@ type AppOptions struct {
 // silenced — per the documented convention that --quiet suppresses
 // non-error output only. The startup banner is printed by Entry
 // before cobra parses arguments, gated by the same arg-level check.
+//
+// --json quiets the UI here too, because stdout then carries a JSON
+// document and human narration would corrupt it. Unlike Quiet it does
+// not lower the stderr log level (issue #112), so --verbose/--debug
+// still surface diagnostics alongside the document.
 func setupGlobals(
 	lc fx.Lifecycle, g *globals.Globals, v *vaultik.Vaultik, opts log.Options,
 ) {
@@ -56,7 +77,7 @@ func setupGlobals(
 		OnStart: func(_ context.Context) error {
 			g.StartTime = time.Now().UTC()
 
-			if opts.Cron || opts.Quiet {
+			if opts.Cron || opts.Quiet || opts.JSON {
 				v.UI.SetQuiet(true)
 			}
 
@@ -137,64 +158,45 @@ func cleanStartupError(err error) error {
 	return &startupError{msg: msg}
 }
 
-// RunApp starts and stops the fx application within the given context.
+// RunApp starts the fx application, blocks until it is asked to stop, and
-// It handles graceful shutdown on interrupt signals (SIGINT, SIGTERM) and
+// then stops it. The app is asked to stop either by an OS interrupt
-// ensures the application stops cleanly. The function blocks until the
+// (SIGINT/SIGTERM — fx installs its own handler when app.Wait is called) or,
-// application completes or is interrupted. Returns an error if startup fails.
+// on normal completion, by the finished operation calling
+// Shutdowner.Shutdown(); both arrive on the app.Wait channel.
+//
+// Stopping runs the fx OnStop hooks, and RunApp does not return until Stop
+// returns. On an interrupt the operation's OnStop hook cancels the running
+// command and waits for it to unwind — removing its decrypted scratch files —
+// so the process cannot proceed to exit mid-cleanup (issue #159). Waiting for
+// Stop before returning is what makes that hook effective: routing the
+// interrupt through app.Stop and not returning until it completes is required,
+// because fx also fires the app.Wait channel on the signal, and an earlier
+// version returned on that alone — unwinding to os.Exit while the concurrent
+// cleanup still ran. The stop is bounded by shutdownTimeout. Returns an error
+// if startup fails.
 func RunApp(ctx context.Context, app *fx.App) error {
-	// Set up signal handling for graceful shutdown
-	sigChan := make(chan os.Signal, 1)
-	signal.Notify(sigChan, os.Interrupt, syscall.SIGTERM)
-
-	// Create a context that will be cancelled on signal
-	ctx, cancel := context.WithCancel(ctx)
-	defer cancel()
-
-	// Start the app
 	err := app.Start(ctx)
 	if err != nil {
 		return cleanStartupError(err)
 	}
 
-	// Handle shutdown
+	// Block until an interrupt or the finished operation's
-	shutdownComplete := make(chan struct{})
+	// Shutdowner.Shutdown() arrives, then stop the app in this goroutine so we
+	// return only after its OnStop hooks — including the operation's cleanup
+	// wait — have run. Detach the stop from ctx's cancellation but keep its
+	// values, and bound it by shutdownTimeout.
+	<-app.Wait()
 
-	go func() {
+	shutdownCtx, cancel := context.WithTimeout(
-		defer close(shutdownComplete)
+		context.WithoutCancel(ctx), shutdownTimeout)
+	defer cancel()
 
-		<-sigChan
+	err = app.Stop(shutdownCtx)
-		log.Notice("Received interrupt signal, shutting down gracefully...")
+	if err != nil {
-
+		log.Error("Error during shutdown", "error", err)
-		// Create a timeout context for shutdown. The parent ctx is being
-		// cancelled, so detach from its cancellation but keep its values.
-		shutdownCtx, shutdownCancel := context.WithTimeout(
-			context.WithoutCancel(ctx), shutdownTimeout)
-		defer shutdownCancel()
-
-		err := app.Stop(shutdownCtx)
-		if err != nil {
-			log.Error("Error during shutdown", "error", err)
-		}
-	}()
-
-	// Wait for the signal handler to complete shutdown or the app to
-	// request shutdown.
-	select {
-	case <-shutdownComplete:
-		// Shutdown completed via signal
-		return nil
-	case <-ctx.Done():
-		// Context cancelled (shouldn't happen in normal operation)
-		err := app.Stop(context.WithoutCancel(ctx))
-		if err != nil {
-			log.Error("Error stopping app", "error", err)
-		}
-
-		return ctx.Err()
-	case <-app.Done():
-		// App finished running (e.g., backup completed)
-		return nil
 	}
+
+	return nil
 }
 
 // errReported marks a failure the operation has already shown the user
@@ -214,7 +216,10 @@ var errReported = errors.New("operation failed")
 //
 // op runs in a goroutine so OnStart returns promptly and an interrupt
 // can still cancel through OnStop; when it finishes, success or failure,
-// it triggers shutdown, which is what lets RunWithApp return. report is
+// it triggers shutdown, which is what lets RunWithApp return. On an
+// interrupt OnStop cancels op and waits for the goroutine to return, so
+// op's cleanup (removing decrypted scratch files) runs before the
+// process exits; the wait is bounded by shutdownTimeout. report is
 // called with a non-canceled failure so the caller can log it (and
 // suppress it under --json) before it becomes errReported. A context
 // cancellation is the interrupt path, not a failure: it is neither
@@ -230,9 +235,11 @@ func RunOperation(
 
 	opts.Invokes = append(opts.Invokes,
 		fx.Invoke(func(v *vaultik.Vaultik, lc fx.Lifecycle) {
+			var stop func(context.Context) bool
+
 			lc.Append(fx.Hook{
 				OnStart: func(_ context.Context) error {
-					go func() {
+					stop = v.StartOperation(func() {
 						err := op(v)
 						if err != nil && !errors.Is(err, context.Canceled) {
 							report(err)
@@ -246,12 +253,20 @@ func RunOperation(
 						if stopErr != nil {
 							log.Error("Failed to shutdown", "error", stopErr)
 						}
-					}()
+					})
 
 					return nil
 				},
-				OnStop: func(_ context.Context) error {
+				// On an interrupt, cancel the operation and wait for it to
-					v.Cancel()
+				// unwind so its cleanup defers (which remove decrypted
+				// scratch files from the temp directory) run before the
+				// process exits. The wait is bounded by ctx, the existing
+				// shutdownTimeout.
+				OnStop: func(ctx context.Context) error {
+					if !stop(ctx) {
+						log.Warn("Shutdown timed out before the operation " +
+							"finished; decrypted temporary files may remain")
+					}
 
 					return nil
 				},
@@ -276,13 +291,15 @@ func RunOperation(
 }
 
 // runVaultikApp runs the standard single-operation command lifecycle
-// shared by the list/purge/verify/remove/remote-info subcommands:
+// shared by the snapshot list/purge/remove and remote nuke subcommands:
 // resolve the config, then run op against the Vaultik instance through
 // RunOperation, reporting a failure prefixed with failMsg (suppressed
-// while suppressErrors is true, e.g. under --json). extraQuiet is OR-ed
+// while suppressErrors is true, e.g. under --json). mode says whether the
-// into LogOptions.Quiet (e.g. --json output modes).
+// command takes the PID lock. jsonOutput marks a command whose stdout is a
+// JSON document: it quiets the UI but, unlike Quiet, leaves the stderr log
+// level alone.
 func runVaultikApp(
-	cmd *cobra.Command, extraQuiet, suppressErrors bool,
+	cmd *cobra.Command, mode lockMode, jsonOutput, suppressErrors bool,
 	failMsg string, op func(v *vaultik.Vaultik) error,
 ) error {
 	configPath, err := ResolveConfigPath()
@@ -297,8 +314,10 @@ func runVaultikApp(
 		LogOptions: log.Options{
 			Verbose: rootFlags.Verbose,
 			Debug:   rootFlags.Debug,
-			Quiet:   rootFlags.Quiet || extraQuiet,
+			Quiet:   rootFlags.Quiet,
+			JSON:    jsonOutput,
 		},
+		Mode: mode,
 	}, op, func(err error) {
 		if suppressErrors {
 			return
@@ -312,28 +331,46 @@ func runVaultikApp(
 // RunWithApp is a helper that creates and runs an fx app with the given options.
 // It combines NewApp and RunApp into a single convenient function. This is the
 // preferred way to run CLI commands that need the full application context.
-// It acquires a PID lock before starting to prevent concurrent instances.
+// A mutating command takes the process-wide PID lock before starting so that
+// only one runs at a time; a read-only command runs without it and is not
+// blocked while a mutator holds the lock (opts.Mode).
 func RunWithApp(ctx context.Context, opts AppOptions) error {
-	// Acquire PID lock to prevent concurrent instances
+	release, err := acquireLockIfMutating(opts.Mode,
-	lockDir := filepath.Join(xdg.DataHome, "vaultik")
+		filepath.Join(xdg.DataHome, "vaultik"))
-
-	lock, err := pidlock.Acquire(lockDir)
 	if err != nil {
-		if errors.Is(err, pidlock.ErrAlreadyRunning) {
+		return err
-			return fmt.Errorf("cannot start: %w", err)
-		}
-
-		return fmt.Errorf("failed to acquire lock: %w", err)
 	}
 
-	defer func() {
+	defer release()
-		err := lock.Release()
-		if err != nil {
-			log.Warn("Failed to release PID lock", "error", err)
-		}
-	}()
 
 	app := NewApp(opts)
 
 	return RunApp(ctx, app)
 }
+
+// acquireLockIfMutating takes the process-wide PID lock in lockDir for a
+// mutating command and returns a function that releases it. A read-only
+// command takes no lock, so it returns a no-op release and is never blocked
+// while a mutator holds the lock. ErrAlreadyRunning (another mutator holds
+// the lock) is surfaced as a "cannot start" error.
+func acquireLockIfMutating(mode lockMode, lockDir string) (func(), error) {
+	if mode != mutating {
+		return func() {}, nil
+	}
+
+	lock, err := pidlock.Acquire(lockDir)
+	if err != nil {
+		if errors.Is(err, pidlock.ErrAlreadyRunning) {
+			return nil, fmt.Errorf("cannot start: %w", err)
+		}
+
+		return nil, fmt.Errorf("failed to acquire lock: %w", err)
+	}
+
+	return func() {
+		err := lock.Release()
+		if err != nil {
+			log.Warn("Failed to release PID lock", "error", err)
+		}
+	}, nil
+}

internal/cli/app_test.go

@@ -2,7 +2,10 @@ package cli //nolint:testpackage // needs access to unexported cleanStartupError
 
 import (
 	"errors"
+	"path/filepath"
 	"testing"
+
+	"sneak.berlin/go/vaultik/internal/pidlock"
 )
 
 func TestCleanStartupError(t *testing.T) {
@@ -53,3 +56,42 @@ func TestCleanStartupError(t *testing.T) {
 		})
 	}
 }
+
+// TestLockScopedToMutatingCommands proves the partition the PID lock now
+// enforces: a read-only command runs while a mutator holds the lock, and
+// two mutating commands still mutually exclude.
+func TestLockScopedToMutatingCommands(t *testing.T) {
+	t.Parallel()
+
+	lockDir := filepath.Join(t.TempDir(), "vaultik")
+
+	// A mutating command takes the process-wide lock.
+	releaseMutator, err := acquireLockIfMutating(mutating, lockDir)
+	if err != nil {
+		t.Fatalf("mutating command could not acquire lock: %v", err)
+	}
+
+	// A read-only command runs to completion even while the lock is held.
+	releaseReader, err := acquireLockIfMutating(readOnly, lockDir)
+	if err != nil {
+		t.Fatalf("read-only command was blocked by held lock: %v", err)
+	}
+
+	releaseReader()
+
+	// A second mutating command is refused while the first holds the lock.
+	_, err = acquireLockIfMutating(mutating, lockDir)
+	if !errors.Is(err, pidlock.ErrAlreadyRunning) {
+		t.Fatalf("second mutating command was not excluded, got: %v", err)
+	}
+
+	// Once the first mutator releases, another mutating command may run.
+	releaseMutator()
+
+	release, err := acquireLockIfMutating(mutating, lockDir)
+	if err != nil {
+		t.Fatalf("mutating command could not acquire released lock: %v", err)
+	}
+
+	release()
+}

internal/cli/config.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"errors"
 	"fmt"
+	"io"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -192,8 +193,8 @@ storage_url: ""
 #   access_key_id: YOUR_ACCESS_KEY
 #   secret_access_key: YOUR_SECRET_KEY
 #   # region: us-east-1        # Default: us-east-1
-#   # use_ssl: true            # Default: true
 #   # part_size: 5MB           # Multipart upload part size. Default: 5MB
+#   # For the s3:// form, disable TLS with ?ssl=false in the URL, not use_ssl.
 
 # ─── OPTIONAL ────────────────────────────────────────────────────────────────
 
@@ -212,6 +213,8 @@ storage_url: ""
 # chunk_size: 10MB
 
 # Maximum blob size before splitting into a new blob.
+# Must be at least four times chunk_size (the largest chunk the chunker can
+# emit); a smaller limit would let a single-chunk blob exceed it.
 # Accepts: 1GB, 10G, 500MB, etc.
 # Default: 10GB
 # blob_size_limit: 10GB
@@ -377,40 +380,53 @@ Examples:
 				return err
 			}
 
-			root, err := loadYAMLFile(path)
+			return writeConfigSet(os.Stdout, path, args[0], args[1])
-			if err != nil {
-				return err
-			}
-
-			err = yamlPathSet(root, strings.Split(args[0], "."), args[1])
-			if err != nil {
-				return err
-			}
-
-			out, err := marshalConfigYAML(root)
-			if err != nil {
-				return fmt.Errorf("marshaling config: %w", err)
-			}
-
-			mode := os.FileMode(configFileMode)
-
-			info, statErr := os.Stat(path)
-			if statErr == nil {
-				mode = info.Mode().Perm()
-			}
-
-			err = os.WriteFile(path, out, mode)
-			if err != nil {
-				return fmt.Errorf("writing config file: %w", err)
-			}
-
-			_, _ = fmt.Fprintf(os.Stdout, "%s = %s\n", args[0], args[1])
-
-			return nil
 		},
 	}
 }
 
+// writeConfigSet applies key=value to the config at path, writes it back
+// owner-only, and confirms the write by printing just the key name to w.
+// The value is never echoed: it may be a secret such as
+// s3.secret_access_key, and captured stdout or a pasted terminal would
+// then leak it.
+func writeConfigSet(w io.Writer, path, key, value string) error {
+	root, err := loadYAMLFile(path)
+	if err != nil {
+		return err
+	}
+
+	err = yamlPathSet(root, strings.Split(key, "."), value)
+	if err != nil {
+		return err
+	}
+
+	out, err := marshalConfigYAML(root)
+	if err != nil {
+		return fmt.Errorf("marshaling config: %w", err)
+	}
+
+	err = os.WriteFile(path, out, configFileMode)
+	if err != nil {
+		return fmt.Errorf("writing config file: %w", err)
+	}
+
+	// os.WriteFile does not change the mode of a file that already exists,
+	// so a config that was group- or world-readable stays that way. As it
+	// may hold S3 credentials, tighten it to owner-only after writing.
+	info, statErr := os.Stat(path)
+	if statErr == nil && info.Mode().Perm()&0o044 != 0 {
+		err = os.Chmod(path, configFileMode)
+		if err != nil {
+			return fmt.Errorf("tightening config file permissions: %w", err)
+		}
+	}
+
+	_, _ = fmt.Fprintln(w, key)
+
+	return nil
+}
+
 // marshalConfigYAML renders a config document tree with 2-space indentation,
 // matching defaultConfigTemplate. yaml.Marshal defaults to 4 spaces, which
 // would reindent the whole file on the first `config set` despite the promise

internal/cli/config_test.go

@@ -1,6 +1,9 @@
 package cli //nolint:testpackage // exercises unexported yamlPathGet/yamlPathSet
 
 import (
+	"bytes"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 
@@ -229,6 +232,68 @@ func TestConfigSetPreservesFormatting(t *testing.T) {
 	}
 }
 
+// TestWriteConfigSetHidesSecret checks that setting a secret key prints
+// only the key name, never the value, to the confirmation output.
+func TestWriteConfigSetHidesSecret(t *testing.T) {
+	t.Parallel()
+
+	const secret = "SUPERSECRETVALUE"
+
+	path := filepath.Join(t.TempDir(), "config.yaml")
+
+	err := os.WriteFile(path, []byte("version: 1\n"), 0o600)
+	if err != nil {
+		t.Fatalf("seed config: %v", err)
+	}
+
+	var out bytes.Buffer
+
+	err = writeConfigSet(&out, path, "s3.secret_access_key", secret)
+	if err != nil {
+		t.Fatalf("writeConfigSet: %v", err)
+	}
+
+	if strings.Contains(out.String(), secret) {
+		t.Errorf("output echoed the secret value: %q", out.String())
+	}
+
+	if !strings.Contains(out.String(), "s3.secret_access_key") {
+		t.Errorf("output did not confirm the key name: %q", out.String())
+	}
+}
+
+// TestWriteConfigSetTightensMode checks that a pre-existing group- or
+// world-readable config is tightened to owner-only after a set, since
+// os.WriteFile leaves an existing file's mode untouched.
+func TestWriteConfigSetTightensMode(t *testing.T) {
+	t.Parallel()
+
+	path := filepath.Join(t.TempDir(), "config.yaml")
+
+	// Seed a world-readable config; the loose mode is the condition under
+	// test, so gosec's G306 is expected here.
+	err := os.WriteFile(path, []byte("version: 1\n"), 0o644) //nolint:gosec // G306
+	if err != nil {
+		t.Fatalf("seed config: %v", err)
+	}
+
+	var out bytes.Buffer
+
+	err = writeConfigSet(&out, path, "compression_level", "9")
+	if err != nil {
+		t.Fatalf("writeConfigSet: %v", err)
+	}
+
+	info, err := os.Stat(path)
+	if err != nil {
+		t.Fatalf("stat config: %v", err)
+	}
+
+	if info.Mode().Perm() != 0o600 {
+		t.Errorf("config mode = %04o, want 0600", info.Mode().Perm())
+	}
+}
+
 func splitPath(s string) []string {
 	return strings.Split(s, ".")
 }

internal/cli/entry_prune_json_stderr_test.go

@@ -0,0 +1,140 @@
+package cli //nolint:testpackage // shares the prune fixtures and capture helpers
+
+import (
+	"bytes"
+	"io"
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// staleRecordLogMessage is the local-cleanup audit line CleanupLocalSnapshots
+// logs for each stale record. It is exactly the signal issue #112 says a
+// machine consumer lost under --json: gated off stdout, and pinned below
+// the log level on stderr because --json used to force Quiet.
+const staleRecordLogMessage = "Removing stale local snapshot record"
+
+// TestEntryPruneJSONStderrHonoursVerbosity is the end-to-end regression
+// guard for issue #112. Under --json the log level must still follow
+// --verbose/--debug rather than being pinned to WARN, so the
+// local-cleanup records reach stderr under --verbose while stdout stays
+// exactly one JSON document; without --verbose they stay below the
+// level, as they do without --json.
+//
+// Both halves are asserted together on the same run, because the fix has
+// to keep the document clean (issue #108) while freeing stderr.
+//
+// Not parallel: it replaces os.Args, os.Stdout, os.Stderr and the xdg
+// globals.
+//
+//nolint:paralleltest // replaces os.Args, os.Stdout, os.Stderr and the xdg globals
+func TestEntryPruneJSONStderrHonoursVerbosity(t *testing.T) {
+	for _, testCase := range []struct {
+		name         string
+		verbose      bool
+		wantOnStderr bool
+	}{
+		{
+			name:         "verbose json surfaces the cleanup record on stderr",
+			verbose:      true,
+			wantOnStderr: true,
+		},
+		{
+			name:         "json alone keeps the cleanup record below the level",
+			verbose:      false,
+			wantOnStderr: false,
+		},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			configPath := writeHermeticPruneConfig(t, true)
+
+			previousArgs := os.Args
+
+			t.Cleanup(func() {
+				os.Args = previousArgs
+				rootFlags = RootFlags{}
+			})
+
+			args := []string{
+				programName, flagConfig, configPath, cmdPrune, flagJSON,
+			}
+			if testCase.verbose {
+				args = append(args, "--verbose")
+			}
+
+			os.Args = args
+
+			stdout, stderr := captureProcessStdoutAndStderr(t,
+				func() { _ = Entry() })
+
+			// The document stays clean in both cases: freeing stderr must
+			// not regress issue #108.
+			requireExactlyOneJSONDocument(t, stdout)
+
+			if testCase.wantOnStderr {
+				assert.Contains(t, stderr, staleRecordLogMessage,
+					"--verbose --json must emit the cleanup record on stderr")
+				assert.Contains(t, stderr, stalePruneSnapshotID,
+					"the record must name the snapshot it removed")
+			} else {
+				assert.NotContains(t, stderr, staleRecordLogMessage,
+					"without --verbose the record stays below the log level")
+			}
+		})
+	}
+}
+
+// captureProcessStdoutAndStderr redirects both of the process's own
+// standard streams to pipes for the duration of fn and returns what was
+// written to each. The redirection is at the file-descriptor level
+// because the logger binds os.Stderr when it initializes inside fn, and
+// the JSON document reaches os.Stdout independently; the point is to see
+// where each actually lands.
+//
+// Not parallel-safe: os.Stdout and os.Stderr are process-global.
+func captureProcessStdoutAndStderr(t *testing.T, fn func()) (string, string) {
+	t.Helper()
+
+	outReader, outWriter, err := os.Pipe()
+	require.NoError(t, err)
+
+	errReader, errWriter, err := os.Pipe()
+	require.NoError(t, err)
+
+	previousOut, previousErr := os.Stdout, os.Stderr
+	os.Stdout, os.Stderr = outWriter, errWriter
+
+	capturedOut := drain(outReader)
+	capturedErr := drain(errReader)
+
+	fn()
+
+	os.Stdout, os.Stderr = previousOut, previousErr
+
+	require.NoError(t, outWriter.Close())
+	require.NoError(t, errWriter.Close())
+
+	out, errOut := <-capturedOut, <-capturedErr
+
+	require.NoError(t, outReader.Close())
+	require.NoError(t, errReader.Close())
+
+	return out, errOut
+}
+
+// drain copies a reader to a string on a goroutine and delivers the
+// result once the writer end is closed.
+func drain(reader io.Reader) <-chan string {
+	captured := make(chan string, 1)
+
+	go func() {
+		var buf bytes.Buffer
+
+		_, _ = io.Copy(&buf, reader)
+		captured <- buf.String()
+	}()
+
+	return captured
+}

internal/cli/info.go

@@ -35,6 +35,7 @@ func NewInfoCommand() *cobra.Command {
 					Debug:   rootFlags.Debug,
 					Quiet:   rootFlags.Quiet,
 				},
+				Mode: readOnly,
 			}, func(v *vaultik.Vaultik) error {
 				return v.ShowInfo()
 			}, func(err error) {

internal/cli/prune.go

@@ -41,8 +41,10 @@ work (e.g. after a crashed backup or to reclaim storage).`,
 				LogOptions: log.Options{
 					Verbose: rootFlags.Verbose,
 					Debug:   rootFlags.Debug,
-					Quiet:   rootFlags.Quiet || opts.JSON,
+					Quiet:   rootFlags.Quiet,
+					JSON:    opts.JSON,
 				},
+				Mode: mutating,
 			}, func(v *vaultik.Vaultik) error {
 				return v.Prune(opts)
 			}, func(err error) {

internal/cli/remote.go

@@ -45,7 +45,7 @@ This is destructive and irreversible. Requires --force.`,
 				return errNukeNeedsForce
 			}
 
-			return runVaultikApp(cmd, false, false, "Remote nuke failed",
+			return runVaultikApp(cmd, mutating, false, false, "Remote nuke failed",
 				func(v *vaultik.Vaultik) error {
 					return v.NukeRemote(true)
 				})
@@ -85,8 +85,10 @@ func newRemoteInfoCommand() *cobra.Command {
 				LogOptions: log.Options{
 					Verbose: rootFlags.Verbose,
 					Debug:   rootFlags.Debug,
-					Quiet:   rootFlags.Quiet || jsonOutput,
+					Quiet:   rootFlags.Quiet,
+					JSON:    jsonOutput,
 				},
+				Mode: readOnly,
 			}, func(v *vaultik.Vaultik) error {
 				return v.RemoteInfo(jsonOutput)
 			}, func(err error) {

internal/cli/root.go

@@ -57,8 +57,9 @@ on the source system.`,
 	cmd.PersistentFlags().BoolVarP(&rootFlags.Quiet, "quiet", "q", false,
 		"Suppress non-error output")
 	cmd.PersistentFlags().BoolVar(&rootFlags.SkipErrors, "skip-errors", false,
-		"Continue past per-file errors instead of aborting "+
+		"Skip files that cannot be read when creating a snapshot, or "+
-			"(applies to snapshot create and restore)")
+			"that cannot be restored when restoring, instead of aborting "+
+			"(packing and storage errors still abort)")
 
 	// Add subcommands
 	cmd.AddCommand(

internal/cli/runapp_interrupt_test.go

@@ -0,0 +1,97 @@
+package cli_test
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+	"go.uber.org/fx"
+	"sneak.berlin/go/vaultik/internal/cli"
+)
+
+// TestRunAppWaitsForOperationCleanupOnShutdown drives RunApp with an fx app
+// wired the way RunOperation wires a command: a single lifecycle hook whose
+// OnStart launches the operation in its own goroutine and whose OnStop cancels
+// it and blocks until that goroutine returns. The operation stands in for a
+// restore blocked mid-download — it holds a decrypted "scratch" file and only
+// removes it as it unwinds on cancellation.
+//
+// The app is asked to stop once the operation is running (standing in for an
+// OS interrupt; fx delivers a real signal and Shutdowner.Shutdown() on the
+// same app.Wait channel, so both drive the identical shutdown path). RunApp
+// must not return until app.Stop has run the OnStop hook, so the scratch file
+// must be gone by the time RunApp returns. Before the fix RunApp returned as
+// soon as the app.Wait/Done channel fired, without running app.Stop, so the
+// cleanup never ran and this file would still be on disk (issue #159).
+func TestRunAppWaitsForOperationCleanupOnShutdown(t *testing.T) {
+	t.Parallel()
+
+	scratch := filepath.Join(t.TempDir(), "decrypted-scratch")
+	require.NoError(t, os.WriteFile(scratch, []byte("secret"), 0o600))
+
+	// Cancel and reap the operation even if RunApp returns without doing so
+	// (the buggy path), so the goroutine cannot leak past the test.
+	opCtx, opCancel := context.WithCancel(context.Background())
+	t.Cleanup(opCancel)
+
+	var stop func(context.Context) bool
+
+	app := fx.New(
+		fx.NopLogger,
+		fx.Invoke(func(lc fx.Lifecycle, sh fx.Shutdowner) {
+			lc.Append(fx.Hook{
+				OnStart: func(_ context.Context) error {
+					done := make(chan struct{})
+
+					go func() {
+						defer close(done)
+
+						// Blocked mid-operation until cancelled, then run the
+						// cleanup an interrupted restore would run.
+						<-opCtx.Done()
+
+						_ = os.Remove(scratch)
+					}()
+
+					stop = func(ctx context.Context) bool {
+						opCancel()
+
+						select {
+						case <-done:
+							return true
+						case <-ctx.Done():
+							return false
+						}
+					}
+
+					// Ask the app to stop now that the operation is running.
+					go func() { _ = sh.Shutdown() }()
+
+					return nil
+				},
+				OnStop: func(ctx context.Context) error {
+					stop(ctx)
+
+					return nil
+				},
+			})
+		}),
+	)
+
+	done := make(chan error, 1)
+	go func() { done <- cli.RunApp(context.Background(), app) }()
+
+	select {
+	case err := <-done:
+		require.NoError(t, err)
+	case <-time.After(30 * time.Second):
+		t.Fatal("RunApp did not return after shutdown was requested")
+	}
+
+	_, err := os.Stat(scratch)
+	require.True(t, os.IsNotExist(err),
+		"RunApp returned before the operation removed its decrypted scratch file")
+}

internal/cli/snapshot.go

@@ -92,6 +92,7 @@ specifying a path using --config or by setting VAULTIK_CONFIG to a path.`,
 					Cron:    opts.Cron,
 					Quiet:   rootFlags.Quiet,
 				},
+				Mode: mutating,
 			}, func(v *vaultik.Vaultik) error {
 				return v.CreateSnapshot(opts)
 			}, func(err error) {
@@ -125,7 +126,7 @@ func newSnapshotListCommand() *cobra.Command {
 		Long:    "Lists all snapshots with their ID, timestamp, and compressed size",
 		Args:    cobra.NoArgs,
 		RunE: func(cmd *cobra.Command, _ []string) error {
-			return runVaultikApp(cmd, false, false,
+			return runVaultikApp(cmd, readOnly, false, false,
 				"Failed to list snapshots",
 				func(v *vaultik.Vaultik) error {
 					return v.ListSnapshots(jsonOutput)
@@ -161,7 +162,7 @@ restrict the operation to specific snapshot names.`,
 				return errPurgeCriteriaBoth
 			}
 
-			return runVaultikApp(cmd, false, false,
+			return runVaultikApp(cmd, mutating, false, false,
 				"Failed to purge snapshots",
 				func(v *vaultik.Vaultik) error {
 					return v.PurgeSnapshotsWithOptions(opts)
@@ -188,8 +189,11 @@ func newSnapshotVerifyCommand() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:   "verify <snapshot-id>",
 		Short: "Verify snapshot integrity",
-		Long:  "Verifies that all blobs referenced in a snapshot exist",
+		Long: "Verifies that all blobs referenced in a snapshot exist.\n\n" +
-		Args:  requireSnapshotIDArg,
+			"The snapshot may be named by its ID or, on a host with no local\n" +
+			"index, by the remote key that 'snapshot list' prints for a\n" +
+			"remote-only snapshot (an unambiguous leading part is enough).",
+		Args: requireSnapshotIDArg,
 		RunE: func(cmd *cobra.Command, args []string) error {
 			snapshotID := args[0]
 
@@ -206,8 +210,10 @@ func newSnapshotVerifyCommand() *cobra.Command {
 				LogOptions: log.Options{
 					Verbose: rootFlags.Verbose,
 					Debug:   rootFlags.Debug,
-					Quiet:   rootFlags.Quiet || opts.JSON,
+					Quiet:   rootFlags.Quiet,
+					JSON:    opts.JSON,
 				},
+				Mode: readOnly,
 			}, func(v *vaultik.Vaultik) error {
 				return v.VerifySnapshotWithOptions(snapshotID, opts)
 			}, func(err error) {
@@ -255,7 +261,7 @@ To wipe the entire destination store and start over, use 'vaultik remote
 nuke --force' — it is the single supported entry point for that.`,
 		Args: requireSnapshotIDArg,
 		RunE: func(cmd *cobra.Command, args []string) error {
-			return runVaultikApp(cmd, opts.JSON, opts.JSON,
+			return runVaultikApp(cmd, mutating, opts.JSON, opts.JSON,
 				"Failed to remove snapshot",
 				func(v *vaultik.Vaultik) error {
 					_, err := v.RemoveSnapshot(args[0], opts)

internal/cli/snapshot_restore.go

@@ -31,6 +31,10 @@ target directory.
 If no paths are specified, all files are restored.
 If paths are specified, only matching files/directories are restored.
 
+The snapshot may be named by its ID or, when restoring on a host with no
+local index, by the remote key that 'snapshot list' prints for a
+remote-only snapshot (an unambiguous leading part is enough).
+
 Requires the VAULTIK_AGE_SECRET_KEY environment variable to be set with
 the age private key.
 
@@ -84,6 +88,7 @@ func runRestore(cmd *cobra.Command, args []string, opts *RestoreOptions) error {
 			Debug:   rootFlags.Debug,
 			Quiet:   rootFlags.Quiet,
 		},
+		Mode: readOnly,
 	}, func(v *vaultik.Vaultik) error {
 		return v.Restore(&vaultik.RestoreOptions{
 			SnapshotID: snapshotID,

internal/config/config.go

@@ -16,11 +16,19 @@ import (
 	"github.com/adrg/xdg"
 	"go.uber.org/fx"
 	"gopkg.in/yaml.v3"
+	"sneak.berlin/go/vaultik/internal/chunker"
 	"sneak.berlin/go/vaultik/internal/log"
 )
 
 const appName = "vaultik"
 
+// secretKeyPrefix marks an age secret (private) key. It is compared
+// case-insensitively so a recipient entry that is actually a private key is
+// caught and never passed to age or echoed back.
+//
+//nolint:gosec // G101: marker for detecting a pasted secret key, not a credential
+const secretKeyPrefix = "AGE-SECRET-KEY-"
+
 // Defaults and validation bounds for tunable settings.
 const (
 	defaultBlobSizeLimit    = Size(10 * 1024 * 1024 * 1024) // 10GB
@@ -37,13 +45,19 @@ var (
 	errNoConfigPath    = errors.New("config path not provided")
 	errNoAgeRecipients = errors.New(
 		"at least one age_recipient is required (generate with: age-keygen)")
+	errRecipientIsSecretKey = errors.New(
+		"an age secret key was given where a public key (age1...) belongs")
+	errRecipientNotX25519 = errors.New(
+		"not a valid recipient; only X25519 age1... public keys are supported")
 	errNoSnapshots = errors.New(
 		"at least one snapshot must be configured (see config.example.yml)")
 	errSnapshotNoPaths   = errors.New("snapshot must have at least one path")
 	errChunkSizeTooSmall = errors.New("chunk_size must be at least 1MB")
-	errBlobSizeTooSmall  = errors.New("blob_size_limit must be at least chunk_size")
+	errBlobSizeTooSmall  = errors.New(
-	errBadCompression    = errors.New("compression_level must be between 1 and 19")
+		"blob_size_limit must be at least the largest chunk the chunker can " +
-	errBadStorageScheme  = errors.New(
+			"emit (chunk_size times the FastCDC size spread)")
+	errBadCompression   = errors.New("compression_level must be between 1 and 19")
+	errBadStorageScheme = errors.New(
 		"storage_url must start with s3://, file://, or rclone://")
 	errStorageNotConfigured = errors.New(
 		"storage not configured; set storage_url or provide s3.endpoint + " +
@@ -162,8 +176,10 @@ type S3Config struct {
 	AccessKeyID     string `yaml:"access_key_id"`
 	SecretAccessKey string `yaml:"secret_access_key"`
 	Region          string `yaml:"region"`
-	UseSSL          bool   `yaml:"use_ssl"`
+	// UseSSL selects HTTPS for a scheme-less endpoint. Omitted (nil) means
-	PartSize        Size   `yaml:"part_size"`
+	// the default, TLS; set it to false only to force plain HTTP.
+	UseSSL   *bool `yaml:"use_ssl"`
+	PartSize Size  `yaml:"part_size"`
 }
 
 // Path wraps the config file path for fx dependency injection.
@@ -285,18 +301,30 @@ func Load(path string) (*Config, error) {
 
 // Validate checks if the configuration is valid and complete.
 // It ensures all required fields are present and have valid values:
-// - At least one age recipient must be specified
+//   - At least one age recipient must be specified, and every recipient must
-// - At least one snapshot must be configured with at least one path
+//     parse as an X25519 age1... public key (so a bad entry fails at load, not
-// - Storage must be configured (either storage_url or s3.* fields)
+//     mid-backup); errors name the position, never the value
-// - Chunk size must be at least 1MB
+//   - At least one snapshot must be configured with at least one path
-// - Blob size limit must be at least the chunk size
+//   - Storage must be configured (either storage_url or s3.* fields)
-// - Compression level must be between 1 and 19
+//   - Chunk size must be at least 1MB
+//   - Blob size limit must be at least the largest chunk the chunker can emit
+//     (chunk_size times chunker.ChunkSizeSpread), so a single-chunk blob never
+//     exceeds the configured limit
+//   - Compression level must be between 1 and 19
+//
 // Returns an error describing the first validation failure encountered.
 func (c *Config) Validate() error {
 	if len(c.AgeRecipients) == 0 {
 		return errNoAgeRecipients
 	}
 
+	for i, recipient := range c.AgeRecipients {
+		err := validateAgeRecipient(recipient)
+		if err != nil {
+			return fmt.Errorf("age_recipients[%d]: %w", i, err)
+		}
+	}
+
 	if len(c.Snapshots) == 0 {
 		return errNoSnapshots
 	}
@@ -317,8 +345,13 @@ func (c *Config) Validate() error {
 		return errChunkSizeTooSmall
 	}
 
-	if c.BlobSizeLimit.Int64() < c.ChunkSize.Int64() {
+	// The chunker can emit chunks up to chunk_size * ChunkSizeSpread, and the
-		return errBlobSizeTooSmall
+	// packer places a single such chunk into an otherwise empty blob. A limit
+	// below that bound would let a blob exceed it, so reject it.
+	largestChunk := c.ChunkSize.Int64() * chunker.ChunkSizeSpread
+	if c.BlobSizeLimit.Int64() < largestChunk {
+		return fmt.Errorf("%w: need at least %d bytes",
+			errBlobSizeTooSmall, largestChunk)
 	}
 
 	if c.CompressionLevel < minCompressionLevel ||
@@ -329,6 +362,23 @@ func (c *Config) Validate() error {
 	return nil
 }
 
+// validateAgeRecipient parses one age_recipients entry with the age library
+// and returns a value-free error on failure. A recipient string can be
+// sensitive (an operator may paste a secret key by mistake), so neither the
+// entry nor age's own error (which quotes its input) is ever included.
+func validateAgeRecipient(recipient string) error {
+	if strings.HasPrefix(strings.ToUpper(recipient), secretKeyPrefix) {
+		return errRecipientIsSecretKey
+	}
+
+	_, err := age.ParseX25519Recipient(recipient)
+	if err != nil {
+		return errRecipientNotX25519
+	}
+
+	return nil
+}
+
 // validateStorage validates storage configuration.
 // If StorageURL is set, it takes precedence. S3 URLs require credentials.
 // File URLs don't require any S3 configuration.

internal/config/config_test.go

@@ -1,9 +1,13 @@
 package config //nolint:testpackage // exercises unexported extractAgeSecretKey
 
 import (
+	"errors"
 	"os"
 	"path/filepath"
+	"strings"
 	"testing"
+
+	"sneak.berlin/go/vaultik/internal/chunker"
 )
 
 const (
@@ -83,6 +87,48 @@ func TestConfigLoad(t *testing.T) {
 	}
 }
 
+// TestExampleConfigIsScrubbedAndLoads checks that the shipped
+// config.example.yml carries only neutral placeholders (no real credentials,
+// private addresses, or internal host names) and still parses.
+func TestExampleConfigIsScrubbedAndLoads(t *testing.T) {
+	t.Parallel()
+
+	examplePath := filepath.Join("..", "..", "config.example.yml")
+
+	cfg, err := Load(examplePath)
+	if err != nil {
+		t.Fatalf("Failed to load config.example.yml: %v", err)
+	}
+
+	if cfg.StorageURL != "rclone://myremote/path/to/backups" {
+		t.Errorf("Expected neutral storage_url, got '%s'", cfg.StorageURL)
+	}
+
+	//nolint:gosec // G304: examplePath is a fixed in-repo path, not user input
+	raw, err := os.ReadFile(examplePath)
+	if err != nil {
+		t.Fatalf("Failed to read config.example.yml: %v", err)
+	}
+
+	text := string(raw)
+
+	wantSubstrings := []string{
+		"YOUR_ACCESS_KEY",
+		"YOUR_SECRET_KEY",
+		"endpoint: https://",
+	}
+	for _, want := range wantSubstrings {
+		if !strings.Contains(text, want) {
+			t.Errorf("Expected config.example.yml to contain %q", want)
+		}
+	}
+
+	// A raw "http://" scheme would mean a plaintext, likely private endpoint.
+	if strings.Contains(text, "http://") {
+		t.Error("config.example.yml should not contain an http:// endpoint")
+	}
+}
+
 // TestConfigFromEnv tests loading config path from environment variable
 func TestConfigFromEnv(t *testing.T) {
 	t.Parallel()
@@ -101,6 +147,157 @@ func TestConfigFromEnv(t *testing.T) {
 	}
 }
 
+// TestValidateBlobSizeLimit checks the blob_size_limit boundary: it must be at
+// least the largest chunk the chunker can emit (chunk_size times
+// chunker.ChunkSizeSpread), because the packer places a single such chunk into
+// an otherwise empty blob. A limit between chunk_size and that bound is rejected.
+func TestValidateBlobSizeLimit(t *testing.T) {
+	t.Parallel()
+
+	const chunkSize = Size(10 * 1024 * 1024) // 10MB
+
+	largestChunk := chunkSize.Int64() * chunker.ChunkSizeSpread
+
+	newConfig := func(blobLimit Size) *Config {
+		return &Config{
+			AgeRecipients:    []string{testSneakAgePublicKey},
+			Snapshots:        map[string]SnapshotConfig{"test": {Paths: []string{"/tmp/src"}}},
+			StorageURL:       "file:///tmp/vaultik-test-store",
+			ChunkSize:        chunkSize,
+			BlobSizeLimit:    blobLimit,
+			CompressionLevel: 3,
+		}
+	}
+
+	tests := []struct {
+		name      string
+		blobLimit Size
+		wantErr   bool
+	}{
+		{
+			name:      "at chunk_size but below largest chunk is rejected",
+			blobLimit: chunkSize,
+			wantErr:   true,
+		},
+		{
+			name:      "between chunk_size and largest chunk is rejected",
+			blobLimit: Size(chunkSize.Int64() * 2),
+			wantErr:   true,
+		},
+		{
+			name:      "one byte below largest chunk is rejected",
+			blobLimit: Size(largestChunk - 1),
+			wantErr:   true,
+		},
+		{
+			name:      "exactly at largest chunk is accepted",
+			blobLimit: Size(largestChunk),
+			wantErr:   false,
+		},
+		{
+			name:      "above largest chunk is accepted",
+			blobLimit: Size(largestChunk * 100),
+			wantErr:   false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			err := newConfig(tt.blobLimit).Validate()
+			if tt.wantErr {
+				if !errors.Is(err, errBlobSizeTooSmall) {
+					t.Fatalf("Validate() error = %v, want errBlobSizeTooSmall", err)
+				}
+
+				return
+			}
+
+			if err != nil {
+				t.Fatalf("Validate() unexpected error: %v", err)
+			}
+		})
+	}
+}
+
+// TestValidateAgeRecipients checks that recipients are parsed at config load
+// (a bad entry fails immediately, not mid-backup) and that no invalid entry —
+// least of all a pasted secret key — is echoed in the error.
+func TestValidateAgeRecipients(t *testing.T) {
+	t.Parallel()
+
+	baseConfig := func(recipients []string) *Config {
+		return &Config{
+			AgeRecipients:    recipients,
+			Snapshots:        map[string]SnapshotConfig{"test": {Paths: []string{"/tmp/src"}}},
+			StorageURL:       "file:///tmp/vaultik-test-store",
+			ChunkSize:        Size(10 * 1024 * 1024),
+			BlobSizeLimit:    Size(10 * 1024 * 1024 * 1024),
+			CompressionLevel: 3,
+		}
+	}
+
+	tests := []struct {
+		name       string
+		recipients []string
+		wantErr    bool
+	}{
+		{
+			name:       "config init placeholder is rejected",
+			recipients: []string{"age1REPLACE_WITH_YOUR_PUBLIC_KEY"},
+			wantErr:    true,
+		},
+		{
+			name:       "ssh-ed25519 recipient is rejected",
+			recipients: []string{"ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIexamplekeydata"},
+			wantErr:    true,
+		},
+		{
+			name:       "truncated age1 string is rejected",
+			recipients: []string{"age1short"},
+			wantErr:    true,
+		},
+		{
+			name:       "secret key passed as recipient is rejected",
+			recipients: []string{testIntegrationAgePrivateKey},
+			wantErr:    true,
+		},
+		{
+			name:       "two valid recipients are accepted",
+			recipients: []string{testSneakAgePublicKey, testIntegrationAgePublicKey},
+			wantErr:    false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			err := baseConfig(tt.recipients).Validate()
+			if !tt.wantErr {
+				if err != nil {
+					t.Fatalf("Validate() unexpected error: %v", err)
+				}
+
+				return
+			}
+
+			if err == nil {
+				t.Fatal("Validate() returned nil, want error")
+			}
+
+			// The entry itself must never appear in the error, since a
+			// recipient string can be a secret key.
+			for _, recipient := range tt.recipients {
+				if strings.Contains(err.Error(), recipient) {
+					t.Fatalf("Validate() error echoed the recipient value: %v", err)
+				}
+			}
+		})
+	}
+}
+
 // TestExtractAgeSecretKey tests extraction of AGE-SECRET-KEY from various inputs
 func TestExtractAgeSecretKey(t *testing.T) {
 	t.Parallel()

internal/crypto/encryption.go

@@ -1,224 +0,0 @@
-// Package crypto provides thread-safe age encryption and decryption
-// helpers used to protect blob and metadata content.
-package crypto //nolint:revive,nolintlint // stdlib crypto unused; see #76
-
-import (
-	"bytes"
-	"errors"
-	"fmt"
-	"io"
-	"sync"
-
-	"filippo.io/age"
-	"go.uber.org/fx"
-)
-
-// ErrNoRecipients is returned when an encryptor is created or updated
-// without any recipient public keys.
-var ErrNoRecipients = errors.New("at least one recipient is required")
-
-// Encryptor provides thread-safe encryption using the age encryption library.
-// It supports encrypting data for multiple recipients simultaneously, allowing
-// any of the corresponding private keys to decrypt the data. This is useful
-// for backup scenarios where multiple parties should be able to decrypt the data.
-type Encryptor struct {
-	recipients []age.Recipient
-	mu         sync.RWMutex
-}
-
-// NewEncryptor creates a new encryptor with the given age public keys.
-// Each public key should be a valid age X25519 recipient string (e.g., "age1...")
-// At least one recipient must be provided. Returns an error if any of the
-// public keys are invalid or if no recipients are specified.
-func NewEncryptor(publicKeys []string) (*Encryptor, error) {
-	if len(publicKeys) == 0 {
-		return nil, ErrNoRecipients
-	}
-
-	recipients := make([]age.Recipient, 0, len(publicKeys))
-	for _, key := range publicKeys {
-		recipient, err := age.ParseX25519Recipient(key)
-		if err != nil {
-			return nil, fmt.Errorf("parsing age recipient %s: %w", key, err)
-		}
-
-		recipients = append(recipients, recipient)
-	}
-
-	return &Encryptor{
-		recipients: recipients,
-	}, nil
-}
-
-// Encrypt encrypts data using age encryption for all configured recipients.
-// The encrypted data can be decrypted by any of the corresponding private keys.
-// This method is suitable for small to medium amounts of data that fit in memory.
-// For large data streams, use EncryptStream or EncryptWriter instead.
-func (e *Encryptor) Encrypt(data []byte) ([]byte, error) {
-	e.mu.RLock()
-	recipients := e.recipients
-	e.mu.RUnlock()
-
-	var buf bytes.Buffer
-
-	// Create encrypted writer for all recipients
-	w, err := age.Encrypt(&buf, recipients...)
-	if err != nil {
-		return nil, fmt.Errorf("creating encrypted writer: %w", err)
-	}
-
-	// Write data
-	_, err = w.Write(data)
-	if err != nil {
-		return nil, fmt.Errorf("writing encrypted data: %w", err)
-	}
-
-	// Close to flush
-	err = w.Close()
-	if err != nil {
-		return nil, fmt.Errorf("closing encrypted writer: %w", err)
-	}
-
-	return buf.Bytes(), nil
-}
-
-// EncryptStream encrypts data from reader to writer using age encryption.
-// This method is suitable for encrypting large files or streams as it processes
-// data in a streaming fashion without loading everything into memory.
-// The encrypted data is written directly to the destination writer.
-func (e *Encryptor) EncryptStream(dst io.Writer, src io.Reader) error {
-	e.mu.RLock()
-	recipients := e.recipients
-	e.mu.RUnlock()
-
-	// Create encrypted writer for all recipients
-	w, err := age.Encrypt(dst, recipients...)
-	if err != nil {
-		return fmt.Errorf("creating encrypted writer: %w", err)
-	}
-
-	// Copy data
-	_, err = io.Copy(w, src)
-	if err != nil {
-		return fmt.Errorf("copying encrypted data: %w", err)
-	}
-
-	// Close to flush
-	err = w.Close()
-	if err != nil {
-		return fmt.Errorf("closing encrypted writer: %w", err)
-	}
-
-	return nil
-}
-
-// EncryptWriter creates a writer that encrypts data written to it.
-// All data written to the returned WriteCloser will be encrypted and written
-// to the destination writer. The caller must call Close() on the returned
-// writer to ensure all encrypted data is properly flushed and finalized.
-// This is useful for integrating encryption into existing writer-based pipelines.
-func (e *Encryptor) EncryptWriter(dst io.Writer) (io.WriteCloser, error) {
-	e.mu.RLock()
-	recipients := e.recipients
-	e.mu.RUnlock()
-
-	// Create encrypted writer for all recipients
-	w, err := age.Encrypt(dst, recipients...)
-	if err != nil {
-		return nil, fmt.Errorf("creating encrypted writer: %w", err)
-	}
-
-	return w, nil
-}
-
-// UpdateRecipients updates the recipients for future encryption operations.
-// This method is thread-safe and can be called while other encryption operations
-// are in progress. Existing encryption operations will continue with the old
-// recipients. At least one recipient must be provided. Returns an error if any
-// of the public keys are invalid or if no recipients are specified.
-func (e *Encryptor) UpdateRecipients(publicKeys []string) error {
-	if len(publicKeys) == 0 {
-		return ErrNoRecipients
-	}
-
-	recipients := make([]age.Recipient, 0, len(publicKeys))
-	for _, key := range publicKeys {
-		recipient, err := age.ParseX25519Recipient(key)
-		if err != nil {
-			return fmt.Errorf("parsing age recipient %s: %w", key, err)
-		}
-
-		recipients = append(recipients, recipient)
-	}
-
-	e.mu.Lock()
-	e.recipients = recipients
-	e.mu.Unlock()
-
-	return nil
-}
-
-// Decryptor provides thread-safe decryption using the age encryption library.
-// It uses a private key to decrypt data that was encrypted for the corresponding
-// public key.
-type Decryptor struct {
-	identity age.Identity
-	mu       sync.RWMutex
-}
-
-// NewDecryptor creates a new decryptor with the given age private key.
-// The private key should be a valid age X25519 identity string.
-// Returns an error if the private key is invalid.
-func NewDecryptor(privateKey string) (*Decryptor, error) {
-	identity, err := age.ParseX25519Identity(privateKey)
-	if err != nil {
-		return nil, fmt.Errorf("parsing age identity: %w", err)
-	}
-
-	return &Decryptor{
-		identity: identity,
-	}, nil
-}
-
-// Decrypt decrypts data using age decryption.
-// This method is suitable for small to medium amounts of data that fit in memory.
-// For large data streams, use DecryptStream instead.
-func (d *Decryptor) Decrypt(data []byte) ([]byte, error) {
-	d.mu.RLock()
-	identity := d.identity
-	d.mu.RUnlock()
-
-	r, err := age.Decrypt(bytes.NewReader(data), identity)
-	if err != nil {
-		return nil, fmt.Errorf("creating decrypted reader: %w", err)
-	}
-
-	decrypted, err := io.ReadAll(r)
-	if err != nil {
-		return nil, fmt.Errorf("reading decrypted data: %w", err)
-	}
-
-	return decrypted, nil
-}
-
-// DecryptStream returns a reader that decrypts data from the provided reader.
-// This method is suitable for decrypting large files or streams as it processes
-// data in a streaming fashion without loading everything into memory.
-// The caller should close the input reader when done.
-func (d *Decryptor) DecryptStream(src io.Reader) (io.Reader, error) {
-	d.mu.RLock()
-	identity := d.identity
-	d.mu.RUnlock()
-
-	r, err := age.Decrypt(src, identity)
-	if err != nil {
-		return nil, fmt.Errorf("creating decrypted reader: %w", err)
-	}
-
-	return r, nil
-}
-
-// Module exports the crypto module for fx dependency injection.
-//
-//nolint:gochecknoglobals // fx module definitions are package globals
-var Module = fx.Module("crypto")

internal/crypto/encryption_test.go

@@ -1,178 +0,0 @@
-package crypto_test
-
-import (
-	"bytes"
-	"testing"
-
-	"filippo.io/age"
-	"sneak.berlin/go/vaultik/internal/crypto"
-)
-
-func TestEncryptor(t *testing.T) {
-	t.Parallel()
-
-	// Generate a test key pair
-	identity, err := age.GenerateX25519Identity()
-	if err != nil {
-		t.Fatalf("failed to generate identity: %v", err)
-	}
-
-	publicKey := identity.Recipient().String()
-
-	// Create encryptor
-	enc, err := crypto.NewEncryptor([]string{publicKey})
-	if err != nil {
-		t.Fatalf("failed to create encryptor: %v", err)
-	}
-
-	// Test data
-	plaintext := []byte("Hello, World! This is a test message.")
-
-	// Encrypt
-	ciphertext, err := enc.Encrypt(plaintext)
-	if err != nil {
-		t.Fatalf("failed to encrypt: %v", err)
-	}
-
-	// Verify it's actually encrypted (should be larger and different)
-	if bytes.Equal(plaintext, ciphertext) {
-		t.Error("ciphertext equals plaintext")
-	}
-
-	// Decrypt to verify
-	r, err := age.Decrypt(bytes.NewReader(ciphertext), identity)
-	if err != nil {
-		t.Fatalf("failed to decrypt: %v", err)
-	}
-
-	var decrypted bytes.Buffer
-
-	_, err = decrypted.ReadFrom(r)
-	if err != nil {
-		t.Fatalf("failed to read decrypted data: %v", err)
-	}
-
-	if !bytes.Equal(plaintext, decrypted.Bytes()) {
-		t.Error("decrypted data doesn't match original")
-	}
-}
-
-func TestEncryptorMultipleRecipients(t *testing.T) {
-	t.Parallel()
-
-	// Generate three test key pairs
-	identity1, err := age.GenerateX25519Identity()
-	if err != nil {
-		t.Fatalf("failed to generate identity1: %v", err)
-	}
-
-	identity2, err := age.GenerateX25519Identity()
-	if err != nil {
-		t.Fatalf("failed to generate identity2: %v", err)
-	}
-
-	identity3, err := age.GenerateX25519Identity()
-	if err != nil {
-		t.Fatalf("failed to generate identity3: %v", err)
-	}
-
-	publicKeys := []string{
-		identity1.Recipient().String(),
-		identity2.Recipient().String(),
-		identity3.Recipient().String(),
-	}
-
-	// Create encryptor with multiple recipients
-	enc, err := crypto.NewEncryptor(publicKeys)
-	if err != nil {
-		t.Fatalf("failed to create encryptor: %v", err)
-	}
-
-	// Test data
-	plaintext := []byte("Secret message for multiple recipients")
-
-	// Encrypt
-	ciphertext, err := enc.Encrypt(plaintext)
-	if err != nil {
-		t.Fatalf("failed to encrypt: %v", err)
-	}
-
-	// Verify each recipient can decrypt
-	identities := []age.Identity{identity1, identity2, identity3}
-	for i, identity := range identities {
-		r, err := age.Decrypt(bytes.NewReader(ciphertext), identity)
-		if err != nil {
-			t.Fatalf("recipient %d failed to decrypt: %v", i+1, err)
-		}
-
-		var decrypted bytes.Buffer
-
-		_, err = decrypted.ReadFrom(r)
-		if err != nil {
-			t.Fatalf("recipient %d failed to read decrypted data: %v", i+1, err)
-		}
-
-		if !bytes.Equal(plaintext, decrypted.Bytes()) {
-			t.Errorf("recipient %d: decrypted data doesn't match original", i+1)
-		}
-	}
-}
-
-func TestEncryptorUpdateRecipients(t *testing.T) {
-	t.Parallel()
-
-	// Generate two identities
-	identity1, _ := age.GenerateX25519Identity()
-	identity2, _ := age.GenerateX25519Identity()
-
-	publicKey1 := identity1.Recipient().String()
-	publicKey2 := identity2.Recipient().String()
-
-	// Create encryptor with first key
-	enc, err := crypto.NewEncryptor([]string{publicKey1})
-	if err != nil {
-		t.Fatalf("failed to create encryptor: %v", err)
-	}
-
-	// Encrypt with first key
-	plaintext := []byte("test data")
-
-	ciphertext1, err := enc.Encrypt(plaintext)
-	if err != nil {
-		t.Fatalf("failed to encrypt: %v", err)
-	}
-
-	// Update to second key
-	err = enc.UpdateRecipients([]string{publicKey2})
-	if err != nil {
-		t.Fatalf("failed to update recipients: %v", err)
-	}
-
-	// Encrypt with second key
-	ciphertext2, err := enc.Encrypt(plaintext)
-	if err != nil {
-		t.Fatalf("failed to encrypt: %v", err)
-	}
-
-	// First ciphertext should only decrypt with first identity
-	_, err = age.Decrypt(bytes.NewReader(ciphertext1), identity1)
-	if err != nil {
-		t.Error("failed to decrypt with identity1")
-	}
-
-	_, err = age.Decrypt(bytes.NewReader(ciphertext1), identity2)
-	if err == nil {
-		t.Error("should not decrypt with identity2")
-	}
-
-	// Second ciphertext should only decrypt with second identity
-	_, err = age.Decrypt(bytes.NewReader(ciphertext2), identity2)
-	if err != nil {
-		t.Error("failed to decrypt with identity2")
-	}
-
-	_, err = age.Decrypt(bytes.NewReader(ciphertext2), identity1)
-	if err == nil {
-		t.Error("should not decrypt with identity1")
-	}
-}

internal/database/blobs.go

@@ -208,6 +208,30 @@ func (r *BlobRepository) DeleteOrphaned(ctx context.Context) error {
 	return nil
 }
 
+// DeleteUnuploaded deletes blob rows whose upload never completed
+// (uploaded_ts IS NULL) and returns how many were removed. Their
+// blob_chunks rows are removed by the ON DELETE CASCADE foreign key.
+// A blob is only ever attached to a snapshot once its upload has been
+// recorded, so an un-uploaded blob is never referenced by a completed
+// snapshot: dropping it discards chunk rows that point at data which
+// was never stored remotely, so the affected content is re-chunked and
+// re-uploaded on the next run.
+func (r *BlobRepository) DeleteUnuploaded(ctx context.Context) (int64, error) {
+	query := `DELETE FROM blobs WHERE uploaded_ts IS NULL`
+
+	result, err := r.db.ExecWithLog(ctx, query)
+	if err != nil {
+		return 0, fmt.Errorf("deleting un-uploaded blobs: %w", err)
+	}
+
+	rowsAffected, _ := result.RowsAffected()
+	if rowsAffected > 0 {
+		log.Debug("Deleted un-uploaded blobs", "count", rowsAffected)
+	}
+
+	return rowsAffected, nil
+}
+
 // getOne fetches a single blob row matched on the given column, or
 // (nil, nil) when no row matches.
 func (r *BlobRepository) getOne(

internal/database/chunks_ext.go

@@ -7,12 +7,32 @@ import (
 
 // List returns every chunk in the index, ordered by chunk hash.
 func (r *ChunkRepository) List(ctx context.Context) ([]*Chunk, error) {
-	query := `
+	return r.list(ctx, `
 		SELECT chunk_hash, size
 		FROM chunks
 		ORDER BY chunk_hash
-	`
+	`)
+}
 
+// ListInUploadedBlobs returns the chunks that are stored in a blob whose
+// upload has completed (uploaded_ts set), ordered by chunk hash. These
+// are the only chunks a backup may safely deduplicate against: a chunk
+// recorded solely in a blob that was never uploaded refers to data that
+// is not in remote storage, so trusting it would silently drop that data
+// from later snapshots.
+func (r *ChunkRepository) ListInUploadedBlobs(ctx context.Context) ([]*Chunk, error) {
+	return r.list(ctx, `
+		SELECT DISTINCT c.chunk_hash, c.size
+		FROM chunks c
+		JOIN blob_chunks bc ON c.chunk_hash = bc.chunk_hash
+		JOIN blobs b ON bc.blob_id = b.id
+		WHERE b.uploaded_ts IS NOT NULL
+		ORDER BY c.chunk_hash
+	`)
+}
+
+// list runs a chunk-selecting query and scans the (chunk_hash, size) rows.
+func (r *ChunkRepository) list(ctx context.Context, query string) ([]*Chunk, error) {
 	rows, err := r.db.conn.QueryContext(ctx, query)
 	if err != nil {
 		return nil, fmt.Errorf("querying chunks: %w", err)

internal/database/database.go

@@ -17,6 +17,7 @@ import (
 	"embed"
 	"errors"
 	"fmt"
+	"net/url"
 	"os"
 	"path/filepath"
 	"sort"
@@ -219,6 +220,135 @@ func openWithRecovery(ctx context.Context, path string) (*DB, error) {
 	return db, nil
 }
 
+// errUntrustedSnapshotSchema is returned when a downloaded snapshot
+// database carries schema objects the real schema never defines, or is
+// missing a table the restore and deep-verify queries read.
+var errUntrustedSnapshotSchema = errors.New(
+	"downloaded snapshot database has an untrusted schema")
+
+// snapshotReadOnlyDSN builds the driver DSN that opens a materialized
+// snapshot database file read-only. mode=ro opens the file read-only at
+// the OS level, query_only rejects any write the engine is asked to make,
+// and trusted_schema=OFF refuses to run application code named in the
+// schema. The file: URI form is required for the driver to honour the
+// mode parameter.
+func snapshotReadOnlyDSN(path string) string {
+	u := url.URL{
+		Scheme:   "file",
+		Path:     path,
+		RawQuery: "mode=ro&_pragma=query_only(true)&_pragma=trusted_schema(false)",
+	}
+
+	return u.String()
+}
+
+// OpenReadOnly opens an already-materialized SQLite file for read-only
+// querying of a snapshot database downloaded from the store, used by
+// restore and deep verify. Unlike New it never applies schema migrations
+// and never writes: the connection is opened read-only with query_only
+// and trusted_schema=OFF. It refuses any file whose schema carries a
+// trigger, view or virtual table, or lacks an expected table, so a forged
+// file cannot redefine what the restore queries return. The caller owns
+// the file and must remove it.
+func OpenReadOnly(ctx context.Context, path string) (*DB, error) {
+	conn, err := sql.Open("sqlite", snapshotReadOnlyDSN(path))
+	if err != nil {
+		return nil, fmt.Errorf("opening read-only database: %w", err)
+	}
+
+	configureConnPool(conn)
+
+	err = conn.PingContext(ctx)
+	if err != nil {
+		_ = conn.Close()
+
+		return nil, fmt.Errorf("opening read-only database: %w", err)
+	}
+
+	err = verifySnapshotSchema(ctx, conn)
+	if err != nil {
+		_ = conn.Close()
+
+		return nil, err
+	}
+
+	return &DB{conn: conn, path: path}, nil
+}
+
+// verifySnapshotSchema rejects a downloaded database whose schema is not
+// the plain table set the real schema defines. Any trigger, view or
+// virtual table, or a missing expected table, fails the open.
+func verifySnapshotSchema(ctx context.Context, conn *sql.DB) error {
+	// expectedSnapshotTables are the tables the restore and deep-verify
+	// queries read. A downloaded database missing any of them is not a
+	// genuine snapshot database and is refused.
+	expectedSnapshotTables := []string{
+		"blob_chunks",
+		"blobs",
+		"chunks",
+		"file_chunks",
+		"files",
+	}
+
+	rows, err := conn.QueryContext(
+		ctx, "SELECT type, name, sql FROM sqlite_master")
+	if err != nil {
+		return fmt.Errorf("reading snapshot schema: %w", err)
+	}
+
+	defer func() { _ = rows.Close() }()
+
+	present := make(map[string]struct{})
+
+	for rows.Next() {
+		var objType, name string
+
+		var objSQL sql.NullString
+
+		err = rows.Scan(&objType, &name, &objSQL)
+		if err != nil {
+			return fmt.Errorf("reading snapshot schema: %w", err)
+		}
+
+		switch objType {
+		case "trigger", "view":
+			return fmt.Errorf(
+				"%w: unexpected %s %q", errUntrustedSnapshotSchema, objType, name)
+		case "table":
+			if isVirtualTableSQL(objSQL.String) {
+				return fmt.Errorf(
+					"%w: unexpected virtual table %q",
+					errUntrustedSnapshotSchema, name)
+			}
+
+			present[name] = struct{}{}
+		}
+	}
+
+	err = rows.Err()
+	if err != nil {
+		return fmt.Errorf("reading snapshot schema: %w", err)
+	}
+
+	for _, table := range expectedSnapshotTables {
+		if _, ok := present[table]; !ok {
+			return fmt.Errorf(
+				"%w: missing table %q", errUntrustedSnapshotSchema, table)
+		}
+	}
+
+	return nil
+}
+
+// isVirtualTableSQL reports whether a sqlite_master row's SQL defines a
+// virtual table. Virtual tables are recorded with type 'table' but a
+// "CREATE VIRTUAL TABLE" definition and can run module code, so they are
+// refused alongside triggers and views.
+func isVirtualTableSQL(createSQL string) bool {
+	return strings.HasPrefix(
+		strings.ToUpper(strings.TrimSpace(createSQL)), "CREATE VIRTUAL TABLE")
+}
+
 // NewTestDB creates an in-memory SQLite database for testing purposes.
 // The database is automatically initialized with the schema and is ready
 // for use. Each call creates a new independent database instance.

internal/database/module.go

@@ -2,6 +2,7 @@ package database
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -15,6 +16,44 @@ import (
 // the index describes the backed-up file tree and must stay private.
 const indexDirPerm = 0o700
 
+// indexFilePerm restricts the index file to the owning user; it lists every
+// backed-up path and chunk hash and must stay private.
+const indexFilePerm = 0o600
+
+// ensureIndexFileMode makes the index file owner-only before the SQLite
+// driver opens it: it creates the file 0600 if absent, or chmods an existing
+// one to 0600. Doing this first matters because SQLite creates its -wal and
+// -shm side files with the mode of the main database file, so a private main
+// file yields private side files. The driver treats a zero-byte file as an
+// empty database, so pre-creating it here is safe.
+func ensureIndexFileMode(path string) error {
+	info, err := os.Stat(path)
+
+	switch {
+	case err == nil:
+		if info.Mode().Perm() == indexFilePerm {
+			return nil
+		}
+
+		err = os.Chmod(path, indexFilePerm)
+		if err != nil {
+			return fmt.Errorf("restricting index file permissions: %w", err)
+		}
+
+		return nil
+	case errors.Is(err, os.ErrNotExist):
+		//nolint:gosec // G304: the index path is operator-configured by design
+		f, err := os.OpenFile(path, os.O_CREATE|os.O_WRONLY, indexFilePerm)
+		if err != nil {
+			return fmt.Errorf("creating index file: %w", err)
+		}
+
+		return f.Close()
+	default:
+		return fmt.Errorf("checking index file: %w", err)
+	}
+}
+
 // Module provides database dependencies
 //
 //nolint:gochecknoglobals // fx module definitions are package globals by convention
@@ -34,6 +73,11 @@ func provideDatabase(lc fx.Lifecycle, cfg *config.Config) (*DB, error) {
 		return nil, fmt.Errorf("creating index directory: %w", err)
 	}
 
+	err = ensureIndexFileMode(cfg.IndexPath)
+	if err != nil {
+		return nil, err
+	}
+
 	db, err := New(context.Background(), cfg.IndexPath)
 	if err != nil {
 		return nil, fmt.Errorf("opening database: %w", err)

internal/database/module_internal_test.go

@@ -0,0 +1,100 @@
+package database
+
+import (
+	"os"
+	"path/filepath"
+	"syscall"
+	"testing"
+
+	"go.uber.org/fx/fxtest"
+	"sneak.berlin/go/vaultik/internal/config"
+)
+
+// TestProvideDatabaseFreshIndexMode verifies that provideDatabase creates a
+// missing index file owner-only (0600), even under a lenient 022 umask that
+// would otherwise leave a freshly created file world-readable.
+//
+//nolint:paralleltest // syscall.Umask is process-global; parallel tests would clash
+func TestProvideDatabaseFreshIndexMode(t *testing.T) {
+	restore := syscall.Umask(0o022)
+	defer syscall.Umask(restore)
+
+	indexPath := filepath.Join(t.TempDir(), "index.sqlite")
+
+	openIndex(t, indexPath)
+	assertPerm(t, indexPath, 0o600)
+}
+
+// TestProvideDatabaseExistingIndexMode verifies that provideDatabase tightens
+// an existing world-readable index (0644) in a group/other-readable directory
+// down to owner-only (0600).
+//
+//nolint:paralleltest // syscall.Umask is process-global; parallel tests would clash
+func TestProvideDatabaseExistingIndexMode(t *testing.T) {
+	restore := syscall.Umask(0o022)
+	defer syscall.Umask(restore)
+
+	dir := filepath.Join(t.TempDir(), "data")
+
+	//nolint:gosec // G301: the test intentionally uses a 0755 directory
+	err := os.MkdirAll(dir, 0o755)
+	if err != nil {
+		t.Fatalf("creating index directory: %v", err)
+	}
+
+	//nolint:gosec // G302: the test intentionally uses a 0755 directory
+	err = os.Chmod(dir, 0o755)
+	if err != nil {
+		t.Fatalf("relaxing index directory permissions: %v", err)
+	}
+
+	indexPath := filepath.Join(dir, "index.sqlite")
+
+	//nolint:gosec // G306: the test intentionally starts from a 0644 index
+	err = os.WriteFile(indexPath, nil, 0o644)
+	if err != nil {
+		t.Fatalf("creating pre-existing index: %v", err)
+	}
+
+	//nolint:gosec // G302: the test intentionally starts from a 0644 index
+	err = os.Chmod(indexPath, 0o644)
+	if err != nil {
+		t.Fatalf("relaxing pre-existing index permissions: %v", err)
+	}
+
+	openIndex(t, indexPath)
+	assertPerm(t, indexPath, 0o600)
+}
+
+// openIndex runs provideDatabase against indexPath and closes the resulting
+// database before returning.
+func openIndex(t *testing.T, indexPath string) {
+	t.Helper()
+
+	cfg := &config.Config{IndexPath: indexPath}
+
+	db, err := provideDatabase(fxtest.NewLifecycle(t), cfg)
+	if err != nil {
+		t.Fatalf("provideDatabase: %v", err)
+	}
+
+	err = db.Close()
+	if err != nil {
+		t.Fatalf("closing database: %v", err)
+	}
+}
+
+// assertPerm fails the test unless path has exactly the given permission bits.
+func assertPerm(t *testing.T, path string, want os.FileMode) {
+	t.Helper()
+
+	info, err := os.Stat(path)
+	if err != nil {
+		t.Fatalf("stat %s: %v", path, err)
+	}
+
+	got := info.Mode().Perm()
+	if got != want {
+		t.Fatalf("permissions of %s = %#o, want %#o", path, got, want)
+	}
+}

internal/database/read_only_test.go

@@ -0,0 +1,145 @@
+//nolint:testpackage // exercises unexported read-only open internals
+package database
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"path/filepath"
+	"testing"
+)
+
+// genuineSnapshotDB writes a real snapshot database (the full schema
+// applied) to a fresh file and returns its path.
+func genuineSnapshotDB(t *testing.T) string {
+	t.Helper()
+
+	path := filepath.Join(t.TempDir(), "snapshot.db")
+
+	db, err := New(context.Background(), path)
+	if err != nil {
+		t.Fatalf("creating snapshot database: %v", err)
+	}
+
+	err = db.Close()
+	if err != nil {
+		t.Fatalf("closing snapshot database: %v", err)
+	}
+
+	return path
+}
+
+// forgedDB creates an empty database file and runs the given statements
+// against it read-write, so a test can plant schema objects the real
+// schema never defines.
+func forgedDB(t *testing.T, stmts ...string) string {
+	t.Helper()
+
+	path := filepath.Join(t.TempDir(), "forged.db")
+
+	db, err := sql.Open("sqlite", path)
+	if err != nil {
+		t.Fatalf("opening forged database: %v", err)
+	}
+
+	for _, stmt := range stmts {
+		_, err = db.ExecContext(context.Background(), stmt)
+		if err != nil {
+			t.Fatalf("executing %q: %v", stmt, err)
+		}
+	}
+
+	err = db.Close()
+	if err != nil {
+		t.Fatalf("closing forged database: %v", err)
+	}
+
+	return path
+}
+
+func TestOpenReadOnlyAcceptsGenuineSnapshot(t *testing.T) {
+	t.Parallel()
+
+	db, err := OpenReadOnly(context.Background(), genuineSnapshotDB(t))
+	if err != nil {
+		t.Fatalf("OpenReadOnly refused a genuine snapshot database: %v", err)
+	}
+
+	t.Cleanup(func() { _ = db.Close() })
+}
+
+func TestOpenReadOnlyRefusesWrites(t *testing.T) {
+	t.Parallel()
+
+	db, err := OpenReadOnly(context.Background(), genuineSnapshotDB(t))
+	if err != nil {
+		t.Fatalf("OpenReadOnly: %v", err)
+	}
+
+	t.Cleanup(func() { _ = db.Close() })
+
+	// A schema write depends on no table columns, so the only reason it
+	// can fail is that the database is open read-only.
+	_, err = db.Conn().ExecContext(context.Background(),
+		"CREATE TABLE probe_readonly (x)")
+	if err == nil {
+		t.Fatal("expected a write to a read-only snapshot database to fail")
+	}
+}
+
+func TestOpenReadOnlyRejectsView(t *testing.T) {
+	t.Parallel()
+
+	path := forgedDB(t, "CREATE VIEW files AS SELECT 1 AS path")
+
+	_, err := OpenReadOnly(context.Background(), path)
+	if !errors.Is(err, errUntrustedSnapshotSchema) {
+		t.Fatalf("expected a view named files to be refused, got %v", err)
+	}
+}
+
+func TestOpenReadOnlyRejectsTrigger(t *testing.T) {
+	t.Parallel()
+
+	path := forgedDB(t,
+		"CREATE TABLE files (path TEXT)",
+		"CREATE TRIGGER t AFTER INSERT ON files BEGIN SELECT 1; END")
+
+	_, err := OpenReadOnly(context.Background(), path)
+	if !errors.Is(err, errUntrustedSnapshotSchema) {
+		t.Fatalf("expected a trigger to be refused, got %v", err)
+	}
+}
+
+func TestOpenReadOnlyRejectsMissingTable(t *testing.T) {
+	t.Parallel()
+
+	// Only one of the expected tables is present.
+	path := forgedDB(t, "CREATE TABLE files (path TEXT)")
+
+	_, err := OpenReadOnly(context.Background(), path)
+	if !errors.Is(err, errUntrustedSnapshotSchema) {
+		t.Fatalf("expected a missing expected table to be refused, got %v", err)
+	}
+}
+
+func TestIsVirtualTableSQL(t *testing.T) {
+	t.Parallel()
+
+	cases := []struct {
+		sql  string
+		want bool
+	}{
+		{"CREATE VIRTUAL TABLE t USING fts5(x)", true},
+		{"  create virtual table t using fts5(x)", true},
+		{"CREATE TABLE t (x)", false},
+		{"CREATE VIEW t AS SELECT 1", false},
+		{"", false},
+	}
+
+	for _, c := range cases {
+		if got := isVirtualTableSQL(c.sql); got != c.want {
+			t.Errorf("isVirtualTableSQL(%q) = %v, want %v", c.sql, got, c.want)
+		}
+	}
+}

internal/log/module.go

@@ -14,8 +14,18 @@ var Module = fx.Module("log",
 )
 
 // New creates a new logger configuration from provided options.
+//
+// JSON is intentionally not carried into Config: a command emitting a
+// JSON document on stdout must keep its stderr log level under
+// --verbose/--debug, so --json must not lower it (issue #112). JSON
+// silences the stdout UI in setupGlobals instead.
 func New(opts Options) Config {
-	return Config(opts)
+	return Config{
+		Verbose: opts.Verbose,
+		Debug:   opts.Debug,
+		Cron:    opts.Cron,
+		Quiet:   opts.Quiet,
+	}
 }
 
 // Options are provided by the CLI.
@@ -24,4 +34,8 @@ type Options struct {
 	Debug   bool
 	Cron    bool
 	Quiet   bool
+	// JSON marks a command whose stdout carries a machine-readable
+	// document. It silences the human UI on stdout (see setupGlobals),
+	// but unlike Quiet it leaves the stderr log level alone.
+	JSON bool
 }

internal/snapshot/export_perm_test.go

@@ -0,0 +1,49 @@
+//nolint:testpackage // exercises the unexported copyFile helper
+package snapshot
+
+import (
+	"os"
+	"path/filepath"
+	"syscall"
+	"testing"
+
+	"github.com/spf13/afero"
+)
+
+// TestCopyFileExportCopyMode verifies that the exported snapshot database
+// copy is created owner-only (0600), even under a lenient 022 umask that
+// would otherwise leave a fresh file world-readable.
+//
+//nolint:paralleltest // syscall.Umask is process-global; parallel tests would clash
+func TestCopyFileExportCopyMode(t *testing.T) {
+	restore := syscall.Umask(0o022)
+	defer syscall.Umask(restore)
+
+	dir := t.TempDir()
+
+	src := filepath.Join(dir, "index.sqlite")
+
+	err := os.WriteFile(src, []byte("index data"), 0o600)
+	if err != nil {
+		t.Fatalf("creating source index: %v", err)
+	}
+
+	dst := filepath.Join(dir, "snapshot.db")
+
+	sm := &SnapshotManager{fs: afero.NewOsFs()}
+
+	err = sm.copyFile(src, dst)
+	if err != nil {
+		t.Fatalf("copyFile: %v", err)
+	}
+
+	info, err := os.Stat(dst)
+	if err != nil {
+		t.Fatalf("stat export copy: %v", err)
+	}
+
+	got := info.Mode().Perm()
+	if got != 0o600 {
+		t.Fatalf("export copy permissions = %#o, want %#o", got, 0o600)
+	}
+}

internal/snapshot/manifest_gen_test.go

@@ -0,0 +1,64 @@
+//nolint:testpackage // exercises the unexported generateBlobManifest
+package snapshot
+
+import (
+	"context"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/types"
+)
+
+// TestGenerateBlobManifest_MissingBlobFails is the regression guard for
+// issue #157: a blob the snapshot references but that is absent from the
+// blobs table used to be logged and skipped, yielding a manifest with
+// fewer blobs than the snapshot needs. Since prune trusts the manifest
+// alone, that omitted blob would be deleted at the next prune. Manifest
+// generation must fail instead.
+func TestGenerateBlobManifest_MissingBlobFails(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ctx := context.Background()
+	dbPath := filepath.Join(t.TempDir(), "snapshot.db")
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	// A real blob row satisfies the snapshot_blobs foreign key on
+	// blob_id; the snapshot then references a different, absent hash.
+	presentBlob := &database.Blob{
+		ID:        types.NewBlobID(),
+		Hash:      types.BlobHash("present-blob-hash"),
+		CreatedTS: time.Now().Truncate(time.Second),
+	}
+	require.NoError(t, repos.Blobs.Create(ctx, nil, presentBlob))
+
+	snap := &database.Snapshot{
+		ID:       "testhost_home_2026-05-01T00:00:00Z",
+		Hostname: "testhost",
+	}
+	require.NoError(t, repos.Snapshots.Create(ctx, nil, snap))
+	require.NoError(t, repos.Snapshots.AddBlob(ctx, nil,
+		snap.ID.String(), presentBlob.ID, types.BlobHash("absent-blob-hash")))
+
+	require.NoError(t, db.Close())
+
+	sm := &SnapshotManager{
+		config: &config.Config{CompressionLevel: 3},
+		fs:     afero.NewOsFs(),
+	}
+
+	_, err = sm.generateBlobManifest(ctx, dbPath, snap.ID.String())
+	require.Error(t, err, "manifest generation must fail on a missing blob")
+	assert.Contains(t, err.Error(), "absent-blob-hash")
+}

internal/snapshot/scanner.go

@@ -63,7 +63,9 @@ type Scanner struct {
 	exclude           []string          // Glob patterns for files/directories to exclude
 	compiledExclude   []compiledPattern // Compiled glob patterns
 	progress          *ProgressReporter
-	skipErrors        bool // Skip file read errors (log loudly but continue)
+	// skipErrors skips files that cannot be opened or read (logged loudly);
+	// packer, database, encryption, and upload errors still abort the run.
+	skipErrors bool
 	// ui is the user-facing output; never nil (defaults to a discarding writer).
 	ui *ui.Writer
 
@@ -121,7 +123,9 @@ type ScannerConfig struct {
 	EnableProgress   bool       // Enable the live progress reporter (ETAs, throughput)
 	UI               *ui.Writer // Where user-facing scanner messages go; nil = discard
 	Exclude          []string   // Glob patterns for files/directories to exclude
-	SkipErrors       bool       // Skip file read errors (log loudly but continue)
+	// SkipErrors skips files that cannot be opened or read (log loudly but
+	// continue); packer, database, encryption, and upload errors still abort.
+	SkipErrors bool
 }
 
 // ScanResult contains the results of a scan operation
@@ -220,7 +224,14 @@ func (s *Scanner) Scan(
 		defer s.progress.Stop()
 	}
 
-	// Phase 0: Load known files and chunks from database into memory for fast lookup
+	// Phase 0: Repair any state left by an interrupted previous run, then
+	// load known files and chunks from the database into memory for fast
+	// lookup.
+	err := s.repairInterruptedBlobs(ctx)
+	if err != nil {
+		return nil, err
+	}
+
 	knownFiles, err := s.loadDatabaseState(ctx, path)
 	if err != nil {
 		return nil, err
@@ -317,6 +328,38 @@ func (s *Scanner) loadDatabaseState(
 	return knownFiles, nil
 }
 
+// repairInterruptedBlobs discards blob rows left by a previous run whose
+// upload never completed. Such a blob has its chunks, blob_chunks, and
+// blobs rows committed to the local index before the upload is attempted,
+// so a crash or dropped connection mid-upload leaves them behind while the
+// data never reaches remote storage. Deduplicating against those chunks on
+// a later run would produce a snapshot that reports success but cannot be
+// restored. Dropping the un-uploaded blobs (their blob_chunks cascade) and
+// then any chunks left unreferenced forces the affected data to be
+// re-chunked and re-uploaded this run. A blob is attached to a snapshot
+// only once its upload is recorded, so this never touches a completed
+// snapshot's data.
+func (s *Scanner) repairInterruptedBlobs(ctx context.Context) error {
+	removed, err := s.repos.Blobs.DeleteUnuploaded(ctx)
+	if err != nil {
+		return fmt.Errorf("removing un-uploaded blob records: %w", err)
+	}
+
+	if removed == 0 {
+		return nil
+	}
+
+	log.Warn("Discarded blob records from an interrupted previous run; "+
+		"their data will be re-uploaded", "blobs", removed)
+
+	err = s.repos.Chunks.DeleteOrphaned(ctx)
+	if err != nil {
+		return fmt.Errorf("removing orphaned chunks: %w", err)
+	}
+
+	return nil
+}
+
 // summarizeScanPhase calculates total size to process, updates progress tracking,
 // and prints the scan phase summary with file counts and sizes
 func (s *Scanner) summarizeScanPhase(
@@ -392,11 +435,14 @@ func (s *Scanner) loadKnownFiles(
 	return result, nil
 }
 
-// loadKnownChunks loads all known chunk hashes from the database into a
+// loadKnownChunks loads the chunk hashes safe to deduplicate against into
-// map for fast lookup. This avoids per-chunk database queries during file
+// an in-memory map for fast lookup, avoiding per-chunk database queries
-// processing.
+// during file processing. Only chunks held by a blob whose upload
+// completed are loaded: a chunk left behind by an interrupted upload
+// refers to data that never reached remote storage, and deduplicating
+// against it would silently produce an unrestorable snapshot.
 func (s *Scanner) loadKnownChunks(ctx context.Context) error {
-	chunks, err := s.repos.Chunks.List(ctx)
+	chunks, err := s.repos.Chunks.ListInUploadedBlobs(ctx)
 	if err != nil {
 		return fmt.Errorf("listing chunks: %w", err)
 	}
@@ -1294,6 +1340,15 @@ func (s *Scanner) processFileWithErrorHandling(
 ) (bool, error) {
 	err := s.processFileStreaming(ctx, fileToProcess, result)
 	if err != nil {
+		// A packer/database/encryption/upload failure means the chunk's data
+		// may not have been stored. Skipping the file would let the snapshot
+		// record a file whose chunk is in no blob and cannot be restored, so
+		// abort the run even under --skip-errors. Only open and read errors
+		// are skipped below.
+		var pErr *packerError
+		if errors.As(err, &pErr) {
+			return false, fmt.Errorf("processing file %s: %w", fileToProcess.Path, err)
+		}
 		// Handle files that were deleted between scan and process phases
 		if errors.Is(err, os.ErrNotExist) {
 			log.Warn("File was deleted during backup, skipping",
@@ -1303,7 +1358,7 @@ func (s *Scanner) processFileWithErrorHandling(
 
 			return true, nil
 		}
-		// Skip file read errors if --skip-errors is enabled
+		// Skip open/read errors if --skip-errors is enabled
 		if s.skipErrors {
 			log.Error("Failed to process file (skipping due to --skip-errors)",
 				"path", fileToProcess.Path, "error", err)
@@ -1401,7 +1456,17 @@ func (s *Scanner) finalizeProcessPhase(ctx context.Context, result *ScanResult)
 				return fmt.Errorf("parsing blob ID: %w", err)
 			}
 
+			// With no remote backend the blob's lifecycle ends here, so
+			// mark it uploaded in the same transaction that attaches it to
+			// the snapshot. This keeps the invariant that any blob a
+			// snapshot references has uploaded_ts set, so deduplication and
+			// interrupted-run repair treat these blobs as trustworthy.
 			err = s.repos.WithTx(ctx, func(ctx context.Context, tx *sql.Tx) error {
+				err := s.repos.Blobs.UpdateUploaded(ctx, tx, b.ID)
+				if err != nil {
+					return fmt.Errorf("marking blob uploaded: %w", err)
+				}
+
 				return s.repos.Snapshots.AddBlob(ctx, tx, s.snapshotID, blobID,
 					types.BlobHash(b.Hash))
 			})
@@ -1660,6 +1725,20 @@ type streamingChunkInfo struct {
 	size      int64
 }
 
+// packerError marks an error that came from adding a chunk to the packer
+// (packing, database, encryption, or upload). Such an error means the chunk's
+// data may not have been stored, so the run must abort even under --skip-errors:
+// skipping the file would leave the chunk recorded as backed up while it lives
+// in no blob, and a later snapshot could record a file that cannot be restored.
+// Only open and read errors are safe to skip.
+type packerError struct {
+	err error
+}
+
+func (e *packerError) Error() string { return e.err.Error() }
+
+func (e *packerError) Unwrap() error { return e.err }
+
 // processFileStreaming processes a file by streaming chunks directly to the packer
 func (s *Scanner) processFileStreaming(
 	ctx context.Context, fileToProcess *FileToProcess, result *ScanResult,
@@ -1710,7 +1789,11 @@ func (s *Scanner) processFileStreaming(
 		if !chunkExists {
 			err := s.addChunkToPacker(ctx, chunk)
 			if err != nil {
-				return err
+				// Mark as a packer error so --skip-errors cannot swallow it:
+				// the chunk was registered as pending before packing, so a
+				// skipped file here would be recorded as backed up while its
+				// data was never stored.
+				return &packerError{err: err}
 			}
 		}
 

internal/snapshot/skip_errors_test.go

@@ -0,0 +1,216 @@
+package snapshot_test
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/spf13/afero"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+)
+
+// errSimTempFail is the one-time temp-file creation failure blobTempFailFs
+// injects, mirroring a full temp filesystem.
+var errSimTempFail = errors.New("simulated temp-file creation failure")
+
+// errSimRead is the read failure readFailFile injects for a file that opens
+// but cannot be read.
+var errSimRead = errors.New("simulated read failure")
+
+// blobTempFailFs fails the first temp-file creation for a packer blob, then
+// behaves normally, simulating a one-time failure to start a new blob.
+type blobTempFailFs struct {
+	afero.Fs
+
+	mu     sync.Mutex
+	failed bool
+}
+
+//nolint:ireturn // afero.Fs.OpenFile is defined to return the interface.
+func (f *blobTempFailFs) OpenFile(
+	name string, flag int, perm os.FileMode,
+) (afero.File, error) {
+	if strings.Contains(name, "vaultik-blob-") {
+		f.mu.Lock()
+		firstTime := !f.failed
+		f.failed = true
+		f.mu.Unlock()
+
+		if firstTime {
+			return nil, errSimTempFail
+		}
+	}
+
+	return f.Fs.OpenFile(name, flag, perm)
+}
+
+// readFailFile wraps an afero.File whose Read always fails.
+type readFailFile struct {
+	afero.File
+}
+
+func (readFailFile) Read([]byte) (int, error) {
+	return 0, errSimRead
+}
+
+// readFailFs fails reads of one target path after a successful open.
+type readFailFs struct {
+	afero.Fs
+
+	target string
+}
+
+//nolint:ireturn // afero.Fs.Open is defined to return the interface.
+func (f *readFailFs) Open(name string) (afero.File, error) {
+	file, err := f.Fs.Open(name)
+	if err != nil {
+		return nil, err
+	}
+
+	if name == f.target {
+		return readFailFile{File: file}, nil
+	}
+
+	return file, nil
+}
+
+// writeSkipErrorTestFile writes one file into fs with a fixed mtime.
+func writeSkipErrorTestFile(t *testing.T, fs afero.Fs, path, content string) {
+	t.Helper()
+
+	err := fs.MkdirAll(filepath.Dir(path), 0755)
+	if err != nil {
+		t.Fatalf("mkdir: %v", err)
+	}
+
+	err = afero.WriteFile(fs, path, []byte(content), 0644)
+	if err != nil {
+		t.Fatalf("write %s: %v", path, err)
+	}
+
+	when := time.Date(2024, 1, 1, 12, 0, 0, 0, time.UTC)
+
+	err = fs.Chtimes(path, when, when)
+	if err != nil {
+		t.Fatalf("chtimes %s: %v", path, err)
+	}
+}
+
+// runSkipErrorScan scans /source on fs with the given skip-errors setting and
+// returns the repositories (for inspection) and the scan error.
+func runSkipErrorScan(
+	t *testing.T, fs afero.Fs, skipErrors bool,
+) (*database.Repositories, error) {
+	t.Helper()
+
+	db, err := database.NewTestDB()
+	if err != nil {
+		t.Fatalf("create test db: %v", err)
+	}
+
+	t.Cleanup(func() {
+		cerr := db.Close()
+		if cerr != nil {
+			t.Errorf("close db: %v", cerr)
+		}
+	})
+
+	repos := database.NewRepositories(db)
+
+	scanner := snapshot.NewScanner(snapshot.ScannerConfig{
+		FS:               fs,
+		ChunkSize:        int64(1024 * 16),
+		Repositories:     repos,
+		MaxBlobSize:      int64(1024 * 1024),
+		CompressionLevel: 3,
+		AgeRecipients:    []string{testAgePublicKey},
+		SkipErrors:       skipErrors,
+	})
+
+	ctx := context.Background()
+	snapshotID := "test-snapshot-skip-errors"
+	createTestSnapshotRecord(ctx, t, repos, snapshotID)
+
+	_, err = scanner.Scan(ctx, "/source", snapshotID)
+
+	return repos, err
+}
+
+// TestScannerPackingFailureAbortsUnderSkipErrors checks that a failure to start
+// a new blob aborts the run even with --skip-errors. Otherwise the file would
+// be skipped while its chunk had already been registered as pending, letting a
+// later blob record that chunk in the chunks table with no blob to back it —
+// a snapshot that completes with a file that cannot be restored.
+func TestScannerPackingFailureAbortsUnderSkipErrors(t *testing.T) {
+	t.Parallel()
+
+	// Two files with distinct content so each yields a distinct chunk: the
+	// first fails to start a blob, and without the fix the second's blob would
+	// commit the first's orphaned chunk row.
+	fs := &blobTempFailFs{Fs: afero.NewMemMapFs()}
+	writeSkipErrorTestFile(t, fs, "/source/file1.txt", "first file content")
+	writeSkipErrorTestFile(t, fs, "/source/file2.txt", "second file content")
+
+	repos, err := runSkipErrorScan(t, fs, true)
+	if err == nil {
+		t.Fatal("expected scan to abort on the packer error, got nil")
+	}
+
+	// ListUnpacked returns chunks recorded with no blob_chunks row: exactly the
+	// unrestorable state this fix prevents.
+	unpacked, err := repos.Chunks.ListUnpacked(context.Background(), 10)
+	if err != nil {
+		t.Fatalf("listing unpacked chunks: %v", err)
+	}
+
+	if len(unpacked) != 0 {
+		t.Fatalf("expected no chunk recorded without a blob, got %d", len(unpacked))
+	}
+}
+
+// TestScannerReadErrorAbortsWithoutSkipErrors checks that a file read error
+// aborts the run when --skip-errors is not set.
+func TestScannerReadErrorAbortsWithoutSkipErrors(t *testing.T) {
+	t.Parallel()
+
+	const target = "/source/unreadable.txt"
+
+	fs := &readFailFs{Fs: afero.NewMemMapFs(), target: target}
+	writeSkipErrorTestFile(t, fs, target, "content that cannot be read")
+
+	_, err := runSkipErrorScan(t, fs, false)
+	if err == nil {
+		t.Fatal("expected scan to fail on the read error, got nil")
+	}
+}
+
+// TestScannerReadErrorSkippedWithSkipErrors checks that a file read error is
+// skipped and the run completes when --skip-errors is set.
+func TestScannerReadErrorSkippedWithSkipErrors(t *testing.T) {
+	t.Parallel()
+
+	const target = "/source/unreadable.txt"
+
+	fs := &readFailFs{Fs: afero.NewMemMapFs(), target: target}
+	writeSkipErrorTestFile(t, fs, target, "content that cannot be read")
+
+	repos, err := runSkipErrorScan(t, fs, true)
+	if err != nil {
+		t.Fatalf("expected scan to complete with --skip-errors, got %v", err)
+	}
+
+	chunks, err := repos.FileChunks.GetByFile(context.Background(), target)
+	if err != nil {
+		t.Fatalf("getting file chunks: %v", err)
+	}
+
+	if len(chunks) != 0 {
+		t.Fatalf("expected unreadable file skipped, got %d chunks", len(chunks))
+	}
+}

internal/snapshot/snapshot.go

@@ -44,6 +44,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"os"
 	"path/filepath"
 	"strings"
 	"time"
@@ -294,68 +295,6 @@ func (sm *SnapshotManager) ExportSnapshotMetadata(
 	return nil
 }
 
-// CleanupIncompleteSnapshots removes incomplete snapshots that don't have
-// metadata in S3. This is critical for data safety: incomplete snapshots
-// can cause deduplication to skip files that were never successfully
-// backed up, resulting in data loss.
-func (sm *SnapshotManager) CleanupIncompleteSnapshots(
-	ctx context.Context, hostname string,
-) error {
-	log.Info("Checking for incomplete snapshots", "hostname", hostname)
-
-	// Get all incomplete snapshots for this hostname
-	incompleteSnapshots, err := sm.repos.Snapshots.GetIncompleteByHostname(ctx, hostname)
-	if err != nil {
-		return fmt.Errorf("getting incomplete snapshots: %w", err)
-	}
-
-	if len(incompleteSnapshots) == 0 {
-		log.Debug("No incomplete snapshots found")
-
-		return nil
-	}
-
-	log.Info("Found incomplete snapshots", "count", len(incompleteSnapshots))
-
-	// Check each incomplete snapshot for metadata in storage
-	for _, snapshot := range incompleteSnapshots {
-		// Check if metadata exists in storage (paths use the hashed
-		// remote key so we don't leak host info to the listing).
-		metadataKey := fmt.Sprintf("metadata/%s/db.zst",
-			RemoteSnapshotKey(snapshot.ID.String()))
-
-		_, err := sm.storage.Stat(ctx, metadataKey)
-		if err != nil {
-			// Metadata doesn't exist in S3 - this is an incomplete snapshot
-			log.Info("Cleaning up incomplete snapshot record",
-				"snapshot_id", snapshot.ID, "started_at", snapshot.StartedAt)
-
-			// Delete the snapshot and all its associations
-			err := sm.deleteSnapshot(ctx, snapshot.ID.String())
-			if err != nil {
-				return fmt.Errorf("deleting incomplete snapshot %s: %w",
-					snapshot.ID, err)
-			}
-
-			log.Info("Deleted incomplete snapshot record and associated data",
-				"snapshot_id", snapshot.ID)
-		} else {
-			// Metadata exists - this snapshot was completed but database wasn't updated
-			// This shouldn't happen in normal operation, but mark it complete
-			log.Warn("Found snapshot with remote metadata but incomplete in database",
-				"snapshot_id", snapshot.ID)
-
-			err := sm.repos.Snapshots.MarkComplete(ctx, nil, snapshot.ID.String())
-			if err != nil {
-				log.Error("Failed to mark snapshot as complete in database",
-					"snapshot_id", snapshot.ID, "error", err)
-			}
-		}
-	}
-
-	return nil
-}
-
 // CleanupOrphanedData removes files, chunks, and blobs that are no longer
 // referenced by any snapshot. This should be called periodically to clean
 // up data from deleted or incomplete snapshots.
@@ -758,12 +697,18 @@ func (sm *SnapshotManager) compressFile(inputPath, outputPath string) error {
 
 	writerClosed = true
 
-	log.Debug("Compression complete", "hash", hex.EncodeToString(writer.Sum256()))
+	log.Debug("Compression complete", "hash", hex.EncodeToString(writer.ContentID()))
 
 	return nil
 }
 
-// copyFile copies a file from src to dst
+// exportCopyPerm restricts the exported snapshot database copy to the owning
+// user; it holds the same private index data as the local index file.
+const exportCopyPerm = 0o600
+
+// copyFile copies a file from src to dst. The destination is the exported
+// snapshot database, so it is created owner-only rather than with the
+// umask-dependent default.
 func (sm *SnapshotManager) copyFile(src, dst string) error {
 	log.Debug("Opening source file for copy", "path", src)
 
@@ -783,7 +728,9 @@ func (sm *SnapshotManager) copyFile(src, dst string) error {
 
 	log.Debug("Creating destination file", "path", dst)
 
-	destFile, err := sm.fs.Create(dst)
+	destFile, err := sm.fs.OpenFile(
+		dst, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, exportCopyPerm,
+	)
 	if err != nil {
 		return err
 	}
@@ -809,6 +756,11 @@ func (sm *SnapshotManager) copyFile(src, dst string) error {
 	return nil
 }
 
+// errBlobMissingFromDatabase means a snapshot references a blob that is
+// absent from the blobs table, so a complete manifest cannot be built.
+var errBlobMissingFromDatabase = errors.New(
+	"blob referenced by snapshot is not in the database")
+
 // generateBlobManifest creates a compressed JSON list of all blobs in the snapshot
 func (sm *SnapshotManager) generateBlobManifest(
 	ctx context.Context, dbPath string, snapshotID string,
@@ -839,20 +791,26 @@ func (sm *SnapshotManager) generateBlobManifest(
 	totalCompressedSize := int64(0)
 
 	for _, hash := range blobHashes {
+		// Every blob the snapshot references must appear in the manifest.
+		// Prune consults only the manifest to decide what is still in use,
+		// so silently dropping a blob here would let a later prune delete
+		// it while this snapshot still needs it. A lookup failure or a
+		// missing blob row therefore fails manifest generation.
 		blob, err := repos.Blobs.GetByHash(ctx, hash)
 		if err != nil {
-			log.Warn("Failed to get blob details", "hash", hash, "error", err)
+			return nil, fmt.Errorf("getting blob details for %s: %w", hash, err)
-
-			continue
 		}
 
-		if blob != nil {
+		if blob == nil {
-			blobs = append(blobs, BlobInfo{
+			return nil, fmt.Errorf("%w: blob %s, snapshot %s",
-				Hash:           hash,
+				errBlobMissingFromDatabase, hash, snapshotID)
-				CompressedSize: blob.CompressedSize,
-			})
-			totalCompressedSize += blob.CompressedSize
 		}
+
+		blobs = append(blobs, BlobInfo{
+			Hash:           hash,
+			CompressedSize: blob.CompressedSize,
+		})
+		totalCompressedSize += blob.CompressedSize
 	}
 
 	// Create manifest. SnapshotID in the unencrypted manifest is the
@@ -913,45 +871,6 @@ type ExtendedBackupStats struct {
 	UploadDurationMs     int64 // Total milliseconds spent uploading to S3
 }
 
-// deleteSnapshot removes a snapshot and all its associations from the database
-func (sm *SnapshotManager) deleteSnapshot(
-	ctx context.Context, snapshotID string,
-) error {
-	// Delete snapshot_files entries
-	err := sm.repos.Snapshots.DeleteSnapshotFiles(ctx, snapshotID)
-	if err != nil {
-		return fmt.Errorf("deleting snapshot files: %w", err)
-	}
-
-	// Delete snapshot_blobs entries
-	err = sm.repos.Snapshots.DeleteSnapshotBlobs(ctx, snapshotID)
-	if err != nil {
-		return fmt.Errorf("deleting snapshot blobs: %w", err)
-	}
-
-	// Delete uploads entries (has foreign key to snapshots without CASCADE)
-	err = sm.repos.Snapshots.DeleteSnapshotUploads(ctx, snapshotID)
-	if err != nil {
-		return fmt.Errorf("deleting snapshot uploads: %w", err)
-	}
-
-	// Delete the snapshot itself
-	err = sm.repos.Snapshots.Delete(ctx, snapshotID)
-	if err != nil {
-		return fmt.Errorf("deleting snapshot: %w", err)
-	}
-
-	// Clean up orphaned data
-	log.Debug("Cleaning up orphaned records in main database")
-
-	err = sm.CleanupOrphanedData(ctx)
-	if err != nil {
-		return fmt.Errorf("cleaning up orphaned data: %w", err)
-	}
-
-	return nil
-}
-
 // deleteOtherSnapshots deletes all snapshots except the current one
 func (sm *SnapshotManager) deleteOtherSnapshots(
 	ctx context.Context, tx *sql.Tx, currentSnapshotID string,

internal/storage/conformance_test.go

@@ -0,0 +1,198 @@
+package storage_test
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"io"
+	"reflect"
+	"sort"
+	"testing"
+
+	"sneak.berlin/go/vaultik/internal/storage"
+)
+
+// runStorerConformance is the shared Storer contract. Every backend that
+// can run in-process is expected to pass it: TestFileStorer runs it against
+// file://, TestS3Storer against s3://. A new backend inherits this coverage
+// by passing its own constructor, so the contract is defined once.
+//
+// It exercises the public Storer interface: round-trip, stat, list with
+// prefix filtering, overwrite, delete, delete-of-missing, and not-found on
+// Get and Stat. Each section takes its own fresh backend instance, so the
+// order of sections never matters and no section sees another's objects.
+func runStorerConformance(t *testing.T, newStorer func(*testing.T) storage.Storer) {
+	t.Helper()
+
+	conformanceRoundTrip(t, newStorer(t))
+	conformanceOverwrite(t, newStorer(t))
+	conformanceList(t, newStorer(t))
+	conformanceDelete(t, newStorer(t))
+	conformanceNotFound(t, newStorer(t))
+}
+
+// conformanceRoundTrip stores a nested key, then reads it back and stats it.
+func conformanceRoundTrip(t *testing.T, s storage.Storer) {
+	t.Helper()
+
+	ctx := context.Background()
+	key := "blobs/aa/bb/object.bin"
+	want := []byte("round-trip payload")
+
+	err := s.Put(ctx, key, bytes.NewReader(want))
+	if err != nil {
+		t.Fatalf("Put: %v", err)
+	}
+
+	got := getBytes(t, s, key)
+	if !bytes.Equal(got, want) {
+		t.Errorf("Get returned %q, want %q", got, want)
+	}
+
+	info, err := s.Stat(ctx, key)
+	if err != nil {
+		t.Fatalf("Stat: %v", err)
+	}
+
+	if info.Key != key {
+		t.Errorf("Stat key = %q, want %q", info.Key, key)
+	}
+
+	if info.Size != int64(len(want)) {
+		t.Errorf("Stat size = %d, want %d", info.Size, len(want))
+	}
+}
+
+// conformanceOverwrite checks that a second Put replaces the first.
+func conformanceOverwrite(t *testing.T, s storage.Storer) {
+	t.Helper()
+
+	ctx := context.Background()
+	key := "meta/snapshot.json"
+
+	err := s.Put(ctx, key, bytes.NewReader([]byte("first")))
+	if err != nil {
+		t.Fatalf("first Put: %v", err)
+	}
+
+	want := []byte("second and longer payload")
+
+	err = s.Put(ctx, key, bytes.NewReader(want))
+	if err != nil {
+		t.Fatalf("second Put: %v", err)
+	}
+
+	got := getBytes(t, s, key)
+	if !bytes.Equal(got, want) {
+		t.Errorf("after overwrite Get returned %q, want %q", got, want)
+	}
+}
+
+// conformanceList checks prefix filtering and the empty result for a
+// prefix that matches nothing.
+func conformanceList(t *testing.T, s storage.Storer) {
+	t.Helper()
+
+	ctx := context.Background()
+	keys := []string{"blobs/aa/one", "blobs/bb/two", "meta/three"}
+
+	for _, k := range keys {
+		err := s.Put(ctx, k, bytes.NewReader([]byte("data")))
+		if err != nil {
+			t.Fatalf("Put %q: %v", k, err)
+		}
+	}
+
+	if got := listSorted(t, s, ""); !reflect.DeepEqual(got, keys) {
+		t.Errorf("List(\"\") = %v, want %v", got, keys)
+	}
+
+	wantBlobs := []string{"blobs/aa/one", "blobs/bb/two"}
+	if got := listSorted(t, s, "blobs/"); !reflect.DeepEqual(got, wantBlobs) {
+		t.Errorf("List(\"blobs/\") = %v, want %v", got, wantBlobs)
+	}
+
+	if got := listSorted(t, s, "absent/"); len(got) != 0 {
+		t.Errorf("List(\"absent/\") = %v, want empty", got)
+	}
+}
+
+// conformanceDelete checks that Delete removes an object and that deleting
+// a missing key is not an error.
+func conformanceDelete(t *testing.T, s storage.Storer) {
+	t.Helper()
+
+	ctx := context.Background()
+	key := "blobs/cc/gone.bin"
+
+	err := s.Put(ctx, key, bytes.NewReader([]byte("temporary")))
+	if err != nil {
+		t.Fatalf("Put: %v", err)
+	}
+
+	err = s.Delete(ctx, key)
+	if err != nil {
+		t.Fatalf("Delete: %v", err)
+	}
+
+	_, err = s.Get(ctx, key)
+	if !errors.Is(err, storage.ErrNotFound) {
+		t.Errorf("Get after Delete error = %v, want ErrNotFound", err)
+	}
+
+	err = s.Delete(ctx, key)
+	if err != nil {
+		t.Errorf("Delete of missing key = %v, want nil", err)
+	}
+}
+
+// conformanceNotFound checks Get and Stat on an absent key.
+func conformanceNotFound(t *testing.T, s storage.Storer) {
+	t.Helper()
+
+	ctx := context.Background()
+	key := "never/written"
+
+	_, err := s.Get(ctx, key)
+	if !errors.Is(err, storage.ErrNotFound) {
+		t.Errorf("Get error = %v, want ErrNotFound", err)
+	}
+
+	_, err = s.Stat(ctx, key)
+	if !errors.Is(err, storage.ErrNotFound) {
+		t.Errorf("Stat error = %v, want ErrNotFound", err)
+	}
+}
+
+// getBytes reads a key fully and closes the reader.
+func getBytes(t *testing.T, s storage.Storer, key string) []byte {
+	t.Helper()
+
+	rc, err := s.Get(context.Background(), key)
+	if err != nil {
+		t.Fatalf("Get %q: %v", key, err)
+	}
+
+	defer func() { _ = rc.Close() }()
+
+	data, err := io.ReadAll(rc)
+	if err != nil {
+		t.Fatalf("read %q: %v", key, err)
+	}
+
+	return data
+}
+
+// listSorted returns the keys under a prefix in a stable order.
+func listSorted(t *testing.T, s storage.Storer, prefix string) []string {
+	t.Helper()
+
+	keys, err := s.List(context.Background(), prefix)
+	if err != nil {
+		t.Fatalf("List %q: %v", prefix, err)
+	}
+
+	sort.Strings(keys)
+
+	return keys
+}

internal/storage/faultstore/faultstore.go

@@ -0,0 +1,209 @@
+// Package faultstore provides a storage.Storer wrapper that injects
+// faults on demand, so tests can reproduce the failure modes a real
+// backend exhibits: an upload that fails partway, a backend that reports
+// success while storing nothing, and reads that return corrupt or
+// truncated bytes. It is the seam called for by the fault-injection
+// tests (sneak/vaultik issue 72) and is meant to be reused by future
+// tests rather than re-implemented per case.
+//
+// The wrapper delegates every method to the inner Storer. Two hooks
+// change that: OnPut decides the fate of each write, and OnGet decides
+// how each read's bytes are returned. Both are keyed by the object key,
+// so a test can fault only blobs, only metadata, or a single object.
+package faultstore
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+
+	"sneak.berlin/go/vaultik/internal/storage"
+)
+
+// ErrInjectedUpload is returned by a Put the OnPut hook chose to fail.
+var ErrInjectedUpload = errors.New("faultstore: injected upload failure")
+
+// PutAction is the disposition OnPut assigns to a write.
+type PutAction int
+
+const (
+	// PutNormal writes through to the inner Storer.
+	PutNormal PutAction = iota
+	// PutFail reads part of the stream, then fails without storing the
+	// object — a network upload that dies partway through.
+	PutFail
+	// PutSwallow reports success but stores nothing — a backend that
+	// lies about durability.
+	PutSwallow
+)
+
+// GetFault is how OnGet chooses to damage a read.
+type GetFault int
+
+const (
+	// GetNormal returns the stored bytes unchanged.
+	GetNormal GetFault = iota
+	// GetCorrupt flips a byte so the returned object no longer matches
+	// what was stored.
+	GetCorrupt
+	// GetTruncate returns a short read: the object's bytes cut off
+	// before the end.
+	GetTruncate
+)
+
+// Storer wraps an inner storage.Storer with fault-injection hooks. A
+// zero-valued hook means "no fault": construct with New and set only the
+// hook a test needs.
+type Storer struct {
+	inner storage.Storer
+
+	// OnPut, when set, is consulted before every Put and
+	// PutWithProgress with the object key.
+	OnPut func(key string) PutAction
+
+	// OnGet, when set, is consulted for every Get with the object key
+	// and damages the returned bytes accordingly.
+	OnGet func(key string) GetFault
+}
+
+// New wraps inner. inner must be non-nil.
+func New(inner storage.Storer) *Storer {
+	return &Storer{inner: inner}
+}
+
+// midStreamBytes is how far a PutFail reads before failing, enough to be
+// past the start of any real blob without depending on the blob's size.
+const midStreamBytes = 512
+
+// Put stores data unless OnPut faults the write.
+func (f *Storer) Put(ctx context.Context, key string, data io.Reader) error {
+	handled, err := f.injectPut(key, data)
+	if handled {
+		return err
+	}
+
+	return f.inner.Put(ctx, key, data)
+}
+
+// PutWithProgress stores data unless OnPut faults the write.
+func (f *Storer) PutWithProgress(
+	ctx context.Context, key string, data io.Reader,
+	size int64, progress storage.ProgressCallback,
+) error {
+	handled, err := f.injectPut(key, data)
+	if handled {
+		return err
+	}
+
+	return f.inner.PutWithProgress(ctx, key, data, size, progress)
+}
+
+// Get retrieves data, damaging it if OnGet faults the read.
+func (f *Storer) Get(ctx context.Context, key string) (io.ReadCloser, error) {
+	rc, err := f.inner.Get(ctx, key)
+	if err != nil {
+		return nil, err
+	}
+
+	fault := GetNormal
+	if f.OnGet != nil {
+		fault = f.OnGet(key)
+	}
+
+	if fault == GetNormal {
+		return rc, nil
+	}
+
+	data, err := io.ReadAll(rc)
+	_ = rc.Close()
+
+	if err != nil {
+		return nil, err
+	}
+
+	return io.NopCloser(bytes.NewReader(damage(fault, data))), nil
+}
+
+// damage returns a faulted copy of the stored bytes. GetCorrupt flips a
+// byte in the middle so decryption authentication fails; GetTruncate
+// drops the final byte so the read ends short. Both are no-ops on empty
+// input, which cannot be damaged into something distinguishable.
+func damage(fault GetFault, data []byte) []byte {
+	out := make([]byte, len(data))
+	copy(out, data)
+
+	if len(out) == 0 {
+		return out
+	}
+
+	switch fault {
+	case GetCorrupt:
+		out[len(out)/2] ^= 0xff
+	case GetTruncate:
+		out = out[:len(out)-1]
+	case GetNormal:
+	}
+
+	return out
+}
+
+// Stat delegates unchanged.
+func (f *Storer) Stat(ctx context.Context, key string) (*storage.ObjectInfo, error) {
+	return f.inner.Stat(ctx, key)
+}
+
+// Delete delegates unchanged.
+func (f *Storer) Delete(ctx context.Context, key string) error {
+	return f.inner.Delete(ctx, key)
+}
+
+// List delegates unchanged.
+func (f *Storer) List(ctx context.Context, prefix string) ([]string, error) {
+	return f.inner.List(ctx, prefix)
+}
+
+// ListStream delegates unchanged.
+func (f *Storer) ListStream(
+	ctx context.Context, prefix string,
+) <-chan storage.ObjectInfo {
+	return f.inner.ListStream(ctx, prefix)
+}
+
+// Info delegates unchanged.
+func (f *Storer) Info() storage.Info {
+	return f.inner.Info()
+}
+
+func (f *Storer) putAction(key string) PutAction {
+	if f.OnPut == nil {
+		return PutNormal
+	}
+
+	return f.OnPut(key)
+}
+
+// injectPut handles the non-normal write dispositions. It reports
+// whether it handled the write and, if so, with what error.
+func (f *Storer) injectPut(key string, data io.Reader) (bool, error) {
+	switch f.putAction(key) {
+	case PutFail:
+		// Consume part of the stream so the failure lands mid-transfer,
+		// the way a dropped connection would, then error without
+		// storing anything.
+		_, _ = io.CopyN(io.Discard, data, midStreamBytes)
+
+		return true, fmt.Errorf("%w for %q", ErrInjectedUpload, key)
+	case PutSwallow:
+		// A lying backend still drains the request body, then keeps
+		// nothing.
+		_, _ = io.Copy(io.Discard, data)
+
+		return true, nil
+	case PutNormal:
+		return false, nil
+	default:
+		return false, nil
+	}
+}

internal/storage/file.go

@@ -46,31 +46,18 @@ func (f *FileStorer) SetFilesystem(fs afero.Fs) {
 // storage base path.
 const storageDirPerm = 0o755
 
+// tempSuffix marks a partially written object. writeAtomic streams into a
+// temp file carrying this suffix and only renames it onto the real key once
+// the whole object is on disk, so an interrupted write can never leave a
+// truncated object at the key a later run would Stat and trust as a complete
+// blob. List and ListStream skip these files, so a leftover from an
+// interrupted write is never listed or trusted as a blob; it is otherwise
+// harmless and is overwritten when the same key is written again.
+const tempSuffix = ".partial"
+
 // Put stores data at the specified key.
 func (f *FileStorer) Put(_ context.Context, key string, data io.Reader) error {
-	path := f.fullPath(key)
+	return f.writeAtomic(key, data, nil)
-
-	// Create parent directories
-	dir := filepath.Dir(path)
-
-	err := f.fs.MkdirAll(dir, storageDirPerm)
-	if err != nil {
-		return fmt.Errorf("creating directories: %w", err)
-	}
-
-	file, err := f.fs.Create(path)
-	if err != nil {
-		return fmt.Errorf("creating file: %w", err)
-	}
-
-	defer func() { _ = file.Close() }()
-
-	_, err = io.Copy(file, data)
-	if err != nil {
-		return fmt.Errorf("writing file: %w", err)
-	}
-
-	return nil
 }
 
 // PutWithProgress stores data with progress reporting.
@@ -78,35 +65,7 @@ func (f *FileStorer) PutWithProgress(
 	_ context.Context, key string, data io.Reader,
 	_ int64, progress ProgressCallback,
 ) error {
-	path := f.fullPath(key)
+	return f.writeAtomic(key, data, progress)
-
-	// Create parent directories
-	dir := filepath.Dir(path)
-
-	err := f.fs.MkdirAll(dir, storageDirPerm)
-	if err != nil {
-		return fmt.Errorf("creating directories: %w", err)
-	}
-
-	file, err := f.fs.Create(path)
-	if err != nil {
-		return fmt.Errorf("creating file: %w", err)
-	}
-
-	defer func() { _ = file.Close() }()
-
-	// Wrap with progress tracking
-	pw := &progressWriter{
-		writer:   file,
-		callback: progress,
-	}
-
-	_, err = io.Copy(pw, data)
-	if err != nil {
-		return fmt.Errorf("writing file: %w", err)
-	}
-
-	return nil
 }
 
 // Get retrieves data from the specified key.
@@ -188,7 +147,7 @@ func (f *FileStorer) List(ctx context.Context, prefix string) ([]string, error)
 		default:
 		}
 
-		if !info.IsDir() {
+		if !info.IsDir() && !strings.HasSuffix(info.Name(), tempSuffix) {
 			// Convert back to key (relative path from basePath)
 			relPath, err := filepath.Rel(f.basePath, path)
 			if err != nil {
@@ -245,7 +204,7 @@ func (f *FileStorer) ListStream(ctx context.Context, prefix string) <-chan Objec
 				return nil //nolint:nilerr // continue walking despite errors
 			}
 
-			if !info.IsDir() {
+			if !info.IsDir() && !strings.HasSuffix(info.Name(), tempSuffix) {
 				relPath, err := filepath.Rel(f.basePath, path)
 				if err != nil {
 					ch <- ObjectInfo{Err: fmt.Errorf("computing relative path: %w", err)}
@@ -275,6 +234,72 @@ func (f *FileStorer) Info() Info {
 	}
 }
 
+// writeAtomic streams data into a temp file in the destination directory,
+// fsyncs it, and renames it onto the final key. The key therefore appears
+// only once the whole object has been durably written; a failure part-way
+// leaves a temp file (removed here on the failing path) rather than a
+// truncated object at the key.
+func (f *FileStorer) writeAtomic(
+	key string, data io.Reader, progress ProgressCallback,
+) error {
+	path := f.fullPath(key)
+	dir := filepath.Dir(path)
+
+	err := f.fs.MkdirAll(dir, storageDirPerm)
+	if err != nil {
+		return fmt.Errorf("creating directories: %w", err)
+	}
+
+	tmp, err := afero.TempFile(f.fs, dir, filepath.Base(path)+"-*"+tempSuffix)
+	if err != nil {
+		return fmt.Errorf("creating temp file: %w", err)
+	}
+
+	tmpPath := tmp.Name()
+
+	// Remove the temp file unless the rename below claims it. On the success
+	// path renamed is true, so the deferred Close and Remove are harmless
+	// no-ops on a name that no longer exists.
+	renamed := false
+
+	defer func() {
+		_ = tmp.Close()
+
+		if !renamed {
+			_ = f.fs.Remove(tmpPath)
+		}
+	}()
+
+	var w io.Writer = tmp
+	if progress != nil {
+		w = &progressWriter{writer: tmp, callback: progress}
+	}
+
+	_, err = io.Copy(w, data)
+	if err != nil {
+		return fmt.Errorf("writing file: %w", err)
+	}
+
+	err = tmp.Sync()
+	if err != nil {
+		return fmt.Errorf("syncing temp file: %w", err)
+	}
+
+	err = tmp.Close()
+	if err != nil {
+		return fmt.Errorf("closing temp file: %w", err)
+	}
+
+	err = f.fs.Rename(tmpPath, path)
+	if err != nil {
+		return fmt.Errorf("renaming temp file: %w", err)
+	}
+
+	renamed = true
+
+	return nil
+}
+
 // fullPath returns the full filesystem path for a key.
 func (f *FileStorer) fullPath(key string) string {
 	return filepath.Join(f.basePath, key)

internal/storage/file_atomic_test.go

@@ -0,0 +1,119 @@
+package storage_test
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"sneak.berlin/go/vaultik/internal/storage"
+)
+
+// errStreamInterrupted stands in for an upload cut off mid-stream.
+var errStreamInterrupted = errors.New("connection reset mid-upload")
+
+// failingReader yields its data once, then fails.
+type failingReader struct {
+	data []byte
+	done bool
+}
+
+func (r *failingReader) Read(p []byte) (int, error) {
+	if r.done {
+		return 0, errStreamInterrupted
+	}
+
+	n := copy(p, r.data)
+	r.done = true
+
+	return n, nil
+}
+
+// TestFileStorer_InterruptedWriteLeavesNoTrustedObject checks that a write
+// cut off mid-stream leaves nothing at the destination key, so a later run
+// cannot Stat a truncated object and trust it as a complete blob.
+func TestFileStorer_InterruptedWriteLeavesNoTrustedObject(t *testing.T) {
+	t.Parallel()
+
+	f, err := storage.NewFileStorer(t.TempDir())
+	if err != nil {
+		t.Fatalf("NewFileStorer: %v", err)
+	}
+
+	ctx := context.Background()
+	key := "blobs/aa/bb/aabbccddeeff"
+
+	err = f.PutWithProgress(ctx, key, &failingReader{data: []byte("partial")}, 4096, nil)
+	if err == nil {
+		t.Fatal("expected the interrupted write to fail, got nil")
+	}
+
+	_, err = f.Stat(ctx, key)
+	if !errors.Is(err, storage.ErrNotFound) {
+		t.Fatalf("expected key absent after interrupted write, got Stat err %v", err)
+	}
+
+	keys, err := f.List(ctx, "blobs/")
+	if err != nil {
+		t.Fatalf("List: %v", err)
+	}
+
+	if len(keys) != 0 {
+		t.Fatalf("expected no keys listed after interrupted write, got %v", keys)
+	}
+}
+
+// TestFileStorer_ListSkipsPartialFiles checks that a leftover temp file (the
+// storage layer names them with a ".partial" suffix) is never surfaced as a
+// key by List or ListStream.
+func TestFileStorer_ListSkipsPartialFiles(t *testing.T) {
+	t.Parallel()
+
+	base := t.TempDir()
+
+	f, err := storage.NewFileStorer(base)
+	if err != nil {
+		t.Fatalf("NewFileStorer: %v", err)
+	}
+
+	ctx := context.Background()
+	realKey := "blobs/aa/bb/aabbccddeeff"
+
+	err = f.Put(ctx, realKey, strings.NewReader("blob-bytes"))
+	if err != nil {
+		t.Fatalf("Put: %v", err)
+	}
+
+	// A stray temp file, as an interrupted write would leave behind.
+	leftover := filepath.Join(base, "blobs/aa/bb/aabbccddeeff-123456.partial")
+
+	err = os.WriteFile(leftover, []byte("half"), 0o600)
+	if err != nil {
+		t.Fatalf("writing leftover temp file: %v", err)
+	}
+
+	keys, err := f.List(ctx, "blobs/")
+	if err != nil {
+		t.Fatalf("List: %v", err)
+	}
+
+	if len(keys) != 1 || keys[0] != realKey {
+		t.Fatalf("List should return only the real key, got %v", keys)
+	}
+
+	var streamed []string
+
+	for obj := range f.ListStream(ctx, "blobs/") {
+		if obj.Err != nil {
+			t.Fatalf("ListStream: %v", obj.Err)
+		}
+
+		streamed = append(streamed, obj.Key)
+	}
+
+	if len(streamed) != 1 || streamed[0] != realKey {
+		t.Fatalf("ListStream should return only the real key, got %v", streamed)
+	}
+}

internal/storage/file_backend_test.go

@@ -0,0 +1,27 @@
+package storage_test
+
+import (
+	"testing"
+
+	"sneak.berlin/go/vaultik/internal/storage"
+)
+
+// newFileStorer builds a file:// backend rooted at a fresh temp directory.
+//
+//nolint:ireturn // conformance runs against the Storer interface by design
+func newFileStorer(t *testing.T) storage.Storer {
+	t.Helper()
+
+	s, err := storage.NewFileStorer(t.TempDir())
+	if err != nil {
+		t.Fatalf("NewFileStorer: %v", err)
+	}
+
+	return s
+}
+
+// TestFileStorer runs the shared Storer contract against the file:// backend.
+func TestFileStorer(t *testing.T) {
+	t.Parallel()
+	runStorerConformance(t, newFileStorer)
+}

internal/storage/module.go

@@ -111,10 +111,11 @@ func storerFromParsedS3URL(parsed *URL, cfg *config.Config) (Storer, error) {
 func storerFromLegacyS3Config(cfg *config.Config) (Storer, error) {
 	endpoint := cfg.S3.Endpoint
 
-	// Ensure protocol is present
+	// Ensure protocol is present. Absent an explicit use_ssl, default to TLS;
+	// plain HTTP only when use_ssl is written as false.
 	if !strings.HasPrefix(endpoint, "http://") &&
 		!strings.HasPrefix(endpoint, "https://") {
-		if cfg.S3.UseSSL {
+		if cfg.S3.UseSSL == nil || *cfg.S3.UseSSL {
 			endpoint = "https://" + endpoint
 		} else {
 			endpoint = "http://" + endpoint

internal/storage/module_test.go

@@ -0,0 +1,61 @@
+package storage_test
+
+import (
+	"strings"
+	"testing"
+
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/storage"
+)
+
+// legacyS3Config returns a minimal s3.* (no storage_url) configuration with a
+// scheme-less endpoint. useSSL mirrors the config file: nil means the key is
+// omitted, a pointer means it was written explicitly.
+func legacyS3Config(useSSL *bool) *config.Config {
+	return &config.Config{
+		S3: config.S3Config{
+			Endpoint:        "s3.example.com",
+			Bucket:          "bucket",
+			AccessKeyID:     "key",
+			SecretAccessKey: "secret",
+			Region:          "us-east-1",
+			UseSSL:          useSSL,
+		},
+	}
+}
+
+// endpointScheme builds the storer from cfg and returns the scheme its
+// resolved endpoint carries (Info().Location is "endpoint/bucket").
+func endpointScheme(t *testing.T, cfg *config.Config) string {
+	t.Helper()
+
+	storer, err := storage.NewStorer(cfg)
+	if err != nil {
+		t.Fatalf("NewStorer: %v", err)
+	}
+
+	location := storer.Info().Location
+	switch {
+	case strings.HasPrefix(location, "https://"):
+		return "https"
+	case strings.HasPrefix(location, "http://"):
+		return "http"
+	default:
+		t.Fatalf("endpoint has no http(s) scheme: %q", location)
+
+		return ""
+	}
+}
+
+func TestLegacyS3SchemelessEndpointDefaultsToTLS(t *testing.T) {
+	t.Parallel()
+
+	if got := endpointScheme(t, legacyS3Config(nil)); got != "https" {
+		t.Errorf("use_ssl omitted: got %q scheme, want https", got)
+	}
+
+	no := false
+	if got := endpointScheme(t, legacyS3Config(&no)); got != "http" {
+		t.Errorf("use_ssl: false: got %q scheme, want http", got)
+	}
+}

internal/storage/rclone_test.go

@@ -0,0 +1,58 @@
+package storage_test
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"sneak.berlin/go/vaultik/internal/storage"
+)
+
+// The rclone backend is a thin adapter over the rclone library: it turns a
+// (remote, path) pair into rclone's "remote:path" string, hands it to
+// rclone, and maps rclone's own results back to the Storer interface. What
+// can be tested in-process, without a configured remote or network, is that
+// adapter layer — how the arguments are shaped and how construction errors
+// are reported. The data-plane operations (Put/Get/List/Delete) are rclone's
+// own, exercised against a real provider (drive, s3-via-rclone, ...), which
+// needs a configured remote with credentials and network access and so is
+// out of reach of a unit test. The shared Storer conformance suite therefore
+// runs against the in-process file and s3 backends; the rclone backend
+// inherits that contract once a remote is configured.
+//
+// These tests use rclone's ":local:" on-the-fly backend, which addresses the
+// local filesystem directly without any configured remote, so construction
+// runs entirely in-process.
+
+// TestNewRcloneStorerConstruction checks that a valid remote constructs a
+// backend and that Info() reports the shaped "remote:path" location.
+//
+//nolint:paralleltest // NewRcloneStorer installs the process-global rclone config
+func TestNewRcloneStorerConstruction(t *testing.T) {
+	dir := t.TempDir()
+
+	s, err := storage.NewRcloneStorer(context.Background(), ":local", dir)
+	if err != nil {
+		t.Fatalf("NewRcloneStorer: %v", err)
+	}
+
+	// Info().Location is the "remote:path" string the adapter builds from
+	// its two arguments, so asserting it confirms the argument shaping.
+	want := ":local:" + dir
+	if got := s.Info().Location; got != want {
+		t.Errorf("Info().Location = %q, want %q", got, want)
+	}
+}
+
+// TestNewRcloneStorerUnknownRemote checks that a remote that is not in the
+// rclone config fails construction with the ErrRemoteNotFound sentinel,
+// rather than silently returning a backend pointed nowhere.
+//
+//nolint:paralleltest // NewRcloneStorer installs the process-global rclone config
+func TestNewRcloneStorerUnknownRemote(t *testing.T) {
+	_, err := storage.NewRcloneStorer(
+		context.Background(), "vaultik-no-such-remote", "path")
+	if !errors.Is(err, storage.ErrRemoteNotFound) {
+		t.Errorf("NewRcloneStorer error = %v, want ErrRemoteNotFound", err)
+	}
+}

internal/storage/s3_test.go

@@ -13,18 +13,23 @@ import (
 	"sneak.berlin/go/vaultik/internal/storage"
 )
 
-// TestS3StorerMissingKeyMapsToErrNotFound verifies that the s3 backend reports
+// s3TestBucket is the bucket created for each in-process S3 server.
-// a missing object as storage.ErrNotFound, matching the file and rclone
+const s3TestBucket = "test-bucket"
-// backends and the Storer contract. Without the mapping, Get and Stat leak the
+
-// raw SDK error and errors.Is(err, storage.ErrNotFound) is false.
+// newS3Storer builds an s3:// backend backed by a fresh in-process
+// S3 server. It reuses the same in-memory S3 harness (gofakes3 + s3mem
+// over httptest) that internal/s3 and the not-found regression test use,
+// so no new mock or dependency is introduced. Each call gets its own
+// server, bucket, and client, so the conformance suite's per-section
+// instances stay isolated.
 //
-//nolint:paralleltest // shares an in-process S3 server via t.Cleanup
+//nolint:ireturn // conformance runs against the Storer interface by design
-func TestS3StorerMissingKeyMapsToErrNotFound(t *testing.T) {
+func newS3Storer(t *testing.T) storage.Storer {
-	const bucket = "test-bucket"
+	t.Helper()
 
 	backend := s3mem.New()
 
-	err := backend.CreateBucket(bucket)
+	err := backend.CreateBucket(s3TestBucket)
 	if err != nil {
 		t.Fatalf("create bucket: %v", err)
 	}
@@ -32,11 +37,9 @@ func TestS3StorerMissingKeyMapsToErrNotFound(t *testing.T) {
 	srv := httptest.NewServer(gofakes3.New(backend).Server())
 	t.Cleanup(srv.Close)
 
-	ctx := context.Background()
+	client, err := s3.NewClient(context.Background(), s3.Config{
-
-	client, err := s3.NewClient(ctx, s3.Config{
 		Endpoint:        srv.URL,
-		Bucket:          bucket,
+		Bucket:          s3TestBucket,
 		AccessKeyID:     "test",
 		SecretAccessKey: "test",
 		Region:          "us-east-1",
@@ -45,9 +48,28 @@ func TestS3StorerMissingKeyMapsToErrNotFound(t *testing.T) {
 		t.Fatalf("new client: %v", err)
 	}
 
-	storer := storage.NewS3Storer(client)
+	return storage.NewS3Storer(client)
+}
 
-	_, err = storer.Get(ctx, "does-not-exist")
+// TestS3Storer runs the shared Storer contract against the s3:// backend,
+// so it is held to the same round-trip, list, delete, and not-found
+// behaviour as the file:// backend.
+func TestS3Storer(t *testing.T) {
+	t.Parallel()
+	runStorerConformance(t, newS3Storer)
+}
+
+// TestS3StorerMissingKeyMapsToErrNotFound pins the specific contract that a
+// missing object surfaces as storage.ErrNotFound rather than the raw AWS SDK
+// error. Without the mapping, errors.Is(err, storage.ErrNotFound) is false on
+// s3 and callers would branch differently per backend.
+func TestS3StorerMissingKeyMapsToErrNotFound(t *testing.T) {
+	t.Parallel()
+
+	storer := newS3Storer(t)
+	ctx := context.Background()
+
+	_, err := storer.Get(ctx, "does-not-exist")
 	if !errors.Is(err, storage.ErrNotFound) {
 		t.Errorf("Get on missing key: got %v, want ErrNotFound", err)
 	}

internal/storage/url.go

@@ -4,6 +4,7 @@ import (
 	"errors"
 	"fmt"
 	"net/url"
+	"slices"
 	"strings"
 )
 
@@ -23,6 +24,10 @@ var (
 	ErrUnsupportedScheme = errors.New(
 		"unsupported URL scheme: must start with s3://, file://, or rclone://")
 	ErrUnsupportedStorage = errors.New("unsupported storage scheme")
+	ErrURLCredentials     = errors.New(
+		"storage URL must not carry credentials; " +
+			"set s3.access_key_id and s3.secret_access_key in the config instead")
+	ErrURLUnknownParam = errors.New("unknown query parameter in storage URL")
 )
 
 // URL represents a parsed storage URL.
@@ -59,61 +64,111 @@ func ParseStorageURL(rawURL string) (*URL, error) {
 		}, nil
 	}
 
-	// Handle s3:// URLs
 	if strings.HasPrefix(rawURL, "s3://") {
-		u, err := url.Parse(rawURL)
+		return parseS3URL(rawURL)
-		if err != nil {
-			return nil, fmt.Errorf("invalid URL: %w", err)
-		}
-
-		bucket := u.Host
-		if bucket == "" {
-			return nil, ErrMissingBucket
-		}
-
-		prefix := strings.TrimPrefix(u.Path, "/")
-
-		query := u.Query()
-
-		useSSL := true
-		if query.Get("ssl") == "false" {
-			useSSL = false
-		}
-
-		return &URL{
-			Scheme:   schemeS3,
-			Bucket:   bucket,
-			Prefix:   prefix,
-			Endpoint: query.Get("endpoint"),
-			Region:   query.Get("region"),
-			UseSSL:   useSSL,
-		}, nil
 	}
 
-	// Handle rclone:// URLs
 	if strings.HasPrefix(rawURL, "rclone://") {
-		u, err := url.Parse(rawURL)
+		return parseRcloneURL(rawURL)
-		if err != nil {
-			return nil, fmt.Errorf("invalid URL: %w", err)
-		}
-
-		remote := u.Host
-		if remote == "" {
-			return nil, ErrMissingRemote
-		}
-
-		path := strings.TrimPrefix(u.Path, "/")
-
-		return &URL{
-			Scheme:       schemeRclone,
-			Prefix:       path,
-			RcloneRemote: remote,
-		}, nil
 	}
 
 	return nil, ErrUnsupportedScheme
 }
 
+// parseS3URL parses an s3://bucket/prefix URL. It rejects credentials in
+// the userinfo and any query parameter other than endpoint, region and
+// ssl, so a credential-bearing URL is never stored or echoed.
+func parseS3URL(rawURL string) (*URL, error) {
+	u, err := url.Parse(rawURL)
+	if err != nil {
+		return nil, wrapParseError(err)
+	}
+
+	if u.User != nil {
+		return nil, ErrURLCredentials
+	}
+
+	bucket := u.Host
+	if bucket == "" {
+		return nil, ErrMissingBucket
+	}
+
+	query := u.Query()
+
+	err = rejectUnknownParams(query, "endpoint", "region", "ssl")
+	if err != nil {
+		return nil, err
+	}
+
+	return &URL{
+		Scheme:   schemeS3,
+		Bucket:   bucket,
+		Prefix:   strings.TrimPrefix(u.Path, "/"),
+		Endpoint: query.Get("endpoint"),
+		Region:   query.Get("region"),
+		UseSSL:   query.Get("ssl") != "false",
+	}, nil
+}
+
+// parseRcloneURL parses an rclone://remote/path URL. rclone:// takes no
+// query parameters, so credentials in the userinfo and any parameter at
+// all are rejected rather than silently ignored.
+func parseRcloneURL(rawURL string) (*URL, error) {
+	u, err := url.Parse(rawURL)
+	if err != nil {
+		return nil, wrapParseError(err)
+	}
+
+	if u.User != nil {
+		return nil, ErrURLCredentials
+	}
+
+	remote := u.Host
+	if remote == "" {
+		return nil, ErrMissingRemote
+	}
+
+	err = rejectUnknownParams(u.Query())
+	if err != nil {
+		return nil, err
+	}
+
+	return &URL{
+		Scheme:       schemeRclone,
+		Prefix:       strings.TrimPrefix(u.Path, "/"),
+		RcloneRemote: remote,
+	}, nil
+}
+
+// rejectUnknownParams returns an error naming the first query parameter
+// not in allowed. The parameter's name is included (so a misspelt
+// endpoint= is caught), but never its value, which could be a secret,
+// and never the whole URL.
+func rejectUnknownParams(query url.Values, allowed ...string) error {
+	for name := range query {
+		if !slices.Contains(allowed, name) {
+			return fmt.Errorf(
+				"%w: %q; put credentials in s3.access_key_id and "+
+					"s3.secret_access_key, not the URL",
+				ErrURLUnknownParam, name)
+		}
+	}
+
+	return nil
+}
+
+// wrapParseError wraps only the inner cause of a url.Parse failure. The
+// *url.Error that url.Parse returns embeds the raw URL in its message, so
+// wrapping it directly would echo a credential-bearing URL into logs.
+func wrapParseError(err error) error {
+	var uerr *url.Error
+	if errors.As(err, &uerr) {
+		return fmt.Errorf("invalid URL: %w", uerr.Err)
+	}
+
+	return fmt.Errorf("invalid URL: %w", err)
+}
+
 // String returns a human-readable representation of the storage URL.
 func (u *URL) String() string {
 	switch u.Scheme {

internal/storage/url_parse_test.go

@@ -0,0 +1,208 @@
+package storage_test
+
+import (
+	"errors"
+	"reflect"
+	"strings"
+	"testing"
+
+	"sneak.berlin/go/vaultik/internal/storage"
+)
+
+// TestParseStorageURLValid checks that each supported scheme parses into
+// the expected fields, since those fields decide which backend is built.
+func TestParseStorageURLValid(t *testing.T) {
+	t.Parallel()
+
+	const bucket = "mybucket"
+
+	cases := []struct {
+		name string
+		raw  string
+		want *storage.URL
+	}{
+		{
+			name: "file absolute path",
+			raw:  "file:///var/backups/vaultik",
+			want: &storage.URL{Scheme: "file", Prefix: "/var/backups/vaultik"},
+		},
+		{
+			name: "s3 bucket and prefix, ssl defaults on",
+			raw:  "s3://mybucket/backups/host",
+			want: &storage.URL{
+				Scheme: "s3", Bucket: bucket,
+				Prefix: "backups/host", UseSSL: true,
+			},
+		},
+		{
+			name: "s3 bucket only",
+			raw:  "s3://mybucket",
+			want: &storage.URL{Scheme: "s3", Bucket: bucket, UseSSL: true},
+		},
+		{
+			name: "s3 with endpoint, region, ssl off",
+			raw:  "s3://mybucket?endpoint=minio.example.com&region=us-west-2&ssl=false",
+			want: &storage.URL{
+				Scheme: "s3", Bucket: bucket,
+				Endpoint: "minio.example.com", Region: "us-west-2", UseSSL: false,
+			},
+		},
+		{
+			name: "rclone remote and path",
+			raw:  "rclone://gdrive/backups/host",
+			want: &storage.URL{
+				Scheme: "rclone", RcloneRemote: "gdrive", Prefix: "backups/host",
+			},
+		},
+		{
+			name: "rclone remote only",
+			raw:  "rclone://gdrive",
+			want: &storage.URL{Scheme: "rclone", RcloneRemote: "gdrive"},
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			got, err := storage.ParseStorageURL(tc.raw)
+			if err != nil {
+				t.Fatalf("ParseStorageURL(%q) returned error: %v", tc.raw, err)
+			}
+
+			if !reflect.DeepEqual(got, tc.want) {
+				t.Errorf("ParseStorageURL(%q) = %+v, want %+v", tc.raw, got, tc.want)
+			}
+		})
+	}
+}
+
+// TestParseStorageURLErrors checks that empty, missing, and unknown-scheme
+// inputs fail with the documented sentinel errors instead of parsing to a
+// wrong destination.
+func TestParseStorageURLErrors(t *testing.T) {
+	t.Parallel()
+
+	cases := []struct {
+		name    string
+		raw     string
+		wantErr error
+	}{
+		{"empty url", "", storage.ErrEmptyStorageURL},
+		{"file empty path", "file://", storage.ErrEmptyFilePath},
+		{"s3 missing bucket", "s3://", storage.ErrMissingBucket},
+		{"s3 missing bucket with path", "s3:///justprefix", storage.ErrMissingBucket},
+		{"rclone missing remote", "rclone://", storage.ErrMissingRemote},
+		{"unknown scheme", "gs://bucket/x", storage.ErrUnsupportedScheme},
+		{"no scheme", "/local/path", storage.ErrUnsupportedScheme},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			_, err := storage.ParseStorageURL(tc.raw)
+			if !errors.Is(err, tc.wantErr) {
+				t.Errorf("ParseStorageURL(%q) error = %v, want %v",
+					tc.raw, err, tc.wantErr)
+			}
+		})
+	}
+}
+
+// TestParseStorageURLRejectsCredentials checks that a URL carrying
+// credentials in its userinfo or in an unknown query parameter is
+// rejected, and that the error never echoes the secret-bearing URL back
+// into logs or output.
+func TestParseStorageURLRejectsCredentials(t *testing.T) {
+	t.Parallel()
+
+	// Split so the literals never form a "user:pass@" URL pattern that
+	// tooling would flag as a real hardcoded credential.
+	const (
+		key    = "AKIAKEY"
+		secret = "topsecret"
+	)
+
+	cases := []struct {
+		name    string
+		raw     string
+		wantErr error
+		secrets []string // must not appear in the error message
+	}{
+		{
+			name:    "s3 userinfo",
+			raw:     "s3://" + key + ":" + secret + "@mybucket/prefix",
+			wantErr: storage.ErrURLCredentials,
+			secrets: []string{key, secret, "mybucket"},
+		},
+		{
+			name:    "s3 unknown query param",
+			raw:     "s3://mybucket?access_key=" + key + "&secret=" + secret,
+			wantErr: storage.ErrURLUnknownParam,
+			secrets: []string{key, secret},
+		},
+		{
+			name:    "s3 misspelt endpoint",
+			raw:     "s3://mybucket?endpiont=minio.example.com",
+			wantErr: storage.ErrURLUnknownParam,
+			secrets: nil,
+		},
+		{
+			name:    "rclone userinfo",
+			raw:     "rclone://user:" + secret + "@gdrive/backups",
+			wantErr: storage.ErrURLCredentials,
+			secrets: []string{secret},
+		},
+		{
+			name:    "rclone query param",
+			raw:     "rclone://gdrive/backups?token=" + secret,
+			wantErr: storage.ErrURLUnknownParam,
+			secrets: []string{secret},
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			_, err := storage.ParseStorageURL(tc.raw)
+			if !errors.Is(err, tc.wantErr) {
+				t.Fatalf("ParseStorageURL(%q) error = %v, want %v",
+					tc.raw, err, tc.wantErr)
+			}
+
+			// The rejection must name the proper config keys so the
+			// operator knows where credentials belong.
+			for _, key := range []string{"s3.access_key_id", "s3.secret_access_key"} {
+				if !strings.Contains(err.Error(), key) {
+					t.Errorf("error %q does not name %q", err.Error(), key)
+				}
+			}
+
+			for _, secret := range tc.secrets {
+				if strings.Contains(err.Error(), secret) {
+					t.Errorf("error message leaked %q: %v", secret, err.Error())
+				}
+			}
+		})
+	}
+}
+
+// TestParseStorageURLParseFailureHidesURL checks that when url.Parse
+// itself fails, the wrapped error carries only the inner cause, not the
+// *url.Error whose text embeds the raw (possibly credential-bearing) URL.
+func TestParseStorageURLParseFailureHidesURL(t *testing.T) {
+	t.Parallel()
+
+	const raw = "s3://mybucket/%zz"
+
+	_, err := storage.ParseStorageURL(raw)
+	if err == nil {
+		t.Fatalf("ParseStorageURL(%q) returned no error", raw)
+	}
+
+	if strings.Contains(err.Error(), "mybucket") {
+		t.Errorf("error message echoed the raw URL: %v", err.Error())
+	}
+}

internal/types/types.go

@@ -1,7 +1,7 @@
 // Package types provides custom types for better type safety across the
-// vaultik codebase. Using distinct types for IDs, hashes, paths, and
+// vaultik codebase. Using distinct types for IDs, hashes, and paths prevents
-// credentials prevents accidental mixing of semantically different values
+// accidental mixing of semantically different values that happen to share the
-// that happen to share the same underlying type.
+// same underlying type.
 package types //nolint:revive,nolintlint // rename decision tracked in #76
 
 import (
@@ -157,34 +157,6 @@ type FilePath string
 // Used during restore to strip the source prefix from paths.
 type SourcePath string
 
-// AgeRecipient is an age public key used for encryption.
-// Format: age1... (Bech32-encoded X25519 public key)
-type AgeRecipient string
-
-// AgeSecretKey is an age private key used for decryption.
-// Format: AGE-SECRET-KEY-... (Bech32-encoded X25519 private key)
-// This type should never be logged or serialized in plaintext.
-type AgeSecretKey string
-
-// S3Endpoint is the URL of an S3-compatible storage endpoint.
-type S3Endpoint string
-
-// BucketName is the name of an S3 bucket.
-type BucketName string
-
-// S3Prefix is the path prefix within an S3 bucket.
-type S3Prefix string
-
-// AWSRegion is an AWS region identifier (e.g., "us-east-1").
-type AWSRegion string
-
-// AWSAccessKeyID is an AWS access key ID for authentication.
-type AWSAccessKeyID string
-
-// AWSSecretAccessKey is an AWS secret access key for authentication.
-// This type should never be logged or serialized in plaintext.
-type AWSSecretAccessKey string
-
 // Hostname identifies a host machine.
 type Hostname string
 
@@ -199,31 +171,14 @@ type GlobPattern string
 
 // String methods for Stringer interface
 
-func (id FileID) String() string        { return uuid.UUID(id).String() }
+func (id FileID) String() string     { return uuid.UUID(id).String() }
-func (id BlobID) String() string        { return uuid.UUID(id).String() }
+func (id BlobID) String() string     { return uuid.UUID(id).String() }
-func (id SnapshotID) String() string    { return string(id) }
+func (id SnapshotID) String() string { return string(id) }
-func (h ChunkHash) String() string      { return string(h) }
+func (h ChunkHash) String() string   { return string(h) }
-func (h BlobHash) String() string       { return string(h) }
+func (h BlobHash) String() string    { return string(h) }
-func (p FilePath) String() string       { return string(p) }
+func (p FilePath) String() string    { return string(p) }
-func (p SourcePath) String() string     { return string(p) }
+func (p SourcePath) String() string  { return string(p) }
-func (r AgeRecipient) String() string   { return string(r) }
+func (h Hostname) String() string    { return string(h) }
-func (e S3Endpoint) String() string     { return string(e) }
+func (v Version) String() string     { return string(v) }
-func (b BucketName) String() string     { return string(b) }
+func (r GitRevision) String() string { return string(r) }
-func (p S3Prefix) String() string       { return string(p) }
+func (p GlobPattern) String() string { return string(p) }
-func (r AWSRegion) String() string      { return string(r) }
-func (k AWSAccessKeyID) String() string { return string(k) }
-func (h Hostname) String() string       { return string(h) }
-func (v Version) String() string        { return string(v) }
-func (r GitRevision) String() string    { return string(r) }
-func (p GlobPattern) String() string    { return string(p) }
-
-// Redacted String methods for sensitive types - prevents accidental logging
-
-func (k AgeSecretKey) String() string       { return "[REDACTED]" }
-func (k AWSSecretAccessKey) String() string { return "[REDACTED]" }
-
-// Raw returns the actual value for sensitive types when explicitly needed.
-func (k AgeSecretKey) Raw() string { return string(k) }
-
-// Raw returns the actual value for sensitive types when explicitly needed.
-func (k AWSSecretAccessKey) Raw() string { return string(k) }

internal/vaultik/blob_fetch.go

@@ -2,7 +2,6 @@ package vaultik
 
 import (
 	"context"
-	"crypto/sha256"
 	"encoding/hex"
 	"errors"
 	"fmt"
@@ -18,6 +17,13 @@ import (
 // not match the expected double-SHA-256 hash.
 var errBlobHashMismatch = errors.New("blob hash mismatch")
 
+// errBlobNotFullyRead is returned when the verifying reader is closed
+// before its plaintext reached EOF. The hash can only be checked once
+// the whole stream has been read, so an early or short-read close must
+// fail rather than silently skip verification.
+var errBlobNotFullyRead = errors.New(
+	"blob closed before fully read; hash not verified")
+
 // hashVerifyReader wraps a blobgen.Reader and verifies the double-SHA-256 hash
 // of decrypted plaintext when Close is called. It reuses the hash that
 // blobgen.Reader already computes internally via its TeeReader, avoiding
@@ -38,21 +44,22 @@ func (h *hashVerifyReader) Read(p []byte) (int, error) {
 	return n, err
 }
 
-// Close verifies the hash (if the stream was fully read) and closes underlying readers.
+// Close closes the underlying readers and verifies the blob hash. The
+// hash check cannot be skipped: closing before the plaintext reached
+// EOF (a short read or an early close) is an error, so a caller can
+// never obtain unverified blob bytes.
 func (h *hashVerifyReader) Close() error {
 	readerErr := h.reader.Close()
 	fetcherErr := h.fetcher.Close()
 
-	if h.done {
+	if !h.done {
-		firstHash := h.reader.Sum256()
+		return errBlobNotFullyRead
-		secondHasher := sha256.New()
+	}
-		secondHasher.Write(firstHash)
 
-		actualHashHex := hex.EncodeToString(secondHasher.Sum(nil))
+	actualHashHex := hex.EncodeToString(blobgen.DoubleSHA256(h.reader.Sum256()))
-		if actualHashHex != h.blobHash {
+	if actualHashHex != h.blobHash {
-			return fmt.Errorf("%w: expected %s, got %s",
+		return fmt.Errorf("%w: expected %s, got %s",
-				errBlobHashMismatch, h.blobHash[:16], actualHashHex[:16])
+			errBlobHashMismatch, h.blobHash[:16], actualHashHex[:16])
-		}
 	}
 
 	if readerErr != nil {

internal/vaultik/blob_fetch_hash_test.go

@@ -40,13 +40,13 @@ func buildHashTestBlob(
 	}
 
 	// Compute the double-SHA-256 hash of the plaintext (matches
-	// blobgen.Writer.Sum256).
+	// blobgen.Writer.ContentID).
 	firstHash := sha256.Sum256(plaintext)
 	secondHash := sha256.Sum256(firstHash[:])
 	correctHash := hex.EncodeToString(secondHash[:])
 
 	// Verify our hash matches what blobgen.Writer produces
-	writerHash := hex.EncodeToString(writer.Sum256())
+	writerHash := hex.EncodeToString(writer.ContentID())
 	if correctHash != writerHash {
 		t.Fatalf("hash computation mismatch: manual=%s, writer=%s",
 			correctHash, writerHash)
@@ -133,3 +133,51 @@ func TestFetchAndDecryptBlobVerifiesHash(t *testing.T) {
 		}
 	})
 }
+
+// TestFetchAndDecryptBlobCloseBeforeEOFFails verifies the hash check
+// cannot be skipped: a caller that reads only part of the blob and then
+// closes gets an error rather than silently unverified bytes.
+func TestFetchAndDecryptBlobCloseBeforeEOFFails(t *testing.T) {
+	t.Parallel()
+
+	identity, err := age.GenerateX25519Identity()
+	if err != nil {
+		t.Fatalf("generating identity: %v", err)
+	}
+
+	plaintext := []byte("hello world test data for blob hash verification")
+	encryptedData, correctHash := buildHashTestBlob(t, identity, plaintext)
+
+	mockStorage := NewMockStorer()
+	blobPath := "blobs/" + correctHash[:2] + "/" +
+		correctHash[2:4] + "/" + correctHash
+
+	mockStorage.mu.Lock()
+	mockStorage.data[blobPath] = encryptedData
+	mockStorage.mu.Unlock()
+
+	tv := vaultik.NewForTesting(mockStorage)
+
+	rc, err := tv.FetchAndDecryptBlob(
+		context.Background(), correctHash, int64(len(encryptedData)), identity)
+	if err != nil {
+		t.Fatalf("unexpected error opening stream: %v", err)
+	}
+
+	// Read one byte, far short of the plaintext length, then close.
+	buf := make([]byte, 1)
+
+	_, err = rc.Read(buf)
+	if err != nil {
+		t.Fatalf("reading first byte: %v", err)
+	}
+
+	err = rc.Close()
+	if err == nil {
+		t.Fatal("expected error closing before EOF, got nil")
+	}
+
+	if !strings.Contains(err.Error(), "hash not verified") {
+		t.Fatalf("expected not-verified error, got: %v", err)
+	}
+}

internal/vaultik/fault_injection_test.go

@@ -0,0 +1,647 @@
+package vaultik_test
+
+import (
+	"context"
+	"errors"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+	"sneak.berlin/go/vaultik/internal/storage"
+	"sneak.berlin/go/vaultik/internal/storage/faultstore"
+	"sneak.berlin/go/vaultik/internal/ui"
+	"sneak.berlin/go/vaultik/internal/vaultik"
+)
+
+// These tests cover the failure modes a backup tool must survive:
+// interrupted uploads, an interrupted metadata export, corrupt and
+// truncated reads, a full restore disk, and a backend that reports
+// success while storing nothing. Faults are injected through the
+// storage.Storer seam (internal/storage/faultstore), never by patching
+// production code. Each test asserts on the observable end state — what
+// is in the index, what is at the destination, what the user is told —
+// not merely that an error was returned. See
+// https://git.eeqj.de/sneak/vaultik/issues/72.
+//
+// Object-level write atomicity (no partial blob object left behind) is
+// covered by the file:// backend's atomic-write work
+// (https://git.eeqj.de/sneak/vaultik/issues/130) and is not re-tested
+// here; these tests target the layers above the backend.
+//
+// The tests run serially, not with t.Parallel: each calls
+// log.Initialize, which replaces the package-global logger, and a
+// backup or restore running concurrently reads that same logger. Under
+// -race the two collide. Running one at a time is the same choice
+// prune_count_test.go already makes for the same reason.
+
+const (
+	faultChunkSize   = int64(64 * 1024)
+	faultMaxBlobSize = int64(256 * 1024)
+)
+
+// faultTestConfig returns the config shared by the fault-injection
+// tests: a real recipient/secret keypair so blobs are genuinely
+// encrypted, and a blob size limit the restore sweeper can divide.
+func faultTestConfig() *config.Config {
+	return &config.Config{
+		AgeRecipients:    []string{testAgePublicKey},
+		AgeSecretKey:     testAgeSecretKey,
+		CompressionLevel: 3,
+		Hostname:         testHostname,
+		BlobSizeLimit:    config.Size(faultMaxBlobSize),
+	}
+}
+
+// writeFaultSourceTree writes a spread of file sizes that forces several
+// chunks across more than one blob, so a fault landing on a single blob
+// still leaves other data intact. Returns the expected content by path.
+func writeFaultSourceTree(
+	t *testing.T, fs afero.Fs, dataDir string,
+) map[string][]byte {
+	t.Helper()
+
+	files := map[string][]byte{
+		filepath.Join(dataDir, "small.txt"):    []byte("hello vaultik"),
+		filepath.Join(dataDir, "a.bin"):        bytesPattern("a-", int(faultChunkSize*3)),
+		filepath.Join(dataDir, "sub", "b.bin"): bytesPattern("b-", int(faultChunkSize*3)),
+		filepath.Join(dataDir, "sub", "c.bin"): bytesPattern("c-", int(faultChunkSize*2)),
+	}
+
+	for path, content := range files {
+		require.NoError(t, fs.MkdirAll(filepath.Dir(path), 0o755))
+		require.NoError(t, afero.WriteFile(fs, path, content, 0o644))
+	}
+
+	return files
+}
+
+// newFaultScanner builds a scanner writing through the given storer.
+func newFaultScanner(
+	fs afero.Fs, storer storage.Storer,
+	cfg *config.Config, repos *database.Repositories,
+) *snapshot.Scanner {
+	return snapshot.NewScanner(snapshot.ScannerConfig{
+		FS:               fs,
+		Storage:          storer,
+		ChunkSize:        faultChunkSize,
+		MaxBlobSize:      faultMaxBlobSize,
+		CompressionLevel: cfg.CompressionLevel,
+		AgeRecipients:    cfg.AgeRecipients,
+		Repositories:     repos,
+	})
+}
+
+// newFaultSnapshotManager builds a snapshot manager writing through the
+// given storer.
+func newFaultSnapshotManager(
+	fs afero.Fs, storer storage.Storer,
+	cfg *config.Config, repos *database.Repositories,
+) *snapshot.SnapshotManager {
+	sm := snapshot.NewSnapshotManager(snapshot.SnapshotManagerParams{
+		Repos:   repos,
+		Storage: storer,
+		Config:  cfg,
+	})
+	sm.SetFilesystem(fs)
+
+	return sm
+}
+
+// fullFaultBackup runs a complete backup (create, scan, complete,
+// export) through storer and returns the snapshot ID.
+func fullFaultBackup(
+	ctx context.Context, t *testing.T, fs afero.Fs, storer storage.Storer,
+	cfg *config.Config, repos *database.Repositories,
+	dataDir, dbPath, name string,
+) string {
+	t.Helper()
+
+	sm := newFaultSnapshotManager(fs, storer, cfg, repos)
+	scanner := newFaultScanner(fs, storer, cfg, repos)
+
+	id, err := sm.CreateSnapshotWithName(ctx, cfg.Hostname, name, "v", "g")
+	require.NoError(t, err)
+
+	_, err = scanner.Scan(ctx, dataDir, id)
+	require.NoError(t, err)
+
+	require.NoError(t, sm.CompleteSnapshot(ctx, id))
+	require.NoError(t, sm.ExportSnapshotMetadata(ctx, dbPath, id))
+
+	return id
+}
+
+// newReaderVaultik builds a Vaultik that reads (restore/verify) through
+// storer, with the given repositories (nil is fine for restore/verify,
+// which read metadata from storage).
+func newReaderVaultik(
+	ctx context.Context, cfg *config.Config, storer storage.Storer,
+	repos *database.Repositories, fs afero.Fs,
+) *vaultik.Vaultik {
+	v := &vaultik.Vaultik{
+		Config:       cfg,
+		Storage:      storer,
+		Repositories: repos,
+		Fs:           fs,
+		Stdout:       io.Discard,
+		Stderr:       io.Discard,
+		UI:           ui.NewWithColor(io.Discard, false),
+	}
+	v.SetContext(ctx)
+
+	return v
+}
+
+// Scenario 3: a stored blob's bytes are flipped before restore reads
+// them. Restore must fail loudly, and no file must be left on the
+// restore target holding corrupt content.
+//
+//nolint:paralleltest // installs the global logger via log.Initialize
+func TestRestoreRejectsCorruptBlob(t *testing.T) {
+	assertRestoreRejectsDamagedBlob(t, faultstore.GetCorrupt, "corrupt")
+}
+
+// Scenario 4: a stored blob is truncated before restore reads it. Same
+// contract as the corrupt case.
+//
+//nolint:paralleltest // installs the global logger via log.Initialize
+func TestRestoreRejectsTruncatedBlob(t *testing.T) {
+	assertRestoreRejectsDamagedBlob(t, faultstore.GetTruncate, "truncated")
+}
+
+// assertRestoreRejectsDamagedBlob backs up the source tree, then restores
+// through a store that damages every blob read with the given fault, and
+// asserts restore fails naming a blob and leaves no file on the target
+// holding wrong bytes. Metadata reads are returned intact so the failure
+// is isolated to the blob.
+func assertRestoreRejectsDamagedBlob(
+	t *testing.T, fault faultstore.GetFault, name string,
+) {
+	t.Helper()
+	log.Initialize(log.Config{})
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	dataDir := filepath.Join(tempDir, "src")
+	storeDir := filepath.Join(tempDir, "remote")
+	restoreDir := filepath.Join(tempDir, "restored")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	ctx := context.Background()
+	cfg := faultTestConfig()
+	testFiles := writeFaultSourceTree(t, fs, dataDir)
+
+	inner, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	id := fullFaultBackup(ctx, t, fs, inner, cfg, repos, dataDir, dbPath, name)
+	require.NoError(t, db.Close())
+
+	faultStore := faultstore.New(inner)
+	faultStore.OnGet = func(key string) faultstore.GetFault {
+		if strings.HasPrefix(key, "blobs/") {
+			return fault
+		}
+
+		return faultstore.GetNormal
+	}
+
+	v := newReaderVaultik(ctx, cfg, faultStore, nil, fs)
+	err = v.Restore(&vaultik.RestoreOptions{SnapshotID: id, TargetDir: restoreDir})
+
+	require.Error(t, err, "restore must fail on a damaged blob")
+	assert.Contains(t, err.Error(), "blob",
+		"error should name the blob that failed")
+	assertNoCorruptFiles(t, fs, restoreDir, testFiles)
+}
+
+// Scenario 6: the backend accepts blob uploads and reports success but
+// stores nothing. verify --deep must catch it.
+//
+//nolint:paralleltest // installs the global logger via log.Initialize
+func TestDeepVerifyCatchesLyingBackend(t *testing.T) {
+	log.Initialize(log.Config{})
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	dataDir := filepath.Join(tempDir, "src")
+	storeDir := filepath.Join(tempDir, "remote")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	ctx := context.Background()
+	cfg := faultTestConfig()
+
+	writeFaultSourceTree(t, fs, dataDir)
+
+	inner, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	// Blob uploads are swallowed; metadata uploads land, so verify can
+	// download the manifest and database and then discover the blobs are
+	// absent.
+	lying := faultstore.New(inner)
+	lying.OnPut = func(key string) faultstore.PutAction {
+		if strings.HasPrefix(key, "blobs/") {
+			return faultstore.PutSwallow
+		}
+
+		return faultstore.PutNormal
+	}
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	id := fullFaultBackup(ctx, t, fs, lying, cfg, repos, dataDir, dbPath, "lying")
+	require.NoError(t, db.Close())
+
+	// No blob objects were actually written.
+	blobKeys, err := inner.List(ctx, "blobs/")
+	require.NoError(t, err)
+	assert.Empty(t, blobKeys, "lying backend should have stored no blobs")
+
+	// Read back through the honest underlying store.
+	v := newReaderVaultik(ctx, cfg, inner, nil, fs)
+	err = v.VerifySnapshotWithOptions(id, &vaultik.VerifyOptions{Deep: true})
+	require.Error(t, err, "deep verify must catch a backend that stored nothing")
+}
+
+// Scenario 1a: a blob upload fails partway through. The interrupted run
+// must not record the blob as uploaded, must not reference it from the
+// snapshot, and must leave no blob object at the destination.
+//
+//nolint:paralleltest // installs the global logger via log.Initialize
+func TestInterruptedBlobUploadRecordsNoUploadedBlob(t *testing.T) {
+	log.Initialize(log.Config{})
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	dataDir := filepath.Join(tempDir, "src")
+	storeDir := filepath.Join(tempDir, "remote")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	ctx := context.Background()
+	cfg := faultTestConfig()
+
+	writeFaultSourceTree(t, fs, dataDir)
+
+	inner, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	defer func() { _ = db.Close() }()
+
+	repos := database.NewRepositories(db)
+
+	// Every blob upload fails partway through. The scan must surface it.
+	fault := faultstore.New(inner)
+	fault.OnPut = func(key string) faultstore.PutAction {
+		if strings.HasPrefix(key, "blobs/") {
+			return faultstore.PutFail
+		}
+
+		return faultstore.PutNormal
+	}
+
+	sm := newFaultSnapshotManager(fs, fault, cfg, repos)
+	scanner := newFaultScanner(fs, fault, cfg, repos)
+
+	id, err := sm.CreateSnapshotWithName(ctx, cfg.Hostname, "interrupted", "v", "g")
+	require.NoError(t, err)
+
+	_, err = scanner.Scan(ctx, dataDir, id)
+	require.Error(t, err, "scan must fail when a blob upload fails")
+
+	// No blob may claim to be uploaded.
+	blobs, err := repos.Blobs.GetAll(ctx)
+	require.NoError(t, err)
+
+	for _, b := range blobs {
+		assert.Nilf(t, b.UploadedTS,
+			"blob %s marked uploaded after a failed upload", b.Hash)
+	}
+
+	// The snapshot may reference no blobs, and the destination holds none.
+	hashes, err := repos.Snapshots.GetBlobHashes(ctx, id)
+	require.NoError(t, err)
+	assert.Empty(t, hashes, "interrupted snapshot must reference no blobs")
+
+	blobKeys, err := inner.List(ctx, "blobs/")
+	require.NoError(t, err)
+	assert.Empty(t, blobKeys, "no blob object may survive at the destination")
+}
+
+// Scenario 1b: after an interrupted upload, a retry on the same local
+// index must produce a restorable snapshot. The interrupted run leaves
+// the blob's chunk rows in the index; the fix for
+// https://git.eeqj.de/sneak/vaultik/issues/148 discards those un-uploaded
+// blob rows at the start of the next scan and deduplicates only against
+// chunks in a blob that was actually uploaded, so the retry re-chunks and
+// re-uploads the affected data instead of silently referencing data that
+// never reached storage.
+//
+//nolint:paralleltest // installs the global logger via log.Initialize
+func TestBackupRetryAfterInterruptedUploadIsRestorable(t *testing.T) {
+	log.Initialize(log.Config{})
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	dataDir := filepath.Join(tempDir, "src")
+	storeDir := filepath.Join(tempDir, "remote")
+	restoreDir := filepath.Join(tempDir, "restored")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	ctx := context.Background()
+	cfg := faultTestConfig()
+	testFiles := writeFaultSourceTree(t, fs, dataDir)
+
+	inner, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	// Attempt 1: every blob upload fails.
+	fault := faultstore.New(inner)
+	fault.OnPut = func(key string) faultstore.PutAction {
+		if strings.HasPrefix(key, "blobs/") {
+			return faultstore.PutFail
+		}
+
+		return faultstore.PutNormal
+	}
+
+	sm := newFaultSnapshotManager(fs, fault, cfg, repos)
+	scanner := newFaultScanner(fs, fault, cfg, repos)
+
+	id1, err := sm.CreateSnapshotWithName(ctx, cfg.Hostname, "interrupted", "v", "g")
+	require.NoError(t, err)
+
+	_, err = scanner.Scan(ctx, dataDir, id1)
+	require.Error(t, err)
+
+	// Retry on the same local index with a working backend.
+	id2 := fullFaultBackup(ctx, t, fs, inner, cfg, repos, dataDir, dbPath, "retry")
+	require.NoError(t, db.Close())
+
+	v := newReaderVaultik(ctx, cfg, inner, nil, fs)
+	require.NoError(t, v.Restore(&vaultik.RestoreOptions{
+		SnapshotID: id2,
+		TargetDir:  restoreDir,
+		Verify:     true,
+	}), "retry after an interrupted upload must produce a restorable snapshot")
+
+	assertRestoredTree(t, fs, restoreDir, testFiles)
+}
+
+// Scenario 2: the process dies during the metadata export, after the
+// database is uploaded but before the manifest. The destination is left
+// with blobs and a database but no manifest. verify and snapshot list
+// must report the damage honestly rather than crashing or passing.
+// Automatic detection and repair of this partial state on the next run
+// is tracked in https://git.eeqj.de/sneak/vaultik/issues/177 and is not
+// asserted here.
+//
+//nolint:paralleltest // installs the global logger via log.Initialize
+func TestBackupSurvivesMetadataExportInterruption(t *testing.T) {
+	log.Initialize(log.Config{})
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	dataDir := filepath.Join(tempDir, "src")
+	storeDir := filepath.Join(tempDir, "remote")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	ctx := context.Background()
+	cfg := faultTestConfig()
+
+	writeFaultSourceTree(t, fs, dataDir)
+
+	inner, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	// Back up and complete with a working backend.
+	sm := newFaultSnapshotManager(fs, inner, cfg, repos)
+	scanner := newFaultScanner(fs, inner, cfg, repos)
+
+	id, err := sm.CreateSnapshotWithName(ctx, cfg.Hostname, "export", "v", "g")
+	require.NoError(t, err)
+
+	_, err = scanner.Scan(ctx, dataDir, id)
+	require.NoError(t, err)
+	require.NoError(t, sm.CompleteSnapshot(ctx, id))
+
+	// Export through a backend that fails only the manifest upload. The
+	// database uploads first and lands; the manifest does not.
+	fault := faultstore.New(inner)
+	fault.OnPut = func(key string) faultstore.PutAction {
+		if strings.HasSuffix(key, "manifest.json.zst") {
+			return faultstore.PutFail
+		}
+
+		return faultstore.PutNormal
+	}
+
+	smFault := newFaultSnapshotManager(fs, fault, cfg, repos)
+
+	err = smFault.ExportSnapshotMetadata(ctx, dbPath, id)
+	require.Error(t, err, "export must fail when the manifest upload fails")
+
+	// The destination is in the partial state the scenario describes.
+	key := snapshot.RemoteSnapshotKey(id)
+
+	_, err = inner.Stat(ctx, "metadata/"+key+"/db.zst.age")
+	require.NoError(t, err, "database should have been uploaded before the manifest")
+
+	_, err = inner.Stat(ctx, "metadata/"+key+"/manifest.json.zst")
+	require.ErrorIs(t, err, storage.ErrNotFound, "manifest upload should not have landed")
+
+	// verify must fail loudly for this snapshot, in both modes.
+	reader := newReaderVaultik(ctx, cfg, inner, repos, fs)
+
+	deepOpts := &vaultik.VerifyOptions{Deep: true}
+	require.Error(t, reader.VerifySnapshotWithOptions(id, deepOpts),
+		"deep verify must report the missing manifest")
+
+	shallowOpts := &vaultik.VerifyOptions{Deep: false}
+	require.Error(t, reader.VerifySnapshotWithOptions(id, shallowOpts),
+		"shallow verify must report the missing manifest")
+
+	// snapshot list must not crash on the partial snapshot.
+	require.NoError(t, reader.ListSnapshots(false),
+		"snapshot list must tolerate a partially-exported snapshot")
+}
+
+// Scenario 5: the restore target runs out of space mid-file. Restore
+// must fail with an out-of-space error, and must not leave a truncated
+// file at the target path presenting as a complete restore. Restore
+// today writes each file straight to its final path and does not remove
+// it when a write fails, so the truncated file survives; deleting it is
+// tracked by https://git.eeqj.de/sneak/vaultik/issues/163. Skipped until
+// that lands, so the destination assertion below is recorded rather than
+// dropped.
+//
+//nolint:paralleltest // installs the global logger via log.Initialize
+func TestRestoreReportsDiskFull(t *testing.T) {
+	t.Skip("blocked on https://git.eeqj.de/sneak/vaultik/issues/163: " +
+		"a disk-full write leaves a truncated file at the target path " +
+		"instead of removing it")
+	log.Initialize(log.Config{})
+
+	osFS := afero.NewOsFs()
+	tempDir := t.TempDir()
+	dataDir := filepath.Join(tempDir, "src")
+	storeDir := filepath.Join(tempDir, "remote")
+	restoreDir := filepath.Join(tempDir, "restored")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	ctx := context.Background()
+	cfg := faultTestConfig()
+
+	testFiles := writeFaultSourceTree(t, osFS, dataDir)
+
+	inner, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	id := fullFaultBackup(ctx, t, osFS, inner, cfg, repos, dataDir, dbPath, "diskfull")
+	require.NoError(t, db.Close())
+
+	// Restore onto a filesystem that allows only a few bytes of file
+	// content: enough to create files, far too little to hold them.
+	budget := int64(8)
+	quota := &quotaFS{Fs: osFS, remaining: &budget}
+
+	v := newReaderVaultik(ctx, cfg, inner, nil, quota)
+	err = v.Restore(&vaultik.RestoreOptions{SnapshotID: id, TargetDir: restoreDir})
+
+	require.Error(t, err, "restore must fail when the target disk is full")
+	assert.Contains(t, err.Error(), errNoSpace.Error(),
+		"restore error should surface the out-of-space cause")
+
+	// The failure must not leave a truncated file behind presenting as a
+	// complete restore: any file at the target must hold the original
+	// bytes, or be absent.
+	assertNoCorruptFiles(t, osFS, restoreDir, testFiles)
+}
+
+// assertRestoredTree byte-compares every restored file against the
+// original.
+func assertRestoredTree(
+	t *testing.T, fs afero.Fs, restoreDir string, testFiles map[string][]byte,
+) {
+	t.Helper()
+
+	for origPath, expected := range testFiles {
+		restoredPath := filepath.Join(restoreDir, origPath)
+		got, err := afero.ReadFile(fs, restoredPath)
+		require.NoErrorf(t, err, "restored file missing: %s", origPath)
+		require.Equalf(t, expected, got, "restored content mismatch for %s", origPath)
+	}
+}
+
+// errNoSpace is the out-of-space error quotaFS returns once its byte
+// budget is exhausted, mirroring a real ENOSPC.
+var errNoSpace = errors.New("no space left on device")
+
+// quotaFS is an afero.Fs whose files may write only a fixed total number
+// of content bytes before failing, simulating a full restore target. It
+// wraps the interface so every method except Create delegates to the
+// real filesystem; only file writes are capped.
+type quotaFS struct {
+	afero.Fs
+
+	remaining *int64
+}
+
+//nolint:ireturn // afero.Fs.Create's signature requires returning afero.File.
+func (q *quotaFS) Create(name string) (afero.File, error) {
+	f, err := q.Fs.Create(name)
+	if err != nil {
+		return nil, err
+	}
+
+	return &quotaFile{File: f, remaining: q.remaining}, nil
+}
+
+// quotaFile fails writes once the shared byte budget is exhausted.
+type quotaFile struct {
+	afero.File
+
+	remaining *int64
+}
+
+func (q *quotaFile) Write(p []byte) (int, error) {
+	if *q.remaining <= 0 {
+		return 0, errNoSpace
+	}
+
+	allowed := min(int64(len(p)), *q.remaining)
+
+	n, err := q.File.Write(p[:allowed])
+	*q.remaining -= int64(n)
+
+	if err != nil {
+		return n, err
+	}
+
+	if int64(n) < int64(len(p)) {
+		return n, errNoSpace
+	}
+
+	return n, nil
+}
+
+// assertNoCorruptFiles fails if any file that made it to the restore
+// target holds content that differs from the original: a failed restore
+// may leave a file absent, but must never leave wrong bytes presenting
+// as the real file.
+func assertNoCorruptFiles(
+	t *testing.T, fs afero.Fs, restoreDir string, testFiles map[string][]byte,
+) {
+	t.Helper()
+
+	for origPath, expected := range testFiles {
+		restoredPath := filepath.Join(restoreDir, origPath)
+
+		got, err := afero.ReadFile(fs, restoredPath)
+		if err != nil {
+			if os.IsNotExist(err) {
+				continue
+			}
+
+			require.NoError(t, err)
+		}
+
+		assert.Equalf(t, expected, got,
+			"restored file %s holds corrupt content", origPath)
+	}
+}

internal/vaultik/prune.go

@@ -167,6 +167,12 @@ func (v *Vaultik) PruneBlobs(opts *PruneOptions) error {
 
 // collectReferencedBlobs downloads all manifests and returns the set of
 // referenced blob hashes.
+//
+// Every manifest must be read successfully. A manifest that cannot be
+// downloaded or decoded means its snapshot's blobs are unknown, so
+// treating them as unreferenced would let prune delete data a snapshot
+// still needs. Rather than risk that silent loss, any failure returns an
+// error naming the remote key and prune deletes nothing.
 func (v *Vaultik) collectReferencedBlobs() (map[string]bool, error) {
 	log.Info("Listing remote snapshots")
 	// IDs returned by listUniqueSnapshotIDs are remote keys (hashed
@@ -179,27 +185,22 @@ func (v *Vaultik) collectReferencedBlobs() (map[string]bool, error) {
 	log.Info("Found manifests in remote storage", "count", len(remoteKeys))
 
 	allBlobsReferenced := make(map[string]bool)
-	manifestCount := 0
 
 	for _, remoteKey := range remoteKeys {
 		log.Debug("Processing manifest", "remote_key", remoteKey)
 
 		manifest, err := v.downloadManifestByKey(remoteKey)
 		if err != nil {
-			log.Error("Failed to download manifest", "remote_key", remoteKey, "error", err)
+			return nil, fmt.Errorf("reading manifest %s: %w", remoteKey, err)
-
-			continue
 		}
 
 		for _, blob := range manifest.Blobs {
 			allBlobsReferenced[blob.Hash] = true
 		}
-
-		manifestCount++
 	}
 
 	log.Info("Processed manifests",
-		"count", manifestCount, "unique_blobs_referenced", len(allBlobsReferenced))
+		"count", len(remoteKeys), "unique_blobs_referenced", len(allBlobsReferenced))
 
 	return allBlobsReferenced, nil
 }

internal/vaultik/prune_count_test.go

@@ -0,0 +1,79 @@
+package vaultik //nolint:testpackage // exercises unexported count helpers
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/log"
+)
+
+// TestTableCountForReportSurfacesReadFailure is the regression guard for
+// the discarded-error bug: getTableCount for a table its query cannot
+// resolve must not silently become 0. A count that could not be read is
+// reported as unknown, which a reader can tell apart from an empty table.
+//
+//nolint:paralleltest // installs the global logger via log.Initialize
+func TestTableCountForReportSurfacesReadFailure(t *testing.T) {
+	log.Initialize(log.Config{})
+
+	ctx := context.Background()
+
+	db, err := database.New(ctx, ":memory:")
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = db.Close() })
+
+	v := &Vaultik{DB: db}
+	v.SetContext(ctx)
+
+	// A table present in the schema reads as a real count.
+	blobs := v.tableCountForReport("blobs")
+	require.NotNil(t, blobs, "an existing table must read as a real count")
+	assert.Equal(t, int64(0), *blobs)
+
+	// A syntactically valid name the sanitizer accepts but whose table
+	// the query cannot resolve is the exact shape #96 describes: a
+	// would-be loud failure that used to be discarded into a 0.
+	_, err = v.getTableCount("snapshots_missing")
+	require.Error(t, err, "a query against a nonexistent table must fail")
+
+	missing := v.tableCountForReport("snapshots_missing")
+	assert.Nil(t, missing, "a failed read is unknown, not a count")
+
+	// The rendered count for a failed read must say unknown, never 0.
+	assert.Equal(t, countUnknown, countText(missing))
+	assert.NotEqual(t, "0", countText(missing))
+}
+
+// TestCountTextDistinguishesEmptyFromUnknown pins the distinction the
+// output has to preserve: 0 means the table was empty, "unknown" means
+// the count could not be read.
+func TestCountTextDistinguishesEmptyFromUnknown(t *testing.T) {
+	t.Parallel()
+
+	zero := int64(0)
+	seven := int64(7)
+
+	assert.Equal(t, "0", countText(&zero))
+	assert.Equal(t, "7", countText(&seven))
+	assert.Equal(t, countUnknown, countText(nil))
+}
+
+// TestCountDiffUnknownWhenEitherSideUnknown checks that a delta computed
+// from an unreadable count is itself unknown rather than a plausible
+// number.
+func TestCountDiffUnknownWhenEitherSideUnknown(t *testing.T) {
+	t.Parallel()
+
+	before := int64(10)
+	after := int64(3)
+
+	require.NotNil(t, countDiff(&before, &after))
+	assert.Equal(t, int64(7), *countDiff(&before, &after))
+
+	assert.Nil(t, countDiff(nil, &after), "unknown before yields unknown delta")
+	assert.Nil(t, countDiff(&before, nil), "unknown after yields unknown delta")
+	assert.Nil(t, countDiff(nil, nil))
+}

internal/vaultik/prune_manifest_safety_test.go

@@ -0,0 +1,47 @@
+package vaultik_test
+
+import (
+	"bytes"
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/vaultik"
+)
+
+// TestPruneBlobs_UnreadableManifestDeletesNothing is the regression guard
+// for issue #157: prune identifies referenced blobs by reading every
+// snapshot's manifest, and a manifest it cannot decode used to be logged
+// and skipped. Blobs referenced only by that snapshot then looked
+// unreferenced and were deleted, with a zero exit — silent backup loss,
+// made worse by `snapshot create --prune` running unattended with force.
+//
+// The single blob here is referenced only by the snapshot whose manifest
+// is corrupt, so the old behaviour would delete it and succeed. Prune
+// must instead delete nothing and return an error.
+func TestPruneBlobs_UnreadableManifestDeletesNothing(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	env := newListEnv(t)
+	ctx := context.Background()
+
+	blobKey := "blobs/" + testBlobHashA[:2] + "/" + testBlobHashA[2:4] +
+		"/" + testBlobHashA
+	require.NoError(t, env.store.Put(ctx, blobKey,
+		bytes.NewReader([]byte("blob-bytes"))))
+
+	// A manifest at the path prune reads, but with contents it cannot
+	// decode.
+	require.NoError(t, env.store.Put(ctx,
+		"metadata/corruptkey/manifest.json.zst",
+		bytes.NewReader([]byte("not a valid manifest"))))
+
+	err := env.v.PruneBlobs(&vaultik.PruneOptions{Force: true})
+
+	require.Error(t, err, "prune must fail when a manifest cannot be read")
+	assert.True(t, env.store.hasKey(blobKey),
+		"no blob may be deleted when a manifest is unreadable")
+}

internal/vaultik/purge_local_remote_consistency_test.go

@@ -0,0 +1,146 @@
+package vaultik_test
+
+import (
+	"bytes"
+	"context"
+	"database/sql"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+	"sneak.berlin/go/vaultik/internal/types"
+	"sneak.berlin/go/vaultik/internal/vaultik"
+)
+
+// setupConsistencyTest builds a Vaultik whose local database and mock
+// remote both hold the given snapshots. Remote metadata is stored under
+// the production layout, metadata/<RemoteSnapshotKey(id)>/manifest.json.zst.
+// It returns the instance and the mock so a test can inspect the remote.
+func setupConsistencyTest(
+	t *testing.T, snapshotIDs []string,
+) (*vaultik.Vaultik, *MockStorer) {
+	t.Helper()
+
+	ctx := context.Background()
+	db, err := database.New(ctx, ":memory:")
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = db.Close() })
+
+	repos := database.NewRepositories(db)
+	mockStorage := NewMockStorer()
+
+	for _, id := range snapshotIDs {
+		parts := strings.Split(id, "_")
+		startedAt, err := time.Parse(time.RFC3339, parts[len(parts)-1])
+		require.NoError(t, err, "parsing timestamp from snapshot ID %q", id)
+
+		completedAt := startedAt.Add(5 * time.Minute)
+		snap := &database.Snapshot{
+			ID:             types.SnapshotID(id),
+			Hostname:       testHostname,
+			VaultikVersion: testLabel,
+			StartedAt:      startedAt,
+			CompletedAt:    &completedAt,
+		}
+		err = repos.WithTx(ctx, func(ctx context.Context, tx *sql.Tx) error {
+			return repos.Snapshots.Create(ctx, tx, snap)
+		})
+		require.NoError(t, err, "creating snapshot %s", id)
+
+		metadataKey := "metadata/" + snapshot.RemoteSnapshotKey(id) +
+			"/manifest.json.zst"
+		err = mockStorage.Put(ctx, metadataKey, strings.NewReader("stub"))
+		require.NoError(t, err)
+	}
+
+	v := &vaultik.Vaultik{
+		Storage:      mockStorage,
+		Repositories: repos,
+		DB:           db,
+		Stdout:       &bytes.Buffer{},
+		Stderr:       &bytes.Buffer{},
+		Stdin:        &bytes.Buffer{},
+	}
+	v.SetContext(ctx)
+
+	return v, mockStorage
+}
+
+func remoteHasSnapshot(t *testing.T, m *MockStorer, id string) bool {
+	t.Helper()
+
+	prefix := "metadata/" + snapshot.RemoteSnapshotKey(id) + "/"
+	keys, err := m.List(context.Background(), prefix)
+	require.NoError(t, err)
+
+	return len(keys) > 0
+}
+
+// TestPurgeKeepsRemotelyBackedLocalRows guards against issue #160
+// (https://git.eeqj.de/sneak/vaultik/issues/160): purge reconciles local
+// rows against the remote first, and that step compared human snapshot IDs
+// against the hashed remote directory names, which never match — so it
+// deleted every local record and the purge itself then removed nothing.
+//
+// With every snapshot still present remotely and nothing old enough to
+// purge, all local rows must survive the reconcile untouched.
+func TestPurgeKeepsRemotelyBackedLocalRows(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ids := []string{snapHomeT0, snapHomeT1, snapSystemT0}
+
+	v, _ := setupConsistencyTest(t, ids)
+
+	err := v.PurgeSnapshotsWithOptions(&vaultik.SnapshotPurgeOptions{
+		// 100 years: nothing is old enough to delete, so the reconcile
+		// is the only thing that touches the rows.
+		OlderThan: "36500d",
+		Force:     true,
+	})
+	require.NoError(t, err)
+
+	remaining := listRemainingSnapshots(t, v)
+	assert.Len(t, remaining, len(ids),
+		"remotely-backed local rows must survive the reconcile")
+	assert.Contains(t, remaining, snapHomeT0)
+	assert.Contains(t, remaining, snapHomeT1)
+	assert.Contains(t, remaining, snapSystemT0)
+}
+
+// TestPurgeRemovesLocalAndRemoteTogether proves the two halves stay
+// consistent: a purged snapshot is gone both locally and remotely, while a
+// retained one keeps both. Before the fix, the reconcile dropped every
+// local row yet the remote metadata was left in place.
+func TestPurgeRemovesLocalAndRemoteTogether(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ids := []string{snapHomeT0, snapHomeT1, snapSystemT0}
+
+	v, mock := setupConsistencyTest(t, ids)
+
+	err := v.PurgeSnapshotsWithOptions(&vaultik.SnapshotPurgeOptions{
+		KeepLatest: true,
+		Force:      true,
+	})
+	require.NoError(t, err)
+
+	// Keep latest per name: newest home and the lone system are kept.
+	remaining := listRemainingSnapshots(t, v)
+	assert.ElementsMatch(t, []string{snapHomeT1, snapSystemT0}, remaining)
+
+	// Local and remote agree: the older home snapshot is gone from both,
+	// the retained ones are present in both.
+	assert.False(t, remoteHasSnapshot(t, mock, snapHomeT0),
+		"purged snapshot must also be removed remotely")
+	assert.True(t, remoteHasSnapshot(t, mock, snapHomeT1),
+		"retained snapshot must remain remotely")
+	assert.True(t, remoteHasSnapshot(t, mock, snapSystemT0),
+		"retained snapshot must remain remotely")
+}

internal/vaultik/purge_per_name_test.go

@@ -12,6 +12,7 @@ import (
 	"github.com/stretchr/testify/require"
 	"sneak.berlin/go/vaultik/internal/database"
 	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
 	"sneak.berlin/go/vaultik/internal/types"
 	"sneak.berlin/go/vaultik/internal/vaultik"
 )
@@ -60,8 +61,11 @@ func setupPurgeTest(t *testing.T, snapshotIDs []string) *vaultik.Vaultik {
 		})
 		require.NoError(t, err, "creating snapshot %s", id)
 
-		// Create remote metadata stub so syncWithRemote keeps it
+		// Create the remote metadata stub under the production layout so
-		metadataKey := "metadata/" + id + "/manifest.json.zst"
+		// syncWithRemote keeps the local row. Production stores metadata
+		// under the hashed remote key, not the human snapshot ID.
+		metadataKey := "metadata/" + snapshot.RemoteSnapshotKey(id) +
+			"/manifest.json.zst"
 		err = mockStorage.Put(ctx, metadataKey, strings.NewReader("stub"))
 		require.NoError(t, err)
 	}

internal/vaultik/restore.go

@@ -11,6 +11,7 @@ import (
 	"math"
 	"os"
 	"path/filepath"
+	"strings"
 	"time"
 
 	"filippo.io/age"
@@ -18,7 +19,6 @@ import (
 	"sneak.berlin/go/vaultik/internal/blobgen"
 	"sneak.berlin/go/vaultik/internal/database"
 	"sneak.berlin/go/vaultik/internal/log"
-	"sneak.berlin/go/vaultik/internal/snapshot"
 	"sneak.berlin/go/vaultik/internal/types"
 )
 
@@ -31,17 +31,34 @@ var (
 			"Set the VAULTIK_AGE_SECRET_KEY environment variable to your " +
 			"age private key:\n" +
 			"  export VAULTIK_AGE_SECRET_KEY='AGE-SECRET-KEY-...'")
-	errBlobMissingFromIndex = errors.New("blob hash missing from blob index")
+	errBlobMissingFromIndex     = errors.New("blob hash missing from blob index")
-	errChunkNotInAnyBlob    = errors.New("chunk not found in any blob")
+	errChunkNotInAnyBlob        = errors.New("chunk not found in any blob")
-	errBlobIDNotInHashIndex = errors.New("blob id missing from hash index")
+	errBlobIDNotInHashIndex     = errors.New("blob id missing from hash index")
-	errShortChunkRead       = errors.New("short read")
+	errShortChunkRead           = errors.New("short read")
+	errRestorePathEscapesTarget = errors.New(
+		"refusing to restore path outside the target directory")
+	errTrailingRestoreData = errors.New(
+		"restored file has trailing data after its last chunk")
+	errRestoreIncomplete = errors.New(
+		"restore loop ended with files still pending")
 )
 
+// snapshotDBFilename is the name the decrypted snapshot database is
+// written under inside its private temp directory.
+const snapshotDBFilename = "snapshot.db"
+
 // restoreDirMode is the permission mode for directories created while
 // restoring (parent directories and the target root; restored
 // directories themselves get their stored mode).
 const restoreDirMode = 0o755
 
+// restoreFileMode is the restrictive mode a regular file is created with
+// during restore. Content is written while the file holds this mode; the
+// stored mode is applied only after the file is fully written and closed,
+// so a file whose stored mode is restrictive is never briefly readable by
+// other local users while its content is being written.
+const restoreFileMode = 0o600
+
 // sweepIntervalDivisor sets the sweeper threshold to one N-th of the
 // configured blob size limit.
 const sweepIntervalDivisor = 100
@@ -91,7 +108,7 @@ func (v *Vaultik) Restore(opts *RestoreOptions) error {
 	// Step 1: Download and decrypt the snapshot metadata database
 	log.Info("Downloading snapshot metadata...")
 
-	tempDB, err := v.downloadSnapshotDB(opts.SnapshotID, identity)
+	tempDB, tempDir, err := v.downloadSnapshotDB(opts.SnapshotID, identity)
 	if err != nil {
 		return fmt.Errorf("downloading snapshot database: %w", err)
 	}
@@ -101,10 +118,11 @@ func (v *Vaultik) Restore(opts *RestoreOptions) error {
 		if err != nil {
 			log.Debug("Failed to close temp database", "error", err)
 		}
-		// Clean up temp file
+		// Remove the whole private directory, so the decrypted database
-		err = v.Fs.Remove(tempDB.Path())
+		// and any SQLite side files it produced are gone on every path.
+		err = v.Fs.RemoveAll(tempDir)
 		if err != nil {
-			log.Debug("Failed to remove temp database", "error", err)
+			log.Debug("Failed to remove temp database directory", "error", err)
 		}
 	}()
 
@@ -357,6 +375,13 @@ func (v *Vaultik) runRestoreLoop(
 			totalBytesExpected, startTime, &lastStatusTime)
 	}
 
+	// The loop above stops as soon as nothing is ready and nothing more
+	// can be downloaded. If files still remain, they were abandoned
+	// rather than restored; fail loudly instead of reporting success.
+	if plan.hasPending() {
+		return errRestoreIncomplete
+	}
+
 	return nil
 }
 
@@ -371,12 +396,18 @@ func (v *Vaultik) runRestoreLoop(
 func (s *restoreSession) downloadNextBlobSet(plan *restorePlan) (bool, error) {
 	s.sweeper.sweep()
 
-	next := plan.pickNextDownload()
+	next, ok := plan.pickNextDownload()
-	if next.IsZero() {
+	if !ok {
 		return false, nil
 	}
 
 	for _, hash := range plan.blobsNeeded(next) {
+		// Stop between blobs on cancel so an interrupt ends the download
+		// phase promptly rather than fetching the rest of the set.
+		if s.ctx.Err() != nil {
+			return false, s.ctx.Err()
+		}
+
 		blob, ok := s.blobByHash[hash]
 		if !ok {
 			return false, fmt.Errorf("%w: %s", errBlobMissingFromIndex, hash[:16])
@@ -577,18 +608,24 @@ func (v *Vaultik) handleRestoreVerification(
 }
 
 // downloadSnapshotDB downloads and decrypts the snapshot metadata
-// database. The snapshotID is the human ID; we hash it to the remote
+// database. The identifier is resolved to the snapshot's remote key: a
-// key for the storage path.
+// human ID is hashed, and a remote key (or its abbreviation, as printed
+// for a remote-only snapshot) is used as-is, so a host with no local
+// index can restore the snapshots it can only see on the store.
 func (v *Vaultik) downloadSnapshotDB(
 	snapshotID string, identity age.Identity,
-) (*database.DB, error) {
+) (*database.DB, string, error) {
+	remoteKey, err := v.resolveSnapshotRemoteKey(snapshotID)
+	if err != nil {
+		return nil, "", err
+	}
+
 	// Download encrypted database from storage
-	dbKey := fmt.Sprintf("metadata/%s/db.zst.age",
+	dbKey := fmt.Sprintf("metadata/%s/db.zst.age", remoteKey)
-		snapshot.RemoteSnapshotKey(snapshotID))
 
 	reader, err := v.Storage.Get(v.ctx, dbKey)
 	if err != nil {
-		return nil, fmt.Errorf("downloading %s: %w", dbKey, err)
+		return nil, "", fmt.Errorf("downloading %s: %w", dbKey, err)
 	}
 
 	defer func() { _ = reader.Close() }()
@@ -596,7 +633,7 @@ func (v *Vaultik) downloadSnapshotDB(
 	// Read all data
 	encryptedData, err := io.ReadAll(reader)
 	if err != nil {
-		return nil, fmt.Errorf("reading encrypted data: %w", err)
+		return nil, "", fmt.Errorf("reading encrypted data: %w", err)
 	}
 
 	log.Debug("Downloaded encrypted database",
@@ -605,7 +642,7 @@ func (v *Vaultik) downloadSnapshotDB(
 	// Decrypt and decompress using blobgen.Reader
 	blobReader, err := blobgen.NewReader(bytes.NewReader(encryptedData), identity)
 	if err != nil {
-		return nil, fmt.Errorf("creating decryption reader: %w", err)
+		return nil, "", fmt.Errorf("creating decryption reader: %w", err)
 	}
 
 	defer func() { _ = blobReader.Close() }()
@@ -613,44 +650,52 @@ func (v *Vaultik) downloadSnapshotDB(
 	// Read the binary SQLite database
 	dbData, err := io.ReadAll(blobReader)
 	if err != nil {
-		return nil, fmt.Errorf("decrypting and decompressing: %w", err)
+		return nil, "", fmt.Errorf("decrypting and decompressing: %w", err)
 	}
 
 	log.Debug("Decrypted database", "size", ubytes(int64(len(dbData))))
 
-	// Create a temporary database file and write the binary SQLite data directly
+	return v.materializeSnapshotDB(dbData)
-	tempFile, err := afero.TempFile(v.Fs, "", "vaultik-restore-*.db")
+}
+
+// materializeSnapshotDB writes the decrypted snapshot database bytes into
+// a fresh private (0700) temp directory and opens the file read-only. On
+// any failure it removes the directory before returning, so no decrypted
+// metadata is left on disk when the open is interrupted or the payload is
+// damaged. On success the returned directory is the caller's to remove.
+func (v *Vaultik) materializeSnapshotDB(
+	dbData []byte,
+) (*database.DB, string, error) {
+	tempDir, err := afero.TempDir(v.Fs, "", "vaultik-restore-")
 	if err != nil {
-		return nil, fmt.Errorf("creating temp file: %w", err)
+		return nil, "", fmt.Errorf("creating temp directory: %w", err)
 	}
 
-	tempPath := tempFile.Name()
+	success := false
 
-	// Write the binary SQLite database directly
+	defer func() {
-	_, err = tempFile.Write(dbData)
+		if !success {
+			_ = v.Fs.RemoveAll(tempDir)
+		}
+	}()
+
+	dbPath := filepath.Join(tempDir, snapshotDBFilename)
+
+	err = afero.WriteFile(v.Fs, dbPath, dbData, restoreFileMode)
 	if err != nil {
-		_ = tempFile.Close()
+		return nil, "", fmt.Errorf("writing database file: %w", err)
-		_ = v.Fs.Remove(tempPath)
-
-		return nil, fmt.Errorf("writing database file: %w", err)
 	}
 
-	err = tempFile.Close()
+	log.Debug("Created restore database", "path", dbPath)
-	if err != nil {
-		_ = v.Fs.Remove(tempPath)
 
-		return nil, fmt.Errorf("closing temp file: %w", err)
+	db, err := database.OpenReadOnly(v.ctx, dbPath)
+	if err != nil {
+		return nil, "", fmt.Errorf("opening restore database: %w", err)
 	}
 
-	log.Debug("Created restore database", "path", tempPath)
+	success = true
 
-	// Open the database
+	return db, tempDir, nil
-	db, err := database.New(v.ctx, tempPath)
-	if err != nil {
-		return nil, fmt.Errorf("opening restore database: %w", err)
-	}
-
-	return db, nil
 }
 
 // getFilesToRestore returns the list of files to restore based on path filters
@@ -755,13 +800,85 @@ type restoreSession struct {
 	runningAsRoot bool
 }
 
+// containedRestorePath resolves rel — a path read from the snapshot
+// database — to its location under targetDir and confirms the write will
+// stay inside the target.
+//
+// age decryption proves a snapshot is readable, not that it is honest, so
+// every stored path is treated as hostile. rel is rejected unless
+// filepath.IsLocal accepts it once the leading separator is stripped:
+// stored paths are absolute and the join to targetDir drops that
+// separator, so "/etc/passwd" is judged as the relative "etc/passwd" it
+// becomes on disk. This bars "..", absolute, and empty paths.
+//
+// A stored symlink whose target points outside the tree is still honest
+// (and restored verbatim), but a later entry must not be written through
+// it. Each existing ancestor directory below the target is therefore
+// Lstat'ed and a symlink among them is refused. The leaf itself is not
+// traversed: honest snapshots restore symlinks at leaf positions, and the
+// unique-path constraint keeps a leaf from being both a symlink and a
+// regular file. The target directory itself may be a symlink; only
+// components below it are checked.
+func containedRestorePath(fs afero.Fs, targetDir, rel string) (string, error) {
+	local := strings.TrimPrefix(rel, string(filepath.Separator))
+	if !filepath.IsLocal(local) {
+		return "", fmt.Errorf("%w: %s", errRestorePathEscapesTarget, rel)
+	}
+
+	local = filepath.Clean(local)
+	targetPath := filepath.Join(targetDir, local)
+
+	relDir := filepath.Dir(local)
+	if relDir == "." {
+		return targetPath, nil
+	}
+
+	current := targetDir
+	for component := range strings.SplitSeq(relDir, string(filepath.Separator)) {
+		current = filepath.Join(current, component)
+
+		info, err := lstatIfPossible(fs, current)
+		if err != nil {
+			if os.IsNotExist(err) {
+				continue
+			}
+
+			return "", fmt.Errorf("checking restore path %s: %w", current, err)
+		}
+
+		if info.Mode()&os.ModeSymlink != 0 {
+			return "", fmt.Errorf("%w: %s descends through symlink %s",
+				errRestorePathEscapesTarget, rel, current)
+		}
+	}
+
+	return targetPath, nil
+}
+
+// lstatIfPossible performs a symlink-aware stat when the filesystem
+// supports it. afero.OsFs does; MemMapFs, which has no symlinks, reports
+// that Lstat was not used and its result never carries ModeSymlink.
+func lstatIfPossible(fs afero.Fs, name string) (os.FileInfo, error) {
+	if lstater, ok := fs.(afero.Lstater); ok {
+		info, _, err := lstater.LstatIfPossible(name)
+
+		return info, err
+	}
+
+	return fs.Stat(name)
+}
+
 // restoreFile dispatches to the right per-kind restorer.
 func (s *restoreSession) restoreFile(file *database.File) error {
-	targetPath := filepath.Join(s.opts.TargetDir, file.Path.String())
+	targetPath, err := containedRestorePath(
+		s.v.Fs, s.opts.TargetDir, file.Path.String())
+	if err != nil {
+		return err
+	}
 
 	parentDir := filepath.Dir(targetPath)
 
-	err := s.v.Fs.MkdirAll(parentDir, restoreDirMode)
+	err = s.v.Fs.MkdirAll(parentDir, restoreDirMode)
 	if err != nil {
 		return fmt.Errorf("creating parent directory: %w", err)
 	}
@@ -809,6 +926,13 @@ func (s *restoreSession) restoreDirectory(
 		return fmt.Errorf("creating directory: %w", err)
 	}
 
+	// MkdirAll applies the process umask, so chmod to the exact stored
+	// mode. A failure here is non-fatal.
+	err = s.v.Fs.Chmod(targetPath, os.FileMode(file.Mode))
+	if err != nil {
+		log.Debug("Failed to set permissions", "path", targetPath, "error", err)
+	}
+
 	s.applyFileMetadata(file, targetPath)
 
 	s.result.FilesRestored++
@@ -816,25 +940,22 @@ func (s *restoreSession) restoreDirectory(
 	return nil
 }
 
-// applyFileMetadata applies stored permissions, ownership (when running
+// applyFileMetadata applies ownership (when running as root on a real
-// as root on a real filesystem), and mtime to a restored path. Failures
+// filesystem) and mtime to a restored path. Permission mode is applied
-// are logged at debug level and do not abort the restore.
+// separately by each caller, with different failure handling, so it is
+// not touched here. Failures are logged at debug level and do not abort
+// the restore.
 func (s *restoreSession) applyFileMetadata(file *database.File, targetPath string) {
-	err := s.v.Fs.Chmod(targetPath, os.FileMode(file.Mode))
-	if err != nil {
-		log.Debug("Failed to set permissions", "path", targetPath, "error", err)
-	}
-
 	if s.runningAsRoot {
 		if _, ok := s.v.Fs.(*afero.OsFs); ok {
-			err = os.Chown(targetPath, int(file.UID), int(file.GID))
+			err := os.Chown(targetPath, int(file.UID), int(file.GID))
 			if err != nil {
 				log.Debug("Failed to set ownership", "path", targetPath, "error", err)
 			}
 		}
 	}
 
-	err = s.v.Fs.Chtimes(targetPath, file.MTime, file.MTime)
+	err := s.v.Fs.Chtimes(targetPath, file.MTime, file.MTime)
 	if err != nil {
 		log.Debug("Failed to set mtime", "path", targetPath, "error", err)
 	}
@@ -868,17 +989,30 @@ func (s *restoreSession) restoreRegularFile(
 
 	t0 = time.Now()
 
-	outFile, err := s.v.Fs.Create(targetPath)
+	// Remove any existing entry, then create the file with a restrictive
+	// mode via O_EXCL. The stored mode is applied only after the content
+	// is written and the file closed, so a file whose stored mode is
+	// restrictive is never briefly readable by other local users while
+	// its content is written. Removing first (rather than failing on a
+	// leftover file) matches the documented behaviour that re-running
+	// restore overwrites partial output.
+	_ = s.v.Fs.Remove(targetPath)
+
+	outFile, err := s.v.Fs.OpenFile(
+		targetPath, os.O_CREATE|os.O_EXCL|os.O_WRONLY, restoreFileMode)
 	createDur := time.Since(t0)
 
 	if err != nil {
 		return fmt.Errorf("creating output file: %w", err)
 	}
 
-	defer func() { _ = outFile.Close() }()
-
 	bytesWritten, timings, err := s.writeFileChunks(outFile, fileChunks)
 	if err != nil {
+		// Do not leave a partial file behind.
+		_ = outFile.Close()
+
+		s.removePartialRestore(targetPath)
+
 		return err
 	}
 
@@ -896,9 +1030,12 @@ func (s *restoreSession) restoreRegularFile(
 
 	err = outFile.Close()
 	if err != nil {
+		s.removePartialRestore(targetPath)
+
 		return fmt.Errorf("closing output file: %w", err)
 	}
 
+	s.applyRestoredFileMode(file, targetPath)
 	s.applyFileMetadata(file, targetPath)
 
 	s.result.FilesRestored++
@@ -909,6 +1046,31 @@ func (s *restoreSession) restoreRegularFile(
 	return nil
 }
 
+// applyRestoredFileMode applies the stored permission bits to a
+// just-written regular file (created with restoreFileMode). A failure is
+// a user-visible warning, not a fatal error: the file's content is
+// intact and it remains at the restrictive create-time mode, so the
+// restore is not aborted or discarded over it.
+func (s *restoreSession) applyRestoredFileMode(
+	file *database.File, targetPath string,
+) {
+	err := s.v.Fs.Chmod(targetPath, os.FileMode(file.Mode))
+	if err != nil {
+		s.v.UI.Warningf("Failed to set mode %s on %s: %v",
+			os.FileMode(file.Mode).Perm(), s.v.UI.Path(targetPath), err)
+	}
+}
+
+// removePartialRestore deletes a restore output file whose write did not
+// complete, so a failed restore never leaves a partial file behind.
+func (s *restoreSession) removePartialRestore(targetPath string) {
+	err := s.v.Fs.Remove(targetPath)
+	if err != nil {
+		log.Debug("Failed to remove partial restore file",
+			"path", targetPath, "error", err)
+	}
+}
+
 // writeFileChunks streams each of the file's chunks from the blob disk
 // cache into outFile, crediting restored bytes to the sweeper as it
 // goes. Returns the bytes written plus per-phase timing accumulators.
@@ -921,6 +1083,12 @@ func (s *restoreSession) writeFileChunks(
 	)
 
 	for _, fc := range fileChunks {
+		// Stop between chunks on cancel so an interrupt does not keep
+		// writing a large file after the operation has been told to stop.
+		if s.ctx.Err() != nil {
+			return bytesWritten, timings, s.ctx.Err()
+		}
+
 		chunkHashStr := fc.ChunkHash.String()
 
 		blobChunk, ok := s.chunkToBlobMap[chunkHashStr]
@@ -990,11 +1158,19 @@ func (s *restoreSession) downloadBlobToCache(
 	streamDur := time.Since(t0)
 	closeErr := rc.Close()
 
+	// closeErr carries the blob's hash-verification result (a mismatch,
+	// or the stream not being fully read). On any failure, drop the
+	// cache entry so a blob that failed verification is never read back
+	// as if it were valid.
 	if copyErr != nil {
+		s.blobCache.Delete(blobHash)
+
 		return copyErr
 	}
 
 	if closeErr != nil {
+		s.blobCache.Delete(blobHash)
+
 		return closeErr
 	}
 
@@ -1057,17 +1233,22 @@ func (v *Vaultik) verifyRestoredFiles(
 			return ctx.Err()
 		}
 
-		targetPath := filepath.Join(targetDir, file.Path.String())
+		targetPath, err := containedRestorePath(v.Fs, targetDir, file.Path.String())
+		if err == nil {
+			var bytesVerified int64
+
+			bytesVerified, err = v.verifyFile(ctx, repos, file, targetPath)
+			if err == nil {
+				result.FilesVerified++
+				result.BytesVerified += bytesVerified
+			}
+		}
 
-		bytesVerified, err := v.verifyFile(ctx, repos, file, targetPath)
 		if err != nil {
 			log.Error("File verification failed", "path", file.Path, "error", err)
 
 			result.FilesFailed++
 			result.FailedFiles = append(result.FailedFiles, file.Path.String())
-		} else {
-			result.FilesVerified++
-			result.BytesVerified += bytesVerified
 		}
 
 		bytesProcessed += file.Size
@@ -1157,6 +1338,17 @@ func (v *Vaultik) verifyFile(
 		bytesVerified += int64(n)
 	}
 
+	// The stored chunks account for the whole file, so the reader must
+	// be at EOF now. Trailing bytes past the last chunk are corruption
+	// the per-chunk loop cannot see.
+	extra := make([]byte, 1)
+
+	n, err := f.Read(extra)
+	if n != 0 || !errors.Is(err, io.EOF) {
+		return bytesVerified, fmt.Errorf("%w: file longer than its %d chunk(s)",
+			errTrailingRestoreData, len(fileChunks))
+	}
+
 	log.Debug("File verified",
 		"path", file.Path, "bytes", bytesVerified, "chunks", len(fileChunks))
 

internal/vaultik/restore_another_machine_test.go

@@ -0,0 +1,167 @@
+package vaultik_test
+
+import (
+	"bytes"
+	"context"
+	"io"
+	"path/filepath"
+	"testing"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+	"sneak.berlin/go/vaultik/internal/storage"
+	"sneak.berlin/go/vaultik/internal/ui"
+	"sneak.berlin/go/vaultik/internal/vaultik"
+)
+
+// TestRestoreOnAnotherMachine proves the disaster-recovery path: a host
+// that has only the vaultik binary, the age secret key, and the storage
+// credentials — no local index, a different hostname, and no
+// age_recipients configured — can list, restore, and verify a snapshot
+// straight from the destination store.
+//
+// The backup half writes a snapshot with one index and hostname. The
+// restore half throws that index away entirely: a fresh, empty index and
+// a config that shares nothing with the original but the storage location
+// and the secret key. If restore or verify needed the original local
+// index — or the human snapshot ID that only that index holds — this test
+// could not run, because the recovery host can know neither.
+func TestRestoreOnAnotherMachine(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+
+	dataDir := filepath.Join(tempDir, "source")
+	storeDir := filepath.Join(tempDir, "remote")
+	restoreDir := filepath.Join(tempDir, "restored")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	chunkSize := int64(64 * 1024)
+	maxBlobSize := int64(512 * 1024)
+
+	sourceFiles := writeRecoverySourceTree(t, fs, dataDir, chunkSize)
+
+	ctx := context.Background()
+
+	// Backup host: one index, hostname test-host, age_recipients set.
+	// runFileStorageBackup closes the index before returning, so nothing
+	// below can lean on it.
+	_, storer, originalID := runFileStorageBackup(
+		ctx, t, fs, dataDir, storeDir, dbPath, chunkSize, maxBlobSize)
+
+	// Recovery host: a fresh empty index, a different hostname, and no
+	// age_recipients — only the secret key and the same storage location.
+	recovery, stdout := newRecoveryHost(ctx, t, fs, storer)
+
+	// The recovery index really is empty. This is the assertion that makes
+	// the test a guard against restore quietly depending on the original
+	// index: if it did, an empty index would make restore fail.
+	localSnaps, err := recovery.Repositories.Snapshots.ListRecent(ctx, 100)
+	require.NoError(t, err)
+	require.Empty(t, localSnaps, "recovery host must start with no local index")
+
+	// List: the snapshot shows up as remote-only, identified by its remote
+	// key, with no recoverable human ID.
+	require.NoError(t, recovery.ListSnapshots(true))
+
+	rows := decodeListJSON(t, stdout.String())
+	require.Len(t, rows, 1)
+
+	remote := rows[0]
+	assert.False(t, remote.LocallyTracked, "snapshot must be remote-only here")
+	assert.Empty(t, remote.ID, "the human ID is unknown to the recovery host")
+	require.Len(t, remote.RemoteKey, 64)
+	assert.Equal(t, snapshot.RemoteSnapshotKey(originalID), remote.RemoteKey,
+		"the listed key is the hashed snapshot ID")
+
+	// Restore driven by the abbreviated identifier the table prints (the
+	// first 12 hex of the remote key), then deep-verify from the store
+	// keyed by the full remote key. Both are what a recovery host can know.
+	require.NoError(t, recovery.Restore(&vaultik.RestoreOptions{
+		SnapshotID: remote.RemoteKey[:12],
+		TargetDir:  restoreDir,
+		Verify:     true,
+	}))
+	require.NoError(t, recovery.RunDeepVerify(
+		remote.RemoteKey, &vaultik.VerifyOptions{Deep: true}))
+
+	assertRestoredTreeMatches(t, fs, restoreDir, sourceFiles)
+}
+
+// writeRecoverySourceTree writes a small source tree spanning several
+// chunks (so restore reassembles real multi-chunk files) and returns the
+// content keyed by absolute path.
+func writeRecoverySourceTree(
+	t *testing.T, fs afero.Fs, dataDir string, chunkSize int64,
+) map[string][]byte {
+	t.Helper()
+
+	sourceFiles := map[string][]byte{
+		filepath.Join(dataDir, "notes.txt"):        []byte("recover me"),
+		filepath.Join(dataDir, "sub", "big.bin"):   bytesPattern("big-", int(chunkSize*3)),
+		filepath.Join(dataDir, "sub", "small.bin"): bytesPattern("small-", 128),
+	}
+
+	for path, content := range sourceFiles {
+		require.NoError(t, fs.MkdirAll(filepath.Dir(path), 0o755))
+		require.NoError(t, afero.WriteFile(fs, path, content, 0o644))
+	}
+
+	return sourceFiles
+}
+
+// newRecoveryHost builds the Vaultik a replacement machine would run: an
+// empty in-memory index, a hostname different from the backup host, no
+// age_recipients, and only the secret key plus the shared storer. It
+// returns the instance and the buffer its stdout is wired to.
+func newRecoveryHost(
+	ctx context.Context, t *testing.T, fs afero.Fs, storer storage.Storer,
+) (*vaultik.Vaultik, *bytes.Buffer) {
+	t.Helper()
+
+	recoveryDB, err := database.New(ctx, ":memory:")
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = recoveryDB.Close() })
+
+	stdout := &bytes.Buffer{}
+
+	recovery := &vaultik.Vaultik{
+		Config: &config.Config{
+			AgeSecretKey: testAgeSecretKey,
+			Hostname:     "recovery-host",
+		},
+		Storage:      storer,
+		Fs:           fs,
+		Repositories: database.NewRepositories(recoveryDB),
+		DB:           recoveryDB,
+		Stdout:       stdout,
+		Stderr:       io.Discard,
+		UI:           ui.NewWithColor(io.Discard, false),
+	}
+	recovery.SetContext(ctx)
+
+	return recovery, stdout
+}
+
+// assertRestoredTreeMatches byte-compares every restored file against its
+// source content.
+func assertRestoredTreeMatches(
+	t *testing.T, fs afero.Fs, restoreDir string, sourceFiles map[string][]byte,
+) {
+	t.Helper()
+
+	for origPath, expected := range sourceFiles {
+		restored := filepath.Join(restoreDir, origPath)
+		got, err := afero.ReadFile(fs, restored)
+		require.NoErrorf(t, err, "restored file missing: %s", restored)
+		require.Truef(t, bytes.Equal(got, expected),
+			"byte mismatch for %s", origPath)
+	}
+}

internal/vaultik/restore_containment_test.go

@@ -0,0 +1,175 @@
+package vaultik //nolint:testpackage // drives unexported restore internals
+
+import (
+	"context"
+	"io"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/types"
+	"sneak.berlin/go/vaultik/internal/ui"
+)
+
+// These tests exercise the path-containment guard that keeps restore from
+// writing outside its target directory. age decryption proves only that a
+// snapshot is readable, not that its recorded paths are honest, so restore
+// treats every stored path as hostile: a compromised backed-up host could
+// forge a snapshot that decrypts cleanly, and restore usually runs as root.
+//
+// They drive restoreAllFiles directly (rather than the full Restore, which
+// downloads and decrypts the metadata database from storage) so a snapshot
+// database with adversarial rows can be handed to the restore loop without
+// the surrounding blob/storage machinery. Directory and symlink entries
+// carry no chunks, so no blobs are needed.
+
+// containmentDirMode marks a File row as a directory for the restore loop.
+const containmentDirMode = uint32(os.ModeDir | 0o755)
+
+// newContainmentVaultik builds the minimal Vaultik needed to run
+// restoreAllFiles against fs.
+func newContainmentVaultik(ctx context.Context, fs afero.Fs) *Vaultik {
+	v := &Vaultik{
+		Config: &config.Config{
+			BlobSizeLimit: config.Size(10 * 1024 * 1024),
+		},
+		Fs:     fs,
+		Stdout: io.Discard,
+		Stderr: io.Discard,
+		UI:     ui.NewWithColor(io.Discard, false),
+	}
+	v.SetContext(ctx)
+
+	return v
+}
+
+// makeFiles inserts the given rows into a fresh in-memory snapshot database
+// and returns them (with IDs assigned) plus the repositories.
+func makeFiles(
+	ctx context.Context, t *testing.T, rows []*database.File,
+) ([]*database.File, *database.Repositories) {
+	t.Helper()
+
+	db, err := database.New(ctx, filepath.Join(t.TempDir(), "index.sqlite"))
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = db.Close() })
+
+	repos := database.NewRepositories(db)
+	for _, f := range rows {
+		require.NoError(t, repos.Files.Create(ctx, nil, f))
+	}
+
+	return rows, repos
+}
+
+func TestRestoreRejectsPathTraversal(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	tests := []struct {
+		name string
+		// rows are inserted in order; the escape entry is restored after
+		// any entry it depends on (the symlink case needs its link first).
+		rows func(outsideDir string) []*database.File
+		// escaped is the path, outside the target, that must not appear.
+		escaped func(tempDir, outsideDir string) string
+	}{
+		{
+			name: "relative dotdot",
+			rows: func(_ string) []*database.File {
+				return []*database.File{{
+					Path: "../escaped-relative",
+					Mode: containmentDirMode,
+				}}
+			},
+			escaped: func(tempDir, _ string) string {
+				return filepath.Join(tempDir, "escaped-relative")
+			},
+		},
+		{
+			name: "absolute with dotdot",
+			rows: func(_ string) []*database.File {
+				return []*database.File{{
+					Path: "/a/../../escaped-absolute",
+					Mode: containmentDirMode,
+				}}
+			},
+			escaped: func(tempDir, _ string) string {
+				return filepath.Join(tempDir, "escaped-absolute")
+			},
+		},
+		{
+			name: "child through symlink",
+			rows: func(outsideDir string) []*database.File {
+				return []*database.File{
+					// Restored first: an in-target symlink pointing out.
+					{Path: "linkdir", LinkTarget: types.FilePath(outsideDir)},
+					// Restored second: a child written through that link.
+					{Path: "linkdir/child", Mode: containmentDirMode},
+				}
+			},
+			escaped: func(_, outsideDir string) string {
+				return filepath.Join(outsideDir, "child")
+			},
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			ctx := context.Background()
+			fs := afero.NewOsFs()
+			tempDir := t.TempDir()
+			targetDir := filepath.Join(tempDir, "target")
+			outsideDir := filepath.Join(tempDir, "outside")
+			require.NoError(t, fs.MkdirAll(outsideDir, 0o755))
+
+			rows, repos := makeFiles(ctx, t, tc.rows(outsideDir))
+			v := newContainmentVaultik(ctx, fs)
+
+			_, err := v.restoreAllFiles(rows, repos,
+				&RestoreOptions{TargetDir: targetDir}, nil, nil)
+
+			require.ErrorIs(t, err, errRestorePathEscapesTarget)
+
+			escaped := tc.escaped(tempDir, outsideDir)
+			_, statErr := os.Lstat(escaped)
+			require.Truef(t, os.IsNotExist(statErr),
+				"restore wrote outside the target at %s", escaped)
+		})
+	}
+}
+
+// TestRestoreAllowsSymlinkPointingOutsideTree confirms the guard does not
+// over-block: an honest snapshot may contain a symlink whose target lies
+// outside the restored tree, and it must still be restored verbatim.
+func TestRestoreAllowsSymlinkPointingOutsideTree(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	ctx := context.Background()
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	targetDir := filepath.Join(tempDir, "target")
+	linkTarget := filepath.Join(tempDir, "outside", "data")
+
+	rows, repos := makeFiles(ctx, t, []*database.File{
+		{Path: "goodlink", LinkTarget: types.FilePath(linkTarget), MTime: time.Unix(0, 0)},
+	})
+	v := newContainmentVaultik(ctx, fs)
+
+	_, err := v.restoreAllFiles(rows, repos,
+		&RestoreOptions{TargetDir: targetDir}, nil, nil)
+	require.NoError(t, err)
+
+	got, err := os.Readlink(filepath.Join(targetDir, "goodlink"))
+	require.NoError(t, err)
+	require.Equal(t, linkTarget, got)
+}

internal/vaultik/restore_interrupt_test.go

@@ -0,0 +1,159 @@
+package vaultik //nolint:testpackage // sets ctx/cancel and inspects scratch files
+
+import (
+	"context"
+	"io"
+	"path/filepath"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/storage"
+	"sneak.berlin/go/vaultik/internal/ui"
+)
+
+// blockingBlobStorer wraps a Storer and blocks the first blob download
+// until its context is cancelled, so a test can catch a restore while it
+// is mid-download. Metadata reads pass straight through, so the restore
+// reaches the blob-download phase — having already written its decrypted
+// scratch files — before it blocks.
+type blockingBlobStorer struct {
+	storage.Storer
+
+	once    sync.Once
+	entered chan struct{}
+}
+
+func newBlockingBlobStorer(inner storage.Storer) *blockingBlobStorer {
+	return &blockingBlobStorer{Storer: inner, entered: make(chan struct{})}
+}
+
+func (b *blockingBlobStorer) Get(
+	ctx context.Context, key string,
+) (io.ReadCloser, error) {
+	if strings.HasPrefix(key, "blobs/") {
+		b.once.Do(func() { close(b.entered) })
+		<-ctx.Done()
+
+		return nil, ctx.Err()
+	}
+
+	return b.Storer.Get(ctx, key)
+}
+
+// TestRestoreCleansTempDirOnInterrupt drives a restore through the stop
+// path (v.StartOperation, which is what the fx OnStop hook uses) instead
+// of calling Restore directly, catches it mid-download, and asserts that
+// stopping waits for the operation to unwind and removes its decrypted
+// scratch files — the blob cache and the temporary snapshot database —
+// from the temp directory. Without the wait a SIGINT exits the process
+// before those defers run, leaving decrypted data on disk (issue #159).
+//
+// Not parallel: it points TMPDIR at a private directory (via t.Setenv)
+// so it can assert on exactly the scratch files this restore created.
+func TestRestoreCleansTempDirOnInterrupt(t *testing.T) {
+	log.Initialize(log.Config{})
+
+	fs := afero.NewOsFs()
+	root := t.TempDir()
+
+	dataDir := filepath.Join(root, "source")
+	storeDir := filepath.Join(root, "remote")
+	restoreDir := filepath.Join(root, "restored")
+	dbPath := filepath.Join(root, "index.sqlite")
+
+	require.NoError(t, fs.MkdirAll(dataDir, 0o755))
+
+	buildLocalityFixture(t, fs, dataDir)
+
+	cfg, storer, snapshotID := setupLocalityBackup(
+		context.Background(), t, fs, dataDir, storeDir, dbPath)
+
+	// Point the "" temp paths (the blob cache directory and the
+	// snapshot-database directory) at a private directory so the test can
+	// assert on exactly the scratch this restore creates.
+	scratch := filepath.Join(root, "scratch")
+	require.NoError(t, fs.MkdirAll(scratch, 0o755))
+	t.Setenv("TMPDIR", scratch)
+
+	gate := newBlockingBlobStorer(storer)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	v := &Vaultik{
+		Config:  cfg,
+		Storage: gate,
+		Fs:      fs,
+		Stdout:  io.Discard,
+		Stderr:  io.Discard,
+		UI:      ui.NewWithColor(io.Discard, false),
+		ctx:     ctx,
+		cancel:  cancel,
+	}
+
+	var (
+		opReturned atomic.Bool
+		restoreErr error
+	)
+
+	stop := v.StartOperation(func() {
+		defer opReturned.Store(true)
+
+		restoreErr = v.Restore(&RestoreOptions{
+			SnapshotID: snapshotID,
+			TargetDir:  restoreDir,
+		})
+	})
+
+	// Wait until the restore is blocked mid-download; its decrypted
+	// scratch files exist by now.
+	select {
+	case <-gate.entered:
+	case <-time.After(30 * time.Second):
+		t.Fatal("restore never reached the blob-download phase")
+	}
+
+	require.NotEmpty(t, scratchEntries(t, scratch),
+		"expected decrypted scratch files to exist mid-restore")
+
+	// Stop the operation the way the fx OnStop hook does.
+	stopCtx, stopCancel := context.WithTimeout(
+		context.Background(), 30*time.Second)
+	defer stopCancel()
+
+	require.True(t, stop(stopCtx),
+		"stop timed out; the operation goroutine did not return")
+
+	// stop returns only once the operation goroutine has returned, so its
+	// cleanup defers have run by the time we read these.
+	require.True(t, opReturned.Load(),
+		"stop returned before the operation goroutine finished")
+	require.ErrorIs(t, restoreErr, context.Canceled)
+	require.Empty(t, scratchEntries(t, scratch),
+		"decrypted scratch files remained after the interrupt")
+}
+
+// scratchEntries returns the vaultik blob-cache and snapshot-database
+// scratch entries currently present in dir.
+func scratchEntries(t *testing.T, dir string) []string {
+	t.Helper()
+
+	var matches []string
+
+	for _, pattern := range []string{
+		"vaultik-blobcache-*", "vaultik-restore-*",
+	} {
+		found, err := filepath.Glob(filepath.Join(dir, pattern))
+		require.NoError(t, err)
+
+		matches = append(matches, found...)
+	}
+
+	return matches
+}

internal/vaultik/restore_perms_test.go

@@ -0,0 +1,306 @@
+package vaultik //nolint:testpackage // drives restore through unexported session
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"syscall"
+	"testing"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/config"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+	"sneak.berlin/go/vaultik/internal/storage"
+	"sneak.berlin/go/vaultik/internal/ui"
+)
+
+// errSpyWrite is the injected write failure used to exercise the
+// partial-file cleanup path.
+var errSpyWrite = errors.New("injected write failure")
+
+// modeSpyFs wraps a real filesystem so restore tests can observe and
+// perturb the single output file whose path contains watch. It records
+// the on-disk permission bits seen at the moment content is first
+// written (the window during which another local user could read it),
+// and can inject a write failure or append trailing bytes on close.
+type modeSpyFs struct {
+	afero.Fs
+
+	watch string
+
+	mu         sync.Mutex
+	writeModes []os.FileMode
+	failWrite  bool
+	trailing   int
+}
+
+//nolint:ireturn // afero.Fs.OpenFile is defined to return the interface
+func (m *modeSpyFs) OpenFile(
+	name string, flag int, perm os.FileMode,
+) (afero.File, error) {
+	f, err := m.Fs.OpenFile(name, flag, perm)
+	if err != nil || !strings.Contains(name, m.watch) {
+		return f, err
+	}
+
+	return &modeSpyFile{File: f, fs: m, path: name}, nil
+}
+
+type modeSpyFile struct {
+	afero.File
+
+	fs      *modeSpyFs
+	path    string
+	written bool
+}
+
+func (f *modeSpyFile) Write(p []byte) (int, error) {
+	if !f.written {
+		f.written = true
+
+		info, err := f.fs.Stat(f.path)
+		if err == nil {
+			f.fs.mu.Lock()
+			f.fs.writeModes = append(f.fs.writeModes, info.Mode().Perm())
+			f.fs.mu.Unlock()
+		}
+	}
+
+	if f.fs.failWrite {
+		return 0, errSpyWrite
+	}
+
+	return f.File.Write(p)
+}
+
+func (f *modeSpyFile) Close() error {
+	if f.fs.trailing > 0 {
+		_, _ = f.File.Write(bytes.Repeat([]byte{'x'}, f.fs.trailing))
+	}
+
+	return f.File.Close()
+}
+
+// backupOneFile writes a single source file with the given mode and
+// backs it up into a fresh file storer, returning everything a restore
+// needs. The index database is closed before returning so the restore
+// half runs from the exported metadata and remote bytes only.
+func backupOneFile(
+	ctx context.Context, t *testing.T, fs afero.Fs, tempDir, name string,
+	content []byte, mode os.FileMode,
+) (*config.Config, *storage.FileStorer, string, string) {
+	t.Helper()
+
+	dataDir := filepath.Join(tempDir, "src")
+	require.NoError(t, fs.MkdirAll(dataDir, 0o755))
+
+	srcPath := filepath.Join(dataDir, name)
+	require.NoError(t, afero.WriteFile(fs, srcPath, content, mode))
+	require.NoError(t, fs.Chmod(srcPath, mode))
+
+	storeDir := filepath.Join(tempDir, "remote")
+	dbPath := filepath.Join(tempDir, "index.sqlite")
+
+	storer, err := storage.NewFileStorer(storeDir)
+	require.NoError(t, err)
+
+	cfg := &config.Config{
+		AgeRecipients: []string{
+			"age1ezrjmfpwsc95svdg0y54mums3zevgzu0x0ecq2f7tp8a05gl0sjq9q9wjg",
+		},
+		AgeSecretKey: "AGE-SECRET-KEY-19CR5YSFW59HM4TLD6GXVEDMZFTVVF7PPHKU" +
+			"T68TXSFPK7APHXA2QS2NJA5",
+		CompressionLevel: 3,
+		Hostname:         "test-host",
+		BlobSizeLimit:    config.Size(5 * 1024 * 1024),
+	}
+
+	db, err := database.New(ctx, dbPath)
+	require.NoError(t, err)
+
+	repos := database.NewRepositories(db)
+
+	sm := snapshot.NewSnapshotManager(snapshot.SnapshotManagerParams{
+		Repos:   repos,
+		Storage: storer,
+		Config:  cfg,
+	})
+	sm.SetFilesystem(fs)
+
+	scanner := snapshot.NewScanner(snapshot.ScannerConfig{
+		FS:               fs,
+		Storage:          storer,
+		ChunkSize:        4 * 1024 * 1024,
+		MaxBlobSize:      5 * 1024 * 1024,
+		CompressionLevel: cfg.CompressionLevel,
+		AgeRecipients:    cfg.AgeRecipients,
+		Repositories:     repos,
+	})
+
+	snapshotID, err := sm.CreateSnapshotWithName(
+		ctx, cfg.Hostname, "perms", "test-version", "test-git")
+	require.NoError(t, err)
+
+	_, err = scanner.Scan(ctx, dataDir, snapshotID)
+	require.NoError(t, err)
+
+	require.NoError(t, sm.CompleteSnapshot(ctx, snapshotID))
+	require.NoError(t, sm.ExportSnapshotMetadata(ctx, dbPath, snapshotID))
+	require.NoError(t, db.Close())
+
+	return cfg, storer, snapshotID, srcPath
+}
+
+// restoredPathFor returns where backupOneFile's source lands under a
+// restore target: restore recreates each file at its original absolute
+// path beneath TargetDir.
+func restoredPathFor(restoreDir, srcPath string) string {
+	return filepath.Join(restoreDir, srcPath)
+}
+
+// withUmask022 forces the process umask to 022 for the duration of a
+// test, so the difference between a 0600 create and a default create is
+// observable. Restored serially (no t.Parallel) so it does not race
+// other tests.
+func withUmask022(t *testing.T) {
+	t.Helper()
+
+	old := syscall.Umask(0o022)
+
+	t.Cleanup(func() { syscall.Umask(old) })
+}
+
+// TestRestoreCreatesFileNeverWiderThanStoredMode checks that a file with
+// a restrictive stored mode (0600) is never observable with a wider mode
+// while its content is being written, and ends at its stored mode.
+//
+//nolint:paralleltest // sets the process umask; must run serially
+func TestRestoreCreatesFileNeverWiderThanStoredMode(t *testing.T) {
+	log.Initialize(log.Config{})
+	withUmask022(t)
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	ctx := context.Background()
+
+	content := randomBytes(t, 4096)
+
+	cfg, storer, snapshotID, srcPath := backupOneFile(
+		ctx, t, fs, tempDir, "secret.bin", content, 0o600)
+
+	restoreDir := filepath.Join(tempDir, "restored")
+	spy := &modeSpyFs{Fs: fs, watch: "secret.bin"}
+
+	v := newRestoreVaultik(ctx, cfg, storer, spy)
+	require.NoError(t, v.Restore(&RestoreOptions{
+		SnapshotID: snapshotID,
+		TargetDir:  restoreDir,
+	}))
+
+	spy.mu.Lock()
+	observed := append([]os.FileMode(nil), spy.writeModes...)
+	spy.mu.Unlock()
+
+	require.NotEmpty(t, observed,
+		"spy never saw the output file being written")
+
+	for _, m := range observed {
+		assert.Equalf(t, os.FileMode(0o600), m,
+			"file was observable at mode %o during write; must be 0600", m)
+	}
+
+	// The stored mode is applied after the content is written.
+	info, err := fs.Stat(restoredPathFor(restoreDir, srcPath))
+	require.NoError(t, err)
+	assert.Equal(t, os.FileMode(0o600), info.Mode().Perm())
+
+	got, err := afero.ReadFile(fs, restoredPathFor(restoreDir, srcPath))
+	require.NoError(t, err)
+	require.True(t, bytes.Equal(got, content))
+}
+
+// TestRestoreRemovesPartialFileOnWriteFailure checks that a file whose
+// content write fails is not left behind.
+//
+//nolint:paralleltest // sets the process umask; must run serially
+func TestRestoreRemovesPartialFileOnWriteFailure(t *testing.T) {
+	log.Initialize(log.Config{})
+	withUmask022(t)
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	ctx := context.Background()
+
+	cfg, storer, snapshotID, srcPath := backupOneFile(
+		ctx, t, fs, tempDir, "doomed.bin", randomBytes(t, 4096), 0o600)
+
+	restoreDir := filepath.Join(tempDir, "restored")
+	spy := &modeSpyFs{Fs: fs, watch: "doomed.bin", failWrite: true}
+
+	v := newRestoreVaultik(ctx, cfg, storer, spy)
+	err := v.Restore(&RestoreOptions{
+		SnapshotID: snapshotID,
+		TargetDir:  restoreDir,
+	})
+	require.Error(t, err, "restore should fail when the write fails")
+
+	exists, err := afero.Exists(fs, restoredPathFor(restoreDir, srcPath))
+	require.NoError(t, err)
+	assert.False(t, exists, "partial file must be removed after a failed write")
+}
+
+// TestVerifyRejectsTrailingBytes checks that --verify fails a restored
+// file that has bytes past its last chunk.
+//
+//nolint:paralleltest // sets the process umask; must run serially
+func TestVerifyRejectsTrailingBytes(t *testing.T) {
+	log.Initialize(log.Config{})
+	withUmask022(t)
+
+	fs := afero.NewOsFs()
+	tempDir := t.TempDir()
+	ctx := context.Background()
+
+	cfg, storer, snapshotID, _ := backupOneFile(
+		ctx, t, fs, tempDir, "padded.bin", randomBytes(t, 4096), 0o600)
+
+	restoreDir := filepath.Join(tempDir, "restored")
+	// Append one byte to the file as it is written, so its content still
+	// matches the stored chunks but it is one byte too long.
+	spy := &modeSpyFs{Fs: fs, watch: "padded.bin", trailing: 1}
+
+	v := newRestoreVaultik(ctx, cfg, storer, spy)
+	err := v.Restore(&RestoreOptions{
+		SnapshotID: snapshotID,
+		TargetDir:  restoreDir,
+		Verify:     true,
+	})
+	require.Error(t, err, "verify should fail on a file with trailing bytes")
+	assert.ErrorIs(t, err, errFilesFailedVerify)
+}
+
+// newRestoreVaultik builds a Vaultik wired for a restore-only test.
+func newRestoreVaultik(
+	ctx context.Context, cfg *config.Config, storer storage.Storer, fs afero.Fs,
+) *Vaultik {
+	v := &Vaultik{
+		Config:  cfg,
+		Storage: storer,
+		Fs:      fs,
+		Stdout:  io.Discard,
+		Stderr:  io.Discard,
+		UI:      ui.NewWithColor(io.Discard, false),
+	}
+	v.SetContext(ctx)
+
+	return v
+}

internal/vaultik/restore_plan.go

@@ -171,10 +171,13 @@ func (p *restorePlan) finishFile(fileID types.FileID) {
 // downloaded next, after which it — together with any other pending
 // files whose blob sets become empty — moves to the ready queue.
 //
-// The zero FileID return means nothing is pending.
+// The second return value is false when no file needs a download, so a
-func (p *restorePlan) pickNextDownload() types.FileID {
+// genuine file carrying the nil UUID is picked rather than mistaken for
+// "nothing left".
+func (p *restorePlan) pickNextDownload() (types.FileID, bool) {
 	var best types.FileID
 
+	found := false
 	bestCount := math.MaxInt
 
 	var bestID string
@@ -188,14 +191,15 @@ func (p *restorePlan) pickNextDownload() types.FileID {
 		}
 
 		idStr := id.String()
-		if n < bestCount || (n == bestCount && (best.IsZero() || idStr < bestID)) {
+		if !found || n < bestCount || (n == bestCount && idStr < bestID) {
 			best = id
+			found = true
 			bestCount = n
 			bestID = idStr
 		}
 	}
 
-	return best
+	return best, found
 }
 
 // blobsNeeded returns the uncached blob hashes for fileID in any order.

internal/vaultik/restore_plan_test.go

@@ -0,0 +1,88 @@
+package vaultik //nolint:testpackage // inspects unexported restore plan internals
+
+import (
+	"context"
+	"math"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/database"
+	"sneak.berlin/go/vaultik/internal/types"
+)
+
+// TestPickNextDownloadReturnsNilUUIDFile proves a genuine pending file
+// carrying the nil UUID is picked for download rather than mistaken for
+// "nothing left" — the bug that could abandon every remaining file.
+func TestPickNextDownloadReturnsNilUUIDFile(t *testing.T) {
+	t.Parallel()
+
+	var nilID types.FileID // zero value is the nil UUID
+
+	plan := &restorePlan{
+		fileBlobs: map[types.FileID]map[string]struct{}{
+			nilID: {"blobhash": {}},
+		},
+	}
+
+	id, ok := plan.pickNextDownload()
+	require.True(t, ok,
+		"pickNextDownload treated a pending nil-UUID file as nothing to do")
+	require.True(t, id.IsZero(), "expected the nil-UUID file to be picked")
+}
+
+// TestPickNextDownloadEmptyPlan confirms the second return value is false
+// only when no file needs a download.
+func TestPickNextDownloadEmptyPlan(t *testing.T) {
+	t.Parallel()
+
+	plan := &restorePlan{
+		fileBlobs: map[types.FileID]map[string]struct{}{},
+	}
+
+	_, ok := plan.pickNextDownload()
+	require.False(t, ok, "pickNextDownload reported work on an empty plan")
+}
+
+// TestRunRestoreLoopFailsOnAbandonedFiles proves the loop returns an
+// error rather than silent success when files remain pending after it
+// can make no further progress.
+func TestRunRestoreLoopFailsOnAbandonedFiles(t *testing.T) {
+	t.Parallel()
+
+	ctx := context.Background()
+
+	db, err := database.NewTestDB()
+	require.NoError(t, err)
+
+	t.Cleanup(func() { _ = db.Close() })
+
+	repos := database.NewRepositories(db)
+
+	cache, err := newBlobDiskCache(math.MaxInt64)
+	require.NoError(t, err)
+
+	t.Cleanup(func() { _ = cache.Close() })
+
+	v := &Vaultik{ctx: ctx}
+	session := &restoreSession{
+		v:       v,
+		ctx:     ctx,
+		repos:   repos,
+		sweeper: newRestoreSweeper(ctx, repos, cache, 1),
+		result:  &RestoreResult{},
+	}
+
+	// A file that is still pending but whose uncached-blob set is empty
+	// and which was never queued as ready: the loop can neither restore
+	// nor download it. This is the abandonment the guard must catch.
+	var stuck types.FileID
+
+	plan := &restorePlan{
+		fileBlobs: map[types.FileID]map[string]struct{}{stuck: {}},
+		blobFiles: map[string]map[types.FileID]struct{}{},
+		cached:    map[string]struct{}{},
+	}
+
+	err = v.runRestoreLoop(session, plan, map[types.FileID]*database.File{}, 0)
+	require.ErrorIs(t, err, errRestoreIncomplete)
+}

internal/vaultik/restore_snapshotdb_test.go

@@ -0,0 +1,73 @@
+package vaultik //nolint:testpackage // inspects unexported snapshot-db materialization
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/spf13/afero"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/database"
+)
+
+// genuineSnapshotDBBytes returns the on-disk bytes of a real snapshot
+// database (the full schema applied).
+func genuineSnapshotDBBytes(t *testing.T) []byte {
+	t.Helper()
+
+	path := filepath.Join(t.TempDir(), "snapshot.db")
+
+	db, err := database.New(context.Background(), path)
+	require.NoError(t, err)
+	require.NoError(t, db.Close())
+
+	data, err := os.ReadFile(path) //nolint:gosec // G304: test-controlled temp path
+	require.NoError(t, err)
+
+	return data
+}
+
+// TestMaterializeSnapshotDBPrivateDir proves the decrypted database lands
+// in a private (0700) directory and opens read-only.
+func TestMaterializeSnapshotDBPrivateDir(t *testing.T) {
+	dbData := genuineSnapshotDBBytes(t)
+
+	t.Setenv("TMPDIR", t.TempDir())
+
+	v := &Vaultik{ctx: context.Background(), Fs: afero.NewOsFs()}
+
+	db, dir, err := v.materializeSnapshotDB(dbData)
+	require.NoError(t, err)
+
+	t.Cleanup(func() {
+		_ = db.Close()
+		_ = os.RemoveAll(dir)
+	})
+
+	info, err := os.Stat(dir)
+	require.NoError(t, err)
+	require.Equal(t, os.FileMode(0o700), info.Mode().Perm(),
+		"snapshot database directory must not be world-readable")
+
+	_, err = db.Conn().ExecContext(context.Background(),
+		"CREATE TABLE probe_readonly (x)")
+	require.Error(t, err, "materialized snapshot database must be read-only")
+}
+
+// TestMaterializeSnapshotDBRemovesDirOnOpenFailure proves a failed open
+// leaves no temp directory behind.
+func TestMaterializeSnapshotDBRemovesDirOnOpenFailure(t *testing.T) {
+	base := t.TempDir()
+
+	t.Setenv("TMPDIR", base)
+
+	v := &Vaultik{ctx: context.Background(), Fs: afero.NewOsFs()}
+
+	_, _, err := v.materializeSnapshotDB([]byte("this is not a sqlite database"))
+	require.Error(t, err)
+
+	entries, rerr := os.ReadDir(base)
+	require.NoError(t, rerr)
+	require.Empty(t, entries, "temp directory left behind after open failure")
+}

internal/vaultik/snapshot.go

@@ -8,6 +8,7 @@ import (
 	"path/filepath"
 	"regexp"
 	"sort"
+	"strconv"
 	"strings"
 	"time"
 
@@ -55,8 +56,8 @@ func (v *Vaultik) CreateSnapshot(opts *SnapshotCreateOptions) error {
 		return err
 	}
 
-	// Clean up incomplete snapshots FIRST, before any scanning
+	// Clean up incomplete snapshots FIRST, before any scanning.
-	// This is critical for data safety - see CleanupIncompleteSnapshots for details
+	// This is critical for data safety; PruneDatabase below does it.
 	hostname := v.Config.Hostname
 	if hostname == "" {
 		hostname, _ = os.Hostname()
@@ -669,9 +670,11 @@ func (v *Vaultik) VerifySnapshotWithOptions(
 
 	v.printVerifyHeader(snapshotID, opts)
 
-	// Download and parse manifest. The caller supplies a human
+	// Resolve the identifier to the snapshot's remote key and download the
-	// snapshot ID; we hash it to address remote storage.
+	// manifest. A human ID is hashed; a remote key (or its abbreviation,
-	manifest, err := v.downloadManifestByKey(snapshot.RemoteSnapshotKey(snapshotID))
+	// as printed for a remote-only snapshot) is used as-is, so a host with
+	// no local index can verify a snapshot it can only see on the store.
+	manifest, err := v.resolveAndDownloadManifest(snapshotID)
 	if err != nil {
 		if opts.JSON {
 			result.Status = verifyStatusFailed
@@ -932,29 +935,23 @@ func (v *Vaultik) downloadManifestByKey(remoteKey string) (*snapshot.Manifest, e
 func (v *Vaultik) syncWithRemote() error {
 	log.Info("Syncing with remote snapshots")
 
-	// Get all remote snapshot IDs
+	// Remote metadata lives under metadata/<remote-key>/, where the
-	remoteSnapshots := make(map[string]bool)
+	// directory name is snapshot.RemoteSnapshotKey(id), not the human
-	objectCh := v.Storage.ListStream(v.ctx, "metadata/")
+	// snapshot ID. Compare each local row's hashed key against that set
-
+	// so a row still backed by remote metadata is kept. Comparing human
-	for object := range objectCh {
+	// IDs against the hashed directory names matches nothing and deletes
-		if object.Err != nil {
+	// every local snapshot record (issue #160).
-			return fmt.Errorf("listing remote snapshots: %w", object.Err)
+	remoteKeys, err := v.listAllRemoteSnapshotKeys()
-		}
+	if err != nil {
-
+		return fmt.Errorf("listing remote snapshots: %w", err)
-		// Extract snapshot ID from paths like metadata/hostname-20240115-143052Z/
-		parts := strings.Split(object.Key, "/")
-		if len(parts) >= minSnapshotIDParts &&
-			parts[0] == metadataDirName && parts[1] != "" {
-			// Skip macOS resource fork files (._*) and other hidden files
-			if strings.HasPrefix(parts[1], ".") {
-				continue
-			}
-
-			remoteSnapshots[parts[1]] = true
-		}
 	}
 
-	log.Debug("Found remote snapshots", "count", len(remoteSnapshots))
+	remoteKeySet := make(map[string]bool, len(remoteKeys))
+	for _, k := range remoteKeys {
+		remoteKeySet[k] = true
+	}
+
+	log.Debug("Found remote snapshots", "count", len(remoteKeySet))
 
 	// Get all local snapshots (use a high limit to get all)
 	localSnapshots, err := v.Repositories.Snapshots.ListRecent(v.ctx, listRecentLimit)
@@ -962,12 +959,12 @@ func (v *Vaultik) syncWithRemote() error {
 		return fmt.Errorf("listing local snapshots: %w", err)
 	}
 
-	// Remove local snapshots that don't exist remotely
+	// Remove local snapshots whose metadata is absent from the remote.
 	removedCount := 0
 
 	for _, snap := range localSnapshots {
 		snapshotIDStr := snap.ID.String()
-		if !remoteSnapshots[snapshotIDStr] {
+		if !remoteKeySet[snapshot.RemoteSnapshotKey(snapshotIDStr)] {
 			log.Info("Removing local snapshot not found in remote",
 				"snapshot_id", snap.ID)
 
@@ -1540,12 +1537,17 @@ func (v *Vaultik) outputRemoveJSON(result *RemoveResult) error {
 	return encoder.Encode(result)
 }
 
-// PruneResult contains statistics about the prune operation
+// PruneResult contains statistics about the prune operation.
+// SnapshotsDeleted counts snapshots actually deleted. FilesDeleted,
+// ChunksDeleted, and BlobsDeleted are derived from before/after row
+// counts of the local index; each is nil when a count could not be read,
+// so an unreadable count is reported as unknown rather than silently
+// as 0.
 type PruneResult struct {
 	SnapshotsDeleted int64
-	FilesDeleted     int64
+	FilesDeleted     *int64
-	ChunksDeleted    int64
+	ChunksDeleted    *int64
-	BlobsDeleted     int64
+	BlobsDeleted     *int64
 }
 
 // PruneDatabase removes incomplete snapshots and orphaned files, chunks,
@@ -1560,7 +1562,7 @@ func (v *Vaultik) PruneDatabase() (*PruneResult, error) {
 	result := &PruneResult{}
 
 	// Snapshot counts before deletion of incompletes.
-	snapshotCountBefore, _ := v.getTableCount("snapshots")
+	snapshotCountBefore := v.tableCountForReport("snapshots")
 
 	// First, delete any incomplete snapshots
 	incompleteSnapshots, err := v.Repositories.Snapshots.GetIncompleteSnapshots(v.ctx)
@@ -1575,9 +1577,9 @@ func (v *Vaultik) PruneDatabase() (*PruneResult, error) {
 	}
 
 	// Get counts before cleanup for reporting
-	fileCountBefore, _ := v.getTableCount("files")
+	fileCountBefore := v.tableCountForReport("files")
-	chunkCountBefore, _ := v.getTableCount("chunks")
+	chunkCountBefore := v.tableCountForReport("chunks")
-	blobCountBefore, _ := v.getTableCount("blobs")
+	blobCountBefore := v.tableCountForReport("blobs")
 
 	// Run the cleanup
 	err = v.SnapshotManager.CleanupOrphanedData(v.ctx)
@@ -1586,36 +1588,83 @@ func (v *Vaultik) PruneDatabase() (*PruneResult, error) {
 	}
 
 	// Get counts after cleanup
-	fileCountAfter, _ := v.getTableCount("files")
+	fileCountAfter := v.tableCountForReport("files")
-	chunkCountAfter, _ := v.getTableCount("chunks")
+	chunkCountAfter := v.tableCountForReport("chunks")
-	blobCountAfter, _ := v.getTableCount("blobs")
+	blobCountAfter := v.tableCountForReport("blobs")
 
-	result.FilesDeleted = fileCountBefore - fileCountAfter
+	result.FilesDeleted = countDiff(fileCountBefore, fileCountAfter)
-	result.ChunksDeleted = chunkCountBefore - chunkCountAfter
+	result.ChunksDeleted = countDiff(chunkCountBefore, chunkCountAfter)
-	result.BlobsDeleted = blobCountBefore - blobCountAfter
+	result.BlobsDeleted = countDiff(blobCountBefore, blobCountAfter)
 
 	log.Info("Local database prune complete",
 		"incomplete_snapshots", result.SnapshotsDeleted,
-		"orphaned_files", result.FilesDeleted,
+		"orphaned_files", countText(result.FilesDeleted),
-		"orphaned_chunks", result.ChunksDeleted,
+		"orphaned_chunks", countText(result.ChunksDeleted),
-		"orphaned_blobs", result.BlobsDeleted,
+		"orphaned_blobs", countText(result.BlobsDeleted),
 	)
 
-	snapshotCountAfter := snapshotCountBefore - result.SnapshotsDeleted
+	// Snapshots remaining after removing the incomplete ones; unknown if
+	// the pre-prune snapshot count could not be read.
+	snapshotsRemain := countDiff(snapshotCountBefore, &result.SnapshotsDeleted)
 
 	v.UI.Completef("Pruned local index database.")
-	v.UI.Detailf("Incomplete snapshots: %d removed (%d remain).",
+	v.UI.Detailf("Incomplete snapshots: %s removed (%s remain).",
-		result.SnapshotsDeleted, snapshotCountAfter)
+		countText(&result.SnapshotsDeleted), countText(snapshotsRemain))
-	v.UI.Detailf("Orphaned files: %d removed (%d remain).",
+	v.UI.Detailf("Orphaned files: %s removed (%s remain).",
-		result.FilesDeleted, fileCountAfter)
+		countText(result.FilesDeleted), countText(fileCountAfter))
-	v.UI.Detailf("Orphaned chunks: %d removed (%d remain).",
+	v.UI.Detailf("Orphaned chunks: %s removed (%s remain).",
-		result.ChunksDeleted, chunkCountAfter)
+		countText(result.ChunksDeleted), countText(chunkCountAfter))
-	v.UI.Detailf("Orphaned blobs: %d removed (%d remain).",
+	v.UI.Detailf("Orphaned blobs: %s removed (%s remain).",
-		result.BlobsDeleted, blobCountAfter)
+		countText(result.BlobsDeleted), countText(blobCountAfter))
 
 	return result, nil
 }
 
+// countUnknown is what a count reads as when its query could not be run,
+// distinct from "0", which means the table really was empty.
+const countUnknown = "unknown"
+
+// tableCountForReport returns the row count of a table for the prune
+// summary, or nil if the count could not be read. A read failure is
+// logged at warn — visible even under --json, which routes warnings to
+// stderr — and then rendered as unknown rather than silently becoming 0,
+// so a broken query is a visible failure instead of a plausible wrong
+// number.
+func (v *Vaultik) tableCountForReport(tableName string) *int64 {
+	count, err := v.getTableCount(tableName)
+	if err != nil {
+		log.Warn("could not read table row count for prune summary",
+			"table", tableName, "error", err)
+
+		return nil
+	}
+
+	return &count
+}
+
+// countDiff returns before-after, or nil if either count is unknown so
+// that an unreadable count does not collapse into a plausible delta.
+func countDiff(before, after *int64) *int64 {
+	if before == nil || after == nil {
+		return nil
+	}
+
+	diff := *before - *after
+
+	return &diff
+}
+
+// countText renders a count that may be unknown: nil (the read failed)
+// becomes "unknown", never "0", so a reader can tell an empty table from
+// one that could not be queried.
+func countText(count *int64) string {
+	if count == nil {
+		return countUnknown
+	}
+
+	return strconv.FormatInt(*count, 10)
+}
+
 // validTableNameRe matches table names containing only lowercase
 // alphanumeric characters and underscores.
 var validTableNameRe = regexp.MustCompile(`^[a-z0-9_]+$`)

internal/vaultik/snapshot_identifier.go

@@ -0,0 +1,101 @@
+package vaultik
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+
+	"sneak.berlin/go/vaultik/internal/snapshot"
+)
+
+// remoteKeyHexLen is the length of a full remote snapshot key: a SHA256
+// digest rendered as lowercase hex.
+const remoteKeyHexLen = 64
+
+// Sentinel errors for resolving a snapshot identifier against the store.
+var (
+	errSnapshotKeyNotFound = errors.New(
+		"no snapshot on the destination store matches this identifier")
+	errSnapshotKeyAmbiguous = errors.New(
+		"identifier matches more than one snapshot on the destination store")
+)
+
+// resolveSnapshotRemoteKey turns a snapshot identifier supplied on the
+// command line into the remote key that names the snapshot's metadata
+// directory on the destination store. Every remote path a restore or
+// verify reads is built from that key.
+//
+// Two forms are accepted, matching the two things a host can know:
+//
+//   - A human snapshot ID (hostname_name_timestamp), which a host holding
+//     the local index has. It is hashed to its remote key; the store is
+//     not consulted.
+//   - A remote key, or the leading part of one, which is all a host with
+//     no local index can know — it is exactly what `snapshot list` prints
+//     for a remote-only snapshot (see formatRemoteOnlyID). It is resolved
+//     against the destination store's metadata listing; an identifier that
+//     matches no snapshot, or more than one, is an error.
+//
+// The two are told apart by shape: a remote key is lowercase hex, and a
+// human snapshot ID never is (it carries a hostname, underscores, and an
+// RFC3339 timestamp).
+func (v *Vaultik) resolveSnapshotRemoteKey(identifier string) (string, error) {
+	if !isRemoteKeyOrPrefix(identifier) {
+		return snapshot.RemoteSnapshotKey(identifier), nil
+	}
+
+	keys, err := v.listAllRemoteSnapshotKeys()
+	if err != nil {
+		return "", fmt.Errorf(
+			"listing destination store to resolve %q: %w", identifier, err)
+	}
+
+	var matches []string
+
+	for _, key := range keys {
+		if strings.HasPrefix(key, identifier) {
+			matches = append(matches, key)
+		}
+	}
+
+	switch len(matches) {
+	case 1:
+		return matches[0], nil
+	case 0:
+		return "", fmt.Errorf("%w: %s", errSnapshotKeyNotFound, identifier)
+	default:
+		return "", fmt.Errorf("%w: %s (%d matches)",
+			errSnapshotKeyAmbiguous, identifier, len(matches))
+	}
+}
+
+// resolveAndDownloadManifest resolves a snapshot identifier to its remote
+// key (see resolveSnapshotRemoteKey) and downloads that snapshot's
+// manifest.
+func (v *Vaultik) resolveAndDownloadManifest(
+	identifier string,
+) (*snapshot.Manifest, error) {
+	remoteKey, err := v.resolveSnapshotRemoteKey(identifier)
+	if err != nil {
+		return nil, err
+	}
+
+	return v.downloadManifestByKey(remoteKey)
+}
+
+// isRemoteKeyOrPrefix reports whether s is a full remote key or the
+// leading part of one: 1 to 64 lowercase hex characters. A human snapshot
+// ID is never all hex, so this shape test is enough to tell the two apart.
+func isRemoteKeyOrPrefix(s string) bool {
+	if s == "" || len(s) > remoteKeyHexLen {
+		return false
+	}
+
+	for _, r := range s {
+		if (r < '0' || r > '9') && (r < 'a' || r > 'f') {
+			return false
+		}
+	}
+
+	return true
+}

internal/vaultik/vaultik.go

@@ -5,7 +5,6 @@ package vaultik
 import (
 	"bytes"
 	"context"
-	"errors"
 	"fmt"
 	"io"
 	"os"
@@ -13,7 +12,6 @@ import (
 	"github.com/spf13/afero"
 	"go.uber.org/fx"
 	"sneak.berlin/go/vaultik/internal/config"
-	"sneak.berlin/go/vaultik/internal/crypto"
 	"sneak.berlin/go/vaultik/internal/database"
 	"sneak.berlin/go/vaultik/internal/globals"
 	"sneak.berlin/go/vaultik/internal/snapshot"
@@ -21,12 +19,6 @@ import (
 	"sneak.berlin/go/vaultik/internal/ui"
 )
 
-// Sentinel errors for misconfigured encryption settings.
-var (
-	errNoAgeRecipients = errors.New("no age recipients configured")
-	errNoAgeSecretKey  = errors.New("no age secret key configured")
-)
-
 // Vaultik contains all dependencies needed for vaultik operations
 type Vaultik struct {
 	Globals         *globals.Globals
@@ -136,31 +128,45 @@ func (v *Vaultik) Cancel() {
 	v.cancel()
 }
 
+// StartOperation runs fn in its own goroutine and returns a stop
+// function. fn is the command being run (a restore, verify, prune, and
+// so on); it observes cancellation through the Vaultik context and
+// removes its decrypted scratch files (the blob cache and the temporary
+// snapshot database) from the temp directory as it unwinds.
+//
+// Calling stop cancels the Vaultik context and then blocks until fn has
+// returned — so that unwinding, and the cleanup it does, completes
+// before the caller proceeds — or until the passed context is done,
+// whichever comes first. It reports whether fn returned before that
+// deadline. A signal-driven shutdown must call stop before the process
+// exits; otherwise the process can exit mid-operation and leave
+// decrypted data behind.
+func (v *Vaultik) StartOperation(fn func()) func(context.Context) bool {
+	done := make(chan struct{})
+
+	go func() {
+		defer close(done)
+
+		fn()
+	}()
+
+	return func(ctx context.Context) bool {
+		v.Cancel()
+
+		select {
+		case <-done:
+			return true
+		case <-ctx.Done():
+			return false
+		}
+	}
+}
+
 // CanDecrypt returns true if this Vaultik instance has decryption capabilities
 func (v *Vaultik) CanDecrypt() bool {
 	return v.Config.AgeSecretKey != ""
 }
 
-// GetEncryptor creates a new Encryptor instance based on the configured age recipients
-// Returns an error if no recipients are configured
-func (v *Vaultik) GetEncryptor() (*crypto.Encryptor, error) {
-	if len(v.Config.AgeRecipients) == 0 {
-		return nil, errNoAgeRecipients
-	}
-
-	return crypto.NewEncryptor(v.Config.AgeRecipients)
-}
-
-// GetDecryptor creates a new Decryptor instance based on the configured age secret key
-// Returns an error if no secret key is configured
-func (v *Vaultik) GetDecryptor() (*crypto.Decryptor, error) {
-	if v.Config.AgeSecretKey == "" {
-		return nil, errNoAgeSecretKey
-	}
-
-	return crypto.NewDecryptor(v.Config.AgeSecretKey)
-}
-
 // GetFilesystem returns the filesystem instance used by Vaultik
 //
 //nolint:ireturn // afero.Fs is the filesystem abstraction by design

internal/vaultik/verify.go

@@ -6,15 +6,15 @@ import (
 	"encoding/hex"
 	"errors"
 	"fmt"
-	"hash"
 	"io"
 	"os"
+	"path/filepath"
 	"time"
 
-	"github.com/klauspost/compress/zstd"
+	"filippo.io/age"
 
-	// Blank import registers the pure-Go sqlite driver for database/sql.
+	"sneak.berlin/go/vaultik/internal/blobgen"
-	_ "modernc.org/sqlite"
+	"sneak.berlin/go/vaultik/internal/database"
 	"sneak.berlin/go/vaultik/internal/log"
 	"sneak.berlin/go/vaultik/internal/snapshot"
 )
@@ -28,8 +28,10 @@ var (
 	errChunkHashMismatch = errors.New("chunk hash mismatch")
 	errTrailingBlobData  = errors.New(
 		"blob has unexpected trailing bytes not covered by chunk list")
-	errManifestExtraBlob = errors.New("manifest contains blob not in database")
+	errManifestExtraBlob   = errors.New("manifest contains blob not in database")
-	errBlobSizeMismatch  = errors.New("blob size mismatch")
+	errManifestMissingBlob = errors.New(
+		"manifest omits blob present in database")
+	errBlobSizeMismatch = errors.New("blob size mismatch")
 )
 
 // verifyStatusFailed is the JSON status value for a failed verification.
@@ -86,13 +88,22 @@ func (v *Vaultik) RunDeepVerify(snapshotID string, opts *VerifyOptions) error {
 			errSecretKeyRequired.Error(), errSecretKeyRequired)
 	}
 
+	// Parse the age secret key once, the same way restore does, and reuse
+	// the identity for the database and every blob.
+	identity, err := v.prepareRestoreIdentity()
+	if err != nil {
+		return v.deepVerifyFailure(result, opts,
+			fmt.Sprintf("parsing age secret key: %v", err), err)
+	}
+
 	log.Info("Starting snapshot verification", "snapshot_id", snapshotID, "mode", "deep")
 
 	if !opts.JSON {
 		v.stdoutf("Deep verification of snapshot: %s\n\n", snapshotID)
 	}
 
-	manifest, tempDB, dbBlobs, err := v.loadVerificationData(snapshotID, opts, result)
+	manifest, tempDB, dbBlobs, err := v.loadVerificationData(
+		snapshotID, opts, result, identity)
 	if err != nil {
 		return err
 	}
@@ -112,7 +123,8 @@ func (v *Vaultik) RunDeepVerify(snapshotID string, opts *VerifyOptions) error {
 
 	result.TotalSize = totalSize
 
-	err = v.runVerificationSteps(manifest, dbBlobs, tempDB, opts, result, totalSize)
+	err = v.runVerificationSteps(
+		manifest, dbBlobs, tempDB, opts, result, totalSize, identity)
 	if err != nil {
 		return err
 	}
@@ -137,9 +149,17 @@ func (v *Vaultik) RunDeepVerify(snapshotID string, opts *VerifyOptions) error {
 // loadVerificationData downloads manifest, database, and blob list for verification
 func (v *Vaultik) loadVerificationData(
 	snapshotID string, opts *VerifyOptions, result *VerifyResult,
+	identity age.Identity,
 ) (*snapshot.Manifest, *tempDB, []snapshot.BlobInfo, error) {
-	// All remote paths use the hashed key derived from the human ID.
+	// Resolve the identifier to the snapshot's remote key. A human ID is
-	remoteKey := snapshot.RemoteSnapshotKey(snapshotID)
+	// hashed; a remote key (or its abbreviation, as printed for a
+	// remote-only snapshot) is used as-is, so a host with no local index
+	// can verify a snapshot it can only see on the store.
+	remoteKey, err := v.resolveSnapshotRemoteKey(snapshotID)
+	if err != nil {
+		return nil, nil, nil, v.deepVerifyFailure(result, opts,
+			fmt.Sprintf("resolving snapshot identifier: %v", err), err)
+	}
 
 	// Download manifest. downloadManifestByKey is the single reader for
 	// remote manifests; see its doc comment.
@@ -179,14 +199,14 @@ func (v *Vaultik) loadVerificationData(
 
 	defer func() { _ = dbReader.Close() }()
 
-	tdb, err := v.decryptAndLoadDatabase(dbReader)
+	tdb, err := v.decryptAndLoadDatabase(dbReader, identity)
 	if err != nil {
 		return nil, nil, nil, v.deepVerifyFailure(result, opts,
 			fmt.Sprintf("failed to decrypt database: %v", err),
 			fmt.Errorf("failed to decrypt database: %w", err))
 	}
 
-	dbBlobs, err := v.getBlobsFromDatabase(snapshotID, tdb.DB)
+	dbBlobs, err := v.getBlobsFromDatabase(tdb.db.Conn())
 	if err != nil {
 		_ = tdb.Close()
 
@@ -221,6 +241,7 @@ func (v *Vaultik) runVerificationSteps(
 	opts *VerifyOptions,
 	result *VerifyResult,
 	totalSize int64,
+	identity age.Identity,
 ) error {
 	if !opts.JSON {
 		v.stdoutf("Verifying manifest against database...\n")
@@ -247,7 +268,7 @@ func (v *Vaultik) runVerificationSteps(
 			len(dbBlobs), ubytes(totalSize))
 	}
 
-	err = v.performDeepVerificationFromDB(dbBlobs, tdb.DB, opts)
+	err = v.performDeepVerificationFromDB(dbBlobs, tdb.db.Conn(), opts, identity)
 	if err != nil {
 		return v.deepVerifyFailure(result, opts, err.Error(), err)
 	}
@@ -255,81 +276,92 @@ func (v *Vaultik) runVerificationSteps(
 	return nil
 }
 
-// tempDB wraps sql.DB with cleanup
+// tempDB is the downloaded snapshot database opened read-only for deep
+// verify, held in a private temp directory removed in full on Close.
 type tempDB struct {
-	*sql.DB
+	db      *database.DB
-
+	tempDir string
-	tempPath string
 }
 
 func (t *tempDB) Close() error {
-	err := t.DB.Close()
+	err := t.db.Close()
-	_ = os.Remove(t.tempPath)
+	// Remove the whole private directory so the decrypted database and
+	// any SQLite side files are gone on every path.
+	_ = os.RemoveAll(t.tempDir)
 
 	return err
 }
 
 // decryptAndLoadDatabase decrypts and loads the binary SQLite database
-// from the encrypted stream.
+// from the encrypted stream. It reads through the same blobgen reader restore
-func (v *Vaultik) decryptAndLoadDatabase(reader io.ReadCloser) (*tempDB, error) {
+// uses, streaming the decrypted, decompressed database to a temp file.
-	// Get decryptor
+func (v *Vaultik) decryptAndLoadDatabase(
-	decryptor, err := v.GetDecryptor()
+	reader io.ReadCloser, identity age.Identity,
+) (*tempDB, error) {
+	// Decrypt and decompress through the shared blobgen reader.
+	blobReader, err := blobgen.NewReader(reader, identity)
 	if err != nil {
-		return nil, fmt.Errorf("failed to get decryptor: %w", err)
+		return nil, fmt.Errorf("failed to create decryption reader: %w", err)
 	}
 
-	// Decrypt the stream
+	defer func() { _ = blobReader.Close() }()
-	decryptedReader, err := decryptor.DecryptStream(reader)
+
+	// Materialize the decrypted database inside a private (0700) temp
+	// directory so it is never world-readable, and remove the whole
+	// directory on any failure below.
+	tempDir, err := os.MkdirTemp("", "vaultik-verify-")
 	if err != nil {
-		return nil, fmt.Errorf("failed to decrypt database: %w", err)
+		return nil, fmt.Errorf("failed to create temp directory: %w", err)
 	}
 
-	// Decompress the binary database
+	success := false
-	decompressor, err := zstd.NewReader(decryptedReader)
-	if err != nil {
-		return nil, fmt.Errorf("failed to create decompressor: %w", err)
-	}
-	defer decompressor.Close()
 
-	// Create temporary file for the database
+	defer func() {
-	tempFile, err := os.CreateTemp("", "vaultik-verify-*.db")
+		if !success {
+			_ = os.RemoveAll(tempDir)
+		}
+	}()
+
+	dbPath := filepath.Join(tempDir, snapshotDBFilename)
+
+	//nolint:gosec // G304: dbPath is our MkdirTemp dir plus a constant filename
+	tempFile, err := os.OpenFile(
+		dbPath, os.O_CREATE|os.O_EXCL|os.O_WRONLY, restoreFileMode)
 	if err != nil {
 		return nil, fmt.Errorf("failed to create temp file: %w", err)
 	}
 
-	tempPath := tempFile.Name()
-
 	// Stream decompress directly to file
 	log.Info("Decompressing database...")
 
-	written, err := io.Copy(tempFile, decompressor)
+	written, err := io.Copy(tempFile, blobReader)
 	if err != nil {
 		_ = tempFile.Close()
-		_ = os.Remove(tempPath)
 
 		return nil, fmt.Errorf("failed to decompress database: %w", err)
 	}
 
-	_ = tempFile.Close()
+	err = tempFile.Close()
+	if err != nil {
+		return nil, fmt.Errorf("failed to close temp database file: %w", err)
+	}
 
 	log.Info("Database decompressed", "size", ubytes(written))
 
-	// Open the database
+	db, err := database.OpenReadOnly(v.ctx, dbPath)
-	db, err := sql.Open("sqlite", tempPath)
 	if err != nil {
-		_ = os.Remove(tempPath)
-
 		return nil, fmt.Errorf("failed to open database: %w", err)
 	}
 
-	return &tempDB{
+	success = true
-		DB:       db,
+
-		tempPath: tempPath,
+	return &tempDB{db: db, tempDir: tempDir}, nil
-	}, nil
 }
 
 // verifyBlob downloads and verifies a single blob
-func (v *Vaultik) verifyBlob(blobInfo snapshot.BlobInfo, db *sql.DB) error {
+func (v *Vaultik) verifyBlob(
+	blobInfo snapshot.BlobInfo, db *sql.DB, identity age.Identity,
+) error {
 	// Download blob using shared fetch method
 	reader, _, err := v.FetchBlob(v.ctx, blobInfo.Hash, blobInfo.CompressedSize)
 	if err != nil {
@@ -338,38 +370,23 @@ func (v *Vaultik) verifyBlob(blobInfo snapshot.BlobInfo, db *sql.DB) error {
 
 	defer func() { _ = reader.Close() }()
 
-	// Get decryptor
+	// Decrypt and decompress through the shared blobgen reader, which hashes
-	decryptor, err := v.GetDecryptor()
+	// the plaintext as it is read. A blob's hash — its remote name — is the
+	// double SHA-256 of that plaintext (see blobgen.DoubleSHA256), not of the
+	// encrypted bytes.
+	blobReader, err := blobgen.NewReader(reader, identity)
 	if err != nil {
-		return fmt.Errorf("failed to get decryptor: %w", err)
+		return fmt.Errorf("failed to create blob reader: %w", err)
 	}
 
-	// Decrypt blob
+	defer func() { _ = blobReader.Close() }()
-	decryptedReader, err := decryptor.DecryptStream(reader)
-	if err != nil {
-		return fmt.Errorf("failed to decrypt: %w", err)
-	}
 
-	// Decompress blob
+	chunkCount, err := v.verifyBlobChunks(db, blobInfo.Hash, blobReader)
-	decompressor, err := zstd.NewReader(decryptedReader)
-	if err != nil {
-		return fmt.Errorf("failed to decompress: %w", err)
-	}
-	defer decompressor.Close()
-
-	// A blob's hash — its remote name — is the double SHA256 of its
-	// decompressed plaintext (see blobgen.Writer.Sum256), not of the
-	// encrypted bytes. Hash the plaintext as chunk verification streams
-	// it, then compare on completion.
-	plaintextHasher := sha256.New()
-	hashedStream := io.TeeReader(decompressor, plaintextHasher)
-
-	chunkCount, err := v.verifyBlobChunks(db, blobInfo.Hash, hashedStream)
 	if err != nil {
 		return err
 	}
 
-	err = v.verifyBlobFinalIntegrity(hashedStream, plaintextHasher, blobInfo.Hash)
+	err = v.verifyBlobFinalIntegrity(blobReader, blobInfo.Hash)
 	if err != nil {
 		return err
 	}
@@ -475,11 +492,12 @@ func (v *Vaultik) verifyBlobChunks(
 // verifyBlobFinalIntegrity checks that no trailing data exists in the
 // decompressed stream and that the blob hash matches the expected value.
 func (v *Vaultik) verifyBlobFinalIntegrity(
-	plaintext io.Reader, plaintextHasher hash.Hash, expectedHash string,
+	blobReader *blobgen.Reader, expectedHash string,
 ) error {
 	// Verify no remaining data in blob - if the chunk list is accurate,
-	// the blob should be fully consumed.
+	// the blob should be fully consumed. Draining to EOF also completes the
-	remaining, err := io.Copy(io.Discard, plaintext)
+	// reader's plaintext hash.
+	remaining, err := io.Copy(io.Discard, blobReader)
 	if err != nil {
 		return fmt.Errorf("failed to check for remaining blob data: %w", err)
 	}
@@ -488,10 +506,9 @@ func (v *Vaultik) verifyBlobFinalIntegrity(
 		return fmt.Errorf("%w: %d bytes", errTrailingBlobData, remaining)
 	}
 
-	// The blob hash is the double SHA256 of its plaintext content.
+	// The blob hash is the double SHA-256 of its plaintext content.
-	firstHash := plaintextHasher.Sum(nil)
+	calculatedBlobHash := hex.EncodeToString(
-	secondHash := sha256.Sum256(firstHash)
+		blobgen.DoubleSHA256(blobReader.Sum256()))
-	calculatedBlobHash := hex.EncodeToString(secondHash[:])
 
 	if calculatedBlobHash != expectedHash {
 		return fmt.Errorf("%w: calculated %s, expected %s",
@@ -501,19 +518,21 @@ func (v *Vaultik) verifyBlobFinalIntegrity(
 	return nil
 }
 
-// getBlobsFromDatabase gets all blobs for the snapshot from the database
+// getBlobsFromDatabase gets all blobs for the snapshot from the database.
-func (v *Vaultik) getBlobsFromDatabase(
+//
-	snapshotID string, db *sql.DB,
+// The exported per-snapshot database holds exactly one snapshot's data
-) ([]snapshot.BlobInfo, error) {
+// (see cleanSnapshotDB), so every row in snapshot_blobs belongs to it.
+// We select them directly rather than filtering by the human snapshot ID,
+// which a host restoring from the store alone does not have.
+func (v *Vaultik) getBlobsFromDatabase(db *sql.DB) ([]snapshot.BlobInfo, error) {
 	query := `
 		SELECT b.blob_hash, b.compressed_size
 		FROM snapshot_blobs sb
 		JOIN blobs b ON sb.blob_hash = b.blob_hash
-		WHERE sb.snapshot_id = ?
 		ORDER BY b.blob_hash
 	`
 
-	rows, err := db.QueryContext(v.ctx, query, snapshotID)
+	rows, err := db.QueryContext(v.ctx, query)
 	if err != nil {
 		return nil, fmt.Errorf("failed to query snapshot blobs: %w", err)
 	}
@@ -566,16 +585,11 @@ func (v *Vaultik) verifyManifestAgainstDatabase(
 		manifestBlobMap[blob.Hash] = blob.CompressedSize
 	}
 
-	// Check counts match
+	// The manifest is the only blob list prune consults, so it must match
-	if len(dbBlobMap) != len(manifestBlobMap) {
+	// the database exactly. A blob in the manifest but not the database
-		log.Warn("Manifest blob count mismatch",
+	// points at a corrupt manifest; a blob in the database but omitted
-			"database_blobs", len(dbBlobMap),
+	// from the manifest would be pruned away while this snapshot still
-			"manifest_blobs", len(manifestBlobMap),
+	// needs it. Either divergence fails verification.
-		)
-		// This is a warning, not an error - database is authoritative
-	}
-
-	// Check each manifest blob exists in database with correct size
 	for hash, manifestSize := range manifestBlobMap {
 		dbSize, exists := dbBlobMap[hash]
 		if !exists {
@@ -589,6 +603,12 @@ func (v *Vaultik) verifyManifestAgainstDatabase(
 		}
 	}
 
+	for hash := range dbBlobMap {
+		if _, exists := manifestBlobMap[hash]; !exists {
+			return fmt.Errorf("%w: %s", errManifestMissingBlob, hash)
+		}
+	}
+
 	log.Info("✓ Manifest verified against database",
 		"manifest_blobs", len(manifestBlobMap),
 		"database_blobs", len(dbBlobMap),
@@ -638,6 +658,7 @@ func (v *Vaultik) verifyBlobExistenceFromDB(blobs []snapshot.BlobInfo) error {
 // each blob using the database as source.
 func (v *Vaultik) performDeepVerificationFromDB(
 	blobs []snapshot.BlobInfo, db *sql.DB, opts *VerifyOptions,
+	identity age.Identity,
 ) error {
 	// Calculate total bytes for ETA
 	var totalBytesExpected int64
@@ -655,7 +676,7 @@ func (v *Vaultik) performDeepVerificationFromDB(
 
 	for i, blobInfo := range blobs {
 		// Verify individual blob
-		err := v.verifyBlob(blobInfo, db)
+		err := v.verifyBlob(blobInfo, db, identity)
 		if err != nil {
 			return fmt.Errorf("blob %s verification failed: %w", blobInfo.Hash, err)
 		}

internal/vaultik/verify_manifest_test.go

@@ -0,0 +1,61 @@
+package vaultik //nolint:testpackage // calls unexported verifyManifestAgainstDatabase
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/vaultik/internal/log"
+	"sneak.berlin/go/vaultik/internal/snapshot"
+)
+
+// Blob hashes shared by the manifest-verification tests below.
+const (
+	manifestTestBlobA = "blob-a"
+	manifestTestBlobB = "blob-b"
+)
+
+// TestVerifyManifestAgainstDatabase_MissingBlobFails is the regression
+// guard for issue #157: deep verify must fail when the manifest omits a
+// blob the database records. The divergence used to be logged as a
+// warning while verification still returned ok, so an incomplete
+// manifest — the exact defect that lets prune later delete a needed blob
+// — passed unnoticed.
+func TestVerifyManifestAgainstDatabase_MissingBlobFails(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	v := &Vaultik{}
+
+	dbBlobs := []snapshot.BlobInfo{
+		{Hash: manifestTestBlobA, CompressedSize: 10},
+		{Hash: manifestTestBlobB, CompressedSize: 20},
+	}
+	manifest := &snapshot.Manifest{
+		Blobs: []snapshot.BlobInfo{
+			{Hash: manifestTestBlobA, CompressedSize: 10},
+		},
+	}
+
+	err := v.verifyManifestAgainstDatabase(manifest, dbBlobs)
+	require.Error(t, err, "verify must fail when the manifest omits a database blob")
+	assert.Contains(t, err.Error(), manifestTestBlobB)
+}
+
+// TestVerifyManifestAgainstDatabase_MatchingSetsPass keeps the other half
+// honest: identical blob sets still verify, so the check above cannot be
+// satisfied by failing everything.
+func TestVerifyManifestAgainstDatabase_MatchingSetsPass(t *testing.T) {
+	log.Initialize(log.Config{})
+	t.Parallel()
+
+	v := &Vaultik{}
+
+	blobs := []snapshot.BlobInfo{
+		{Hash: manifestTestBlobA, CompressedSize: 10},
+		{Hash: manifestTestBlobB, CompressedSize: 20},
+	}
+	manifest := &snapshot.Manifest{Blobs: blobs}
+
+	require.NoError(t, v.verifyManifestAgainstDatabase(manifest, blobs))
+}

internal/vaultik/verify_test.go

@@ -1,100 +0,0 @@
-package vaultik_test
-
-import (
-	"bytes"
-	"crypto/rand"
-	"crypto/sha256"
-	"encoding/hex"
-	"io"
-	"testing"
-
-	"github.com/klauspost/compress/zstd"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-	"sneak.berlin/go/vaultik/internal/crypto"
-)
-
-// TestTeeReaderWithDecryption tests that TeeReader correctly hashes all encrypted
-// bytes when streaming through age decryption and zstd decompression.
-// This validates the verification path: hash encrypted blob -> decrypt -> decompress.
-func TestTeeReaderWithDecryption(t *testing.T) {
-	t.Parallel()
-
-	// Test data - use random data that doesn't compress well (5MB)
-	testData := make([]byte, 5*1024*1024)
-	_, err := rand.Read(testData)
-	require.NoError(t, err)
-
-	// Compress the data
-	var compressedBuf bytes.Buffer
-
-	compressor, err := zstd.NewWriter(&compressedBuf,
-		zstd.WithEncoderLevel(zstd.SpeedDefault))
-	require.NoError(t, err)
-	_, err = compressor.Write(testData)
-	require.NoError(t, err)
-	err = compressor.Close()
-	require.NoError(t, err)
-
-	// Encrypt the compressed data
-	testRecipient := "age1cplgrwj77ta54dnmydvvmzn64ltk83ankxl5sww04mrt" +
-		"mu62kv3s89gmvv"
-	testSecretKey := "AGE-SECRET-KEY-1C77PYNTHXSHNNC6EYR2W52UWYXACXA5J" +
-		"T00J9CCW9986M3XY87PSGP89AQ"
-
-	encryptor, err := crypto.NewEncryptor([]string{testRecipient})
-	require.NoError(t, err)
-
-	var encryptedBuf bytes.Buffer
-
-	err = encryptor.EncryptStream(&encryptedBuf, bytes.NewReader(compressedBuf.Bytes()))
-	require.NoError(t, err)
-
-	encryptedData := encryptedBuf.Bytes()
-
-	// Calculate the expected hash of the encrypted data directly
-	expectedHash := sha256.Sum256(encryptedData)
-	expectedHashStr := hex.EncodeToString(expectedHash[:])
-
-	t.Logf("Encrypted data size: %d bytes", len(encryptedData))
-	t.Logf("Expected hash: %s", expectedHashStr)
-
-	// Now simulate what verifyBlob does: use TeeReader to hash while decrypting
-	decryptor, err := crypto.NewDecryptor(testSecretKey)
-	require.NoError(t, err)
-
-	// Create hasher and tee reader
-	hasher := sha256.New()
-	reader := bytes.NewReader(encryptedData)
-	teeReader := io.TeeReader(reader, hasher)
-
-	// Decrypt through the tee reader
-	decryptedReader, err := decryptor.DecryptStream(teeReader)
-	require.NoError(t, err)
-
-	// Decompress
-	decompressor, err := zstd.NewReader(decryptedReader)
-	require.NoError(t, err)
-
-	defer decompressor.Close()
-
-	// Read all decompressed data (simulating chunk verification)
-	decompressedData, err := io.ReadAll(decompressor)
-	require.NoError(t, err)
-
-	// Verify we got the original data back
-	assert.Equal(t, testData, decompressedData, "Decompressed data should match original")
-
-	// Drain remaining decompressed data (should be 0)
-	remaining, err := io.Copy(io.Discard, decompressor)
-	require.NoError(t, err)
-	assert.Equal(t, int64(0), remaining, "No remaining decompressed data")
-
-	// Calculate hash from tee reader
-	calculatedHashStr := hex.EncodeToString(hasher.Sum(nil))
-	t.Logf("Calculated hash (before drain): %s", calculatedHashStr)
-
-	// Verify the hash matches the direct hash of encrypted data
-	assert.Equal(t, expectedHashStr, calculatedHashStr,
-		"Hash calculated via TeeReader should match direct hash of encrypted data")
-}

test/config.yaml

@@ -1,6 +1,6 @@
 age_recipients:
   - age1278m9q7dp3chsh2dcy82qk27v047zywyvtxwnj4cvt0z65jw6a7q5dqhfj  # sneak's long term age key
-  - age1otherpubkey...  # add additional recipients as needed
+  - age1ezrjmfpwsc95svdg0y54mums3zevgzu0x0ecq2f7tp8a05gl0sjq9q9wjg  # add additional recipients as needed
 snapshots:
   test:
     paths: