Merge branch 'r2/disk' into integration/round-1

app/src-tauri/src/commands/migration_commands.rs

@@ -256,10 +256,25 @@ pub(crate) fn is_migrating(project_id: &str) -> bool {
 struct ActiveGuard(#[allow(dead_code)] crate::project_lock::ProjectGuard);
 
 impl ActiveGuard {
-    /// `None` when a migration — or anything else — already holds this project.
+    /// `Err` with the registry's own refusal when a migration — **or anything
-    fn acquire(project_id: &str) -> Option<Self> {
+    /// else** — already holds this project.
+    ///
+    /// The error string is the point. This returned `Option`, and all three
+    /// callers replaced the discarded reason with a sentence about a migration
+    /// — so a user blocked by a *compaction*, a reset or a cache clear was told
+    /// to wait for a base update that was not running, with nothing in the UI
+    /// that could ever name what actually held the project.
+    /// `project_lock::try_acquire` already composes "what holds it" with "what
+    /// you were trying to do"; there is nothing to add to it here.
+    ///
+    /// The tail it composes is `ProjectOp::Migration`'s — "…before starting a
+    /// base update" — for confirm and rollback as well as for the migration
+    /// itself. That is the class all three belong to, and splitting it would
+    /// mean a `ProjectOp` variant per command: the wrong place to encode a
+    /// verb, for a phrase that is at worst imprecise where the old one was
+    /// simply wrong.
+    fn acquire(project_id: &str) -> Result<Self, String> {
         crate::project_lock::try_acquire(project_id, crate::project_lock::ProjectOp::Migration)
-            .ok()
             .map(Self)
     }
 }
@@ -276,10 +291,9 @@ pub async fn migrate_project_to_base(
     app_handle: tauri::AppHandle,
     state: State<'_, AppState>,
 ) -> Result<MigrationReport, String> {
-    let Some(_guard) = ActiveGuard::acquire(&project_id) else {
+    let _guard = match ActiveGuard::acquire(&project_id) {
-        return Ok(MigrationReport::failed_preflight(
+        Ok(guard) => guard,
-            "A migration is already running for this project.",
+        Err(busy) => return Ok(MigrationReport::failed_preflight(&busy)),
-        ));
     };
 
     let existing = migration_store::load(&project_id)?;
@@ -817,12 +831,7 @@ pub async fn confirm_migration(
     let _ = &state;
     // Confirming drops the only way back. Doing that underneath a running
     // migration would delete the pin it is relying on mid-flight.
-    let Some(_guard) = ActiveGuard::acquire(&project_id) else {
+    let _guard = ActiveGuard::acquire(&project_id)?;
-        return Err(
-            "A container base update is running for this project right now. Wait for it to finish."
-                .to_string(),
-        );
-    };
     let Some(mstate) = migration_store::load(&project_id)? else {
         return Ok(());
     };
@@ -869,12 +878,7 @@ pub async fn rollback_migration(
     app_handle: tauri::AppHandle,
     state: State<'_, AppState>,
 ) -> Result<(), String> {
-    let Some(_guard) = ActiveGuard::acquire(&project_id) else {
+    let _guard = ActiveGuard::acquire(&project_id)?;
-        return Err(
-            "A container base update is running for this project right now. Wait for it to finish."
-                .to_string(),
-        );
-    };
 
     let mut project = state
         .projects_store
@@ -987,7 +991,21 @@ pub async fn get_migration_state(
 pub async fn reconcile_migration(project: &Project, app_handle: &tauri::AppHandle) {
     // A migration running right now is indistinguishable from a crashed one
     // from the outside; only this process knows the difference.
-    if is_migrating(&project.id) {
+    //
+    // **Any holder, not just a migration.** This asked `is_migrating`, which is
+    // `held() == Some(Migration)` — so a compaction, a reset or a destroy
+    // holding the project made this fall straight through and start rewriting
+    // the migration record's phase and untagging its rollback image underneath
+    // whatever was running. `reconcile_project_statuses` is a command, not just
+    // a startup step, so "nothing else can be running yet" is not available as
+    // an argument.
+    if let Some(holder) = crate::project_lock::held(&project.id) {
+        log::debug!(
+            "Skipping migration reconcile for '{}' ({}): {}",
+            project.name,
+            project.id,
+            holder.describe()
+        );
         return;
     }
     let state = match migration_store::load(&project.id) {
@@ -1897,16 +1915,23 @@ mod tests {
         {
             let g = ActiveGuard::acquire(id).expect("first acquire must succeed");
             assert!(is_migrating(id));
+            let refused = ActiveGuard::acquire(id)
+                .err()
+                .expect("a second concurrent migration must be refused");
+            // The refusal has to say what is holding the project, not what the
+            // caller happens to be — the three commands used to substitute
+            // their own sentence for this and lost the distinction.
             assert!(
-                ActiveGuard::acquire(id).is_none(),
+                refused.contains("base update"),
-                "a second concurrent migration must be refused"
+                "the refusal must name the holder: {}",
+                refused
             );
             drop(g);
         }
         assert!(!is_migrating(id), "the guard must release on drop");
         // …including when the migration bailed out through an early return.
         fn early_return(id: &str) -> Option<()> {
-            let _g = ActiveGuard::acquire(id)?;
+            let _g = ActiveGuard::acquire(id).ok()?;
             None
         }
         assert!(early_return(id).is_none());

app/src-tauri/src/docker/disk.rs

@@ -1059,6 +1059,48 @@ pub fn confirmation_matches(expected_project_name: &str, typed: &str) -> bool {
     !expected_project_name.is_empty() && typed.trim() == expected_project_name.trim()
 }
 
+/// Longest project id this will accept. A UUID is 36 characters; the slack is
+/// for a hand-edited `projects.json`, not for anything structural.
+const PROJECT_ID_MAX_LEN: usize = 64;
+
+/// Whether a string may be interpolated into a Docker object name.
+///
+/// ## This is a path-injection guard, not a tidiness check
+///
+/// `project_id` reaches [`destroy`] as a free-form string over IPC, and the
+/// destructive paths interpolate it into `triple-c-snapshot-{id}:{tag}` — which
+/// bollard puts straight into a request path. **bollard does not percent-encode
+/// it.** `bollard::uri::Uri::parse` builds `unix://…/v1.47/images/{reference}`
+/// and then calls `Url::join(path)` on it; `path` starts with `/`, so the join
+/// *replaces* the whole path and RFC 3986 dot-segment removal is applied to the
+/// result. An id of `a/../../v1.47/volumes/{name}?` therefore turns
+/// `DELETE /v1.47/images/triple-c-snapshot-a/../../v1.47/volumes/{name}?…`
+/// into `DELETE /v1.47/volumes/{name}` — the trailing `?` swallowing the tag
+/// into a query string that bollard then overwrites with its own. Reproduced
+/// against the live daemon: it answered 204 and the volume was gone.
+///
+/// The typed confirmation is no defence, because [`confirmation_matches`]
+/// compares the caller's own two strings and an attacker-shaped id can be typed
+/// back verbatim.
+///
+/// ## Why a character class and not a UUID parse
+///
+/// Ids have been `uuid::Uuid::new_v4().to_string()` since the first commit, so
+/// a UUID parse would be correct today — and would silently strand a project
+/// whose id a user or an import tool wrote by hand. What actually has to hold
+/// is that the id cannot leave the path segment it is interpolated into, and
+/// alphanumerics plus `-`/`_` cannot: no `/`, no `.` (so no `..`), no `?`, `#`,
+/// `%`, `:` or `@`. That alphabet is also exactly what
+/// `migration_store::sanitize` leaves untouched, so an id that passes here
+/// names the same tombstone file it did before.
+pub fn is_project_id(candidate: &str) -> bool {
+    !candidate.is_empty()
+        && candidate.len() <= PROJECT_ID_MAX_LEN
+        && candidate
+            .chars()
+            .all(|c| c.is_ascii_alphanumeric() || c == '-' || c == '_')
+}
+
 /// The Windows/WSL2 caveat, in one place so the UI and the logs cannot drift.
 ///
 /// Docker Desktop keeps the whole daemon inside `ext4.vhdx` under
@@ -1112,14 +1154,36 @@ pub fn is_docker_desktop(operating_system: &str) -> bool {
 /// volume looks unclaimed. Deleting them would take the user's credentials,
 /// transcripts and toolchains for every project they have.
 ///
-/// So an empty list is only believed when the file is *also* absent, which is
+/// ## Why an empty list is never believed, and why the shape of the list is not
-/// the genuine fresh-install case and the one where there is nothing on the
+/// the test any more
-/// daemon to mis-attribute anyway. Anything else returns the reason, and orphan
+///
-/// detection is suppressed rather than run optimistically.
+/// The old rule was "an empty list is believed when the file is *also* absent",
+/// on the reasoning that a missing file is the fresh-install case and a fresh
+/// install has nothing on the daemon to mis-attribute. Both halves were wrong:
+///
+/// * **A missing `projects.json` is not the same as a new one.** A data
+///   directory that was moved, restored from a partial backup, or reached with
+///   a different `XDG_DATA_HOME` than the one the app ran under before looks
+///   identical, and every one of those has a daemon full of live volumes behind
+///   it. And a genuinely fresh install loses nothing by being refused, because
+///   there is by definition nothing for it to find — which is what makes
+///   refusing the cheap side of the trade.
+/// * **"The list is empty" stops being true the moment anything is saved.**
+///   `ProjectsStore::new()` swallows a corrupt file into an empty list *without
+///   rewriting it*, and the next `save()` — `update_status()`, i.e. merely
+///   starting a project — writes `[{that one project}]` over it. The list is
+///   then non-empty, this guard passes, and every pre-existing project's
+///   volumes are offered as orphans. Reproduced against a byte-exact replica.
+///
+/// So the corrupt load is recorded on disk by the store itself and read back
+/// here through `corrupt_since`, rather than being inferred from a symptom that
+/// erases itself; see [`crate::storage::projects_store::corrupt_since`]. And an
+/// empty `known` set is refused unconditionally.
 pub fn project_store_trust(
     projects: &[Project],
     json_exists: bool,
     json_ids: Option<&[String]>,
+    corrupt_since: Option<&str>,
 ) -> Result<HashSet<String>, String> {
     let Some(json_ids) = json_ids else {
         return Err(
@@ -1141,19 +1205,39 @@ pub fn project_store_trust(
     let mut known: HashSet<String> = json_ids.iter().cloned().collect();
     known.extend(projects.iter().map(|p| p.id.clone()));
 
-    if known.is_empty() && json_exists {
+    // **Checked before the list is looked at, because the list looks fine.**
-        return Err(
+    // This is the state where `known` is non-empty, parses, and is still
+    // missing every project that was in the file before it stopped parsing.
+    if let Some(since) = corrupt_since {
+        return Err(format!(
+            "This data directory loaded a projects.json it could not parse ({}), so the project \
+             list here is known to be incomplete and a live project's volumes cannot be told \
+             from an orphan. A copy of the unreadable file was kept beside it as \
+             projects.json.bak. Restore it, or delete the projects.json.corrupt marker next to \
+             it to say the loss is accepted.",
+            since
+        ));
+    }
+
+    if known.is_empty() {
+        return Err(if json_exists {
             "The project list loaded empty from a projects.json that exists, which is what a \
              recovered-from-corrupt store looks like. Orphan detection is suppressed rather than \
              treat every project's volumes as unclaimed."
-                .to_string(),
+                .to_string()
-        );
+        } else {
+            "There is no projects.json in this data directory, so there is no list to tell a \
+             live project's volumes from an orphan — a moved or partially restored data \
+             directory looks exactly like a fresh install. Nothing is listed here until one \
+             exists."
+                .to_string()
+        });
     }
     Ok(known)
 }
 
-/// Re-read `projects.json` from disk: whether it exists, and the project ids it
+/// Re-read `projects.json` from disk: whether it exists, the project ids it
-/// actually holds.
+/// actually holds, and whether this data directory has ever failed to parse it.
 ///
 /// The in-memory store cannot answer either question. By the time it is
 /// consulted a corrupt file has already been swallowed into an empty list, and
@@ -1164,14 +1248,19 @@ pub fn project_store_trust(
 /// is not a string is skipped rather than failing the whole read — the file is
 /// still parseable, so the honest answer is the ids it does carry.
 ///
+/// The third element is the sticky corrupt-load marker
+/// (`projects_store::corrupt_since`), read on the same pass because a file that
+/// parses *now* says nothing about whether the ids in it are all of them.
+///
 /// Blocking `std::fs`, so every async caller goes through
 /// [`projects_json_snapshot_async`].
-fn projects_json_snapshot() -> (bool, Option<Vec<String>>) {
+fn projects_json_snapshot() -> (bool, Option<Vec<String>>, Option<String>) {
+    let corrupt_since = crate::storage::projects_store::corrupt_since();
     let Some(path) = dirs::data_dir().map(|d| d.join("triple-c").join("projects.json")) else {
-        return (false, None);
+        return (false, None, corrupt_since);
     };
     if !path.exists() {
-        return (false, Some(Vec::new()));
+        return (false, Some(Vec::new()), corrupt_since);
     }
     let parsed = std::fs::read_to_string(&path)
         .ok()
@@ -1182,7 +1271,7 @@ fn projects_json_snapshot() -> (bool, Option<Vec<String>>) {
                 .filter_map(|entry| entry.get("id")?.as_str().map(str::to_string))
                 .collect::<Vec<_>>()
         });
-    (true, parsed)
+    (true, parsed, corrupt_since)
 }
 
 /// [`projects_json_snapshot`] off the async worker.
@@ -1191,10 +1280,10 @@ fn projects_json_snapshot() -> (bool, Option<Vec<String>>) {
 /// Windows volume behind an antivirus filter blocks the whole tokio worker
 /// thread it lands on, and this one is called from inside [`scan`] — the
 /// longest-running command in the app.
-async fn projects_json_snapshot_async() -> (bool, Option<Vec<String>>) {
+async fn projects_json_snapshot_async() -> (bool, Option<Vec<String>>, Option<String>) {
     tokio::task::spawn_blocking(projects_json_snapshot)
         .await
-        .unwrap_or((false, None))
+        .unwrap_or((false, None, None))
 }
 
 // ---------------------------------------------------------------------------
@@ -1445,9 +1534,14 @@ pub async fn scan(projects: &[Project]) -> Result<DiskUsageReport, String> {
     // as exact.
     let base_images_bytes = base_images.iter().map(|b| b.bytes).sum();
 
-    let (json_exists, json_ids) = projects_json_snapshot_async().await;
+    let (json_exists, json_ids, corrupt_since) = projects_json_snapshot_async().await;
     let (orphan_volumes_list, orphan_volumes_unavailable) =
-        match project_store_trust(projects, json_exists, json_ids.as_deref()) {
+        match project_store_trust(
+            projects,
+            json_exists,
+            json_ids.as_deref(),
+            corrupt_since.as_deref(),
+        ) {
             Ok(known) => {
                 let facts: Vec<VolumeFacts> = volumes.iter().map(volume_facts).collect();
                 (orphan_volumes(&facts, &known, true), None)
@@ -2843,7 +2937,9 @@ async fn reclaim_migration_pins() -> ReclaimResult {
                 continue;
             }
             // Starts the clock when nothing has sighted this pin before, which
-            // is what makes the first pass a no-op for it.
+            // is what makes the first pass a no-op for it. The window between
+            // the `has_record` above and this write is closed inside
+            // `note_ownerless_since` — see its docs.
             let ownerless_since =
                 migration_store::note_ownerless_since(&project_id, &tag, &now);
             if pin_disposition(&tag, has_record, ownerless_since, &now) != PinDisposition::Reapable
@@ -3026,34 +3122,103 @@ async fn reclaim_containers(
     }
 }
 
-/// Remove one orphaned volume, re-checking every safety condition first.
-///
-/// The plan that offered this was computed against a `df()` from some seconds
-/// ago. A project could have been added since — by this app or by a second copy
-/// of it — and a container could have attached. So the store is re-consulted
-/// *from disk*, the name is re-parsed and the live ref count is re-read here:
-/// the typed confirmation is permission to act, not a promise that the world
-/// stood still. Both of those re-checks predate the move to the destructive
-/// path and are deliberately unchanged.
 /// Drop a rollback pin whose project is no longer in `projects.json`.
 ///
-/// The owned case ([`destroy`]'s `RollbackPin` arm) takes the project's claim
+/// ## Both of its strings come from IPC, and both are checked
-/// and clears the ownerless marker. Neither applies here: there is no project
+///
-/// to claim, and nothing else can be mid-operation on an id the store does not
+/// This arm is reached *because* `find_project` failed, which means
-/// know. What *does* still apply is the tag validation — this is the one
+/// `project_id` matched nothing and is an unconstrained string straight off the
-/// destructive variant carrying a free-form string over IPC, and `latest` would
+/// wire — and it was interpolated into `triple-c-snapshot-{id}:{tag}` and handed
-/// name a live snapshot rather than a pin.
+/// to bollard, which does not percent-encode a path. See [`is_project_id`] for
+/// the traversal that bought and the daemon transcript that confirmed it. The
+/// tag check was always here (`latest` would name the project's live snapshot);
+/// the id check is the other half of the same rule and is applied first,
+/// because a malformed id makes every later refusal message a lie about what
+/// would have been touched.
+///
+/// [`confirmation_matches`] is not a barrier for either: it compares the
+/// caller's own two strings, so a crafted id is simply typed back verbatim.
+///
+/// ## And it re-derives the reference from the daemon
+///
+/// Validation says the id *could* name a pin; it does not say one exists. So
+/// the daemon is asked for `triple-c-snapshot-*:pre-migration-*` and the
+/// reference that is actually removed is **the daemon's own string**, matched
+/// on the `(project_id, tag)` pair. Nothing built out of IPC input reaches
+/// `remove_image` at all — the same shape as [`destroy_orphan_volume`], which
+/// discards the frontend's `project_id` and re-derives it from a volume name
+/// the daemon reported.
+///
+/// ## Guards its sibling applies, that this used to skip
+///
+/// Ownership was decided from the *in-memory* project list alone, and that was
+/// the only check: no `project_store_trust`, no `has_record`, no lock. A
+/// corrupt `projects.json` makes every live project's pin look ownerless — and
+/// this arm calls `sweep_orphaned_snapshots()`, which deletes the freshly
+/// dangling image on the same pass. That was the one construction that could
+/// reap a pin whose migration is still **awaiting confirmation**, the invariant
+/// `pin_is_reapable` orders its conditions to protect. So, in order:
+///
+/// 1. The store is re-read *from disk* and put through `project_store_trust`,
+///    which refuses outright when it cannot be trusted.
+/// 2. An id the trusted store *does* know is refused: this is the ownerless
+///    path, and a pin with an owner belongs to [`destroy`]'s owned arm, which
+///    asks for the project's name rather than its id.
+/// 3. The lock is taken **before** anything is read that a decision rests on.
+///    The old comment justified returning before the lock with "nothing else
+///    can be mid-operation on an id the store does not know", which is false:
+///    a migration holds the lock under exactly the id being typed here, and it
+///    is the thing most likely to be writing the record checked next.
+/// 4. `has_record` — filesystem presence, not `load`, the same conservative
+///    question both reapers ask. A record still claims this pin.
 async fn destroy_ownerless_rollback_pin(
     project_id: &str,
     tag: &str,
+    projects: &[Project],
 ) -> Result<ReclaimResult, String> {
+    if !is_project_id(project_id) {
+        return Err(format!(
+            "{:?} is not a project id. Nothing was removed.",
+            project_id
+        ));
+    }
     if migration::parse_rollback_tag(tag).is_none() {
         return Err(format!(
             "{:?} is not a rollback pin tag. Nothing was removed.",
             tag
         ));
     }
-    let reference = format!("triple-c-snapshot-{}:{}", project_id, tag);
+
+    let (json_exists, json_ids, corrupt_since) = projects_json_snapshot_async().await;
+    let known = project_store_trust(
+        projects,
+        json_exists,
+        json_ids.as_deref(),
+        corrupt_since.as_deref(),
+    )?;
+    if known.contains(project_id) {
+        return Err(format!(
+            "Project {} is in your project list after all, so this pin is not ownerless. \
+             Reopen the disk panel and confirm it by project name instead. Nothing was removed.",
+            project_id
+        ));
+    }
+
+    let _guard =
+        crate::project_lock::try_acquire(project_id, crate::project_lock::ProjectOp::Destroy)?;
+
+    if migration_store::has_record(project_id).unwrap_or(true) {
+        return Err(format!(
+            "A migration record for {} is still on disk, so something can still ask for this \
+             rollback. Nothing was removed.",
+            project_id
+        ));
+    }
+
+    // The reference the daemon itself reports for this pin. `None` means the
+    // plan is stale — the pin was reaped or confirmed since the panel drew it.
+    let reference = live_rollback_pin_reference(project_id, tag).await?;
+
     migration::untag_image(&reference).await?;
     // The grace clock is meaningless once the tag is gone, and the marker file
     // would otherwise outlive everything that could ever read it.
@@ -3081,11 +3246,68 @@ async fn destroy_ownerless_rollback_pin(
     })
 }
 
+/// The daemon's own `repo_tag` for one rollback pin, or why it is not there.
+///
+/// The point is the *provenance* of the returned string, not the existence
+/// check: it is a tag Docker listed, so removing it cannot address anything but
+/// an image. A reference assembled here from a caller's id would be a string
+/// the daemon never vouched for, which is the whole of C-2.
+///
+/// The `reference` filter matches on `repo:tag` and is the same one every other
+/// pin path uses, so this asks for the shape `rollback_tag` produces and
+/// nothing else. The match is still made on the parsed `(project_id, tag)` pair
+/// rather than on the filter's word, because a filter is never the only guard
+/// on a removal here.
+async fn live_rollback_pin_reference(project_id: &str, tag: &str) -> Result<String, String> {
+    let docker = get_docker()?;
+    let images = docker
+        .list_images(Some(ListImagesOptions {
+            all: false,
+            filters: HashMap::from([(
+                "reference".to_string(),
+                vec!["triple-c-snapshot-*:pre-migration-*".to_string()],
+            )]),
+            ..Default::default()
+        }))
+        .await
+        .map_err(|e| format!("Could not re-check rollback pins: {}", e))?;
+
+    images
+        .iter()
+        .flat_map(|image| image.repo_tags.iter())
+        .find(|reference| {
+            migration::parse_snapshot_reference(reference)
+                .is_some_and(|(id, t)| id == project_id && t == tag)
+        })
+        .cloned()
+        .ok_or_else(|| {
+            format!(
+                "There is no rollback pin {} for {} on the daemon any more. Nothing was \
+                 removed.",
+                tag, project_id
+            )
+        })
+}
+
+/// Remove one orphaned volume, re-checking every safety condition first.
+///
+/// The plan that offered this was computed against a `df()` from some seconds
+/// ago. A project could have been added since — by this app or by a second copy
+/// of it — and a container could have attached. So the store is re-consulted
+/// *from disk*, the name is re-parsed and the live ref count is re-read here:
+/// the typed confirmation is permission to act, not a promise that the world
+/// stood still. Both of those re-checks predate the move to the destructive
+/// path and are deliberately unchanged.
 async fn destroy_orphan_volume(name: &str, projects: &[Project]) -> Result<ReclaimResult, String> {
     let docker = get_docker()?;
 
-    let (json_exists, json_ids) = projects_json_snapshot_async().await;
+    let (json_exists, json_ids, corrupt_since) = projects_json_snapshot_async().await;
-    let known = project_store_trust(projects, json_exists, json_ids.as_deref())?;
+    let known = project_store_trust(
+        projects,
+        json_exists,
+        json_ids.as_deref(),
+        corrupt_since.as_deref(),
+    )?;
 
     let usage = docker
         .df()
@@ -3154,6 +3376,22 @@ pub async fn compact_snapshot(project: &Project) -> ReclaimResult {
         project_id: project.id.clone(),
     };
 
+    // The same guard [`destroy`] applies, for the same reason: this resolves
+    // `triple-c-snapshot-{id}:latest`, builds `…:compacting`, names a
+    // `triple-c-compact-*` container and commits back over `:latest`. An id
+    // that cannot be safely interpolated into a Docker name is refused before
+    // any of that starts rather than at whichever step happens to notice.
+    if !is_project_id(&project.id) {
+        return failed(
+            target,
+            format!(
+                "Project {:?} has an id that is not a project id, so its snapshot cannot be \
+                 addressed safely on the daemon.",
+                project.name
+            ),
+        );
+    }
+
     // **Acquired, and held for the whole rewrite.** This is the operation the
     // per-project lock was written for. Compaction resolves
     // `triple-c-snapshot-{id}:latest` when its build starts and commits back
@@ -3830,15 +4068,35 @@ pub async fn destroy(
         if find_project(projects, target.project_id()).is_err() {
             if !confirmation_matches(project_id, confirmation) {
                 return Err(format!(
-                    "This pin's project is no longer in Triple-C, so there is no name to type.                      Type the project id ({}) exactly to confirm. Nothing was removed.",
+                    "This pin's project is no longer in Triple-C, so there is no name to type. \
+                     Type the project id ({}) exactly to confirm. Nothing was removed.",
                     project_id
                 ));
             }
-            return destroy_ownerless_rollback_pin(project_id, tag).await;
+            // **The in-memory list is not the ownership decision.** It is only
+            // what routes to this arm; `destroy_ownerless_rollback_pin` re-reads
+            // `projects.json` and refuses if the store cannot be trusted, or if
+            // it turns out to know this id after all.
+            return destroy_ownerless_rollback_pin(project_id, tag, projects).await;
         }
     }
 
     let project = find_project(projects, target.project_id())?;
+    // **Every arm below interpolates this id into a Docker object name** — two
+    // volume names, a snapshot reference, a rollback reference — and bollard
+    // puts those straight into a request path without encoding them. The id
+    // came from the store rather than from IPC here, which is a weaker source
+    // than it sounds: `projects.json` is a plain file a user or an import tool
+    // can write. One check in front of all four is cheaper than four, and the
+    // only thing it can refuse is a project whose id was never one this app
+    // generated. See [`is_project_id`].
+    if !is_project_id(&project.id) {
+        return Err(format!(
+            "Project {:?} has an id that is not a project id, so nothing of its can be addressed \
+             safely on the daemon. Nothing was removed.",
+            project.name
+        ));
+    }
     if !confirmation_matches(&project.name, confirmation) {
         return Err(format!(
             "Type the project name ({}) exactly to confirm. Nothing was removed.",
@@ -3941,13 +4199,19 @@ pub async fn destroy(
             Err("An orphaned volume is not a project object.".to_string())
         }
         DestructiveTarget::RollbackPin { tag, .. } => {
-            // **The one destructive variant carrying a free-form string.**
+            // **The variant carrying a free-form string.** Every other arm
-            // Every other arm builds its target from constants; this one takes
+            // builds its target from constants; this one takes a tag over IPC
-            // a tag over IPC and interpolates it into an image reference that
+            // and interpolates it into an image reference that is then removed.
-            // is then removed. Unvalidated, `tag: "latest"` names the project's
+            // Unvalidated, `tag: "latest"` names the project's live snapshot —
-            // live snapshot — deleted under a dialog that says "rollback pin".
+            // deleted under a dialog that says "rollback pin".
             // `parse_rollback_tag` accepts only `pre-migration-<YYYYmmdd-HHMMSS>`,
             // which is exactly what `rollback_tag` produces and nothing else.
+            //
+            // The *id* in the reference below is `project.id`, taken from the
+            // record `find_project` matched — not the one the caller sent. That
+            // is what makes this arm safe and is the difference from
+            // `destroy_ownerless_rollback_pin`, which has no record to match
+            // and must validate the id itself; see [`is_project_id`].
             if migration::parse_rollback_tag(tag).is_none() {
                 return Err(format!(
                     "{:?} is not a rollback pin tag. Nothing was removed.",

app/src-tauri/src/docker/disk_tests.rs

@@ -482,7 +482,7 @@ fn project(id: &str, name: &str) -> Project {
 
 #[test]
 fn an_unreadable_projects_json_is_never_trusted() {
-    let err = project_store_trust(&[project("a", "api")], true, None).unwrap_err();
+    let err = project_store_trust(&[project("a", "api")], true, None, None).unwrap_err();
     assert!(err.contains("could not be read"), "{}", err);
 }
 
@@ -492,18 +492,53 @@ fn an_empty_list_from_an_existing_file_is_treated_as_a_failed_load() {
     // up and starts empty. That is right for the app and catastrophic here, so
     // the combination "empty list + file present" is refused rather than read as
     // "the user has no projects".
-    let err = project_store_trust(&[], true, Some(&[])).unwrap_err();
+    let err = project_store_trust(&[], true, Some(&[]), None).unwrap_err();
     assert!(err.contains("suppressed"), "{}", err);
+}
 
-    // No file at all is a genuine fresh install, and there is nothing on the
+#[test]
-    // daemon to mis-attribute in that state.
+fn a_missing_projects_json_is_not_evidence_of_a_fresh_install() {
-    assert!(project_store_trust(&[], false, Some(&[])).unwrap().is_empty());
+    // H-3: this used to return `Ok(empty)` here, on the reasoning that no file
+    // means a fresh install and a fresh install has nothing to mis-attribute.
+    // A data directory that was moved, partially restored, or reached under a
+    // different XDG_DATA_HOME is indistinguishable from that — and has a daemon
+    // full of live volumes behind it. The genuinely fresh case loses nothing by
+    // being refused, because there is nothing there for it to find.
+    let err = project_store_trust(&[], false, Some(&[]), None).unwrap_err();
+    assert!(err.contains("no projects.json"), "{}", err);
+}
+
+#[test]
+fn a_recorded_corrupt_load_is_not_undone_by_the_next_write() {
+    // H-3, the shape that made the old guard evaporate: `ProjectsStore::new()`
+    // swallows a corrupt file into an empty list *without rewriting it*, and
+    // the first `save()` after that — `update_status()`, i.e. merely starting a
+    // project — writes `[{that one project}]` over it. `known` is then
+    // non-empty and the "empty list + file present" guard passes, so every
+    // pre-existing project's volumes are offered as orphans.
+    //
+    // The store's sticky marker is what survives that write, so it is checked
+    // before the list is looked at at all.
+    let err = project_store_trust(
+        &[project("the-one-project-started-since", "api")],
+        true,
+        Some(&["the-one-project-started-since".to_string()]),
+        Some("2026-08-23T00:00:00Z"),
+    )
+    .unwrap_err();
+    assert!(err.contains("could not parse"), "{}", err);
+    assert!(
+        err.contains("projects.json.corrupt"),
+        "the refusal must name the marker the user has to delete: {}",
+        err
+    );
 }
 
 #[test]
 fn a_healthy_store_yields_its_ids() {
     let ids =
-        project_store_trust(&[project("a", "api"), project("b", "web")], true, Some(&[])).unwrap();
+        project_store_trust(&[project("a", "api"), project("b", "web")], true, Some(&[]), None)
+            .unwrap();
     assert_eq!(ids, HashSet::from(["a".to_string(), "b".to_string()]));
 }
 
@@ -518,6 +553,7 @@ fn a_project_only_the_file_knows_about_still_counts_as_live() {
         &[project("a", "api")],
         true,
         Some(&["a".to_string(), "b-from-the-other-window".to_string()]),
+        None,
     )
     .unwrap();
     assert!(
@@ -1592,3 +1628,346 @@ async fn an_ownerless_rollback_pin_still_refuses_a_tag_that_is_not_a_pin() {
         err
     );
 }
+
+/// The exact string that was reproduced against the live daemon: an id built to
+/// climb out of `/images/` and land on `/volumes/`, with a trailing `?` to
+/// swallow the tag into a query string bollard then overwrites with its own.
+const C2_TRAVERSAL_ID: &str = "a/../../v1.47/volumes/tcaudit-victim?";
+
+#[test]
+fn a_project_id_cannot_leave_the_path_segment_it_is_interpolated_into() {
+    // C-2. `format!("triple-c-snapshot-{}:{}", project_id, tag)` goes to
+    // bollard, which does not percent-encode: `Uri::parse` joins an absolute
+    // path onto the base URL, which *replaces* the path and applies RFC 3986
+    // dot-segment removal. Verified against the daemon: this id turned a
+    // `DELETE /v1.47/images/…` into `DELETE /v1.47/volumes/tcaudit-victim`,
+    // answered 204, and the volume was gone.
+    assert!(!is_project_id(C2_TRAVERSAL_ID));
+    // Each ingredient on its own, so a future relaxation of the alphabet has to
+    // fail here rather than only in the composed payload.
+    assert!(!is_project_id("a/b"), "a path separator escapes the segment");
+    assert!(!is_project_id(".."), "dot segments are what the join removes");
+    assert!(!is_project_id("a?x=1"), "a query terminates the path");
+    assert!(!is_project_id("a#f"), "a fragment terminates the path");
+    assert!(!is_project_id("a%2f"), "percent-encoding is decoded by the daemon");
+    assert!(!is_project_id("a:b"), "a colon would forge the tag separator");
+    assert!(!is_project_id("a b"));
+    assert!(!is_project_id(""));
+    assert!(!is_project_id(&"a".repeat(PROJECT_ID_MAX_LEN + 1)));
+
+    // The real shape has to survive untouched, or every existing pin becomes
+    // undeletable through the panel that exists to find it.
+    assert!(is_project_id("dead0000-0000-0000-0000-000000000000"));
+    assert!(is_project_id(&Project::new("api".to_string(), Vec::new()).id));
+    // And the alphabet is the one `migration_store::sanitize` leaves alone, so
+    // an id that passes here names the same tombstone file it always did.
+    assert!(is_project_id("legacy_id-2"));
+}
+
+#[tokio::test]
+async fn an_ownerless_rollback_pin_refuses_an_id_that_is_not_a_project_id() {
+    // C-2 end to end, through the public entry point. Note the confirmation:
+    // it is the crafted id typed back verbatim, which `confirmation_matches`
+    // accepts — it compares the caller's own two strings and was never a
+    // barrier here. The refusal has to come from the id check.
+    let projects: Vec<Project> = Vec::new();
+    let target = DestructiveTarget::RollbackPin {
+        project_id: C2_TRAVERSAL_ID.to_string(),
+        tag: "pre-migration-20260101-101500".to_string(),
+    };
+    let err = destroy(&target, C2_TRAVERSAL_ID, &projects)
+        .await
+        .expect_err("a traversal id must never reach the Docker API");
+    // Remove `is_project_id` from `destroy_ownerless_rollback_pin` and this is
+    // the assertion that fails: the call gets as far as the store re-read and
+    // comes back with some other refusal — or, with a store present and a
+    // daemon behind it, with a 204 and somebody's volume gone.
+    assert!(
+        err.contains("not a project id"),
+        "the id must be refused before anything else is consulted, got: {}",
+        err
+    );
+    assert!(err.contains("Nothing was removed"), "{}", err);
+}
+
+// ---------------------------------------------------------------------------
+// C-2 and H-3, against a real daemon and a real data directory
+// ---------------------------------------------------------------------------
+
+/// Serialises the tests that redirect `XDG_DATA_HOME`.
+///
+/// The process environment is process-wide and `cargo test` runs test functions
+/// on a thread pool, so two of these racing would have one reading the other's
+/// data directory. `#[ignore]` keeps them out of the default run; this keeps
+/// them out of each other's way when the whole ignored set is run at once.
+fn env_lock() -> std::sync::MutexGuard<'static, ()> {
+    static LOCK: std::sync::OnceLock<std::sync::Mutex<()>> = std::sync::OnceLock::new();
+    LOCK.get_or_init(|| std::sync::Mutex::new(()))
+        .lock()
+        .unwrap_or_else(|e| e.into_inner())
+}
+
+/// H-3 as it actually unfolds, against a real `ProjectsStore` on a real disk.
+///
+/// `#[ignore]` because it redirects `XDG_DATA_HOME` for the whole process —
+/// `cargo test` has to stay free of that. Run it with
+/// `cargo test -- --ignored a_corrupt_projects_json_stays_untrusted`.
+///
+/// The sequence is the one that was reproduced against a byte-exact replica of
+/// a real data directory, and every step of it is ordinary app behaviour:
+///
+/// 1. `projects.json` holds two live projects and stops parsing.
+/// 2. `ProjectsStore::new()` backs it up, starts empty, and the app runs on.
+/// 3. The user starts a project. `update_status()` calls `save()`, which writes
+///    `[{that one project}]` over the file.
+/// 4. `projects.json` now parses and holds one id — so "the list loaded empty
+///    from a file that exists" is false, the guard passes, and both original
+///    projects' home and config volumes are offered as orphans at
+///    `Safety::Safe`.
+///
+/// Step 4 is what this asserts is no longer reachable.
+#[test]
+#[ignore]
+fn a_corrupt_projects_json_stays_untrusted_after_the_next_write() {
+    let _env = env_lock();
+    let data_home =
+        std::env::temp_dir().join(format!("triple-c-h3-{}", uuid::Uuid::new_v4().simple()));
+    let dir = data_home.join("triple-c");
+    std::fs::create_dir_all(&dir).expect("temp data dir");
+    let previous = std::env::var_os("XDG_DATA_HOME");
+    std::env::set_var("XDG_DATA_HOME", &data_home);
+
+    // 1. Two live projects, and a file that no longer parses.
+    std::fs::write(dir.join("projects.json"), "[{\"id\":\"live-a\"},{\"id\":\"liv")
+        .expect("corrupt projects.json");
+
+    // 2. The app starts and recovers.
+    let store = crate::storage::projects_store::ProjectsStore::new().expect("store");
+    assert!(store.list().is_empty(), "the corrupt file loads as an empty list");
+
+    // 3. The user adds and starts a project — the ordinary write that used to
+    //    erase the only evidence of step 1.
+    let started = store
+        .add(project("new-project-added-after-the-corruption", "api"))
+        .expect("add");
+    store
+        .update_status(&started.id, crate::models::ProjectStatus::Running)
+        .expect("update_status");
+    let rewritten = std::fs::read_to_string(dir.join("projects.json")).unwrap();
+    assert!(
+        rewritten.contains("new-project-added-after-the-corruption")
+            && !rewritten.contains("live-a"),
+        "the precondition for H-3 is that the file now parses and has lost the old ids: {}",
+        rewritten
+    );
+
+    // 4. The state the old guard could not see.
+    let (json_exists, json_ids, corrupt_since) = projects_json_snapshot();
+    assert!(json_exists);
+    assert_eq!(json_ids.as_deref().map(<[String]>::len), Some(1));
+    assert!(
+        corrupt_since.is_some(),
+        "the corrupt load must still be on the record after the rewrite"
+    );
+    let err = project_store_trust(
+        &store.list(),
+        json_exists,
+        json_ids.as_deref(),
+        corrupt_since.as_deref(),
+    )
+    .unwrap_err();
+    assert!(err.contains("could not parse"), "{}", err);
+
+    // And the consequence: `live-a`'s volumes are not offered to anyone.
+    let volumes = vec![VolumeFacts {
+        name: "triple-c-home-live-a".to_string(),
+        bytes: 8_000_000_000,
+        links: 0,
+        created_at: None,
+    }];
+    assert!(
+        orphan_volumes(&volumes, &HashSet::new(), false).is_empty(),
+        "a suppressed store must offer nothing"
+    );
+
+    std::fs::remove_dir_all(&data_home).ok();
+    match previous {
+        Some(v) => std::env::set_var("XDG_DATA_HOME", v),
+        None => std::env::remove_var("XDG_DATA_HOME"),
+    }
+}
+
+/// The traversal, run for real, plus the legitimate case it must not break.
+///
+/// `#[ignore]` for the same reason as the compaction test: `cargo test` has to
+/// stay daemon-free. Run it with
+/// `cargo test -- --ignored a_traversal_project_id_cannot_reach_the_docker_api`.
+///
+/// It exists because the unit test above can only assert a *message*. The thing
+/// that made C-2 critical is not that a string was unvalidated, it is that the
+/// daemon honoured what the string turned the request into — and no pure test
+/// can show that. This one creates a volume, points a rollback-pin removal at
+/// it through `destroy`, and checks the volume is still there afterwards.
+///
+/// It also does the other half, which is the half a validator can easily break:
+/// a genuinely ownerless pin still has to be removable, or the panel that
+/// exists to find multi-gigabyte orphans can no longer act on them.
+///
+/// `XDG_DATA_HOME` is redirected for the duration so the store re-read, the
+/// migration records and the ownerless tombstones all land in a temp directory
+/// rather than the developer's real one.
+#[tokio::test]
+#[ignore]
+async fn a_traversal_project_id_cannot_reach_the_docker_api() {
+    let _env = env_lock();
+    let run = uuid::Uuid::new_v4().simple().to_string();
+    let data_home = std::env::temp_dir().join(format!("triple-c-c2-{}", run));
+    std::fs::create_dir_all(data_home.join("triple-c")).expect("temp data dir");
+    // A store that parses and knows about *some other* project, so
+    // `project_store_trust` is satisfied and the id under test is genuinely not
+    // in it. `projects_json_snapshot` reads ids as opaque JSON strings.
+    std::fs::write(
+        data_home.join("triple-c").join("projects.json"),
+        r#"[{"id":"c2-some-other-project"}]"#,
+    )
+    .expect("temp projects.json");
+    let previous_data_home = std::env::var_os("XDG_DATA_HOME");
+    std::env::set_var("XDG_DATA_HOME", &data_home);
+
+    let docker = get_docker().expect("a daemon");
+
+    // ---- 1. The attack -----------------------------------------------------
+    let victim = format!("triple-c-c2-victim-{}", run);
+    docker
+        .create_volume(bollard::volume::CreateVolumeOptions {
+            name: victim.clone(),
+            ..Default::default()
+        })
+        .await
+        .expect("create the victim volume");
+
+    // `a/` gives the join something to climb out of; the two `..` segments eat
+    // `triple-c-snapshot-a` and `images`; the trailing `?` turns `:{tag}` into
+    // a query string, which bollard then replaces with `force`/`noprune`.
+    let traversal_id = format!("a/../../v1.47/volumes/{}?", victim);
+    let target = DestructiveTarget::RollbackPin {
+        project_id: traversal_id.clone(),
+        tag: "pre-migration-20250101-000000".to_string(),
+    };
+    // The confirmation is the crafted id itself — which is exactly why the
+    // typed confirmation was never a defence here.
+    let err = destroy(&target, &traversal_id, &[])
+        .await
+        .expect_err("a traversal id must be refused");
+    assert!(err.contains("not a project id"), "got: {}", err);
+    assert!(
+        docker.inspect_volume(&victim).await.is_ok(),
+        "the volume was deleted by an image-tag removal — C-2 is not fixed"
+    );
+
+    // ---- 2. The legitimate case, which must still work ---------------------
+    let orphan_id = uuid::Uuid::new_v4().to_string();
+    let tag = "pre-migration-20250101-000000";
+    let pin = format!("triple-c-snapshot-{}:{}", orphan_id, tag);
+    // Deliberately *not* labelled `triple-c.managed`: this test's leftovers
+    // must never be something `sweep_orphaned_snapshots` can act on, and the
+    // sweep is called on the success path.
+    build_from_dockerfile("FROM busybox:1.36
+RUN touch /c2-pin
+", &pin)
+        .await
+        .expect("build the pin image");
+
+    let result = destroy(
+        &DestructiveTarget::RollbackPin {
+            project_id: orphan_id.clone(),
+            tag: tag.to_string(),
+        },
+        &orphan_id,
+        &[],
+    )
+    .await
+    .expect("a well-formed ownerless pin must still be removable");
+    assert!(result.ok, "{}", result.message);
+    assert!(
+        docker.inspect_image(&pin).await.is_err(),
+        "the pin tag survived the removal that reported success"
+    );
+
+    // ---- cleanup -----------------------------------------------------------
+    let _ = docker.remove_volume(&victim, None).await;
+    let _ = docker
+        .remove_image(
+            &pin,
+            Some(RemoveImageOptions {
+                force: true,
+                noprune: false,
+            }),
+            None,
+        )
+        .await;
+    let _ = std::fs::remove_dir_all(&data_home);
+    match previous_data_home {
+        Some(v) => std::env::set_var("XDG_DATA_HOME", v),
+        None => std::env::remove_var("XDG_DATA_HOME"),
+    }
+}
+
+#[tokio::test]
+async fn an_owned_target_is_refused_when_the_stored_id_is_not_a_project_id() {
+    // The owned arms take their id from the record `find_project` matched
+    // rather than from IPC — which is a weaker source than it sounds, because
+    // `projects.json` is a plain file. All four of them (two volume names, the
+    // snapshot reference, the rollback reference) go on to be interpolated into
+    // a Docker request path, so one check stands in front of all four.
+    let mut bad = project(C2_TRAVERSAL_ID, "api");
+    bad.name = "api".to_string();
+    let projects = vec![bad];
+    for target in [
+        DestructiveTarget::HomeVolume {
+            project_id: C2_TRAVERSAL_ID.to_string(),
+        },
+        DestructiveTarget::SnapshotImage {
+            project_id: C2_TRAVERSAL_ID.to_string(),
+        },
+        DestructiveTarget::RollbackPin {
+            project_id: C2_TRAVERSAL_ID.to_string(),
+            tag: "pre-migration-20260101-101500".to_string(),
+        },
+    ] {
+        let err = destroy(&target, "api", &projects)
+            .await
+            .expect_err("a stored id that is not a project id must be refused");
+        assert!(err.contains("not a project id"), "{:?}: {}", target, err);
+        // And before the confirmation is even compared, so the refusal cannot
+        // be walked past by typing the right name.
+        assert!(err.contains("Nothing was removed"), "{}", err);
+    }
+}
+
+#[tokio::test]
+async fn the_ownerless_arm_will_not_touch_a_pin_the_store_still_claims() {
+    // H-1. Ownership was decided from the in-memory list alone, and this arm
+    // then calls `sweep_orphaned_snapshots()`, which deletes the freshly
+    // dangling image on the same pass — so a `projects.json` that failed to
+    // load made every live project's pin "ownerless", including one whose
+    // migration is still awaiting confirmation. The re-read is the fix; here it
+    // is the union with the in-memory list that carries the id.
+    //
+    // Called directly rather than through `destroy`, because `destroy` routes
+    // by the in-memory list and a project that is *in* it never reaches this
+    // arm. The point is that the arm refuses on its own account.
+    let projects = vec![project("dead0000-0000-0000-0000-000000000000", "api")];
+    let err = destroy_ownerless_rollback_pin(
+        "dead0000-0000-0000-0000-000000000000",
+        "pre-migration-20260101-101500",
+        &projects,
+    )
+    .await
+    .expect_err("a pin whose project the store knows is not ownerless");
+    assert!(
+        err.contains("not ownerless"),
+        "the store's answer must override the caller's, got: {}",
+        err
+    );
+}

app/src-tauri/src/docker/migration.rs

@@ -1233,6 +1233,14 @@ pub async fn reap_stale_migration_pins() -> usize {
             }
             // Records the first sighting when there is none, which is why this
             // returns `None` on that pass and the pin survives it.
+            //
+            // A `save` can land between the `has_record` above and this write,
+            // which would plant a tombstone dated *now* behind a perfectly
+            // valid record — invisible until that record is legitimately lost,
+            // at which point the pin is already past its grace period and is
+            // reaped on the first check. `note_ownerless_since` re-asks
+            // `has_record` after the write and removes the marker again; the
+            // reasoning for why that closes the window is on it.
             let ownerless_since =
                 crate::storage::migration_store::note_ownerless_since(&project_id, &tag, &now);
             if !pin_is_reapable(&tag, has_record, ownerless_since, &now) {

app/src-tauri/src/storage/migration_store.rs

@@ -83,9 +83,9 @@ pub fn load(project_id: &str) -> Result<Option<MigrationState>, String> {
         Ok(state) => Ok(Some(state)),
         Err(e) => {
             let backup = corrupt_backup_path(&path, &chrono::Utc::now());
-            let copied = if backup.exists() {
+            let copied = if backup.exists() || corrupt_backups_full(&path) {
-                // Already kept a copy of this exact corruption this second;
+                // Already kept a copy of this exact corruption this second, or
-                // nothing to add.
+                // kept as many as are worth keeping. Either way nothing to add.
                 Ok(())
             } else {
                 fs::copy(&path, &backup).map(|_| ())
@@ -114,6 +114,50 @@ fn corrupt_backup_path(path: &std::path::Path, now: &chrono::DateTime<chrono::Ut
     path.with_extension(format!("json.corrupt-{}.bak", now.format("%Y%m%d-%H%M%S")))
 }
 
+/// How many timestamped copies of one project's corrupt record are kept.
+///
+/// Timestamping fixed the "second corruption overwrote the first" bug and
+/// introduced its opposite: [`load`] runs on every reconcile, every survey and
+/// every reaper pass, so a record that is *persistently* unparseable — the
+/// normal case, since nothing repairs it — mints a new copy every time the
+/// clock's second changes. Nothing ever reads them back and nothing ever
+/// removed them.
+///
+/// Four is enough for the only use there is: a human looking at what the file
+/// held. See [`corrupt_backups_full`] for why the cap is applied before the
+/// copy rather than by pruning after it.
+const MAX_CORRUPT_BACKUPS: usize = 4;
+
+/// Whether [`MAX_CORRUPT_BACKUPS`] copies of this record already exist.
+///
+/// Asked *before* the copy rather than pruning after it, so the cap is not
+/// implemented by writing a file and deleting it again on every pass — and so
+/// the copies that survive are the oldest, which are the ones taken closest to
+/// whatever produced the corruption.
+///
+/// A directory that cannot be listed answers "not full": failing open here
+/// costs at most one extra file, and failing closed would drop the very first
+/// copy of a record nothing else has kept.
+fn corrupt_backups_full(path: &std::path::Path) -> bool {
+    let (Some(dir), Some(stem)) = (path.parent(), path.file_stem()) else {
+        return false;
+    };
+    // `{stem}.json.corrupt-` — the same shape `corrupt_backup_path` builds, so
+    // this can never match another project's copies or an unrelated `.bak`.
+    let prefix = format!("{}.json.corrupt-", stem.to_string_lossy());
+    let Ok(entries) = fs::read_dir(dir) else {
+        return false;
+    };
+    entries
+        .flatten()
+        .filter(|e| {
+            let name = e.file_name().to_string_lossy().to_string();
+            name.starts_with(&prefix) && name.ends_with(".bak")
+        })
+        .count()
+        >= MAX_CORRUPT_BACKUPS
+}
+
 /// Whether a project has a migration record on disk *at all*, without parsing
 /// it.
 ///
@@ -251,6 +295,22 @@ pub fn peek_ownerless_since(
 /// the period below what has actually elapsed on the *marker's* terms, because
 /// there is nothing to compare against but wall time; what it cannot do any
 /// more is make every pin instantly reapable, which dating from the tag did.
+///
+/// ## Why the write re-checks `has_record`
+///
+/// Both reapers ask [`has_record`] and only call this when the answer is no,
+/// which leaves a window: a [`save`] landing between the two runs its
+/// `clear_ownerless_for_project` against a marker that does not exist yet, and
+/// this then plants one — dated *now* — behind a perfectly valid record. The
+/// marker is invisible while the record stands, so nothing notices. It only
+/// matters later, if that record is legitimately lost: the pin is then already
+/// fourteen days ownerless on its very first check and is reaped with **zero**
+/// grace, which is the exact failure the tombstone exists to prevent.
+///
+/// So the write is followed by a second `has_record`, and a marker that turns
+/// out to sit behind a record is removed again. The two orderings that remain
+/// are both safe: a `save` completing *after* this re-check clears the marker
+/// itself, and one completing before it is what the re-check sees.
 pub fn note_ownerless_since(
     project_id: &str,
     tag: &str,
@@ -274,6 +334,19 @@ pub fn note_ownerless_since(
                     tag,
                     e
                 );
+                return None;
+            }
+            // A record that appeared while this was being written owns the pin,
+            // and a tombstone behind an owned pin is a fourteen-day head start
+            // on reaping it the moment that record is next lost.
+            if has_record(project_id).unwrap_or(false) {
+                log::debug!(
+                    "A migration record for {} appeared while marking {} ownerless; \
+                     the marker was dropped again",
+                    project_id,
+                    tag
+                );
+                clear_ownerless(project_id, tag);
             }
             None
         }
@@ -339,6 +412,39 @@ pub fn clear_staging(project_id: &str) -> Result<(), String> {
 mod tests {
     use super::*;
 
+    #[test]
+    fn corrupt_copies_of_one_record_are_capped() {
+        // `load` runs on every reconcile, every survey and every reaper pass,
+        // and nothing repairs an unparseable record — so a persistently corrupt
+        // one minted a new timestamped copy every time the clock's second
+        // changed, and nothing ever removed them.
+        let dir = std::env::temp_dir().join(format!(
+            "triple-c-corrupt-cap-{}",
+            uuid::Uuid::new_v4().simple()
+        ));
+        fs::create_dir_all(&dir).expect("temp dir");
+        let record = dir.join("some-project.json");
+
+        assert!(!corrupt_backups_full(&record), "an empty directory is not full");
+        for n in 0..MAX_CORRUPT_BACKUPS {
+            fs::write(
+                dir.join(format!("some-project.json.corrupt-2026010{}-000000.bak", n)),
+                "x",
+            )
+            .unwrap();
+        }
+        assert!(corrupt_backups_full(&record));
+
+        // Another project's copies, and an unrelated `.bak`, are not this
+        // record's — the prefix is the whole point of the naming.
+        let other = dir.join("other-project.json");
+        assert!(!corrupt_backups_full(&other));
+        fs::write(dir.join("some-project.json.bak"), "x").unwrap();
+        assert!(!corrupt_backups_full(&other));
+
+        fs::remove_dir_all(&dir).ok();
+    }
+
     #[test]
     fn project_ids_cannot_escape_the_migrations_directory() {
         assert_eq!(sanitize("../../etc/passwd"), "______etc_passwd");

app/src-tauri/src/storage/projects_store.rs

@@ -1,9 +1,96 @@
 use std::fs;
-use std::path::PathBuf;
+use std::path::{Path, PathBuf};
 use std::sync::Mutex;
 
 use crate::models::Project;
 
+/// The sticky marker for `projects.json`: `projects.json.corrupt`, beside it.
+///
+/// Derived from the file rather than from `dirs::data_dir()` so the marker
+/// always lands in the directory the store is actually using — and so the
+/// writer can be tested against a temp directory.
+fn corrupt_marker_for(file_path: &Path) -> PathBuf {
+    file_path.with_extension("json.corrupt")
+}
+
+/// `<data_dir>/triple-c/projects.json.corrupt`, whether or not it exists.
+pub fn corrupt_marker_path() -> Option<PathBuf> {
+    dirs::data_dir().map(|d| corrupt_marker_for(&d.join("triple-c").join("projects.json")))
+}
+
+/// When this data directory last loaded a `projects.json` it could not parse,
+/// as the RFC3339 instant recorded in the marker.
+///
+/// ## Why this outlives the load that wrote it
+///
+/// A corrupt load is *recoverable for the app* — the list starts empty and
+/// everything keeps working — and that recovery is precisely what makes it
+/// dangerous for anything that reasons about which projects exist. The
+/// in-memory symptom does not survive: the first [`ProjectsStore::save`] after
+/// the failure, which is as little as starting one project (`update_status`),
+/// writes `[{that one project}]` over the file. From then on `projects.json`
+/// parses, holds one id, and looks exactly like a user with one project — while
+/// every *other* project's home and config volume is on the daemon claimed by
+/// nobody.
+///
+/// The guard in `project_store_trust` keyed on "the list is empty and the file
+/// exists", which that write silently ends. So the fact is recorded on disk
+/// instead of inferred from the list's shape, and it is **sticky**: nothing in
+/// this app clears it, because nothing in this app can reconstruct what the
+/// unreadable file held. The refusal names the marker so a user who has
+/// restored their list — or accepted the loss — can delete it deliberately.
+pub fn corrupt_since() -> Option<String> {
+    let raw = fs::read_to_string(corrupt_marker_path()?).ok()?;
+    let trimmed = raw.trim();
+    if trimmed.is_empty() {
+        // The marker's presence is the signal; an empty one still means a
+        // corrupt load happened, it just cannot say when.
+        return Some("an unknown time".to_string());
+    }
+    Some(trimmed.lines().next().unwrap_or(trimmed).to_string())
+}
+
+/// Keep the bytes of an unparseable `projects.json`, and record that it
+/// happened.
+///
+/// **The existing `.bak` is never overwritten.** A second corruption used to
+/// clobber the first, and the first is the valuable one: it was taken before
+/// the app rewrote the file with whatever it had in memory, so it is the only
+/// copy that can still hold the full project list. Later ones are copies of an
+/// already-degraded file and get a timestamped name.
+fn record_corrupt_load(file_path: &Path, now: &chrono::DateTime<chrono::Utc>) {
+    let first = file_path.with_extension("json.bak");
+    let backup = if first.exists() {
+        file_path.with_extension(format!("json.corrupt-{}.bak", now.format("%Y%m%d-%H%M%S")))
+    } else {
+        first
+    };
+    if !backup.exists() {
+        if let Err(e) = fs::copy(file_path, &backup) {
+            log::error!("Failed to back up corrupted projects.json: {}", e);
+        } else {
+            log::error!(
+                "A copy of the unreadable projects.json was kept at {}",
+                backup.display()
+            );
+        }
+    }
+
+    let marker = corrupt_marker_for(file_path);
+    if marker.exists() {
+        // Sticky: the *first* corruption is the one that dates the loss.
+        return;
+    }
+    if let Err(e) = fs::write(&marker, now.to_rfc3339()) {
+        log::error!(
+            "Could not record the corrupt projects.json load at {}: {} — orphan detection will \
+             not know the project list is incomplete",
+            marker.display(),
+            e
+        );
+    }
+}
+
 pub struct ProjectsStore {
     projects: Mutex<Vec<Project>>,
     file_path: PathBuf,
@@ -43,20 +130,14 @@ impl ProjectsStore {
                                 Ok(parsed) => (parsed, migrated),
                                 Err(e) => {
                                     log::error!("Failed to parse migrated projects.json: {}. Starting with empty list.", e);
-                                    let backup = file_path.with_extension("json.bak");
+                                    record_corrupt_load(&file_path, &chrono::Utc::now());
-                                    if let Err(be) = fs::copy(&file_path, &backup) {
-                                        log::error!("Failed to back up corrupted projects.json: {}", be);
-                                    }
                                     (Vec::new(), false)
                                 }
                             }
                         }
                         Err(e) => {
                             log::error!("Failed to parse projects.json: {}. Starting with empty list.", e);
-                            let backup = file_path.with_extension("json.bak");
+                            record_corrupt_load(&file_path, &chrono::Utc::now());
-                            if let Err(be) = fs::copy(&file_path, &backup) {
-                                log::error!("Failed to back up corrupted projects.json: {}", be);
-                            }
                             (Vec::new(), false)
                         }
                     }
@@ -203,3 +284,89 @@ impl ProjectsStore {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn temp_dir(tag: &str) -> PathBuf {
+        let dir = std::env::temp_dir().join(format!(
+            "triple-c-store-{}-{}",
+            tag,
+            uuid::Uuid::new_v4().simple()
+        ));
+        fs::create_dir_all(&dir).expect("temp dir");
+        dir
+    }
+
+    #[test]
+    fn a_corrupt_load_leaves_a_marker_the_next_write_cannot_erase() {
+        // H-3, the whole chain in one test. `ProjectsStore::new()` swallows an
+        // unparseable file into an empty list *without rewriting it*, and the
+        // first `save()` after that — as little as `update_status()` — writes
+        // `[{one project}]` over it. Everything the old guard keyed on ("the
+        // list is empty and the file exists") is gone at that point, while
+        // every *other* project's volumes are still on the daemon claimed by
+        // nobody.
+        let dir = temp_dir("corrupt");
+        let file = dir.join("projects.json");
+        fs::write(&file, "{ this is not a project list").unwrap();
+
+        let now = chrono::Utc::now();
+        record_corrupt_load(&file, &now);
+
+        let marker = corrupt_marker_for(&file);
+        assert!(marker.exists(), "the corrupt load must be recorded on disk");
+        assert_eq!(fs::read_to_string(&marker).unwrap(), now.to_rfc3339());
+        assert!(
+            dir.join("projects.json.bak").exists(),
+            "the unreadable bytes must be kept"
+        );
+
+        // The write that used to erase the evidence. The marker is a separate
+        // file, so it does not care.
+        fs::write(&file, r#"[{"id":"the-one-project-started-since"}]"#).unwrap();
+        assert!(marker.exists());
+
+        fs::remove_dir_all(&dir).ok();
+    }
+
+    #[test]
+    fn a_second_corruption_keeps_the_first_copy_and_the_first_date() {
+        // The `.bak` used to be a fixed name, so a second corruption clobbered
+        // the first — and the first is the only copy taken before the app
+        // rewrote the file with whatever it had in memory, i.e. the only one
+        // that can still hold the full project list.
+        let dir = temp_dir("second");
+        let file = dir.join("projects.json");
+        fs::write(&file, "original bytes").unwrap();
+        let first = chrono::DateTime::parse_from_rfc3339("2026-01-01T00:00:00Z")
+            .unwrap()
+            .with_timezone(&chrono::Utc);
+        record_corrupt_load(&file, &first);
+
+        fs::write(&file, "degraded bytes").unwrap();
+        let second = chrono::DateTime::parse_from_rfc3339("2026-06-01T00:00:00Z")
+            .unwrap()
+            .with_timezone(&chrono::Utc);
+        record_corrupt_load(&file, &second);
+
+        assert_eq!(
+            fs::read_to_string(dir.join("projects.json.bak")).unwrap(),
+            "original bytes",
+            "the first copy must survive the second corruption"
+        );
+        assert_eq!(
+            fs::read_to_string(dir.join("projects.json.corrupt-20260601-000000.bak")).unwrap(),
+            "degraded bytes"
+        );
+        // And the marker still dates the loss from the first failure, which is
+        // when the project list actually stopped being complete.
+        assert_eq!(
+            fs::read_to_string(corrupt_marker_for(&file)).unwrap(),
+            first.to_rfc3339()
+        );
+
+        fs::remove_dir_all(&dir).ok();
+    }
+}