release(1.8.2)

Cargo.lock

@@ -3538,7 +3538,7 @@ checksum = "35fb2e5f958ec131621fdd531e9fc186ed768cbe395337403ae56c17a74c68ec"
 
 [[package]]
 name = "pdfs-cli"
-version = "1.8.0"
+version = "1.8.2"
 dependencies = [
  "anyhow",
  "clap 4.6.4",
@@ -3553,7 +3553,7 @@ dependencies = [
 
 [[package]]
 name = "pdfs-core"
-version = "1.8.0"
+version = "1.8.2"
 dependencies = [
  "async-trait",
  "directories",
@@ -3574,7 +3574,7 @@ dependencies = [
 
 [[package]]
 name = "pdfs-fuse"
-version = "1.8.0"
+version = "1.8.2"
 dependencies = [
  "fuser",
  "futures",
@@ -3591,7 +3591,7 @@ dependencies = [
 
 [[package]]
 name = "pdfs-gui"
-version = "1.8.0"
+version = "1.8.2"
 dependencies = [
  "anyhow",
  "async-channel",

Cargo.toml

@@ -3,7 +3,7 @@ resolver = "2"
 members = ["crates/pdfs-core", "crates/pdfs-fuse", "crates/pdfs-cli", "crates/pdfs-gui"]
 
 [workspace.package]
-version = "1.8.0"
+version = "1.8.2"
 edition = "2024"
 license = "MIT"
 authors = ["Nils Pukropp <contact@narl.io>"]

crates/pdfs-core/src/db/migrations.rs

@@ -9,7 +9,7 @@ use super::Db;
 use crate::Result;
 
 /// Current schema version. Bump on every forward migration added below.
-pub(super) const SCHEMA_VERSION: i64 = 27;
+pub(super) const SCHEMA_VERSION: i64 = 28;
 
 impl Db {
     pub(super) fn migrate(&self) -> Result<()> {
@@ -201,6 +201,23 @@ impl Db {
                 tx.execute_batch(MIGRATION_V27)?;
             }
         }
+        if current < 28 {
+            // Same guards as V26/V27, for the same two reasons.
+            let has_column: bool = tx.query_row(
+                "SELECT COUNT(*) FROM pragma_table_info('sync_entry') \
+                 WHERE name = 'local_mtime_ns'",
+                [],
+                |row| row.get::<_, i64>(0),
+            )? > 0;
+            let has_entries: bool = tx.query_row(
+                "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = 'sync_entry'",
+                [],
+                |row| row.get::<_, i64>(0),
+            )? > 0;
+            if has_entries && !has_column {
+                tx.execute_batch(MIGRATION_V28)?;
+            }
+        }
         tx.execute(
             "INSERT INTO sync_state (key, value) VALUES ('schema_version', ?1)
              ON CONFLICT(key) DO UPDATE SET value = excluded.value",
@@ -808,3 +825,24 @@ CREATE INDEX IF NOT EXISTS idx_pending_op_claimed
 const MIGRATION_V27: &str = "
 ALTER TABLE pending_op ADD COLUMN access_deferred_since INTEGER NOT NULL DEFAULT 0;
 ";
+
+/// Schema v28: sub-second precision for a mirror folder's local baseline.
+///
+/// A reconcile pass decides whether the local side of a path changed by
+/// comparing `(local_mtime, local_size)` against the baseline. Both sides of
+/// that comparison were whole seconds, so an edit that landed in the same second
+/// as the last sync and left the file the same length — a fixed-size record
+/// rewritten, an image re-exported, a database page updated in place — read as
+/// *unchanged* and was never uploaded. The remote copy silently stayed at the
+/// older content until something else about the file moved (bugs.md B25).
+///
+/// Nullable rather than a converted `local_mtime * 1_000_000_000`: the
+/// nanosecond part of an existing baseline is genuinely unknown, and inventing
+/// zero for it would make every already-synced file compare as changed and
+/// re-upload an entire mirror on first start. `NULL` means "this row predates
+/// the column", which the comparison reads as "fall back to whole seconds" —
+/// exactly the behaviour that row was written under. Each row gains a real value
+/// the next time its path is synced.
+const MIGRATION_V28: &str = "
+ALTER TABLE sync_entry ADD COLUMN local_mtime_ns INTEGER;
+";

crates/pdfs-core/src/db/sync.rs

@@ -35,6 +35,14 @@ pub struct StoredSyncEntry {
     pub remote_uid: Option<String>,
     pub local_mtime: i64,
     pub local_size: i64,
+    /// The same modification time to nanosecond precision, when it is known.
+    ///
+    /// `local_mtime` alone cannot tell an edit that landed inside the same
+    /// second as the last sync — and left the file the same length — from no
+    /// edit at all, so those never uploaded (bugs.md B25). `None` on a row
+    /// written before schema 28, which the comparison reads as "compare whole
+    /// seconds", so upgrading does not make an entire mirror look changed.
+    pub local_mtime_ns: Option<i64>,
     pub remote_rev: Option<String>,
     pub remote_hash: Option<String>,
 }
@@ -168,7 +176,8 @@ impl Db {
     pub fn sync_entries(&self, folder_id: i64) -> Result<HashMap<String, StoredSyncEntry>> {
         let conn = self.read();
         let mut stmt = conn.prepare(
-            "SELECT rel_path, remote_uid, local_mtime, local_size, remote_rev, remote_hash
+            "SELECT rel_path, remote_uid, local_mtime, local_size, remote_rev, remote_hash,
+                    local_mtime_ns
              FROM sync_entry WHERE folder_id = ?1",
         )?;
         let rows = stmt.query_map(params![folder_id], |r| {
@@ -179,6 +188,7 @@ impl Db {
                 local_size: r.get(3)?,
                 remote_rev: r.get(4)?,
                 remote_hash: r.get(5)?,
+                local_mtime_ns: r.get(6)?,
             })
         })?;
         let mut map = HashMap::new();
@@ -216,14 +226,16 @@ impl Db {
         let conn = self.conn.lock();
         conn.execute(
             "INSERT INTO sync_entry
-               (folder_id, rel_path, remote_uid, local_mtime, local_size, remote_rev, remote_hash)
-             VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7)
+               (folder_id, rel_path, remote_uid, local_mtime, local_size, remote_rev,
+                remote_hash, local_mtime_ns)
+             VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8)
              ON CONFLICT(folder_id, rel_path) DO UPDATE SET
-               remote_uid  = excluded.remote_uid,
-               local_mtime = excluded.local_mtime,
-               local_size  = excluded.local_size,
-               remote_rev  = excluded.remote_rev,
-               remote_hash = excluded.remote_hash",
+               remote_uid     = excluded.remote_uid,
+               local_mtime    = excluded.local_mtime,
+               local_size     = excluded.local_size,
+               remote_rev     = excluded.remote_rev,
+               remote_hash    = excluded.remote_hash,
+               local_mtime_ns = excluded.local_mtime_ns",
             params![
                 folder_id,
                 e.rel_path,
@@ -232,6 +244,7 @@ impl Db {
                 e.local_size,
                 e.remote_rev,
                 e.remote_hash,
+                e.local_mtime_ns,
             ],
         )?;
         Ok(())

crates/pdfs-core/src/db/tests.rs

@@ -1895,6 +1895,123 @@ fn failed_superseding_insert_keeps_the_old_pending_op() {
     assert_eq!(ops[0].blob_path.as_deref(), Some("/staging/original"));
 }
 
+/// B53's remaining half. The insert-failure regression above proves the *row*
+/// rolls back; this proves the two things that make the rollback usable — the
+/// old blob is not reported as superseded (so the caller does not delete the
+/// bytes the surviving row still points at), and the rollback survives a
+/// restart rather than living only in the connection that failed.
+#[test]
+fn a_rolled_back_supersede_keeps_its_blob_and_survives_reopen() {
+    let dir = std::env::temp_dir().join(format!("pdfs-db-b53-{}", std::process::id()));
+    let _ = std::fs::remove_dir_all(&dir);
+    std::fs::create_dir_all(&dir).unwrap();
+    let path = dir.join("queue.sqlite");
+    let op = |blob: &str| PendingOp {
+        id: 0,
+        kind: OP_REVISION.to_string(),
+        uid: uid("durable").to_string(),
+        parent_uid: None,
+        name: None,
+        blob_path: Some(blob.to_string()),
+        meta_json: Some("{}".to_string()),
+        created_at: 1,
+        attempts: 0,
+        last_error: None,
+        next_attempt_at: 0,
+    };
+
+    {
+        let db = Db::open(&path).unwrap();
+        db.enqueue_op(&op("/staging/original")).unwrap();
+        db.with_conn(|conn| {
+            conn.execute_batch(
+                "CREATE TRIGGER reject_replacement BEFORE INSERT ON pending_op
+                 WHEN NEW.blob_path = '/staging/replacement'
+                 BEGIN SELECT RAISE(ABORT, 'injected insert failure'); END;",
+            )?;
+            Ok(())
+        })
+        .unwrap();
+
+        // The failure must not hand back the old blob path. That value is the
+        // caller's instruction to delete those bytes, and the row that owns
+        // them is still queued.
+        assert!(db.enqueue_op(&op("/staging/replacement")).is_err());
+        db.with_conn(|conn| {
+            conn.execute_batch("DROP TRIGGER reject_replacement")?;
+            Ok(())
+        })
+        .unwrap();
+    }
+
+    let db = Db::open(&path).unwrap();
+    let ops = db.pending_ops().unwrap();
+    assert_eq!(ops.len(), 1, "old or new — never neither");
+    assert_eq!(ops[0].blob_path.as_deref(), Some("/staging/original"));
+
+    // And once the injected failure is gone, superseding works and *does*
+    // report the blob it retired.
+    let (_, superseded) = db.enqueue_op(&op("/staging/replacement")).unwrap();
+    assert_eq!(superseded.as_deref(), Some("/staging/original"));
+    assert_eq!(db.pending_ops().unwrap().len(), 1);
+
+    drop(db);
+    std::fs::remove_dir_all(&dir).unwrap();
+}
+
+/// A full disk, as SQLite reports it. `max_page_count` makes the database
+/// refuse to grow, which is the same `SQLITE_FULL` an out-of-space filesystem
+/// produces — and the queue must come out of it with the acknowledged work
+/// still in it (B53).
+#[test]
+fn a_full_database_rolls_the_supersede_back_rather_than_erasing_it() {
+    let db = Db::open_in_memory().unwrap();
+    let op = |blob: &str| PendingOp {
+        id: 0,
+        kind: OP_REVISION.to_string(),
+        uid: uid("full").to_string(),
+        parent_uid: None,
+        name: None,
+        blob_path: Some(blob.to_string()),
+        meta_json: Some("{}".to_string()),
+        created_at: 1,
+        attempts: 0,
+        last_error: None,
+        next_attempt_at: 0,
+    };
+    db.enqueue_op(&op("/staging/original")).unwrap();
+
+    // Cap the file at its current size, so any page the insert needs fails.
+    let pages: i64 = db
+        .with_conn(|conn| Ok(conn.query_row("PRAGMA page_count", [], |r| r.get(0))?))
+        .unwrap();
+    db.with_conn(|conn| {
+        conn.execute_batch(&format!("PRAGMA max_page_count = {pages}"))?;
+        Ok(())
+    })
+    .unwrap();
+
+    // A payload far larger than one page, so growth is unavoidable.
+    let mut fat = op("/staging/replacement");
+    fat.meta_json = Some("x".repeat(512 * 1024));
+    let result = db.enqueue_op(&fat);
+
+    db.with_conn(|conn| {
+        conn.execute_batch("PRAGMA max_page_count = 1073741823")?;
+        Ok(())
+    })
+    .unwrap();
+
+    assert!(result.is_err(), "a database that cannot grow must refuse");
+    let ops = db.pending_ops().unwrap();
+    assert_eq!(ops.len(), 1);
+    assert_eq!(
+        ops[0].blob_path.as_deref(),
+        Some("/staging/original"),
+        "the acknowledged upload is still queued and still owns its bytes"
+    );
+}
+
 #[test]
 fn atomic_trash_replacement_returns_owned_blobs_after_commit() {
     let db = Db::open_in_memory().unwrap();
@@ -2408,6 +2525,7 @@ fn mirror_uid_lookup_covers_the_root_and_synced_descendants_only() {
             local_size: 2,
             remote_rev: Some("3".into()),
             remote_hash: Some("2".into()),
+            local_mtime_ns: None,
         },
     )
     .unwrap();
@@ -2424,6 +2542,7 @@ fn mirror_uid_lookup_covers_the_root_and_synced_descendants_only() {
             local_size: 2,
             remote_rev: Some("3".into()),
             remote_hash: Some("2".into()),
+            local_mtime_ns: None,
         },
     )
     .unwrap();
@@ -3545,6 +3664,78 @@ fn migration_v27_adds_the_access_deferral_column_to_a_v26_queue() {
     let _ = std::fs::remove_file(format!("{}.lock", path.display()));
 }
 
+/// A mirror baseline carried over from V27 keeps its rows, and each one reads
+/// back with no nanosecond time. That `None` is the whole point: it makes the
+/// comparison fall back to whole seconds for rows written before the column
+/// existed, instead of asserting a sub-second time of zero and re-uploading
+/// every already-synced file in the folder (bugs.md B25).
+#[test]
+fn migration_v28_adds_sub_second_local_times_without_disturbing_a_v27_baseline() {
+    let path = std::env::temp_dir().join(format!(
+        "pdfs-db-v27-fixture-{}-{}.db",
+        std::process::id(),
+        std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap()
+            .as_nanos()
+    ));
+    {
+        let db = Db::open(&path).unwrap();
+        let folder = db
+            .sync_folder_add("/home/me/Mirror", "vol~root", "share")
+            .unwrap();
+        db.sync_entry_upsert(
+            folder,
+            &StoredSyncEntry {
+                rel_path: "notes.txt".into(),
+                remote_uid: Some("vol~notes".into()),
+                local_mtime: 1_700_000_000,
+                local_mtime_ns: Some(1_700_000_000_123_456_789),
+                local_size: 42,
+                remote_rev: Some("1700000000".into()),
+                remote_hash: Some("42".into()),
+            },
+        )
+        .unwrap();
+    }
+    {
+        // Put the file back in the state a released V27 database was in.
+        let conn = rusqlite::Connection::open(&path).unwrap();
+        conn.execute_batch(
+            "ALTER TABLE sync_entry DROP COLUMN local_mtime_ns;
+             UPDATE sync_state SET value = '27' WHERE key = 'schema_version';",
+        )
+        .unwrap();
+    }
+
+    let db = Db::open(&path).unwrap();
+    let entries = db.sync_entries(1).unwrap();
+    let entry = entries.get("notes.txt").expect("the baseline row survives");
+    assert_eq!(entry.local_mtime, 1_700_000_000);
+    assert_eq!(entry.local_size, 42);
+    assert_eq!(
+        entry.local_mtime_ns, None,
+        "a carried-over row must not claim a sub-second time it never recorded"
+    );
+
+    // And the column is live: a refreshed row keeps the precision it is given.
+    db.sync_entry_upsert(
+        1,
+        &StoredSyncEntry {
+            local_mtime_ns: Some(1_700_000_001_000_000_500),
+            ..entry.clone()
+        },
+    )
+    .unwrap();
+    assert_eq!(
+        db.sync_entries(1).unwrap()["notes.txt"].local_mtime_ns,
+        Some(1_700_000_001_000_000_500)
+    );
+
+    let _ = std::fs::remove_file(&path);
+    let _ = std::fs::remove_file(format!("{}.lock", path.display()));
+}
+
 // --- read-only connection pool (DB2) ---------------------------------------
 
 /// A read must not queue behind whatever happens to be writing.

crates/pdfs-fuse/src/background.rs

@@ -1,6 +1,7 @@
 //! Remote-event invalidation and local-index background workers.
 
 use super::*;
+use crate::shutdown::Shutdown;
 
 /// Drop `uid` from one mount's tree and tell that mount's kernel session the
 /// entry is gone. Shared by the trash and delete paths, which differ only in
@@ -437,6 +438,7 @@ pub(super) fn run_local_index(
     indexing: Arc<AtomicBool>,
     transfers: Arc<TransferRegistry>,
     mountpoint: PathBuf,
+    shutdown: Arc<Shutdown>,
 ) {
     loop {
         let age = db.local_indexed_at().ok().flatten();
@@ -445,7 +447,11 @@ pub(super) fn run_local_index(
         if stale {
             scan_local_once(&db, &indexing, &transfers, &mountpoint);
         }
-        std::thread::sleep(LOCAL_INDEX_CHECK);
+        // Interruptible, so teardown joins this thread instead of leaving it to
+        // keep rewriting the index for a mount that is gone (bugs.md B44).
+        if !shutdown.sleep(LOCAL_INDEX_CHECK) {
+            return;
+        }
     }
 }
 

crates/pdfs-fuse/src/control.rs

@@ -104,6 +104,31 @@ fn read_request_line(reader: &mut impl BufRead) -> std::io::Result<Option<String
         .map_err(|e| std::io::Error::new(std::io::ErrorKind::InvalidData, e))
 }
 
+/// Turn a CLI-supplied path into the mount that owns it and the path relative
+/// to *that* mount's root.
+///
+/// [`rel_to_mount`] is the older, primary-only reading, and it is what every
+/// mutating request still uses. It rejects anything outside the primary
+/// mountpoint, which leaves a secondary on-demand mount — a synced folder
+/// switched out of mirror mode — with no way to be named at all: `pdfs ls` and
+/// `pdfs refresh` both answered "is not under the mountpoint", so a folder
+/// serving a stale listing had no manual escape hatch (`docs/BUGS.md` B86).
+///
+/// A relative path keeps the old meaning, relative to the primary mount, since
+/// there is nothing in it to choose a mount by.
+fn route_to_mount(core: &Core, mountpoint: &Path, path: &str) -> CoreResult<(Core, PathBuf)> {
+    let p = Path::new(path);
+    if !p.is_absolute() {
+        return Ok((core.clone(), p.to_path_buf()));
+    }
+    if let Some(routed) = core.rooted_at(p) {
+        return Ok(routed);
+    }
+    // Fall back to the primary reading so the error text stays the familiar one
+    // when the path is simply outside everything this daemon has mounted.
+    rel_to_mount(mountpoint, path).map(|rel| (core.clone(), rel))
+}
+
 /// Turn a CLI-supplied path into a mountpoint-relative path. An absolute path
 /// must live under `mountpoint`; a relative path is taken as already relative to
 /// the mount root.
@@ -193,8 +218,8 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
         Ok(CtlRequest::ListPins) => CtlResponse::Pins {
             pins: core.cache.list_pins(),
         },
-        Ok(CtlRequest::ListDir { path }) => match rel_to_mount(mountpoint, &path) {
-            Ok(rel) => match core.list_dir(&rel) {
+        Ok(CtlRequest::ListDir { path }) => match route_to_mount(core, mountpoint, &path) {
+            Ok((core, rel)) => match core.list_dir(&rel) {
                 Ok(entries) => CtlResponse::Entries { entries },
                 Err(e) => CtlResponse::error(e),
             },
@@ -202,9 +227,8 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
         },
         Ok(CtlRequest::Refresh { scope }) => {
             let result = match &scope {
-                RefreshScope::Dir { path } => {
-                    rel_to_mount(mountpoint, path).and_then(|rel| core.refresh_dir(&rel))
-                }
+                RefreshScope::Dir { path } => route_to_mount(core, mountpoint, path)
+                    .and_then(|(core, rel)| core.refresh_dir(&rel)),
                 RefreshScope::Trash => {
                     core.invalidate_trash();
                     Ok(())
@@ -1084,6 +1108,12 @@ pub(crate) fn run_control_socket(
 ) {
     info!("control socket listening");
     for conn in listener.incoming() {
+        // Checked on every accept, including the deliberate poke teardown sends
+        // to wake this thread out of a blocking accept (bugs.md B44).
+        if core.shutdown.is_stopping() {
+            info!("control socket closing");
+            return;
+        }
         match conn {
             Ok(stream) => {
                 let Some(permit) = ControlHandlerPermit::acquire() else {

crates/pdfs-fuse/src/drain.rs

@@ -260,6 +260,13 @@ impl Core {
     /// [`Db::claim_next_due_op`]: pdfs_core::db::Db::claim_next_due_op
     pub(crate) fn run_pending_drain(&self, primary: bool) {
         loop {
+            // Between ops, never inside one: an op that has started is either
+            // retired or released, so stopping here cannot leave the queue with
+            // a row claimed by a worker that no longer exists (bugs.md B44).
+            if self.shutdown.is_stopping() {
+                debug!(primary, "drain worker stopping");
+                return;
+            }
             let now = now_millis();
             // One row, chosen and *claimed* by the database. Reading the whole
             // queue to pick one op made a long queue quadratic to drain, and

crates/pdfs-fuse/src/filesystem.rs

@@ -1,6 +1,7 @@
 //! Kernel-facing FUSE callback adapter.
 
 use super::*;
+use crate::state::DirListing;
 
 impl Filesystem for ProtonFs {
     /// Negotiate kernel-side behaviour once, before any other callback.
@@ -162,15 +163,46 @@ impl Filesystem for ProtonFs {
         }
     }
 
+    /// Freeze this directory's listing for the life of the handle.
+    ///
+    /// `readdir` is paged, and its continuation offset only means anything
+    /// against a listing that does not move underneath it. Taking the snapshot
+    /// here — once per `opendir`, which the kernel pairs exactly with one
+    /// `releasedir` — is what makes the offsets stable, so a create or a trash
+    /// during an enumeration can no longer make the caller skip or repeat an
+    /// entry (bugs.md B43).
+    ///
+    /// Enumerating a cold folder is a remote call, so it goes to a worker for
+    /// the same reason `lookup` and `readdir` do.
+    fn opendir(&self, _req: &Request, ino: INodeNo, _flags: OpenFlags, reply: ReplyOpen) {
+        let ino = ino.0;
+        if self.core.children_cached(ino) {
+            self.serve_opendir(ino, reply);
+            return;
+        }
+        let fs = self.clone();
+        self.core
+            .workers
+            .run(Lane::Meta, move || fs.serve_opendir(ino, reply));
+    }
+
     fn readdir(
         &self,
         _req: &Request,
         ino: INodeNo,
-        _fh: FileHandle,
+        fh: FileHandle,
         offset: u64,
         reply: ReplyDirectory,
     ) {
         let ino = ino.0;
+        // The snapshot is the whole point: served from it, this is a pure
+        // in-memory walk with no chance of a remote call, so it never needs a
+        // worker. Only a handle whose `opendir` failed to snapshot (or a kernel
+        // that issued `readdir` without one) falls back to the live path.
+        if let Some(listing) = self.core.state().dir_snapshots.get(&fh.0).cloned() {
+            reply_listing(&listing, offset, reply);
+            return;
+        }
         // Same split as `lookup`: only the cold, remote-enumerating path pays
         // for a worker.
         if self.core.children_cached(ino) {
@@ -183,6 +215,18 @@ impl Filesystem for ProtonFs {
             .run(Lane::Meta, move || fs.serve_readdir(ino, offset, reply));
     }
 
+    fn releasedir(
+        &self,
+        _req: &Request,
+        _ino: INodeNo,
+        fh: FileHandle,
+        _flags: OpenFlags,
+        reply: ReplyEmpty,
+    ) {
+        self.core.state().dir_snapshots.remove(&fh.0);
+        reply.ok();
+    }
+
     /// A read-only open is a map lookup and answers inline; a write open has to
     /// allocate a scratch file and, when a queued revision is the base, duplicate
     /// a blob that is the whole file — so it is served from a worker rather than
@@ -912,6 +956,58 @@ pub struct ProtonFs {
     gid: u32,
 }
 
+/// How many times a write `open` will re-read its base after a `release`
+/// published a newer queued revision underneath it (bugs.md B32).
+///
+/// One retry covers the race, which needs a close and an open of the same file
+/// to overlap on different threads. More than that is not contention, it is a
+/// file being republished continuously, and looping on it would keep a FUSE
+/// worker busy indefinitely.
+const OPEN_BASE_ATTEMPTS: u32 = 3;
+
+/// One directory's entries as `readdir` reports them: `.` and `..` first, then
+/// the interned children in `children` order.
+///
+/// Split out because it is built from two places that must agree — `opendir`
+/// freezes it, and the fallback `readdir` builds it live — and a divergence
+/// between them would show up as entries appearing or vanishing depending on
+/// whether the kernel happened to send an `opendir`.
+fn build_listing(st: &State, ino: u64) -> DirListing {
+    let parent = st.entries.get(&ino).map_or(ROOT_INO, |e| e.parent);
+    // "." and ".." occupy offsets 0 and 1; real children follow.
+    let mut listing = vec![
+        (ino, true, ".".to_string()),
+        (parent, true, "..".to_string()),
+    ];
+    if let Some(kids) = st.children.get(&ino) {
+        for &kid in kids {
+            if let Some(e) = st.entries.get(&kid) {
+                listing.push((kid, e.node.is_folder(), e.node.name.clone()));
+            }
+        }
+    }
+    listing
+}
+
+/// Fill a `readdir` reply from `listing`, resuming at `offset`.
+///
+/// The cookie handed back is the index of the *next* entry, so a resumed call
+/// continues exactly where the last one stopped — which is only true because
+/// the listing behind those indexes is a snapshot (bugs.md B43).
+fn reply_listing(listing: &DirListing, offset: u64, mut reply: ReplyDirectory) {
+    for (i, (ino, is_dir, name)) in listing.iter().enumerate().skip(offset as usize) {
+        let ft = if *is_dir {
+            FileType::Directory
+        } else {
+            FileType::RegularFile
+        };
+        if reply.add(INodeNo(*ino), (i + 1) as u64, ft, name) {
+            break;
+        }
+    }
+    reply.ok();
+}
+
 pub(crate) fn access_allowed(is_dir: bool, access: Access, mask: AccessFlags) -> bool {
     (!mask.contains(AccessFlags::W_OK) || access.writable())
         && (!mask.contains(AccessFlags::X_OK) || is_dir)
@@ -1027,40 +1123,39 @@ impl ProtonFs {
     }
 
     /// The body of [`Filesystem::readdir`], on whichever thread ends up serving it.
-    fn serve_readdir(&self, ino: u64, offset: u64, mut reply: ReplyDirectory) {
+    /// The body of [`Filesystem::opendir`]: enumerate the folder if it is cold,
+    /// then publish one frozen listing under a fresh handle.
+    ///
+    /// A failure to enumerate is reported as a failed `opendir` rather than
+    /// deferred to the first `readdir` — the caller gets the errno at the syscall
+    /// that can still act on it.
+    fn serve_opendir(&self, ino: u64, reply: ReplyOpen) {
+        if let Err(e) = self.core.ensure_children(ino) {
+            reply.error(e);
+            return;
+        }
+        let mut st = self.core.state();
+        let listing = Arc::new(build_listing(&st, ino));
+        // Shares the counter with write handles, so a directory handle can never
+        // collide with a file handle (both are `fh` values in one kernel space).
+        let fh = st.next_fh;
+        st.next_fh += 1;
+        st.dir_snapshots.insert(fh, listing);
+        drop(st);
+        reply.opened(FileHandle(fh), FopenFlags::empty());
+    }
+
+    /// The live, un-snapshotted `readdir`. Reached only by a handle that has no
+    /// snapshot — see [`Filesystem::readdir`].
+    fn serve_readdir(&self, ino: u64, offset: u64, reply: ReplyDirectory) {
         if let Err(e) = self.core.ensure_children(ino) {
             reply.error(e);
             return;
         }
         let st = self.core.state();
-        let parent = st.entries.get(&ino).map_or(ROOT_INO, |e| e.parent);
-
-        // "." and ".." occupy offsets 0 and 1; real children follow.
-        let mut listing: Vec<(u64, FileType, String)> = vec![
-            (ino, FileType::Directory, ".".to_string()),
-            (parent, FileType::Directory, "..".to_string()),
-        ];
-        if let Some(kids) = st.children.get(&ino) {
-            for &kid in kids {
-                if let Some(e) = st.entries.get(&kid) {
-                    let ft = if e.node.is_folder() {
-                        FileType::Directory
-                    } else {
-                        FileType::RegularFile
-                    };
-                    listing.push((kid, ft, e.node.name.clone()));
-                }
-            }
-        }
+        let listing = build_listing(&st, ino);
         drop(st);
-
-        for (i, (ino, ft, name)) in listing.into_iter().enumerate().skip(offset as usize) {
-            // The stored offset is that of the *next* entry to resume at.
-            if reply.add(INodeNo(ino), (i + 1) as u64, ft, &name) {
-                break;
-            }
-        }
-        reply.ok();
+        reply_listing(&listing, offset, reply);
     }
 
     fn attr(&self, ino: u64, node: &Node, access: Access) -> FileAttr {
@@ -1142,66 +1237,101 @@ impl ProtonFs {
         // any network access. An incomplete blob still has gaps referring to
         // the remote base; stacking another write on it cannot be represented
         // safely by WriteHandle, so fail the open rather than risk corruption.
-        let pending_base = self.core.pending.lock().get(&uid).cloned();
-        if pending_base.as_ref().is_some_and(|p| !p.meta.complete) {
-            if let Some(entry) = self.core.state().entries.get_mut(&ino.0) {
-                entry.open_count = entry.open_count.saturating_sub(1);
-            }
-            error!(%uid, "refusing write over incomplete queued revision");
-            reply.error(Errno::EIO);
-            return;
-        }
-        let (file, path) = match self.core.cache.create_scratch() {
-            Ok(x) => x,
-            Err(e) => {
-                if let Some(entry) = self.core.state().entries.get_mut(&ino.0) {
-                    entry.open_count = entry.open_count.saturating_sub(1);
+        //
+        // Reading that blob and installing the handle cannot be done under one
+        // lock — the copy is the size of the file — so the base is re-checked at
+        // the moment the handle goes in, and a `release` that published a newer
+        // revision in between sends this around again. Without the re-check the
+        // new handle carried the *older* base, and its own release published
+        // content that silently discarded the revision closed a moment earlier
+        // (bugs.md B32). A second attempt is already unlikely and a third would
+        // mean something is republishing continuously, so give up rather than
+        // spin: the caller's `open` fails instead of its data being lost.
+        //
+        // `reply` is consumed by whichever arm answers, so the loop yields an
+        // outcome and the answer is sent once, after it.
+        let outcome = 'install: {
+            for attempt in 0..OPEN_BASE_ATTEMPTS {
+                let pending_base = self.core.pending.lock().get(&uid).cloned();
+                if pending_base.as_ref().is_some_and(|p| !p.meta.complete) {
+                    error!(%uid, "refusing write over incomplete queued revision");
+                    break 'install Err(Errno::EIO);
                 }
-                error!(%uid, error = %e, "create scratch file failed");
-                reply.error(Errno::EIO);
-                return;
+                let (file, path) = match self.core.cache.create_scratch() {
+                    Ok(x) => x,
+                    Err(e) => {
+                        error!(%uid, error = %e, "create scratch file failed");
+                        break 'install Err(Errno::EIO);
+                    }
+                };
+                let mut initial_written = Intervals::default();
+                if let Some(pending) = &pending_base {
+                    // Reflink where the filesystem can: this blob is the whole
+                    // file, and a byte-for-byte copy of a multi-GB pending
+                    // revision is the kind of work `open` must not do
+                    // (`ContentCache::clone_or_copy`).
+                    if let Err(e) = ContentCache::clone_or_copy(&pending.path, &path) {
+                        let _ = std::fs::remove_file(&path);
+                        error!(%uid, source = %pending.path.display(), error = %e,
+                            "copy queued revision into write scratch failed");
+                        break 'install Err(Errno::EIO);
+                    }
+                    initial_written.add(0, pending.meta.len);
+                }
+                // Sampled *before* the state lock, never under it: no site in
+                // the daemon holds `pending` and `state` at once (see
+                // `stamp_pending_sizes`) and this must not become the first.
+                // What that costs is a window of a few instructions instead of
+                // a lock-tight comparison — against the multi-gigabyte copy
+                // above, which is the window that actually loses data.
+                let current = self.core.pending_blob(&uid);
+                let mut st = self.core.state();
+                // Only a handle this open is about to *create* can carry a stale
+                // base; if one already exists, that handle is authoritative and
+                // the scratch prepared above is discarded either way.
+                let creating = !st.active_writes.contains_key(&ino.0);
+                if creating && current != pending_base.map(|p| p.path) {
+                    drop(st);
+                    let _ = std::fs::remove_file(&path);
+                    warn!(%uid, attempt, "queued revision changed while opening for write; retrying");
+                    continue;
+                }
+                let fh = st.next_fh;
+                st.next_fh += 1;
+                let aw = st.active_writes.entry(ino.0).or_insert_with(|| {
+                    WriteHandle {
+                        ino: ino.0,
+                        uid: uid.clone(),
+                        file: Arc::new(file),
+                        path,
+                        written: initial_written,
+                        // Starts at the current size; reads in [0, base_size)
+                        // come from the base until overwritten.
+                        len: base_size,
+                        base_size,
+                        base_mtime,
+                        base_revision_id,
+                        dirty: false,
+                        open_count: 0,
+                    }
+                });
+                aw.open_count += 1;
+                st.handles.insert(fh, ino.0);
+                break 'install Ok(fh);
             }
+            error!(%uid, "queued revision kept changing while opening for write");
+            Err(Errno::EIO)
         };
-        let mut initial_written = Intervals::default();
-        if let Some(pending) = &pending_base {
-            // Reflink where the filesystem can: this blob is the whole file, and
-            // a byte-for-byte copy of a multi-GB pending revision is the kind of
-            // work `open` must not do (`ContentCache::clone_or_copy`).
-            if let Err(e) = ContentCache::clone_or_copy(&pending.path, &path) {
+        match outcome {
+            Ok(fh) => reply.opened(FileHandle(fh), FopenFlags::empty()),
+            Err(error) => {
+                // The lookup count was raised before any of this could fail.
                 if let Some(entry) = self.core.state().entries.get_mut(&ino.0) {
                     entry.open_count = entry.open_count.saturating_sub(1);
                 }
-                let _ = std::fs::remove_file(&path);
-                error!(%uid, source = %pending.path.display(), error = %e,
-                    "copy queued revision into write scratch failed");
-                reply.error(Errno::EIO);
-                return;
+                reply.error(error);
             }
-            initial_written.add(0, pending.meta.len);
         }
-        let mut st = self.core.state();
-        let fh = st.next_fh;
-        st.next_fh += 1;
-        let aw = st.active_writes.entry(ino.0).or_insert_with(|| {
-            WriteHandle {
-                ino: ino.0,
-                uid,
-                file: Arc::new(file),
-                path,
-                written: initial_written,
-                // Starts at the current size; reads in [0, base_size) come from
-                // the base until overwritten.
-                len: base_size,
-                base_size,
-                base_mtime,
-                base_revision_id,
-                dirty: false,
-                open_count: 0,
-            }
-        });
-        aw.open_count += 1;
-        st.handles.insert(fh, ino.0);
-        reply.opened(FileHandle(fh), FopenFlags::empty());
     }
 
     /// The body of [`Filesystem::fsync`], run off the dispatch loop.

crates/pdfs-fuse/src/lib.rs

@@ -55,7 +55,7 @@ use fuser::{
 };
 use futures::StreamExt as _;
 use pdfs_core::batch;
-use pdfs_core::cache::{BLOCK_SIZE, Baseline, ContentCache, StagedWrite};
+use pdfs_core::cache::{BLOCK_SIZE, Baseline, BlockGeometry, BlockSpan, ContentCache, StagedWrite};
 use pdfs_core::config::{AppDirs, SweepMode};
 use pdfs_core::control::{
     ActivityEntry, ActivityKind, DirEntry, ErrorKind, LocalHit, PhotoKind, PublicLinkInfo,
@@ -92,6 +92,7 @@ mod profile;
 mod reads;
 mod revisions;
 mod sharing;
+mod shutdown;
 mod state;
 mod sweep;
 mod sync;
@@ -410,6 +411,10 @@ struct Core {
     /// kernel read-ahead, our own prefetch) share a single download and decrypt.
     /// See [`BlockFlight`].
     block_flight: Arc<BlockFlight>,
+    /// One lock per file being repaired by [`Core::repair_block`], so the
+    /// concurrent block reads that all discover the same untrustworthy block
+    /// table share a single whole-file download instead of one each.
+    content_repairs: Arc<tokio::sync::Mutex<HashMap<NodeUid, Arc<tokio::sync::Mutex<()>>>>>,
     /// Per-file sequential-read detection driving [`Core::prefetch`].
     prefetch: Arc<Prefetch>,
     /// Permits bounding speculative block fetches, so prefetch cannot queue in
@@ -450,6 +455,9 @@ struct Core {
     /// Nudges the drain workers: set true and notify to have them re-examine
     /// the queue instead of waiting out their backoff.
     drain_wake: Arc<(Mutex<bool>, Condvar)>,
+    /// The daemon's stop signal, shared with every background loop this `Core`
+    /// was cloned into so teardown can end and join them (bugs.md B44).
+    shutdown: Arc<shutdown::Shutdown>,
     /// How long this node's last upload actually took, in milliseconds, with
     /// the time it was measured for eviction. Feeds
     /// [`Core::revision_debounce`]: a file that takes 30 seconds to send needs
@@ -707,6 +715,26 @@ struct MountedState {
     /// [`Core::notifier`], still empty until its session is spawned.
     notifier: Arc<OnceLock<Notifier>>,
     session_live: Arc<AtomicBool>,
+    /// This mount's in-flight size upgrades, keyed by *its* folder inodes — so
+    /// a `Core` rebuilt onto this inode space by [`Core::rooted_at`] shares the
+    /// batch the mount's own session is already running rather than starting a
+    /// second one against inode numbers that mean nothing to it.
+    size_upgrades: Arc<Mutex<HashMap<u64, Arc<SizeUpgrade>>>>,
+}
+
+/// The per-mount half of a [`Core`] — every field [`Core::fork_state`] gives a
+/// fork a fresh copy of — recovered from the registry.
+///
+/// This is what lets a control request naming a path under a secondary mount be
+/// answered *in that mount's inode space* instead of being rejected for not
+/// being under the primary mountpoint (`docs/BUGS.md` B86). See
+/// [`Core::rooted_at`].
+struct MountParts {
+    mountpoint: PathBuf,
+    state: Arc<Mutex<State>>,
+    notifier: Arc<OnceLock<Notifier>>,
+    session_live: Arc<AtomicBool>,
+    size_upgrades: Arc<Mutex<HashMap<u64, Arc<SizeUpgrade>>>>,
 }
 
 /// Every mounted inode space in the daemon, shared by the primary `Core` and
@@ -729,6 +757,7 @@ impl StateRegistry {
         state: &Arc<Mutex<State>>,
         notifier: Arc<OnceLock<Notifier>>,
         session_live: Arc<AtomicBool>,
+        size_upgrades: Arc<Mutex<HashMap<u64, Arc<SizeUpgrade>>>>,
     ) {
         let mut states = self.0.lock();
         states.retain(|m| m.state.strong_count() > 0);
@@ -743,6 +772,7 @@ impl StateRegistry {
             state: Arc::downgrade(state),
             notifier,
             session_live,
+            size_upgrades,
         });
     }
 
@@ -769,9 +799,9 @@ impl StateRegistry {
     ///
     /// On-demand roots can be nested below another location. Selecting by the
     /// longest component prefix ensures the nested session wins rather than the
-    /// broader primary mount. Callers currently need the selected mount only;
-    /// a future path router can derive a relative suffix with
-    /// `path.strip_prefix(mountpoint)` without storing a second path value here.
+    /// broader primary mount. Callers that need the relative suffix derive it
+    /// with `path.strip_prefix(mountpoint)`; [`StateRegistry::covering_parts`]
+    /// is the variant that also hands back enough to *serve* a request there.
     fn covering(&self, path: &Path) -> Option<LiveMount> {
         self.live()
             .into_iter()
@@ -781,6 +811,48 @@ impl StateRegistry {
             .max_by_key(|(mountpoint, _, _, _)| mountpoint.components().count())
     }
 
+    /// [`StateRegistry::register`] with an empty size-upgrade map, for the
+    /// tests that exercise the reap-and-route rules and nothing else.
+    #[cfg(test)]
+    fn register_bare(
+        &self,
+        mountpoint: &Path,
+        state: &Arc<Mutex<State>>,
+        notifier: Arc<OnceLock<Notifier>>,
+        session_live: Arc<AtomicBool>,
+    ) {
+        self.register(
+            mountpoint,
+            state,
+            notifier,
+            session_live,
+            Arc::new(Mutex::new(HashMap::new())),
+        );
+    }
+
+    /// [`StateRegistry::covering`], but returning every per-mount field a
+    /// [`Core`] needs to be re-rooted onto that inode space.
+    ///
+    /// Kept separate from `covering` so the common callers — which only ask
+    /// *whether* a path is mounted — keep the cheaper tuple.
+    fn covering_parts(&self, path: &Path) -> Option<MountParts> {
+        let mut states = self.0.lock();
+        states.retain(|m| m.state.strong_count() > 0);
+        states
+            .iter()
+            .filter(|m| m.session_live.load(Ordering::Acquire) && path.starts_with(&m.mountpoint))
+            .max_by_key(|m| m.mountpoint.components().count())
+            .and_then(|m| {
+                Some(MountParts {
+                    mountpoint: m.mountpoint.clone(),
+                    state: m.state.upgrade()?,
+                    notifier: m.notifier.clone(),
+                    session_live: m.session_live.clone(),
+                    size_upgrades: m.size_upgrades.clone(),
+                })
+            })
+    }
+
     fn is_mounted_at(&self, path: &Path) -> bool {
         self.covering(path).is_some_and(|(mountpoint, _, _, live)| {
             mountpoint == path && live.load(Ordering::Acquire)
@@ -900,9 +972,41 @@ impl Core {
             &self.state,
             self.notifier.clone(),
             self.session_live.clone(),
+            self.size_upgrades.clone(),
         );
     }
 
+    /// This `Core`, re-rooted onto whichever live mount most specifically covers
+    /// the absolute path `abs`, together with `abs` relative to that mount's
+    /// root.
+    ///
+    /// Everything a `Core` holds is per-daemon except the five fields
+    /// [`Core::fork_state`] replaces, so swapping exactly those turns the
+    /// primary `Core` into the fork that owns the path — sharing the fork's
+    /// inode space, notification channel and size-upgrade batches rather than
+    /// shadowing them. Without this, every path-addressed control request is
+    /// answered against the primary inode space, so anything under a secondary
+    /// on-demand mount is rejected as "not under the mountpoint" and the user
+    /// has no way to re-enumerate a folder they can see (`docs/BUGS.md` B86).
+    ///
+    /// `None` when no live mount covers `abs`.
+    fn rooted_at(&self, abs: &Path) -> Option<(Core, PathBuf)> {
+        let parts = self.states.covering_parts(abs)?;
+        let rel = abs.strip_prefix(&parts.mountpoint).ok()?.to_path_buf();
+        let mut core = self.clone();
+        if !Arc::ptr_eq(&parts.state, &self.state) {
+            // A different inode space, so adopt its per-mount half wholesale.
+            // `primary` is a property of the mount, and the only mount this
+            // `Core` can be the primary of is the one it was built for.
+            core.primary = false;
+            core.state = parts.state;
+            core.notifier = parts.notifier;
+            core.session_live = parts.session_live;
+            core.size_upgrades = parts.size_upgrades;
+        }
+        Some((core, rel))
+    }
+
     /// Run `apply` against every live inode space — this daemon's primary mount
     /// and each on-demand fork — one lock at a time.
     ///
@@ -1708,7 +1812,9 @@ impl Core {
     fn run_online_probe(&self) {
         let mut delay = ONLINE_PROBE_MIN;
         loop {
-            std::thread::sleep(delay);
+            if !self.shutdown.sleep(delay) {
+                return;
+            }
             match self.rt.block_on(self.client.get_my_files_folder()) {
                 Ok(root) => {
                     {
@@ -3225,6 +3331,15 @@ impl Core {
         Ok(())
     }
 
+    /// Which staged blob is currently queued for `uid`, if any.
+    ///
+    /// The path identifies the revision: publication stages a fresh file and
+    /// swaps the map entry to point at it, so an unchanged path means the base a
+    /// caller sampled earlier is still the base now (bugs.md B32).
+    fn pending_blob(&self, uid: &NodeUid) -> Option<PathBuf> {
+        self.pending.lock().get(uid).map(|p| p.path.clone())
+    }
+
     /// Nudge the drain worker to re-examine the queue now.
     fn wake_drain(&self) {
         let (lock, cv) = &*self.drain_wake;
@@ -3915,6 +4030,33 @@ impl Core {
         }
     }
 
+    /// Drop the cached child listing of the folder `parent`, everywhere it is
+    /// held — naming the folder by uid rather than by inode.
+    ///
+    /// [`Core::invalidate_parent_listing`] is the path-based sibling, and it
+    /// only reaches *this* mount's inode space. A remote mutation made by the
+    /// sync engine needs more than that, for two reasons. The folder it changed
+    /// is usually not resident in any inode space at the time — a mirror folder
+    /// has no FUSE session at all — so there is no inode to name. And the stale
+    /// listing that outlives the pass is not a hot-cache map but the `listed`
+    /// flag on the folder's DB row, which survives a daemon restart and is what
+    /// a later `ensure_children` trusts.
+    ///
+    /// So the flag is cleared unconditionally, and the resident inode spaces are
+    /// invalidated on top of it. That is what makes a mirror folder switched to
+    /// on-demand re-enumerate instead of serving a snapshot that predates the
+    /// engine's own uploads — `docs/BUGS.md` B86.
+    pub(crate) fn invalidate_children_of(&self, parent: &NodeUid) {
+        if let Err(e) = self.db.set_listed(parent, false) {
+            warn!(uid = %parent, error = ?e, "could not clear cached listing");
+        }
+        self.for_each_state(|st| {
+            if let Some(&ino) = st.by_uid.get(parent) {
+                st.invalidate_listing(ino);
+            }
+        });
+    }
+
     /// Rename a file or folder to `new_name`. `rel` is mountpoint-relative.
     /// Mirrors the FUSE `rename` write path: rename on the remote, forget the
     /// node so it re-interns under its new name, and drop the parent listing so
@@ -5834,7 +5976,7 @@ mod tests {
             ("/mnt/fork-a", &fork_a),
             ("/mnt/fork-b", &fork_b),
         ] {
-            registry.register(
+            registry.register_bare(
                 std::path::Path::new(mountpoint),
                 state,
                 std::sync::Arc::new(std::sync::OnceLock::new()),
@@ -5885,19 +6027,19 @@ mod tests {
         let nested_live = session_flag(false);
         let sibling_live = session_flag(true);
 
-        registry.register(
+        registry.register_bare(
             primary_path,
             &primary,
             std::sync::Arc::new(std::sync::OnceLock::new()),
             primary_live.clone(),
         );
-        registry.register(
+        registry.register_bare(
             nested_path,
             &nested,
             std::sync::Arc::new(std::sync::OnceLock::new()),
             nested_live.clone(),
         );
-        registry.register(
+        registry.register_bare(
             sibling_path,
             &sibling,
             std::sync::Arc::new(std::sync::OnceLock::new()),
@@ -5975,6 +6117,71 @@ mod tests {
         assert_eq!(registry.live().len(), 2);
     }
 
+    /// The routing half of `docs/BUGS.md` B86: a path under a secondary mount
+    /// has to resolve to *that* mount's inode space and a suffix relative to
+    /// *its* root, or `pdfs ls` and `pdfs refresh` can only ever name the
+    /// primary mount and a stale on-demand folder has no escape hatch.
+    #[test]
+    fn covering_parts_route_a_path_to_the_mount_that_owns_it() {
+        let registry = StateRegistry::default();
+        let (primary, _primary_dir) = rooted_state("primary-volume", "primary");
+        let (nested, _nested_dir) = rooted_state("nested-volume", "nested");
+        let primary = std::sync::Arc::new(parking_lot::Mutex::new(primary));
+        let nested = std::sync::Arc::new(parking_lot::Mutex::new(nested));
+        let primary_path = std::path::Path::new("/home/me/ProtonDrive");
+        let nested_path = std::path::Path::new("/home/me/ProtonDrive/Device");
+        let primary_upgrades = std::sync::Arc::new(parking_lot::Mutex::new(HashMap::new()));
+        let nested_upgrades = std::sync::Arc::new(parking_lot::Mutex::new(HashMap::new()));
+
+        registry.register(
+            primary_path,
+            &primary,
+            std::sync::Arc::new(std::sync::OnceLock::new()),
+            session_flag(true),
+            primary_upgrades.clone(),
+        );
+        registry.register(
+            nested_path,
+            &nested,
+            std::sync::Arc::new(std::sync::OnceLock::new()),
+            session_flag(true),
+            nested_upgrades.clone(),
+        );
+
+        let deep = nested_path.join("sub/file.txt");
+        let parts = registry.covering_parts(&deep).expect("nested mount covers");
+        assert_eq!(parts.mountpoint, nested_path, "the nested mount wins");
+        assert!(
+            std::sync::Arc::ptr_eq(&parts.state, &nested),
+            "the request must be answered in the nested inode space"
+        );
+        assert!(
+            std::sync::Arc::ptr_eq(&parts.size_upgrades, &nested_upgrades),
+            "and share that mount's in-flight size upgrades, not the primary's"
+        );
+        assert_eq!(
+            deep.strip_prefix(&parts.mountpoint).unwrap(),
+            std::path::Path::new("sub/file.txt"),
+            "the suffix is relative to the mount that owns it, not the primary"
+        );
+
+        let shallow = primary_path.join("Documents");
+        let parts = registry
+            .covering_parts(&shallow)
+            .expect("primary mount covers");
+        assert!(
+            std::sync::Arc::ptr_eq(&parts.state, &primary),
+            "a path outside the nested root still routes to the primary"
+        );
+
+        assert!(
+            registry
+                .covering_parts(std::path::Path::new("/tmp/outside"))
+                .is_none(),
+            "a path under no mount routes nowhere, and the caller reports that"
+        );
+    }
+
     #[test]
     fn state_registry_resolves_uids_from_any_resident_mount() {
         let registry = StateRegistry::default();
@@ -5988,13 +6195,13 @@ mod tests {
         fork.children.insert(super::ROOT_INO, vec![fork_ino]);
         let primary = std::sync::Arc::new(parking_lot::Mutex::new(primary));
         let fork = std::sync::Arc::new(parking_lot::Mutex::new(fork));
-        registry.register(
+        registry.register_bare(
             std::path::Path::new("/mnt/primary"),
             &primary,
             std::sync::Arc::new(std::sync::OnceLock::new()),
             session_flag(true),
         );
-        registry.register(
+        registry.register_bare(
             std::path::Path::new("/mnt/device"),
             &fork,
             std::sync::Arc::new(std::sync::OnceLock::new()),
@@ -6022,7 +6229,7 @@ mod tests {
         let child_ino = state.intern(super::ROOT_INO, child_node);
         state.children.insert(super::ROOT_INO, vec![child_ino]);
         let state = std::sync::Arc::new(parking_lot::Mutex::new(state));
-        registry.register(
+        registry.register_bare(
             std::path::Path::new("/mnt/device"),
             &state,
             std::sync::Arc::new(std::sync::OnceLock::new()),
@@ -6085,7 +6292,7 @@ mod tests {
         );
 
         let state = std::sync::Arc::new(parking_lot::Mutex::new(state));
-        registry.register(
+        registry.register_bare(
             std::path::Path::new("/mnt/device"),
             &state,
             std::sync::Arc::new(std::sync::OnceLock::new()),
@@ -6353,7 +6560,7 @@ mod tests {
         state.children.insert(revoked_ino, vec![child_ino]);
 
         let state = std::sync::Arc::new(parking_lot::Mutex::new(state));
-        registry.register(
+        registry.register_bare(
             std::path::Path::new("/mnt/device"),
             &state,
             std::sync::Arc::new(std::sync::OnceLock::new()),
@@ -6407,7 +6614,7 @@ mod tests {
         state.children.insert(cycle_b, vec![cycle_a]);
 
         let state = std::sync::Arc::new(parking_lot::Mutex::new(state));
-        registry.register(
+        registry.register_bare(
             std::path::Path::new("/mnt/device"),
             &state,
             std::sync::Arc::new(std::sync::OnceLock::new()),

crates/pdfs-fuse/src/mount.rs

@@ -365,6 +365,7 @@ pub fn mount(
         readers: Arc::new(Mutex::new(HashMap::new())),
         block_ring: Arc::new(Mutex::new(BlockRing::default())),
         block_flight: Arc::new(BlockFlight::default()),
+        content_repairs: Arc::new(tokio::sync::Mutex::new(HashMap::new())),
         prefetch: Arc::new(Prefetch::default()),
         prefetch_budget: Arc::new(tokio::sync::Semaphore::new(PREFETCH_BUDGET)),
         workers: Arc::new(Workers::new(FUSE_WORKERS)?),
@@ -376,6 +377,7 @@ pub fn mount(
         pending: Arc::new(Mutex::new(HashMap::new())),
         hidden: Arc::new(Mutex::new(HashSet::new())),
         drain_wake: Arc::new((Mutex::new(false), Condvar::new())),
+        shutdown: Arc::new(crate::shutdown::Shutdown::default()),
         upload_times: Arc::new(Mutex::new(HashMap::new())),
         upload_cancel: Arc::new(Mutex::new(HashMap::new())),
         timeline_refreshing: Arc::new(AtomicBool::new(false)),
@@ -420,16 +422,24 @@ pub fn mount(
     // long as it ran; the claim column (`Db::claim_next_due_op`) keeps the
     // workers off each other's nodes, which is the only ordering the queue
     // actually needs. Worker 0 additionally runs the queue's idle chores.
+    //
+    // Every long-lived thread started from here on is kept in `workers` and
+    // joined at the end of this function. Leaving them running turned an
+    // in-process remount into a second generation of drain workers, sync engines
+    // and sweeps all operating on a mount that no longer exists (bugs.md B44).
+    let mut workers: Vec<std::thread::JoinHandle<()>> = Vec::new();
     for worker in 0..DRAIN_WORKERS {
         let core = core.clone();
-        std::thread::Builder::new()
-            .name(format!("pdfs-drain-{worker}"))
-            .spawn(move || core.run_pending_drain(worker == 0))?;
+        workers.push(
+            std::thread::Builder::new()
+                .name(format!("pdfs-drain-{worker}"))
+                .spawn(move || core.run_pending_drain(worker == 0))?,
+        );
     }
 
     // Start the folder-sync engine. It watches every mirror folder, polls the
     // remotes, and reconciles on its own thread — never in front of a FUSE call.
-    sync::spawn(core.clone(), sync_rx);
+    workers.extend(sync::spawn(core.clone(), sync_rx));
 
     // Reconcile leftover `(sync-conflict …)` copies: drop the ones that turned
     // out identical to the live file, surface the ones that genuinely diverge.
@@ -441,18 +451,22 @@ pub fn mount(
     } else {
         info!(mode = ?sweep_mode, "conflict sweep enabled");
         let core = core.clone();
-        std::thread::Builder::new()
-            .name("pdfs-conflict-sweep".into())
-            .spawn(move || core.run_conflict_sweep_loop())?;
+        workers.push(
+            std::thread::Builder::new()
+                .name("pdfs-conflict-sweep".into())
+                .spawn(move || core.run_conflict_sweep_loop())?,
+        );
     }
 
     // Mounted from the cache: watch for the network coming back so the mount can
     // stop being read-only-ish without the user restarting the daemon.
     if !online {
         let core = core.clone();
-        std::thread::Builder::new()
-            .name("pdfs-online-probe".into())
-            .spawn(move || core.run_online_probe())?;
+        workers.push(
+            std::thread::Builder::new()
+                .name("pdfs-online-probe".into())
+                .spawn(move || core.run_online_probe())?,
+        );
     }
 
     // Keep the launcher prompt's "This computer" index fresh. Its own thread:
@@ -462,9 +476,12 @@ pub fn mount(
         let indexing = core.indexing.clone();
         let transfers = core.transfers.clone();
         let mountpoint = mountpoint.to_path_buf();
-        std::thread::Builder::new()
-            .name("pdfs-localindex".into())
-            .spawn(move || run_local_index(db, indexing, transfers, mountpoint))?;
+        let shutdown = core.shutdown.clone();
+        workers.push(
+            std::thread::Builder::new()
+                .name("pdfs-localindex".into())
+                .spawn(move || run_local_index(db, indexing, transfers, mountpoint, shutdown))?,
+        );
     }
 
     // Bind the control socket before the FUSE session takes over the thread. A
@@ -501,23 +518,29 @@ pub fn mount(
         let control_core = core.clone();
         let username = username.clone();
         let mountpoint = mountpoint.to_path_buf();
-        if let Err(error) = std::thread::Builder::new()
+        match std::thread::Builder::new()
             .name("pdfs-control".into())
             .spawn(move || run_control_socket(control_core, username, mountpoint, listener))
         {
-            let secondaries = core.mounts.lock().close_and_drain();
-            for (id, mount) in secondaries {
-                if let Err(unmount_error) = mount.teardown() {
-                    warn!(id, error = %unmount_error, "secondary teardown after control startup failure failed");
+            Ok(handle) => workers.push(handle),
+            Err(error) => {
+                // A failure here abandons the mount, so the background threads
+                // started above have to be told to stop with it.
+                stop_workers(&core, control_socket, workers);
+                let secondaries = core.mounts.lock().close_and_drain();
+                for (id, mount) in secondaries {
+                    if let Err(unmount_error) = mount.teardown() {
+                        warn!(id, error = %unmount_error, "secondary teardown after control startup failure failed");
+                    }
                 }
+                if let Err(unmount_error) = teardown_session(&core.session_live, || {
+                    umount_session_unblocked(bg, main_conn)
+                }) {
+                    warn!(error = %unmount_error, "unmount after control startup failure failed");
+                }
+                let _ = std::fs::remove_file(control_socket);
+                return Err(error);
             }
-            if let Err(unmount_error) = teardown_session(&core.session_live, || {
-                umount_session_unblocked(bg, main_conn)
-            }) {
-                warn!(error = %unmount_error, "unmount after control startup failure failed");
-            }
-            let _ = std::fs::remove_file(control_socket);
-            return Err(error);
         }
     }
     // Re-establish on-demand mounts only after the primary session is live.
@@ -626,10 +649,43 @@ pub fn mount(
         }
     }
 
-    let _ = std::fs::remove_file(control_socket);
+    // Last, because the workers may legitimately be finishing an upload, and
+    // nothing they can still do is harmful once the mounts are down.
+    stop_workers(&core, control_socket, workers);
     Ok(outcome)
 }
 
+/// Signal every background loop, wake the ones that are blocked, and join them.
+///
+/// The mount used to return with its drain workers, sync engine, sweep, online
+/// probe, indexer and control listener all still running on `Core` clones. The
+/// process usually exited straight afterwards and hid it; an in-process remount
+/// did not, and started a second full set on top of the first (bugs.md B44).
+///
+/// Two of the waits need waking rather than merely flagging: the drain workers
+/// block on their own condvar (`wake_drain`), and the control listener blocks
+/// inside `accept`, which only returns when a connection arrives — so it gets
+/// one, from us, after the flag is already set. A join that takes a while is
+/// expected and correct: a worker part-way through an upload finishes it rather
+/// than abandoning the user's bytes mid-flight.
+fn stop_workers(core: &Core, control_socket: &Path, workers: Vec<std::thread::JoinHandle<()>>) {
+    core.shutdown.stop();
+    core.wake_drain();
+    let _ = core.sync_tx.send(sync::SyncMsg::Stop);
+    // Connect while the socket still exists — that is the whole point of the
+    // poke — then unlink it so nothing else can arrive behind us. The listener
+    // re-checks the flag after every accept, so even a connection that races the
+    // unlink cannot put it back to sleep.
+    let _ = std::os::unix::net::UnixStream::connect(control_socket);
+    let _ = std::fs::remove_file(control_socket);
+    for worker in workers {
+        if worker.join().is_err() {
+            warn!("a background worker panicked before shutdown");
+        }
+    }
+    debug!("background workers stopped");
+}
+
 #[cfg(test)]
 mod tests {
     use std::sync::atomic::{AtomicBool, Ordering};

crates/pdfs-fuse/src/reads.rs

@@ -26,6 +26,11 @@ const PREFETCH_MAX: u64 = 8;
 /// Window a sequential read starts at, and returns to after a seek.
 const PREFETCH_MIN: u64 = 2;
 
+/// Largest file [`Core::repair_block`] will pull down whole to work around a
+/// revision whose block table understates it. Past this a single read would cost
+/// gigabytes of download, so the read fails instead and says to pin the file.
+pub(super) const REPAIR_MAX: u64 = 512 * 1024 * 1024;
+
 /// Blocks that may be prefetching at once, across every file. Prefetch takes a
 /// permit and gives up rather than queueing, so it can never sit in front of a
 /// demand read in the SDK's flat block semaphore.
@@ -97,22 +102,38 @@ static NEXT_OPEN_ID: AtomicU64 = AtomicU64::new(0);
 /// Validated by the same `(mtime, size)` pair the on-disk caches use: a node
 /// whose tag no longer matches has its blocks dropped rather than served, so a
 /// new revision can never be stitched together from the old one.
+/// A ring entry's identity: the node, and the plaintext range the block covers.
+///
+/// The range, rather than the block index, because the index alone stopped
+/// naming one range when the geometry started coming from the revision — a
+/// block cached under the [`BlockGeometry::uniform`] fallback and the block at
+/// the same index under the real geometry are different bytes
+/// (`docs/BUGS.md` B85). A range is the same identity under either plan, so a
+/// re-planned read misses rather than being served the wrong bytes.
+type RingKey = (NodeUid, BlockSpan);
+
 #[derive(Default)]
 pub(super) struct BlockRing {
-    blocks: HashMap<(NodeUid, u64), Arc<Vec<u8>>>,
+    blocks: HashMap<RingKey, Arc<Vec<u8>>>,
     /// Keys least-recently-used first, for eviction.
-    order: VecDeque<(NodeUid, u64)>,
+    order: VecDeque<RingKey>,
     /// Per-node validity tag, `(mtime, size)`.
     tags: HashMap<NodeUid, (i64, u64)>,
     bytes: u64,
 }
 
 impl BlockRing {
-    fn get(&mut self, uid: &NodeUid, mtime: i64, size: u64, idx: u64) -> Option<Arc<Vec<u8>>> {
+    fn get(
+        &mut self,
+        uid: &NodeUid,
+        mtime: i64,
+        size: u64,
+        span: BlockSpan,
+    ) -> Option<Arc<Vec<u8>>> {
         if self.tags.get(uid) != Some(&(mtime, size)) {
             return None;
         }
-        let key = (uid.clone(), idx);
+        let key = (uid.clone(), span);
         let hit = self.blocks.get(&key).cloned()?;
         // Move to the young end: a block being read repeatedly (the whole reason
         // this exists) must not age out under a sequential read passing through.
@@ -123,14 +144,21 @@ impl BlockRing {
         Some(hit)
     }
 
-    fn insert(&mut self, uid: &NodeUid, mtime: i64, size: u64, idx: u64, bytes: Arc<Vec<u8>>) {
+    fn insert(
+        &mut self,
+        uid: &NodeUid,
+        mtime: i64,
+        size: u64,
+        span: BlockSpan,
+        bytes: Arc<Vec<u8>>,
+    ) {
         if self.tags.get(uid) != Some(&(mtime, size)) {
             // Revision changed under us (or first sight): anything held for this
             // node describes the old one.
             self.drop_node(uid);
             self.tags.insert(uid.clone(), (mtime, size));
         }
-        let key = (uid.clone(), idx);
+        let key = (uid.clone(), span);
         if self.blocks.contains_key(&key) {
             return;
         }
@@ -171,8 +199,10 @@ impl BlockRing {
 /// The result of fetching one block, shared between everyone who wanted it.
 type BlockResult = Result<Arc<Vec<u8>>, Errno>;
 
-/// Identifies a block *of a revision*: node, validity tag, index.
-type BlockKey = (NodeUid, i64, u64, u64);
+/// Identifies a block *of a revision*: node, validity tag, and the plaintext
+/// range the block covers — see [`RingKey`] for why it is the range and not the
+/// index.
+type BlockKey = (NodeUid, i64, u64, BlockSpan);
 
 /// A claim on a block key: who owns the fetch, and where its result is published.
 type Claim = (u64, tokio::sync::watch::Receiver<Option<BlockResult>>);
@@ -412,11 +442,30 @@ impl Core {
         self.prefetch.streams.lock().remove(uid);
     }
 
+    /// How this revision's plaintext is divided into content blocks.
+    ///
+    /// The revision's own geometry when a previous read has learned and recorded
+    /// it, and the [`BLOCK_SIZE`] assumption when it has not — which is every
+    /// first read, because learning it means opening a `RevisionReader`, and
+    /// opening one per read is precisely the per-file key derivation the block
+    /// cache exists to avoid (bugs.md B12, B85).
+    ///
+    /// Being wrong here is not a correctness problem: `read_at` plans over the
+    /// real sizes and clamps to what was asked for either way. It costs a
+    /// straddling fetch and a cached block the next read will not ask for, and
+    /// both settle once the geometry is known.
+    fn block_geometry(&self, uid: &NodeUid, mtime: i64, fsize: u64) -> BlockGeometry {
+        self.cache
+            .block_geometry(uid, mtime, fsize)
+            .unwrap_or_else(|| BlockGeometry::uniform(fsize))
+    }
+
     /// Serve bytes `[offset, offset + len)` of `uid`'s active revision, hitting
     /// the on-disk caches before the network: a whole-file blob (pinned files)
-    /// first, then the block cache — fetching only the [`BLOCK_SIZE`]-aligned
-    /// blocks that overlap the request and caching each. `mtime`/`fsize` validate
-    /// both caches. Network I/O runs without any lock held.
+    /// first, then the block cache — fetching only the blocks that overlap the
+    /// request, on the revision's own boundaries ([`Core::block_geometry`]), and
+    /// caching each. `mtime`/`fsize` validate both caches. Network I/O runs
+    /// without any lock held.
     pub(super) fn read_range(
         &self,
         uid: &NodeUid,
@@ -495,7 +544,7 @@ impl Core {
         Ok(out)
     }
 
-    /// Fetch block `bidx`, joining a fetch already in flight for it rather than
+    /// Fetch the block at `span`, joining a fetch already in flight rather than
     /// issuing a second one, and publish the result to everyone who joined.
     ///
     /// See [`BlockFlight`]. A follower whose leader disappears (a panic, or a
@@ -507,10 +556,10 @@ impl Core {
         uid: &NodeUid,
         mtime: i64,
         fsize: u64,
-        bidx: u64,
+        span: BlockSpan,
         cache_blocks: bool,
     ) -> BlockResult {
-        let key: BlockKey = (uid.clone(), mtime, fsize, bidx);
+        let key: BlockKey = (uid.clone(), mtime, fsize, span);
         let claim = {
             let mut inflight = self.block_flight.inflight.lock();
             match inflight.get(&key) {
@@ -542,12 +591,12 @@ impl Core {
                     }
                 }
                 return self
-                    .read_block(reader, uid, mtime, fsize, bidx, cache_blocks)
+                    .read_block(reader, uid, mtime, fsize, span, cache_blocks)
                     .await;
             }
         };
         let result = self
-            .read_block(reader, uid, mtime, fsize, bidx, cache_blocks)
+            .read_block(reader, uid, mtime, fsize, span, cache_blocks)
             .await;
         // Retire the claim before publishing, so a racer that wakes and re-enters
         // finds no stale entry to join.
@@ -563,36 +612,193 @@ impl Core {
     /// The disk store is handed to a blocking thread — it is a 4 MiB write plus
     /// a cache-index update, and neither the reactor nor the kernel read waiting
     /// on these bytes should be held up by it.
+    ///
+    /// A block that comes back shorter than the file's size says it must be is
+    /// never served, cached or promoted: it is repaired from the whole-file
+    /// download, or the read fails. See [`Core::repair_block`] (bugs.md B84).
     async fn read_block(
         &self,
         reader: &Arc<RevisionReader>,
         uid: &NodeUid,
         mtime: i64,
         fsize: u64,
-        bidx: u64,
+        span: BlockSpan,
         cache_blocks: bool,
     ) -> BlockResult {
-        let bstart = bidx * BLOCK_SIZE;
-        let blen = BLOCK_SIZE.min(fsize - bstart);
-        let bytes = reader.read_at(bstart, blen).await.map_err(|e| {
-            warn!(%uid, bstart, blen, error = %e, "block read failed");
-            Errno::EIO
-        })?;
+        // The reader derives every block boundary from the revision's recorded
+        // block sizes. If they do not add up to the size the node reports, this
+        // file's boundaries are wrong, and the failure is not confined to short
+        // reads: sizes that *overstate* a block shift every block after it and
+        // return full-length reads of the wrong bytes. So a disagreement of any
+        // kind takes the file off the range path entirely (bugs.md B84).
+        let bytes = if reader.size() != fsize {
+            warn!(
+                %uid, bidx = span.idx, fsize,
+                reader_size = reader.size(),
+                reader_blocks = reader.block_sizes().len(),
+                "the revision's block table disagrees with the node size"
+            );
+            self.repair_block(uid, mtime, fsize, span).await?
+        } else {
+            // Learned here rather than on a path of its own: this is where a
+            // reader is open anyway, and recording the real geometry is what
+            // lets the *next* read plan its blocks on the revision's own
+            // boundaries instead of the 4 MiB fallback (bugs.md B85).
+            if cache_blocks {
+                self.cache
+                    .store_block_geometry(uid, mtime, fsize, reader.block_sizes());
+            }
+            let bytes = reader.read_at(span.start, span.len).await.map_err(|e| {
+                warn!(%uid, start = span.start, len = span.len, error = %e, "block read failed");
+                Errno::EIO
+            })?;
+            if bytes.len() as u64 == span.len {
+                bytes
+            } else {
+                warn!(
+                    %uid, bidx = span.idx, blen = span.len, got = bytes.len(), fsize,
+                    "block came back shorter than the range asked for"
+                );
+                self.repair_block(uid, mtime, fsize, span).await?
+            }
+        };
         let bytes = Arc::new(bytes);
         if cache_blocks {
             let cache = self.cache.clone();
             let uid = uid.clone();
             let bytes = bytes.clone();
             tokio::task::spawn_blocking(move || {
-                let _ = cache.store_block(&uid, mtime, fsize, bidx, &bytes);
+                let _ = cache.store_block(&uid, mtime, fsize, span, &bytes);
             });
         }
         self.block_ring
             .lock()
-            .insert(uid, mtime, fsize, bidx, bytes.clone());
+            .insert(uid, mtime, fsize, span, bytes.clone());
         Ok(bytes)
     }
 
+    /// Recover the block at `span` of a file whose per-block reads cannot be
+    /// by fetching the whole revision through the manifest-verified download and
+    /// serving the block out of that.
+    ///
+    /// The range path derives every block boundary from the revision's recorded
+    /// block sizes. When those understate the file, reads come back short with
+    /// no error anywhere, and the caller hands userspace a truncated file that
+    /// looks like a clean EOF (bugs.md B84). `download_file_to` does not do that
+    /// arithmetic — it streams the block list — so it returns the whole file
+    /// where the range path cannot, which is why pinning was the workaround.
+    /// This does the same thing on the user's behalf, once, and the resulting
+    /// blob answers every later read of the file before the block path is
+    /// consulted at all.
+    ///
+    /// Bounded by [`REPAIR_MAX`]: past that, one unlucky read would pull down
+    /// gigabytes, and a clear `EIO` plus a `pdfs pin` is the better trade.
+    async fn repair_block(
+        &self,
+        uid: &NodeUid,
+        mtime: i64,
+        fsize: u64,
+        span: BlockSpan,
+    ) -> Result<Vec<u8>, Errno> {
+        let (bstart, blen) = (span.start, span.len);
+        if fsize > REPAIR_MAX {
+            warn!(
+                %uid, fsize,
+                "file is too large to repair from a whole-file download; pin it to read it"
+            );
+            return Err(Errno::EIO);
+        }
+        // One download per file, not per block: every block of a file with a bad
+        // block table discovers the same problem at once.
+        let lock = {
+            let mut locks = self.content_repairs.lock().await;
+            locks.entry(uid.clone()).or_default().clone()
+        };
+        let guard = lock.clone().lock_owned().await;
+        let downloaded = match self.cache.read_range(uid, mtime, fsize, bstart, blen) {
+            // The first arrival did the work; the rest just read its blob.
+            Some(bytes) if bytes.len() as u64 == blen => Ok(Some(bytes)),
+            _ => self
+                .download_whole_revision(uid, mtime, fsize)
+                .await
+                .map(|()| None),
+        };
+        // Before releasing, so the map entry is only dropped once no waiter is
+        // still holding it.
+        drop(guard);
+        self.release_repair_lock(uid, lock).await;
+        if let Some(bytes) = downloaded? {
+            return Ok(bytes);
+        }
+        match self.cache.read_range(uid, mtime, fsize, bstart, blen) {
+            Some(bytes) if bytes.len() as u64 == blen => {
+                info!(%uid, bidx = span.idx, "served a short block from a repaired whole-file download");
+                Ok(bytes)
+            }
+            _ => {
+                error!(%uid, bidx = span.idx, fsize, "repairing a short block did not produce the block");
+                Err(Errno::EIO)
+            }
+        }
+    }
+
+    /// Download the whole of `uid`'s active revision to disk and adopt it as the
+    /// cached blob, so the range path is bypassed entirely for this file.
+    ///
+    /// Streamed into a scratch file rather than memory: this runs for files up
+    /// to [`REPAIR_MAX`], and buffering that per concurrent reader is not a cost
+    /// a read should carry. A download that does not produce exactly `fsize`
+    /// bytes is discarded — a short blob would fail the cache's own validity
+    /// check anyway, and storing it would only hide the disagreement one layer
+    /// further down.
+    async fn download_whole_revision(
+        &self,
+        uid: &NodeUid,
+        mtime: i64,
+        fsize: u64,
+    ) -> Result<(), Errno> {
+        let (mut file, path) = self.cache.create_scratch().map_err(|error| {
+            error!(%uid, %error, "creating a scratch file for a repair download failed");
+            Errno::EIO
+        })?;
+        let outcome = self.client.download_file_to(uid, &mut file).await;
+        let stored = (|| {
+            outcome.map_err(|error| {
+                warn!(%uid, %error, "repair download failed");
+                Errno::EIO
+            })?;
+            file.sync_all().map_err(|error| {
+                warn!(%uid, %error, "syncing a repair download failed");
+                Errno::EIO
+            })?;
+            let got = file.metadata().map(|m| m.len()).unwrap_or(0);
+            if got != fsize {
+                warn!(%uid, fsize, got, "repair download did not match the node size");
+                return Err(Errno::EIO);
+            }
+            self.cache
+                .store_file(uid, mtime, fsize, &path)
+                .map_err(|error| {
+                    error!(%uid, %error, "adopting a repair download as cached content failed");
+                    Errno::EIO
+                })
+        })();
+        // The blob is a hardlink to the same inode, so unlinking the scratch
+        // name leaves it intact.
+        let _ = std::fs::remove_file(&path);
+        stored
+    }
+
+    /// Drop a repair lock's map entry once nobody else holds it, so a file
+    /// repaired long ago does not keep a mutex alive forever.
+    async fn release_repair_lock(&self, uid: &NodeUid, lock: Arc<tokio::sync::Mutex<()>>) {
+        let mut locks = self.content_repairs.lock().await;
+        // Two: the map's and ours. Anything more means a waiter still needs it.
+        if Arc::strong_count(&lock) <= 2 {
+            locks.remove(uid);
+        }
+    }
+
     /// Warm the blocks after a sequential read into the caches, in the
     /// background, so the next read finds them instead of paying a round-trip.
     ///
@@ -619,17 +825,19 @@ impl Core {
         if depth == 0 {
             return;
         }
-        let last_block = fsize.saturating_sub(1) / BLOCK_SIZE;
-        let wanted: Vec<u64> = ((last + 1)..=(last + depth).min(last_block))
-            .filter(|&bidx| {
+        let geometry = self.block_geometry(uid, mtime, fsize);
+        let last_block = geometry.blocks().saturating_sub(1);
+        let wanted: Vec<BlockSpan> = ((last + 1)..=(last + depth).min(last_block))
+            .filter_map(|bidx| geometry.span(bidx))
+            .filter(|&span| {
                 // Blocks already in memory need nothing; blocks already on disk
                 // are cheap enough to serve that spending a round-trip to move
                 // them into the ring would be the wrong trade.
                 self.block_ring
                     .lock()
-                    .get(uid, mtime, fsize, bidx)
+                    .get(uid, mtime, fsize, span)
                     .is_none()
-                    && !(cache_blocks && self.cache.has_block(uid, mtime, fsize, bidx))
+                    && !(cache_blocks && self.cache.has_block(uid, mtime, fsize, span))
             })
             .collect();
         if wanted.is_empty() {
@@ -646,7 +854,7 @@ impl Core {
             // round-trips, and the reader catches up with the window before it
             // finishes.
             let mut set = tokio::task::JoinSet::new();
-            for bidx in wanted {
+            for span in wanted {
                 let Ok(permit) = core.prefetch_budget.clone().try_acquire_owned() else {
                     break; // the budget is spent; the demand path can have it
                 };
@@ -654,10 +862,10 @@ impl Core {
                 set.spawn(async move {
                     let _permit = permit;
                     if let Err(error) = core
-                        .fetch_block(&reader, &uid, mtime, fsize, bidx, cache_blocks)
+                        .fetch_block(&reader, &uid, mtime, fsize, span, cache_blocks)
                         .await
                     {
-                        debug!(%uid, bidx, ?error, "prefetch failed");
+                        debug!(%uid, bidx = span.idx, ?error, "prefetch failed");
                     }
                 });
             }
@@ -687,8 +895,13 @@ impl Core {
         }
         let end = offset.saturating_add(len).min(fsize);
         let mut out = Vec::with_capacity((end - offset) as usize);
-        let first = offset / BLOCK_SIZE;
-        let last = (end - 1) / BLOCK_SIZE;
+        // On the revision's own boundaries where they are known, and on the
+        // 4 MiB fallback where they are not (bugs.md B85).
+        let wanted = self.block_geometry(uid, mtime, fsize).spans(offset, end);
+        let (Some(&head), Some(&tail)) = (wanted.first(), wanted.last()) else {
+            return Ok(out);
+        };
+        let (first, last) = (head.idx, tail.idx);
 
         // Collect the blocks overlapping the request, serving any already cached
         // and fetching the rest concurrently. A multi-block read (e.g. a media
@@ -696,31 +909,31 @@ impl Core {
         // would otherwise stall on each block round-trip in turn; downloading the
         // misses in parallel saturates the connection and bounds latency at the
         // slowest single block instead of their sum.
-        let mut blocks: Vec<Option<Arc<Vec<u8>>>> = Vec::with_capacity((last - first + 1) as usize);
-        let mut misses: Vec<u64> = Vec::new();
-        for bidx in first..=last {
+        let mut blocks: Vec<Option<Arc<Vec<u8>>>> = Vec::with_capacity(wanted.len());
+        let mut misses: Vec<BlockSpan> = Vec::new();
+        for &span in &wanted {
             // The in-memory ring comes first on both paths: a streaming read has
             // no other cache at all, and a cached one would otherwise pay two
             // opens, a JSON parse and a 4 MiB read+alloc for every 128 KiB the
             // kernel asks for (audit T3).
-            let hit = self.block_ring.lock().get(uid, mtime, fsize, bidx);
+            let hit = self.block_ring.lock().get(uid, mtime, fsize, span);
             let hit = hit.or_else(|| {
                 if !cache_blocks {
                     return None;
                 }
-                let bytes = Arc::new(self.cache.cached_block(uid, mtime, fsize, bidx)?);
+                let bytes = Arc::new(self.cache.cached_block(uid, mtime, fsize, span)?);
                 // Promote it, so the other ~31 reads of this block are answered
                 // from memory.
                 self.block_ring
                     .lock()
-                    .insert(uid, mtime, fsize, bidx, bytes.clone());
+                    .insert(uid, mtime, fsize, span, bytes.clone());
                 Some(bytes)
             });
             match hit {
                 Some(b) => blocks.push(Some(b)),
                 None => {
                     blocks.push(None);
-                    misses.push(bidx);
+                    misses.push(span);
                 }
             }
         }
@@ -733,12 +946,12 @@ impl Core {
                 let reader = self.revision_reader(uid, mtime, fsize).await?;
 
                 let mut set = tokio::task::JoinSet::new();
-                for &bidx in &misses {
+                for &span in &misses {
                     let (core, reader, uid) = (self.clone(), reader.clone(), uid.clone());
                     set.spawn(async move {
-                        core.fetch_block(&reader, &uid, mtime, fsize, bidx, cache_blocks)
+                        core.fetch_block(&reader, &uid, mtime, fsize, span, cache_blocks)
                             .await
-                            .map(|bytes| (bidx, bytes))
+                            .map(|bytes| (span.idx, bytes))
                     });
                 }
                 let mut out = Vec::with_capacity(misses.len());
@@ -756,11 +969,22 @@ impl Core {
         // ahead of a sequential reader instead of stalling once it catches up.
         self.prefetch_ahead(uid, mtime, fsize, first, last, cache_blocks);
 
-        for (i, block) in blocks.into_iter().enumerate() {
-            let bidx = first + i as u64;
-            let bstart = bidx * BLOCK_SIZE;
+        for (span, block) in wanted.into_iter().zip(blocks) {
+            let bstart = span.start;
             // Every slot is populated: cache hits up front, misses by the fetch above.
             let block = block.expect("block fetched or cached");
+            // Both sources guarantee the length the geometry says the block has
+            // — the cache validates it on read, the network path in `read_block`
+            // — so this refuses rather than quietly contributing fewer bytes
+            // than the range covers, which is what turned a bad block into a
+            // truncated file (bugs.md B84).
+            if block.len() as u64 != span.len {
+                error!(
+                    %uid, bidx = span.idx, len = block.len(), expected = span.len, fsize,
+                    "assembling a read hit a short block"
+                );
+                return Err(Errno::EIO);
+            }
             let s = (offset.max(bstart) - bstart) as usize;
             let e = (end.min(bstart + block.len() as u64) - bstart) as usize;
             if s < e {
@@ -850,13 +1074,13 @@ impl Core {
             return None;
         }
         let mut out = Vec::with_capacity((end - offset) as usize);
-        for bidx in (offset / BLOCK_SIZE)..=((end - 1) / BLOCK_SIZE) {
-            let ring = self.block_ring.lock().get(uid, mtime, fsize, bidx);
+        for span in self.block_geometry(uid, mtime, fsize).spans(offset, end) {
+            let ring = self.block_ring.lock().get(uid, mtime, fsize, span);
             let block = match ring {
                 Some(bytes) => bytes,
-                None => Arc::new(self.cache.cached_block(uid, mtime, fsize, bidx)?),
+                None => Arc::new(self.cache.cached_block(uid, mtime, fsize, span)?),
             };
-            let bstart = bidx * BLOCK_SIZE;
+            let bstart = span.start;
             let s = (offset.max(bstart) - bstart) as usize;
             let e = (end.min(bstart + block.len() as u64) - bstart) as usize;
             if s >= e {
@@ -976,31 +1200,42 @@ mod block_ring_tests {
         RING_BYTES / BLOCK_SIZE
     }
 
+    /// Block `idx` under the uniform geometry — the shape these tests are
+    /// written against, since what they exercise is tagging and eviction rather
+    /// than block layout.
+    fn span(idx: u64) -> BlockSpan {
+        BlockSpan {
+            idx,
+            start: idx * BLOCK_SIZE,
+            len: BLOCK_SIZE,
+        }
+    }
+
     #[test]
     fn a_stored_block_is_served_back() {
         let mut ring = BlockRing::default();
         let node = uid("a");
-        ring.insert(&node, 7, 100, 3, block(0xab));
-        assert_eq!(ring.get(&node, 7, 100, 3).map(|b| b[0]), Some(0xab));
+        ring.insert(&node, 7, 100, span(3), block(0xab));
+        assert_eq!(ring.get(&node, 7, 100, span(3)).map(|b| b[0]), Some(0xab));
     }
 
     #[test]
     fn a_block_of_another_revision_is_never_served() {
         let mut ring = BlockRing::default();
         let node = uid("a");
-        ring.insert(&node, 7, 100, 0, block(1));
+        ring.insert(&node, 7, 100, span(0), block(1));
         // Same node, newer revision: the tag moved, so the old bytes go.
-        assert!(ring.get(&node, 8, 100, 0).is_none());
-        assert!(ring.get(&node, 7, 101, 0).is_none());
+        assert!(ring.get(&node, 8, 100, span(0)).is_none());
+        assert!(ring.get(&node, 7, 101, span(0)).is_none());
     }
 
     #[test]
     fn a_new_revision_drops_what_the_old_one_left() {
         let mut ring = BlockRing::default();
         let node = uid("a");
-        ring.insert(&node, 7, 100, 0, block(1));
-        ring.insert(&node, 8, 100, 0, block(2));
-        assert_eq!(ring.get(&node, 8, 100, 0).map(|b| b[0]), Some(2));
+        ring.insert(&node, 7, 100, span(0), block(1));
+        ring.insert(&node, 8, 100, span(0), block(2));
+        assert_eq!(ring.get(&node, 8, 100, span(0)).map(|b| b[0]), Some(2));
         assert_eq!(ring.bytes, BLOCK_SIZE);
     }
 
@@ -1009,13 +1244,13 @@ mod block_ring_tests {
         let mut ring = BlockRing::default();
         let node = uid("a");
         for idx in 0..capacity() {
-            ring.insert(&node, 1, u64::MAX, idx, block(idx as u8));
+            ring.insert(&node, 1, u64::MAX, span(idx), block(idx as u8));
         }
         // Re-reading block 0 makes block 1 the oldest instead.
-        assert!(ring.get(&node, 1, u64::MAX, 0).is_some());
-        ring.insert(&node, 1, u64::MAX, capacity(), block(0xff));
-        assert!(ring.get(&node, 1, u64::MAX, 0).is_some());
-        assert!(ring.get(&node, 1, u64::MAX, 1).is_none());
+        assert!(ring.get(&node, 1, u64::MAX, span(0)).is_some());
+        ring.insert(&node, 1, u64::MAX, span(capacity()), block(0xff));
+        assert!(ring.get(&node, 1, u64::MAX, span(0)).is_some());
+        assert!(ring.get(&node, 1, u64::MAX, span(1)).is_none());
         assert_eq!(ring.bytes, RING_BYTES);
     }
 
@@ -1023,9 +1258,9 @@ mod block_ring_tests {
     fn a_node_whose_last_block_ages_out_keeps_no_tag() {
         let mut ring = BlockRing::default();
         let (old, new) = (uid("old"), uid("new"));
-        ring.insert(&old, 1, u64::MAX, 0, block(1));
+        ring.insert(&old, 1, u64::MAX, span(0), block(1));
         for idx in 0..capacity() {
-            ring.insert(&new, 1, u64::MAX, idx, block(2));
+            ring.insert(&new, 1, u64::MAX, span(idx), block(2));
         }
         assert!(!ring.tags.contains_key(&old));
         assert!(ring.tags.contains_key(&new));
@@ -1035,11 +1270,11 @@ mod block_ring_tests {
     fn dropping_a_node_frees_only_its_own_bytes() {
         let mut ring = BlockRing::default();
         let (a, b) = (uid("a"), uid("b"));
-        ring.insert(&a, 1, 100, 0, block(1));
-        ring.insert(&b, 1, 100, 0, block(2));
+        ring.insert(&a, 1, 100, span(0), block(1));
+        ring.insert(&b, 1, 100, span(0), block(2));
         ring.drop_node(&a);
-        assert!(ring.get(&a, 1, 100, 0).is_none());
-        assert_eq!(ring.get(&b, 1, 100, 0).map(|x| x[0]), Some(2));
+        assert!(ring.get(&a, 1, 100, span(0)).is_none());
+        assert_eq!(ring.get(&b, 1, 100, span(0)).map(|x| x[0]), Some(2));
         assert_eq!(ring.bytes, BLOCK_SIZE);
     }
 }

crates/pdfs-fuse/src/shutdown.rs

@@ -0,0 +1,96 @@
+//! One shutdown signal shared by every long-lived worker in the daemon.
+//!
+//! The mount used to end by unmounting and returning, leaving the drain
+//! workers, the sync engine, the conflict sweep, the online probe and the local
+//! indexer running on `Core` clones. In a process that exits immediately that is
+//! invisible; in an in-process remount — which is how the daemon recovers from a
+//! mount going away underneath it — it leaks a whole generation of threads that
+//! keep writing to the database and calling the API on behalf of a mount that no
+//! longer exists (bugs.md B44).
+//!
+//! The primitive is deliberately small: a flag plus a condvar. Every loop that
+//! would otherwise call `thread::sleep` calls [`Shutdown::sleep`] instead, which
+//! returns early — and reports `false` — the moment [`Shutdown::stop`] is
+//! called, so teardown does not have to wait out a five-minute sweep interval to
+//! join a thread.
+
+use std::time::Duration;
+
+use parking_lot::{Condvar, Mutex};
+
+/// The daemon's stop signal. Set once, never cleared: a `Core` that has been
+/// told to stop belongs to a mount that is going away.
+#[derive(Default)]
+pub(crate) struct Shutdown {
+    stopping: Mutex<bool>,
+    changed: Condvar,
+}
+
+impl Shutdown {
+    /// Whether teardown has begun. Loops check this where they would otherwise
+    /// start another unit of work.
+    pub(crate) fn is_stopping(&self) -> bool {
+        *self.stopping.lock()
+    }
+
+    /// Begin teardown and wake everything waiting in [`sleep`](Self::sleep).
+    /// Idempotent, because more than one path can reach teardown (a signal, the
+    /// kernel mount ending, a failed startup).
+    pub(crate) fn stop(&self) {
+        let mut stopping = self.stopping.lock();
+        *stopping = true;
+        self.changed.notify_all();
+    }
+
+    /// Wait up to `duration`, or until teardown starts.
+    ///
+    /// Returns `false` if the daemon is stopping — including when it was already
+    /// stopping on entry, so `while shutdown.sleep(interval)` is a complete and
+    /// correct loop condition and no separate check is needed at the top.
+    pub(crate) fn sleep(&self, duration: Duration) -> bool {
+        let mut stopping = self.stopping.lock();
+        if !*stopping {
+            self.changed.wait_for(&mut stopping, duration);
+        }
+        !*stopping
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::sync::Arc;
+    use std::time::Instant;
+
+    #[test]
+    fn sleeping_returns_early_once_stopped() {
+        let shutdown = Arc::new(Shutdown::default());
+        assert!(!shutdown.is_stopping());
+
+        let waker = shutdown.clone();
+        std::thread::spawn(move || {
+            std::thread::sleep(Duration::from_millis(20));
+            waker.stop();
+        });
+
+        let started = Instant::now();
+        // An hour's worth of interval, cut short by the stop.
+        assert!(!shutdown.sleep(Duration::from_secs(3600)));
+        assert!(
+            started.elapsed() < Duration::from_secs(5),
+            "stop must interrupt the wait, not wait it out"
+        );
+        assert!(shutdown.is_stopping());
+        // Already stopping: the next wait does not sleep at all.
+        let started = Instant::now();
+        assert!(!shutdown.sleep(Duration::from_secs(3600)));
+        assert!(started.elapsed() < Duration::from_secs(5));
+    }
+
+    #[test]
+    fn sleeping_runs_to_the_timeout_while_running() {
+        let shutdown = Shutdown::default();
+        assert!(shutdown.sleep(Duration::from_millis(10)));
+        assert!(!shutdown.is_stopping());
+    }
+}

crates/pdfs-fuse/src/state.rs

@@ -141,6 +141,11 @@ pub(crate) struct PendingRevision {
     pub(crate) meta: StagedWrite,
 }
 
+/// One directory's entries as `readdir` reports them: `(ino, is_dir, name)` per
+/// entry, in reply order. The kernel's file-type enum is mapped in at reply
+/// time so this module stays free of `fuser`.
+pub(crate) type DirListing = Vec<(u64, bool, String)>;
+
 /// Mutable inode bookkeeping, guarded by a mutex because fuser drives the
 /// `Filesystem` trait through `&self`.
 pub(crate) struct State {
@@ -158,6 +163,16 @@ pub(crate) struct State {
     /// have no entry here.
     pub(crate) handles: HashMap<u64, u64>,
     pub(crate) next_fh: u64,
+    /// One frozen listing per open directory handle, taken at `opendir`.
+    ///
+    /// `readdir` resumes from a byte offset into the listing it is walking, so
+    /// that listing has to stay the same across the calls of one `getdents`
+    /// sequence. Rebuilding it live meant a create or a trash landing between
+    /// two pages shifted every later entry, and the caller silently skipped or
+    /// repeated whatever crossed the page boundary. Each entry is
+    /// `(ino, is_dir, name)` — the kernel type is mapped at reply time so this
+    /// module stays free of `fuser` (bugs.md B43).
+    pub(crate) dir_snapshots: HashMap<u64, Arc<DirListing>>,
     /// Resident inode attributes whose effective access changed during an
     /// intern/root refresh. Core drains this set and notifies the matching
     /// kernel only after releasing the State lock.
@@ -317,6 +332,7 @@ impl State {
             active_writes: HashMap::new(),
             handles: HashMap::new(),
             next_fh: 1,
+            dir_snapshots: HashMap::new(),
             access_changes: HashSet::new(),
             share_access,
             db,

crates/pdfs-fuse/src/sweep.rs

@@ -67,10 +67,16 @@ impl Core {
     /// Sweep for stale conflict copies forever, on its own thread. See the module
     /// docs for the identical/divergent policy.
     pub(crate) fn run_conflict_sweep_loop(&self) {
-        std::thread::sleep(CONFLICT_SWEEP_WARMUP);
+        // Both waits are interruptible, so teardown does not have to sit out a
+        // five-minute interval to join this thread (bugs.md B44).
+        if !self.shutdown.sleep(CONFLICT_SWEEP_WARMUP) {
+            return;
+        }
         loop {
             self.sweep_conflicts_once();
-            std::thread::sleep(CONFLICT_SWEEP_INTERVAL);
+            if !self.shutdown.sleep(CONFLICT_SWEEP_INTERVAL) {
+                return;
+            }
         }
     }
 

crates/pdfs-fuse/src/sync.rs

@@ -54,7 +54,7 @@ use engine::{classify, settle_with};
 #[cfg(test)]
 use local_scan::is_write_mode;
 use local_scan::open_for_write_set;
-use model::{LocalItem, Outcome, PassAbort, Pending, RemoteItem};
+use model::{LocalItem, LocalSig, Outcome, PassAbort, Pending, RemoteItem};
 pub(crate) use planner::base_name;
 #[cfg(test)]
 use planner::conflict_path;
@@ -525,6 +525,9 @@ impl Core {
                 continue;
             }
             trashed.push(rel.clone());
+            if let Some(parent) = parent_uid_of(rel, &remote_dirs, &baseline) {
+                self.invalidate_children_of(&parent);
+            }
             let _ = self.db.sync_entry_remove(folder_id, rel);
             self.log_activity(
                 ActivityKind::Trash,
@@ -687,6 +690,7 @@ impl Core {
                             rel_path: rel.clone(),
                             remote_uid: Some(remote.uid.to_string()),
                             local_mtime: local.mtime,
+                            local_mtime_ns: local.mtime_ns,
                             local_size: local.size,
                             remote_rev: Some(remote.mtime.to_string()),
                             remote_hash: Some(remote.size.to_string()),
@@ -717,6 +721,9 @@ impl Core {
                         uid: remote.uid.clone(),
                         mtime: remote.mtime,
                         size: remote.size,
+                        // What the walk saw here, so the apply step can tell a
+                        // stale plan from a live one before it overwrites.
+                        local: l.map(LocalSig::from),
                     });
                 }
                 FilePlan::Conflict => {
@@ -739,7 +746,9 @@ impl Core {
                     }
                 },
                 FilePlan::DeleteLocal => {
-                    if let Err(e) = std::fs::remove_file(local_root.join(rel_to_path(rel))) {
+                    if let Err(e) =
+                        removed_locally(std::fs::remove_file(local_root.join(rel_to_path(rel))))
+                    {
                         warn!(rel, error = %e, "sync: remove local file failed");
                         self.log_activity(
                             ActivityKind::Trash,
@@ -771,6 +780,9 @@ impl Core {
                         outcome.errors += 1;
                         continue;
                     }
+                    if let Some(parent) = parent_uid_of(rel, &remote_dirs, &baseline) {
+                        self.invalidate_children_of(&parent);
+                    }
                     let _ = self.db.sync_entry_remove(folder_id, rel);
                     self.log_activity(
                         ActivityKind::Trash,
@@ -797,7 +809,9 @@ impl Core {
         // Deferred folder deletions, deepest first.
         delete_local_dirs.sort_by_key(|p| std::cmp::Reverse(p.matches('/').count()));
         for rel in delete_local_dirs {
-            if let Err(e) = std::fs::remove_dir_all(local_root.join(rel_to_path(&rel))) {
+            if let Err(e) =
+                removed_locally(std::fs::remove_dir_all(local_root.join(rel_to_path(&rel))))
+            {
                 warn!(rel, error = %e, "sync: remove local folder failed");
                 self.log_activity(ActivityKind::Trash, base_name(&rel), e.to_string(), false);
                 outcome.errors += 1;
@@ -824,6 +838,9 @@ impl Core {
                 outcome.errors += 1;
                 continue;
             }
+            if let Some(parent) = parent_uid_of(&rel, &remote_dirs, &baseline) {
+                self.invalidate_children_of(&parent);
+            }
             let _ = self.db.sync_entry_remove(folder_id, &rel);
             self.log_activity(
                 ActivityKind::Trash,
@@ -849,68 +866,9 @@ impl Core {
         writing: &HashSet<PathBuf>,
         out: &mut HashMap<String, LocalItem>,
     ) -> Result<(), String> {
-        let entries = std::fs::read_dir(dir).map_err(|e| format!("read {}: {e}", dir.display()))?;
-        for entry in entries {
-            let entry = entry.map_err(|e| format!("read entry in {}: {e}", dir.display()))?;
-            let path = entry.path();
-            let meta = std::fs::symlink_metadata(&path)
-                .map_err(|e| format!("stat {}: {e}", path.display()))?;
-            let stripped = path.strip_prefix(root).map_err(|e| {
-                format!(
-                    "local path {} escaped sync root {}: {e}",
-                    path.display(),
-                    root.display()
-                )
-            })?;
-            let rel = stripped.to_str().ok_or_else(|| {
-                format!(
-                    "local path {} is not valid UTF-8; refusing a destructive sync pass",
-                    path.display()
-                )
-            })?;
-            // Ignore our own in-flight download temp files.
-            if rel.contains(".pdfs-tmp-") {
-                continue;
-            }
-            // Ignored paths are dropped here as well as from the classification
-            // union — not for correctness (the union filter is what makes this
-            // safe) but so an ignored `node_modules/` is never descended into or
-            // counted against scan progress.
-            if rules.is_ignored(rel, meta.is_dir()) {
-                continue;
-            }
-            if meta.is_dir() {
-                out.insert(
-                    rel.to_string(),
-                    LocalItem {
-                        is_dir: true,
-                        mtime: 0,
-                        size: 0,
-                        open_for_write: false,
-                    },
-                );
-                self.progress_scanned(folder_id, base_name(rel));
-                self.walk_local(folder_id, root, &path, rules, writing, out)?;
-            } else if meta.is_file() {
-                // Check whether any process holds this file open for writing.
-                // The canonical path is used because /proc/*/fd links resolve
-                // to canonical targets.
-                let ofw = std::fs::canonicalize(&path)
-                    .map(|canon| writing.contains(&canon))
-                    .unwrap_or(false);
-                out.insert(
-                    rel.to_string(),
-                    LocalItem {
-                        is_dir: false,
-                        mtime: system_mtime(&meta),
-                        size: meta.len() as i64,
-                        open_for_write: ofw,
-                    },
-                );
-                self.progress_scanned(folder_id, base_name(rel));
-            }
-        }
-        Ok(())
+        walk_local_tree(root, dir, rules, writing, out, &mut |rel| {
+            self.progress_scanned(folder_id, base_name(rel))
+        })
     }
 
     /// Map every remote folder under `folder` to its uid by rel path, ignoring
@@ -1069,6 +1027,7 @@ impl Core {
                     rel_path: rel.to_string(),
                     remote_uid: Some(uid.to_string()),
                     local_mtime: 0,
+                    local_mtime_ns: None,
                     local_size: 0,
                     remote_rev: None,
                     remote_hash: None,
@@ -1175,6 +1134,9 @@ impl Core {
                     .create_folder(parent, base_name(rel), Some(now_secs()))
                     .await
                     .map_err(|e| format!("create remote folder {rel}: {e}"))?;
+                // The parent has gained a child that no cached listing of it
+                // knows about (B86).
+                self.invalidate_children_of(parent);
                 self.baseline_dir(folder_id, rel, &uid)?;
                 Ok(Applied::Dir(rel.clone(), uid))
             }
@@ -1192,8 +1154,9 @@ impl Core {
                 uid,
                 mtime,
                 size,
+                local,
             } => {
-                self.download_file(folder_id, local_root, rel, uid, *mtime, *size)
+                self.download_file(folder_id, local_root, rel, uid, *mtime, *size, *local)
                     .await?;
                 Ok(Applied::Downloaded)
             }
@@ -1211,7 +1174,11 @@ impl Core {
                         .map_err(|e| format!("set aside conflict copy for {rel}: {e}"))?;
                     info!(rel, "sync: kept local changes as a conflict copy");
                 }
-                self.download_file(folder_id, local_root, rel, uid, *mtime, *size)
+                // The local copy has just been moved aside, so the state this
+                // download expects to find at the path is "nothing" — anything
+                // that appears while it runs is a *new* racing write, and gets
+                // its own conflict copy rather than being overwritten (B39).
+                self.download_file(folder_id, local_root, rel, uid, *mtime, *size, None)
                     .await?;
                 Ok(Applied::Conflict)
             }
@@ -1250,6 +1217,9 @@ impl Core {
         let file = std::fs::File::open(&path).map_err(|e| format!("open {rel}: {e}"))?;
         let meta = file.metadata().map_err(|e| format!("stat {rel}: {e}"))?;
         let mtime = system_mtime(&meta);
+        // Taken from the descriptor being streamed, so it names the content that
+        // actually goes up rather than whatever is at the path when we finish.
+        let streamed = LocalSig::from(&meta);
         // Count the bytes as they are read, so this upload shows up in
         // `GetQueueStatus` next to a manual one. The uid isn't known until the
         // draft is sealed, so the transfer registers without one.
@@ -1276,7 +1246,13 @@ impl Core {
             )
             .await
             .map_err(|e| format!("upload {rel}: {e}"))?;
-        self.record_file_baseline(folder_id, rel, &path, &uid).await
+        // The parent has gained a child that no cached listing of it knows
+        // about. Without this, a folder later switched to on-demand serves a
+        // snapshot that predates this upload (B86).
+        self.invalidate_children_of(parent);
+        warn_if_torn(rel, &path, streamed);
+        self.record_file_baseline(folder_id, rel, Some(streamed), &uid)
+            .await
     }
 
     /// Upload a changed local file as a new revision of an existing remote node.
@@ -1291,6 +1267,7 @@ impl Core {
         let file = std::fs::File::open(&path).map_err(|e| format!("open {rel}: {e}"))?;
         let meta = file.metadata().map_err(|e| format!("stat {rel}: {e}"))?;
         let mtime = system_mtime(&meta);
+        let streamed = LocalSig::from(&meta);
         let reader = OwnedCountingReader::new(
             file,
             self.transfers.begin(
@@ -1304,11 +1281,20 @@ impl Core {
             .upload_new_revision_from(uid, reader, meta.len() as i64, Vec::new(), Some(mtime))
             .await
             .map_err(|e| format!("upload revision {rel}: {e}"))?;
-        self.record_file_baseline(folder_id, rel, &path, uid).await
+        warn_if_torn(rel, &path, streamed);
+        self.record_file_baseline(folder_id, rel, Some(streamed), uid)
+            .await
     }
 
     /// Download a remote file to its local path (atomically via a temp file),
     /// stamp the local mtime to match the remote, then record baseline.
+    ///
+    /// `expected_local` is the identity planning saw at the destination. The
+    /// rename below is the destructive step, and the download in between can run
+    /// for minutes, so the destination is re-examined right before it: a local
+    /// file that no longer matches the plan is set aside as a conflict copy
+    /// instead of being replaced (bugs.md B39).
+    #[allow(clippy::too_many_arguments)]
     async fn download_file(
         &self,
         folder_id: i64,
@@ -1317,6 +1303,7 @@ impl Core {
         uid: &NodeUid,
         mtime: i64,
         size: i64,
+        expected_local: Option<LocalSig>,
     ) -> Result<(), String> {
         let path = local_root.join(rel_to_path(rel));
         if let Some(parent) = path.parent() {
@@ -1351,6 +1338,13 @@ impl Core {
             out.flush().map_err(|e| format!("flush tmp {rel}: {e}"))?;
             out.sync_all().map_err(|e| format!("sync tmp {rel}: {e}"))?;
         }
+        // The last moment at which the destination can still be saved. Anything
+        // other than the state planning classified means a writer got there
+        // first, and its bytes are the ones nothing else has a copy of.
+        if let Err(e) = keep_racing_local_edit(&path, expected_local) {
+            let _ = std::fs::remove_file(&tmp);
+            return Err(format!("set aside racing local edit for {rel}: {e}"));
+        }
         std::fs::rename(&tmp, &path).map_err(|e| format!("place {rel}: {e}"))?;
         // Match local mtime to the remote's so neither side looks "changed" next pass.
         let f = std::fs::File::options()
@@ -1366,23 +1360,36 @@ impl Core {
                 .and_then(|dir| dir.sync_all())
                 .map_err(|e| format!("sync directory for {rel}: {e}"))?;
         }
-        self.record_file_baseline(folder_id, rel, &path, uid).await
+        // Here the file on disk *is* the transferred content: it was published
+        // by a rename and stamped just above, so its stat is the identity.
+        self.record_file_baseline(folder_id, rel, local_sig(&path), uid)
+            .await
     }
 
-    /// Refresh a file's baseline from ground truth: the local stat and the
-    /// remote node's reported signature. Called after every upload/download so
-    /// the very next reconcile sees the path as unchanged.
+    /// Refresh a file's baseline from ground truth: the local identity these
+    /// bytes came from and the remote node's reported signature. Called after
+    /// every upload/download so the very next reconcile sees the path as
+    /// unchanged.
+    ///
+    /// `local` is the identity the *transferred content* corresponds to, which
+    /// is not always the file's current stat. An upload streams the live file,
+    /// so it passes the identity it read the first byte at: if a writer changed
+    /// the file mid-stream, the remote now holds a torn revision, and recording
+    /// the pre-upload identity is what makes the next pass see the path as
+    /// locally changed and upload a whole, clean revision over it. Recording the
+    /// post-upload stat instead declared the torn revision synchronized and left
+    /// it as the file's content on Drive (bugs.md B40).
     async fn record_file_baseline(
         &self,
         folder_id: i64,
         rel: &str,
-        local_path: &Path,
+        local: Option<LocalSig>,
         uid: &NodeUid,
     ) -> Result<(), String> {
-        let (lmtime, lsize) = match std::fs::metadata(local_path) {
-            Ok(m) => (system_mtime(&m), m.len() as i64),
-            Err(e) => return Err(format!("stat {rel}: {e}")),
+        let Some(local) = local else {
+            return Err(format!("stat {rel}: no local identity to baseline"));
         };
+        let (lmtime, lsize) = (local.mtime, local.size);
         // Re-fetch the node so the stored remote signature is exactly what a walk
         // will report next time.
         let (rmtime, rsize) = match self.client.enumerate_nodes(std::slice::from_ref(uid)).await {
@@ -1402,6 +1409,7 @@ impl Core {
                     rel_path: rel.to_string(),
                     remote_uid: Some(uid.to_string()),
                     local_mtime: lmtime,
+                    local_mtime_ns: local.mtime_ns,
                     local_size: lsize,
                     remote_rev: Some(rmtime.to_string()),
                     remote_hash: Some(rsize.to_string()),
@@ -1411,6 +1419,177 @@ impl Core {
     }
 }
 
+/// Say so when a file changed under an upload that was streaming it.
+///
+/// Purely diagnostic — the correction is already in place, because the baseline
+/// records the identity the upload *read* rather than the one on disk now, so
+/// the next pass classifies the path as locally changed and replaces the torn
+/// revision. Without the log there is nothing in the journal to explain the
+/// extra revision (bugs.md B40).
+fn warn_if_torn(rel: &str, path: &Path, streamed: LocalSig) {
+    let now = local_sig(path);
+    if !now.is_some_and(|now| now.same_content(&streamed)) {
+        warn!(
+            rel,
+            ?streamed,
+            ?now,
+            "sync: file changed while uploading; the remote revision may be torn \
+             and will be replaced on the next pass"
+        );
+    }
+}
+
+/// The identity a reconcile pass uses to recognise a local file it has already
+/// looked at: `None` when nothing is at the path.
+fn local_sig(path: &Path) -> Option<LocalSig> {
+    std::fs::metadata(path).ok().map(|m| LocalSig {
+        mtime: system_mtime(&m),
+        mtime_ns: system_mtime_ns(&m),
+        size: m.len() as i64,
+    })
+}
+
+/// Guard the destination of a download: if what is at `path` is no longer what
+/// planning classified, move it aside as a conflict copy before the caller
+/// renames over it.
+///
+/// A reconcile pass decides what to do from a walk that finished before the
+/// transfer started, and a download of a large file can run for minutes. Without
+/// this, a local edit completed inside that window was replaced by the remote
+/// version with no copy of it kept anywhere — the one shape of loss a sync
+/// engine must not have (bugs.md B39). Keeping the copy makes it re-upload as a
+/// new file on the next pass, which is the same resolution a both-sides-changed
+/// conflict already gets.
+///
+/// A path that is empty when it was expected to be is not a conflict: the
+/// download is about to recreate it, and there is nothing to preserve.
+fn keep_racing_local_edit(path: &Path, expected: Option<LocalSig>) -> std::io::Result<()> {
+    let actual = local_sig(path);
+    let unchanged = match (actual, expected) {
+        (Some(a), Some(e)) => a.same_content(&e),
+        // Nothing there: the download is about to create it, and there is
+        // nothing to preserve either way.
+        (None, _) => true,
+        (Some(_), None) => false,
+    };
+    if unchanged {
+        return Ok(());
+    }
+    warn!(
+        path = %path.display(),
+        ?expected,
+        ?actual,
+        "sync: local file changed while downloading; keeping it as a conflict copy"
+    );
+    preserve_conflict_copy(path, now_secs()).map(|_| ())
+}
+
+/// Read a local removal as the sync engine has to read it: the *state* it was
+/// asking for is "this path is gone", so a path that was already gone is a
+/// success, not a failure.
+///
+/// It matters because the baseline is only dropped on success. Treating
+/// `NotFound` as an error left the baseline in place for a file that no longer
+/// exists, so the same removal was re-attempted and re-reported as an error on
+/// every pass, forever. Every other error is still an error, and keeping the
+/// baseline is exactly right there: the surviving file must not be mistaken for
+/// a new local creation on the next pass and uploaded back (bugs.md B42).
+fn removed_locally(result: std::io::Result<()>) -> std::io::Result<()> {
+    match result {
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(()),
+        other => other,
+    }
+}
+
+/// The body of [`Core::walk_local`], without a `Core`: recursively collect the
+/// local tree under `dir` into `out`, reporting each entry to `scanned`.
+///
+/// **Every failure here is fatal to the pass, deliberately.** A `read_dir` that
+/// fails, an entry that cannot be stat-ed, a name that is not UTF-8 — each of
+/// them would, if skipped, leave `out` describing a *smaller* tree than the one
+/// on disk. Reconciliation reads a path present in the baseline and absent from
+/// the scan as "the user deleted this", and propagates that to Drive. So an
+/// unreadable subdirectory would trash its remote counterpart, and the deeper
+/// the failure, the more it takes with it (`docs/BUGS.md` B55). Refusing the
+/// whole pass keeps the previous baseline and leaves the remote alone; the next
+/// pass retries once the transient condition clears.
+///
+/// Split out from the method so exactly that can be tested against a real tree
+/// with a real unreadable directory in it, which needs no `Core` at all.
+fn walk_local_tree(
+    root: &Path,
+    dir: &Path,
+    rules: &IgnoreRules,
+    writing: &HashSet<PathBuf>,
+    out: &mut HashMap<String, LocalItem>,
+    scanned: &mut dyn FnMut(&str),
+) -> Result<(), String> {
+    let entries = std::fs::read_dir(dir).map_err(|e| format!("read {}: {e}", dir.display()))?;
+    for entry in entries {
+        let entry = entry.map_err(|e| format!("read entry in {}: {e}", dir.display()))?;
+        let path = entry.path();
+        let meta = std::fs::symlink_metadata(&path)
+            .map_err(|e| format!("stat {}: {e}", path.display()))?;
+        let stripped = path.strip_prefix(root).map_err(|e| {
+            format!(
+                "local path {} escaped sync root {}: {e}",
+                path.display(),
+                root.display()
+            )
+        })?;
+        let rel = stripped.to_str().ok_or_else(|| {
+            format!(
+                "local path {} is not valid UTF-8; refusing a destructive sync pass",
+                path.display()
+            )
+        })?;
+        // Ignore our own in-flight download temp files.
+        if rel.contains(".pdfs-tmp-") {
+            continue;
+        }
+        // Ignored paths are dropped here as well as from the classification
+        // union — not for correctness (the union filter is what makes this
+        // safe) but so an ignored `node_modules/` is never descended into or
+        // counted against scan progress.
+        if rules.is_ignored(rel, meta.is_dir()) {
+            continue;
+        }
+        if meta.is_dir() {
+            out.insert(
+                rel.to_string(),
+                LocalItem {
+                    is_dir: true,
+                    mtime: 0,
+                    mtime_ns: None,
+                    size: 0,
+                    open_for_write: false,
+                },
+            );
+            scanned(rel);
+            walk_local_tree(root, &path, rules, writing, out, scanned)?;
+        } else if meta.is_file() {
+            // Check whether any process holds this file open for writing.
+            // The canonical path is used because /proc/*/fd links resolve
+            // to canonical targets.
+            let ofw = std::fs::canonicalize(&path)
+                .map(|canon| writing.contains(&canon))
+                .unwrap_or(false);
+            out.insert(
+                rel.to_string(),
+                LocalItem {
+                    is_dir: false,
+                    mtime: system_mtime(&meta),
+                    mtime_ns: system_mtime_ns(&meta),
+                    size: meta.len() as i64,
+                    open_for_write: ofw,
+                },
+            );
+            scanned(rel);
+        }
+    }
+    Ok(())
+}
+
 /// Publish a conflict copy without overwriting an earlier conflict bearing the
 /// same timestamp. The original is removed only after the new copy is durable.
 fn preserve_conflict_copy(path: &Path, stamp: i64) -> std::io::Result<PathBuf> {
@@ -1444,6 +1623,26 @@ fn preserve_conflict_copy(path: &Path, stamp: i64) -> std::io::Result<PathBuf> {
     unreachable!("u32 conflict suffix space exhausted")
 }
 
+/// The remote uid of `rel`'s parent folder: from the pass's folder-uid map
+/// first, and from the baseline second for a parent the map never covered.
+///
+/// Used to name the folder whose cached listing a remote mutation has just
+/// invalidated (B86). `None` is not an error — it only means this pass cannot
+/// say which folder to invalidate, and the caller skips it.
+fn parent_uid_of(
+    rel: &str,
+    remote_dirs: &HashMap<String, NodeUid>,
+    baseline: &HashMap<String, StoredSyncEntry>,
+) -> Option<NodeUid> {
+    let parent = parent_rel(rel);
+    remote_dirs.get(parent).cloned().or_else(|| {
+        baseline
+            .get(parent)
+            .and_then(|e| e.remote_uid.as_deref())
+            .and_then(parse_uid)
+    })
+}
+
 fn sync_parent_dir(path: &Path) -> std::io::Result<()> {
     match path.parent() {
         Some(parent) => std::fs::File::open(parent)?.sync_all(),
@@ -1476,10 +1675,77 @@ fn system_mtime(meta: &std::fs::Metadata) -> i64 {
         .map(|d| d.as_secs() as i64)
         .unwrap_or(0)
 }
+
+/// The same modification time in nanoseconds, when the filesystem reports one.
+///
+/// Whole seconds cannot distinguish an edit made in the same second as the last
+/// sync from no edit at all, and if the size did not change either, that file
+/// was never uploaded (bugs.md B25). `None` for a clock or filesystem that
+/// cannot answer, which falls the comparison back to seconds rather than
+/// asserting an mtime of zero.
+fn system_mtime_ns(meta: &std::fs::Metadata) -> Option<i64> {
+    meta.modified()
+        .ok()
+        .and_then(|t| t.duration_since(SystemTime::UNIX_EPOCH).ok())
+        .and_then(|d| i64::try_from(d.as_nanos()).ok())
+}
 #[cfg(test)]
 mod tests {
     use super::*;
 
+    /// A removal of something already gone has produced the state the sync
+    /// engine wanted, so its baseline is dropped rather than retried forever.
+    /// Anything else stays an error and keeps the baseline (bugs.md B42).
+    #[test]
+    fn an_already_absent_path_counts_as_removed() {
+        let root = std::env::temp_dir().join(format!("pdfs-b42-{}", std::process::id()));
+        let missing = root.join("never-existed");
+        assert!(removed_locally(std::fs::remove_file(&missing)).is_ok());
+        assert!(removed_locally(std::fs::remove_dir_all(&missing)).is_ok());
+
+        // A directory in the way of `remove_file` is a real failure, and must
+        // stay one: the file survives, and its baseline has to survive with it.
+        let occupied = root.join("a-directory");
+        std::fs::create_dir_all(&occupied).unwrap();
+        assert!(removed_locally(std::fs::remove_file(&occupied)).is_err());
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
+    /// The download guard: a destination that still matches the plan is
+    /// replaced silently, one that has moved on since is kept (bugs.md B39).
+    #[test]
+    fn a_download_keeps_a_local_edit_it_did_not_plan_for() {
+        let root = std::env::temp_dir().join(format!("pdfs-b39-{}", std::process::id()));
+        std::fs::create_dir_all(&root).unwrap();
+        let path = root.join("report.txt");
+
+        // Nothing there, nothing expected: the download simply creates it.
+        assert!(keep_racing_local_edit(&path, None).is_ok());
+        assert!(!path.exists());
+
+        // Unchanged since planning: no copy, the caller overwrites it.
+        std::fs::write(&path, b"planned").unwrap();
+        let planned = local_sig(&path).unwrap();
+        keep_racing_local_edit(&path, Some(planned)).unwrap();
+        assert_eq!(std::fs::read(&path).unwrap(), b"planned");
+
+        // Changed since planning: the racing bytes are the only copy of
+        // themselves, so they must survive the overwrite as a conflict copy.
+        std::fs::write(&path, b"a much longer local edit").unwrap();
+        keep_racing_local_edit(&path, Some(planned)).unwrap();
+        let kept: Vec<_> = std::fs::read_dir(&root)
+            .unwrap()
+            .filter_map(|e| e.ok())
+            .map(|e| std::fs::read(e.path()).unwrap_or_default())
+            .collect();
+        assert!(
+            kept.iter()
+                .any(|bytes| bytes == b"a much longer local edit"),
+            "the racing edit must still exist somewhere: {kept:?}"
+        );
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
     /// Feed `n` `Reconcile` events `gap` apart on another thread, starting
     /// immediately. Returns the sender's join handle so the test can keep it
     /// alive (dropping the sender early would look like a disconnect).
@@ -1565,6 +1831,7 @@ mod tests {
 
     fn local_item(is_dir: bool) -> LocalItem {
         LocalItem {
+            mtime_ns: None,
             is_dir,
             mtime: 1,
             size: 1,
@@ -1722,6 +1989,44 @@ mod tests {
         assert!(guard_local_wipe(&base, &fresh).is_err());
     }
 
+    /// B54's remaining case, end to end: a folder whose local root is still
+    /// *readable* but is no longer the tree it was — an empty directory left
+    /// where a mount used to be, or a different device mounted over it.
+    ///
+    /// Nothing fails here. The scan succeeds and honestly reports an empty
+    /// directory, so the only thing standing between the user and a trashed
+    /// remote folder is the guard, and it has to fire for a one-file baseline
+    /// exactly as it does for a large one.
+    #[test]
+    fn a_replaced_mountpoint_trips_the_guard_at_every_baseline_size() {
+        let root = sync_test_dir("wipe-replaced-mount");
+        let rules = rules_for(&[]);
+        let writing = HashSet::new();
+        let mut local: HashMap<String, LocalItem> = HashMap::new();
+        let mut progress = no_progress();
+        walk_local_tree(&root, &root, &rules, &writing, &mut local, &mut progress).unwrap();
+        assert!(local.is_empty(), "the replacement root is genuinely empty");
+
+        for size in 1..=3 {
+            let baseline: HashMap<String, ()> =
+                (0..size).map(|i| (format!("file{i}.txt"), ())).collect();
+            assert!(
+                guard_local_wipe(&filter_baseline(&baseline, &rules), &local).is_err(),
+                "a {size}-path baseline against an empty root is not a deletion"
+            );
+        }
+
+        // And the guard steps aside the moment the real tree is back, so a
+        // remounted folder resumes rather than staying wedged.
+        std::fs::write(root.join("file0.txt"), b"back").unwrap();
+        let mut local: HashMap<String, LocalItem> = HashMap::new();
+        walk_local_tree(&root, &root, &rules, &writing, &mut local, &mut progress).unwrap();
+        let baseline: HashMap<String, ()> = [("file0.txt".to_string(), ())].into_iter().collect();
+        assert!(guard_local_wipe(&filter_baseline(&baseline, &rules), &local).is_ok());
+
+        std::fs::remove_dir_all(&root).unwrap();
+    }
+
     /// A restored folder (features.md 5.2) starts with an empty baseline against
     /// an empty local directory and a full remote. That must reconcile as
     /// "download everything" rather than tripping the wipe guard — otherwise
@@ -1764,6 +2069,7 @@ mod tests {
                 remote_uid: Some("v~l".into()),
                 local_mtime: 0,
                 local_size: 0,
+                local_mtime_ns: None,
                 remote_rev: Some(mtime.to_string()),
                 remote_hash: Some(size.to_string()),
             },
@@ -1801,6 +2107,7 @@ mod tests {
                 remote_uid: Some("v~l".into()),
                 local_mtime: 0,
                 local_size: 0,
+                local_mtime_ns: None,
                 remote_rev: None,
                 remote_hash: None,
             },
@@ -1815,6 +2122,7 @@ mod tests {
             remote_uid: Some("v~l".into()),
             local_mtime: 10,
             local_size: 20,
+            local_mtime_ns: None,
             remote_rev: Some("1700".into()),
             remote_hash: Some("4096".into()),
         };
@@ -1828,6 +2136,7 @@ mod tests {
             remote_uid: Some("v~l".into()),
             local_mtime: 0,
             local_size: 0,
+            local_mtime_ns: None,
             remote_rev: None,
             remote_hash: None,
         };
@@ -1863,6 +2172,7 @@ mod tests {
             uid: uid(),
             mtime: 0,
             size: 10,
+            local: None,
         };
         assert_eq!(op.kind(), ActivityKind::Download);
     }
@@ -1918,6 +2228,204 @@ mod tests {
         std::fs::remove_dir_all(dir).unwrap();
     }
 
+    /// B49's remaining cases. Preservation is the step that makes a download
+    /// safe to publish, so every way it can fail has to leave the local file
+    /// where it was — the download is aborted, not the file replaced.
+    #[test]
+    fn conflict_preservation_failures_all_keep_the_local_file() {
+        use std::os::unix::fs::PermissionsExt as _;
+
+        // An unwritable directory: the conflict copy cannot be created at all.
+        let dir = sync_test_dir("conflict-eacces");
+        let source = dir.join("notes.txt");
+        std::fs::write(&source, b"the only copy").unwrap();
+        std::fs::set_permissions(&dir, std::fs::Permissions::from_mode(0o500)).unwrap();
+        let denied = preserve_conflict_copy(&source, 1700);
+        std::fs::set_permissions(&dir, std::fs::Permissions::from_mode(0o700)).unwrap();
+        assert!(denied.is_err(), "a read-only directory must fail the copy");
+        assert_eq!(
+            std::fs::read(&source).unwrap(),
+            b"the only copy",
+            "the local file survives byte-identical"
+        );
+        std::fs::remove_dir_all(&dir).unwrap();
+
+        // A source that vanished between planning and preservation. The pass
+        // must report it rather than publish the download over the gap.
+        let dir = sync_test_dir("conflict-enoent");
+        assert!(preserve_conflict_copy(&dir.join("gone.txt"), 1700).is_err());
+        std::fs::remove_dir_all(&dir).unwrap();
+
+        // A write error partway through the copy — induced by making the
+        // destination name a directory, so `create_new` fails as something
+        // other than `AlreadyExists`.
+        let dir = sync_test_dir("conflict-eisdir");
+        let source = dir.join("notes.txt");
+        std::fs::write(&source, b"still the only copy").unwrap();
+        std::fs::create_dir(conflict_path(&source, 1700)).unwrap();
+        // Suffix 0 is taken by a directory; `create_new` on it reports
+        // `AlreadyExists`, so preservation moves on to suffix 1 and succeeds.
+        // What must not happen is the source being removed without a copy.
+        let preserved = preserve_conflict_copy(&source, 1700).unwrap();
+        assert!(preserved.is_file());
+        assert_eq!(std::fs::read(&preserved).unwrap(), b"still the only copy");
+        assert!(!source.exists());
+        std::fs::remove_dir_all(&dir).unwrap();
+    }
+
+    /// A conflict copy keeps the bytes *and* the metadata that identifies them,
+    /// so the next pass uploads the user's version rather than a file that looks
+    /// freshly written.
+    #[test]
+    fn a_conflict_copy_is_byte_and_mtime_identical() {
+        use std::os::unix::fs::PermissionsExt as _;
+
+        let dir = sync_test_dir("conflict-identity");
+        let source = dir.join("notes.txt");
+        std::fs::write(&source, b"local edit that must survive").unwrap();
+        let before = std::fs::metadata(&source).unwrap();
+
+        let preserved = preserve_conflict_copy(&source, 1700).unwrap();
+        let after = std::fs::metadata(&preserved).unwrap();
+
+        assert_eq!(
+            std::fs::read(&preserved).unwrap(),
+            b"local edit that must survive"
+        );
+        assert_eq!(after.len(), before.len());
+        assert_eq!(after.modified().unwrap(), before.modified().unwrap());
+        assert_eq!(after.permissions().mode(), before.permissions().mode());
+        std::fs::remove_dir_all(&dir).unwrap();
+    }
+
+    // ---- B55: an incomplete local scan is never a set of deletions ---------
+
+    fn no_progress() -> impl FnMut(&str) {
+        |_: &str| {}
+    }
+
+    /// A directory the pass cannot read must fail the pass. Omitting it would
+    /// hand reconciliation a baseline-minus-subtree, which it reads as the user
+    /// having deleted that subtree and propagates to Drive (B55).
+    #[test]
+    fn an_unreadable_subdirectory_fails_the_scan_instead_of_shrinking_it() {
+        use std::os::unix::fs::PermissionsExt as _;
+
+        let root = sync_test_dir("scan-eacces");
+        std::fs::write(root.join("keep.txt"), b"a").unwrap();
+        let locked = root.join("locked");
+        std::fs::create_dir(&locked).unwrap();
+        std::fs::write(locked.join("inner.txt"), b"b").unwrap();
+        std::fs::set_permissions(&locked, std::fs::Permissions::from_mode(0o000)).unwrap();
+
+        let rules = rules_for(&[]);
+        let writing = HashSet::new();
+        let mut out = HashMap::new();
+        let mut progress = no_progress();
+        let result = walk_local_tree(&root, &root, &rules, &writing, &mut out, &mut progress);
+
+        std::fs::set_permissions(&locked, std::fs::Permissions::from_mode(0o700)).unwrap();
+        std::fs::remove_dir_all(&root).unwrap();
+
+        let error = result.expect_err("an unreadable directory must abort the pass");
+        assert!(
+            error.contains("locked"),
+            "the error should name the subtree it could not read: {error}"
+        );
+    }
+
+    /// The same rule one level up: if the sync root itself cannot be read, the
+    /// pass sees nothing, and "nothing" must never mean "delete everything".
+    #[test]
+    fn an_unreadable_root_fails_the_scan() {
+        use std::os::unix::fs::PermissionsExt as _;
+
+        let root = sync_test_dir("scan-root-eacces");
+        std::fs::write(root.join("keep.txt"), b"a").unwrap();
+        std::fs::set_permissions(&root, std::fs::Permissions::from_mode(0o000)).unwrap();
+
+        let rules = rules_for(&[]);
+        let writing = HashSet::new();
+        let mut out = HashMap::new();
+        let mut progress = no_progress();
+        let result = walk_local_tree(&root, &root, &rules, &writing, &mut out, &mut progress);
+
+        std::fs::set_permissions(&root, std::fs::Permissions::from_mode(0o700)).unwrap();
+        std::fs::remove_dir_all(&root).unwrap();
+
+        assert!(result.is_err());
+        assert!(out.is_empty());
+    }
+
+    /// A root that is gone entirely — an unmounted device folder, a removed
+    /// directory — is the sharpest version of the same case: every baseline path
+    /// is "missing", and the scan must refuse rather than report an empty tree.
+    #[test]
+    fn an_absent_root_fails_the_scan() {
+        let root = sync_test_dir("scan-absent");
+        std::fs::remove_dir_all(&root).unwrap();
+
+        let rules = rules_for(&[]);
+        let writing = HashSet::new();
+        let mut out = HashMap::new();
+        let mut progress = no_progress();
+        assert!(walk_local_tree(&root, &root, &rules, &writing, &mut out, &mut progress).is_err());
+        assert!(out.is_empty());
+    }
+
+    /// A name that is not valid UTF-8 cannot be keyed into the scan, so it
+    /// cannot be compared against the baseline either. Refusing the pass is the
+    /// only safe reading: skipping it would classify the file as deleted.
+    #[test]
+    fn a_non_utf8_name_fails_the_scan() {
+        use std::ffi::OsStr;
+        use std::os::unix::ffi::OsStrExt as _;
+
+        let root = sync_test_dir("scan-non-utf8");
+        std::fs::write(root.join(OsStr::from_bytes(b"bad\xffname")), b"x").unwrap();
+
+        let rules = rules_for(&[]);
+        let writing = HashSet::new();
+        let mut out = HashMap::new();
+        let mut progress = no_progress();
+        let result = walk_local_tree(&root, &root, &rules, &writing, &mut out, &mut progress);
+        std::fs::remove_dir_all(&root).unwrap();
+
+        let error = result.expect_err("a name the scan cannot key must abort the pass");
+        assert!(error.contains("UTF-8"), "{error}");
+    }
+
+    /// The complement: a scan that *can* read everything reports everything, so
+    /// the strictness above does not come at the cost of a working pass. A
+    /// symlink is skipped by design rather than by failure, and a dangling one
+    /// stats fine through `symlink_metadata`.
+    #[test]
+    fn a_readable_tree_scans_completely() {
+        let root = sync_test_dir("scan-complete");
+        std::fs::write(root.join("a.txt"), b"aa").unwrap();
+        std::fs::create_dir(root.join("sub")).unwrap();
+        std::fs::write(root.join("sub/b.txt"), b"bbb").unwrap();
+        std::os::unix::fs::symlink("nowhere", root.join("dangling")).unwrap();
+
+        let rules = rules_for(&[]);
+        let writing = HashSet::new();
+        let mut out = HashMap::new();
+        let mut seen: Vec<String> = Vec::new();
+        let result = walk_local_tree(&root, &root, &rules, &writing, &mut out, &mut |rel| {
+            seen.push(rel.to_string())
+        });
+        std::fs::remove_dir_all(&root).unwrap();
+
+        result.unwrap();
+        let mut keys: Vec<&str> = out.keys().map(String::as_str).collect();
+        keys.sort_unstable();
+        assert_eq!(keys, vec!["a.txt", "sub", "sub/b.txt"]);
+        assert_eq!(out["a.txt"].size, 2);
+        assert_eq!(out["sub/b.txt"].size, 3);
+        assert!(out["sub"].is_dir);
+        assert_eq!(seen.len(), 3, "every reported entry ticks scan progress");
+    }
+
     fn sync_test_dir(label: &str) -> PathBuf {
         let nonce = SystemTime::now()
             .duration_since(SystemTime::UNIX_EPOCH)

crates/pdfs-fuse/src/sync/engine.rs

@@ -32,17 +32,24 @@ pub(crate) enum SyncMsg {
     ReconcileAll,
     /// Re-read the folder set and adjust the filesystem watches (after add/remove).
     Rewatch,
+    /// The daemon is tearing down: finish the current pass and end the engine
+    /// thread, so the mount can join it (bugs.md B44).
+    Stop,
 }
 
 /// Start the sync engine: a control thread that owns the filesystem watcher and
 /// serialises reconcile passes, plus a periodic poll thread. The engine receives
 /// on `rx`; senders live in [`Core::sync_tx`].
-pub(crate) fn spawn(core: Core, rx: Receiver<SyncMsg>) {
-    if let Err(e) = std::thread::Builder::new()
+pub(crate) fn spawn(core: Core, rx: Receiver<SyncMsg>) -> Option<std::thread::JoinHandle<()>> {
+    match std::thread::Builder::new()
         .name("pdfs-sync".into())
         .spawn(move || engine_loop(core, rx))
     {
-        warn!(error = %e, "failed to start sync engine");
+        Ok(handle) => Some(handle),
+        Err(e) => {
+            warn!(error = %e, "failed to start sync engine");
+            None
+        }
     }
 }
 
@@ -74,11 +81,11 @@ fn engine_loop(core: Core, rx: Receiver<SyncMsg>) {
     // Periodic remote poll.
     {
         let tx = core.sync_tx.clone();
+        let shutdown = core.shutdown.clone();
         std::thread::Builder::new()
             .name("pdfs-sync-poll".into())
             .spawn(move || {
-                loop {
-                    std::thread::sleep(POLL_INTERVAL);
+                while shutdown.sleep(POLL_INTERVAL) {
                     if tx.send(SyncMsg::ReconcileAll).is_err() {
                         break;
                     }
@@ -94,8 +101,12 @@ fn engine_loop(core: Core, rx: Receiver<SyncMsg>) {
         let mut ids: HashSet<i64> = HashSet::new();
         let mut all = false;
         let mut do_rewatch = false;
-        classify(msg, &mut ids, &mut all, &mut do_rewatch);
-        settle(&rx, &mut ids, &mut all, &mut do_rewatch);
+        if !classify(msg, &mut ids, &mut all, &mut do_rewatch)
+            || !settle(&rx, &mut ids, &mut all, &mut do_rewatch)
+        {
+            info!("sync engine stopping");
+            return;
+        }
 
         if do_rewatch {
             rewatch(&core, &mut watcher, &watched);
@@ -116,9 +127,9 @@ fn engine_loop(core: Core, rx: Receiver<SyncMsg>) {
 fn poll_only_loop(core: &Core, rx: Receiver<SyncMsg>) {
     {
         let tx = core.sync_tx.clone();
+        let shutdown = core.shutdown.clone();
         std::thread::spawn(move || {
-            loop {
-                std::thread::sleep(POLL_INTERVAL);
+            while shutdown.sleep(POLL_INTERVAL) {
                 if tx.send(SyncMsg::ReconcileAll).is_err() {
                     break;
                 }
@@ -130,8 +141,12 @@ fn poll_only_loop(core: &Core, rx: Receiver<SyncMsg>) {
         let mut ids: HashSet<i64> = HashSet::new();
         let mut all = false;
         let mut do_rewatch = false;
-        classify(msg, &mut ids, &mut all, &mut do_rewatch);
-        settle(&rx, &mut ids, &mut all, &mut do_rewatch);
+        if !classify(msg, &mut ids, &mut all, &mut do_rewatch)
+            || !settle(&rx, &mut ids, &mut all, &mut do_rewatch)
+        {
+            info!("sync engine stopping");
+            return;
+        }
         if all {
             reconcile_all(core);
         } else {
@@ -158,7 +173,7 @@ pub(super) fn settle_with(
     do_rewatch: &mut bool,
     quiet: Duration,
     cap: Duration,
-) {
+) -> bool {
     let deadline = Instant::now() + cap;
     loop {
         let now = Instant::now();
@@ -170,28 +185,49 @@ pub(super) fn settle_with(
         match rx.recv_timeout(wait) {
             // Another event: the burst is still going, so the quiet window
             // starts again from here.
-            Ok(m) => classify(m, ids, all, do_rewatch),
+            // A stop lands mid-burst: give up the rest of the burst rather
+            // than settling for it, and let the caller end the loop.
+            Ok(m) => {
+                if !classify(m, ids, all, do_rewatch) {
+                    return false;
+                }
+            }
             // Quiet for a full window: whatever was being written has stopped.
             Err(RecvTimeoutError::Timeout) => break,
             // Every sender is gone; the caller's loop is about to end anyway.
             Err(RecvTimeoutError::Disconnected) => break,
         }
     }
+    true
 }
 
 /// [`settle_with`] at the real timings.
-fn settle(rx: &Receiver<SyncMsg>, ids: &mut HashSet<i64>, all: &mut bool, do_rewatch: &mut bool) {
-    settle_with(rx, ids, all, do_rewatch, DEBOUNCE, MAX_SETTLE);
+fn settle(
+    rx: &Receiver<SyncMsg>,
+    ids: &mut HashSet<i64>,
+    all: &mut bool,
+    do_rewatch: &mut bool,
+) -> bool {
+    settle_with(rx, ids, all, do_rewatch, DEBOUNCE, MAX_SETTLE)
 }
 
-pub(super) fn classify(msg: SyncMsg, ids: &mut HashSet<i64>, all: &mut bool, rewatch: &mut bool) {
+/// Fold one message into the pending work of the current burst. Returns `false`
+/// for [`SyncMsg::Stop`], which is not work but an instruction to end the loop.
+pub(super) fn classify(
+    msg: SyncMsg,
+    ids: &mut HashSet<i64>,
+    all: &mut bool,
+    rewatch: &mut bool,
+) -> bool {
     match msg {
         SyncMsg::Reconcile(id) => {
             ids.insert(id);
         }
         SyncMsg::ReconcileAll => *all = true,
         SyncMsg::Rewatch => *rewatch = true,
+        SyncMsg::Stop => return false,
     }
+    true
 }
 
 /// Reconcile every mirror folder in turn.

crates/pdfs-fuse/src/sync/model.rs

@@ -2,9 +2,73 @@
 
 use super::*;
 
+/// What a reconcile pass remembers about the local side of one path: enough to
+/// recognise the same content again, and no more.
+///
+/// `mtime_ns` is the whole reason this is a struct rather than a pair — see
+/// [`StoredSyncEntry::local_mtime_ns`](pdfs_core::db::StoredSyncEntry).
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(super) struct LocalSig {
+    pub(super) mtime: i64,
+    pub(super) mtime_ns: Option<i64>,
+    pub(super) size: i64,
+}
+
+impl LocalSig {
+    /// Whether these two observations describe the same local content.
+    ///
+    /// Nanoseconds decide it when both sides have them; otherwise the comparison
+    /// falls back to whole seconds, which is what a baseline row written before
+    /// schema 28 — or a filesystem that reports no sub-second time — can support
+    /// (bugs.md B25).
+    pub(super) fn same_content(&self, other: &Self) -> bool {
+        if self.size != other.size {
+            return false;
+        }
+        match (self.mtime_ns, other.mtime_ns) {
+            (Some(a), Some(b)) => a == b,
+            _ => self.mtime == other.mtime,
+        }
+    }
+}
+
+impl From<&LocalItem> for LocalSig {
+    fn from(item: &LocalItem) -> Self {
+        Self {
+            mtime: item.mtime,
+            mtime_ns: item.mtime_ns,
+            size: item.size,
+        }
+    }
+}
+
+impl From<&std::fs::Metadata> for LocalSig {
+    fn from(meta: &std::fs::Metadata) -> Self {
+        Self {
+            mtime: system_mtime(meta),
+            mtime_ns: system_mtime_ns(meta),
+            size: meta.len() as i64,
+        }
+    }
+}
+
+/// The baseline's record of the local side, as [`LocalSig`] compares it.
+impl From<&StoredSyncEntry> for LocalSig {
+    fn from(entry: &StoredSyncEntry) -> Self {
+        Self {
+            mtime: entry.local_mtime,
+            mtime_ns: entry.local_mtime_ns,
+            size: entry.local_size,
+        }
+    }
+}
+
 pub(super) struct LocalItem {
     pub(super) is_dir: bool,
     pub(super) mtime: i64,
+    /// [`LocalItem::mtime`] to nanosecond precision, when the filesystem
+    /// reports one.
+    pub(super) mtime_ns: Option<i64>,
     pub(super) size: i64,
     /// True when another process holds this file open for writing (detected via
     /// `/proc/*/fd`). An open-for-write file is kept in the map — so it is not
@@ -91,11 +155,18 @@ pub(super) enum Pending {
     UploadRevision { rel: String, uid: NodeUid },
     /// Download remote `uid` to the local path, stamping `mtime`. `size` is the
     /// remote's reported size, used as the transfer's expected total.
+    ///
+    /// `local` is what the planning walk saw at this path, or `None` if nothing
+    /// was there. A download can run for minutes, so the apply step revalidates
+    /// against it immediately before it renames over the destination: a writer
+    /// that finished an edit in the meantime must not be overwritten without a
+    /// conflict copy (bugs.md B39).
     Download {
         rel: String,
         uid: NodeUid,
         mtime: i64,
         size: i64,
+        local: Option<LocalSig>,
     },
     /// Both sides changed: set the local copy aside, then download remote `uid`.
     Conflict {

crates/pdfs-fuse/src/sync/planner.rs

@@ -218,7 +218,10 @@ pub(super) fn plan_file(
     }
 
     let local_changed = local.is_some_and(|item| {
-        baseline.is_none_or(|base| base.local_mtime != item.mtime || base.local_size != item.size)
+        // Nanoseconds where both sides have them (bugs.md B25): whole seconds
+        // could not see an edit made in the same second as the last sync that
+        // left the file the same length, so it was never uploaded.
+        baseline.is_none_or(|base| !LocalSig::from(item).same_content(&LocalSig::from(base)))
     });
     let remote_changed = remote.is_some_and(|item| {
         baseline.is_none_or(|base| remote_sig(base) != Some((item.mtime, item.size)))
@@ -252,6 +255,7 @@ mod file_plan_tests {
 
     fn local(mtime: i64, size: i64) -> LocalItem {
         LocalItem {
+            mtime_ns: None,
             is_dir: false,
             mtime,
             size,
@@ -278,12 +282,49 @@ mod file_plan_tests {
             rel_path: "file".into(),
             remote_uid: Some("v~l".into()),
             local_mtime,
+            local_mtime_ns: None,
             local_size,
             remote_rev: Some(remote_mtime.to_string()),
             remote_hash: Some(remote_size.to_string()),
         }
     }
 
+    /// Whole seconds cannot see an edit made inside the same second that leaves the
+    /// file the same length, so a mirror silently stopped uploading it. With a
+    /// nanosecond time on both sides it is an ordinary local change (bugs.md B25).
+    #[test]
+    fn a_same_second_same_size_edit_is_still_a_local_change() {
+        let mut base = baseline(10, 20, 30, 40);
+        base.local_mtime_ns = Some(10_000_000_000);
+        let edited = LocalItem {
+            mtime_ns: Some(10_400_000_000),
+            ..local(10, 20)
+        };
+        let remote = remote(30, 40);
+        assert_eq!(
+            plan_file(Some(&edited), Some(&remote), Some(&base)),
+            FilePlan::UploadRevision
+        );
+
+        // The same file untouched still reads as unchanged.
+        let same = LocalItem {
+            mtime_ns: Some(10_000_000_000),
+            ..local(10, 20)
+        };
+        assert_eq!(
+            plan_file(Some(&same), Some(&remote), Some(&base)),
+            FilePlan::Unchanged
+        );
+
+        // And a row from before the column exists compares on seconds, exactly
+        // as it did when it was written — no mass re-upload on upgrade.
+        let old = baseline(10, 20, 30, 40);
+        assert_eq!(
+            plan_file(Some(&edited), Some(&remote), Some(&old)),
+            FilePlan::Unchanged
+        );
+    }
+
     #[test]
     fn file_plan_covers_every_presence_and_change_shape() {
         let same_local = local(10, 20);

docs/CHANGELOG.md

@@ -9,6 +9,166 @@ release ships. Migrations are forward-only — a database written by a newer cli
 refuse-to-open, not a downgrade, so rolling back a release means restoring the cache from
 scratch (user data in `staging/` and `recovery/` is never touched by this).
 
+## [1.8.2] — 2026-08-17
+
+Closes the two findings 1.8.1's live verification produced (B85, B86) and finishes the data-safety
+audit entries that had been sitting as "fixed in code, verification pending" since 1.0 (B49–B55).
+No schema change. Schema: **28**.
+
+The read path no longer assumes a revision is a series of 4 MiB blocks; it takes the geometry from
+the revision. The B85 investigation also **contradicted its own bug report** — see Notes, because
+the conclusion recorded there was that reads of such files were losing data, and they were not.
+
+### Fixed
+- A folder switched from mirror to on-demand now shows the files the mirror pass itself uploaded.
+  The sync engine created nodes on the remote without invalidating the parent's cached child
+  listing, and that listing is a database flag rather than an in-memory cache — so it survived a
+  daemon restart, and the files stayed invisible until the folder was switched back. Every remote
+  mutation the engine makes (folder create, file upload, and the three trash paths) now drops the
+  parent's cached listing by uid, which reaches a folder that no mount has resident — the case a
+  mirror folder always is (`docs/BUGS.md` B86).
+- `pdfs ls` and `pdfs refresh` accept a path under any mount the daemon owns, not only the primary
+  one. Both previously answered "is not under the mountpoint" for anything inside a secondary
+  on-demand mount, which left a folder serving a stale listing with no manual way to re-enumerate
+  it. A path under nothing this daemon has mounted still gets the same error as before.
+- A file whose blocks are not 4 MiB is no longer read twice over. Block offsets came from
+  `index * 4 MiB`, so on a revision laid out differently every block the client asked for
+  straddled two of the revision's own, and each was fetched and decrypted to answer for one — then
+  cached at boundaries the next read would not ask about. Offsets now come from the revision's
+  recorded block sizes, and the on-disk cache, the in-memory ring and the in-flight fetch key
+  address a block by the byte range it covers rather than by its index (`docs/BUGS.md` B85).
+
+### Changed
+- The block cache records each block's plaintext start alongside its `(mtime, size)` tag, and each
+  revision's block table in a small sidecar beside its blocks. Blocks cached by an earlier release
+  carry no start; they are honoured wherever the old 4 MiB assumption and the new plan agree on
+  the offset, which is every file this client uploaded — so the existing cache survives the
+  upgrade intact and is quietly refetched only for the files it was placing wrongly.
+
+### Notes
+- **B85's diagnosis was wrong, and the entry has been corrected rather than closed.** It read a
+  set of journal lines as proof that this client was requesting wrong byte ranges and being saved
+  only by B84's repair path. `RevisionReader::read_at` does not work that way: it plans over the
+  revision's real block sizes and clamps every block to the range asked for, identically in the
+  0.6.0 the client builds against and in every earlier version. A read can therefore never return
+  more bytes than it requested — yet every quoted line shows exactly that (`blen=4194304
+  got=4939506`), and the one place in the client that logs that shape can only reach it on the
+  opposite condition. The numbers were either transcribed with two fields swapped or captured from
+  a build not in the history; they want re-capturing from the journal before anything is concluded
+  from them. What was real is the assumption itself, and that is what this release removes — a
+  cost fix, not a correctness one. Nothing here changes what a read returns.
+- The first read of a file still plans on the 4 MiB assumption, and learns the real geometry as a
+  side effect of opening the reader it needs anyway. Resolving a reader merely to *plan* would
+  reintroduce the per-read key derivation the block cache exists to avoid, so the first read of a
+  non-4-MiB file pays one straddling fetch and then self-corrects.
+- Seven data-safety audit entries (B49–B55) moved from "fixed in code, verification pending" to
+  verified. The fixes were all genuinely present and needed no change; what they lacked was the
+  injected-failure testing each entry asked for, and that is what was written — unwritable
+  directories and `chmod 000` subtrees for permission failures, a real `/dev/shm`↔`/tmp` device
+  boundary for the cross-filesystem fallbacks, SQLite's `max_page_count` for a full disk, an abort
+  trigger for a failed insert, and truncated and unpublished sidecars for the crash boundaries.
+  Two seams were added to make failure arrangeable at all: `stage_write_at`, mirroring the
+  `preserve_write_at` that already existed, and `walk_local_tree` lifted out of the sync engine so
+  the local scan can be driven against a real tree without a daemon.
+- **Two of those entries are not fully closed, and say so.** B51 and B52 concern crash consistency,
+  and nothing short of a machine actually losing power tests that; the crash *boundaries* are
+  covered by reconstructing on disk exactly what a crash at each point leaves. B54 also diverges
+  from its required fix on purpose: it has no "deliberate delete" confirmation path, so a total
+  local wipe is refused outright rather than offered for confirmation. That is stricter than asked
+  for, and the cost is that a user who really did delete everything must remove and re-add the
+  sync folder.
+- B86's fix routes only `ls` and `refresh` to the owning mount. The other path-addressed control
+  requests — pin, rename, move, delete, sharing — still resolve against the primary mount only.
+  That is a larger change and was not what blocked the repro; it is recorded as an open follow-up
+  (`docs/BUGS.md` B88).
+- Three findings were opened rather than fixed: the geometry above is learned one read late, so a
+  non-4-MiB file pays a single wasted fetch on its first read (B87); the remaining path-addressed
+  control requests are still primary-only (B88); and the container-level `CLAUDE.md` still
+  describes the SDK dependency as 0.1.11 when 1.7.1 bumped it to 0.6.0, which is what sent the
+  B85 investigation to the wrong source tree to begin with (B89).
+
+## [1.8.1] — 2026-08-16
+
+Some files read short through the mount and reported success — an application got a truncated
+copy with nothing anywhere saying so (`docs/BUGS.md` B84). Found because GNOME's file indexer
+kept complaining that PNGs under `~/ProtonDrive/Pictures` were corrupt; it was right. The rest
+of this release is the long-standing architectural backlog from the July concurrency audits
+(B25, B32, B39, B40, B43, B44), each verified against the current code before being fixed.
+Schema: **28** — migration V28 adds `sync_entry.local_mtime_ns`.
+
+### Fixed
+- A read no longer trusts a revision whose recorded block sizes disagree with the file's size.
+  Such a file leaves the block/range path entirely: it is fetched through the whole-file,
+  manifest-verified download — the same path `pdfs pin` uses, and the reason pinning was the
+  workaround — and served from the resulting cached blob. One download per file however many
+  reads discover the problem at once, and files above 512 MiB fail `EIO` with a note to pin
+  them rather than pulling down gigabytes for one read.
+- A block that comes back shorter than the file's size implies is never served, cached or kept
+  in memory. Assembling a read out of short blocks is what quietly produced the truncation:
+  the missing bytes were simply not appended, the kernel handed userspace a short read, and
+  every tool reads a short read as end-of-file.
+- A synced folder no longer reports the same deletion as an error on every pass. Deleting a
+  local file that something else had already deleted counted as a failure, which kept its
+  baseline and made the next pass try again — forever. A path that is already gone is the
+  state the sync pass wanted (`docs/BUGS.md` B42).
+- A mirror folder no longer misses an edit made in the same second as its last sync. Change
+  detection compared whole seconds, so rewriting a file in place without changing its length
+  read as "unchanged" and never uploaded — the remote silently stayed behind until something
+  else about the file moved. Baselines now record the modification time to nanoseconds
+  (`docs/BUGS.md` B25).
+- A sync download no longer overwrites an edit that finished while it was running. The
+  destination is re-checked in the last moment before the download is published, and a local
+  file that has moved on since the pass planned for it is kept as a conflict copy — the same
+  resolution a both-sides-changed conflict gets (`docs/BUGS.md` B39).
+- An upload interrupted by a writer no longer reports itself as synchronized. The baseline now
+  records the content that actually went up rather than whatever is on disk when the transfer
+  ends, so a revision torn by a concurrent write is replaced by a clean one on the next pass
+  instead of standing as the file's content on Drive (`docs/BUGS.md` B40).
+- Listing a directory that changes while it is being read no longer skips or repeats entries.
+  `readdir` resumed by index into a listing rebuilt on every call, so a create or a trash
+  between two pages shifted everything after it. The listing is now frozen for the life of the
+  directory handle (`docs/BUGS.md` B43).
+- Opening a file for writing while a previous close is still publishing no longer discards that
+  close's revision. The base is re-checked at the moment the write handle is installed, rather
+  than trusted across a copy that can be the size of the whole file (`docs/BUGS.md` B32).
+- Unmounting now stops and joins the daemon's background workers instead of leaving them
+  running. The drain, sync engine, conflict sweep, online probe, local indexer and control
+  listener share a stop signal and interruptible waits, so an in-process remount no longer
+  starts a second set on top of the first (`docs/BUGS.md` B44).
+
+### Notes
+- The size `stat` reports was always the correct one; it is the reader's belief about where
+  blocks end that was wrong, so nothing about reported sizes changes.
+- Verified live on the affected account: four multi-block PNGs that had been reading short now
+  deliver their full size, and the journal shows each one repaired. The diagnostics also
+  answered the open question, and not as expected — the revision metadata is fine; those files
+  simply do not have 4 MiB blocks, and this client assumes they do. Filed as `docs/BUGS.md`
+  B85. Until that is fixed, a file with non-4-MiB blocks is downloaded whole on its first read.
+- Migration V28's new column is nullable rather than converted from the existing seconds. A
+  baseline row written before this release has no sub-second time to recover, and inventing
+  zero for it would make every already-synced file compare as changed and re-upload an entire
+  mirror on first start; `NULL` keeps those rows comparing exactly as they did, and each gains
+  a real value the next time its path syncs.
+- Three audit entries closed without a code change, having been fixed already or being correct
+  as they stand: sync-folder removal is serialized against reconciliation (B41), the drain's
+  database contention was removed by the read-only connection pool and the claim column (B30),
+  and the refusals for symlinks, hard links, device nodes and `chmod` are deliberate — Drive
+  cannot represent them, and a local-only representation no other Proton client can read would
+  be worse than a clean refusal (B31).
+- Six of this release's fixes were then driven live against the installed 1.8.1 on a real
+  account: the same-second edit uploads (B25), three rapid rewrites of a 64 MiB file publish the
+  third one byte for byte (B32), a write racing a download survives as a conflict copy (B39), an
+  upload torn by a concurrent write heals itself on the following pass (B40), a `getdents64`
+  walk with a 160-byte buffer against a directory mutated between pages returns every entry
+  exactly once (B43), and a restart of a daemon with six mounts and a busy drain stops in 110 ms
+  and comes back at the same thread count (B44). B42 could not be driven from a single machine —
+  see B86 below.
+- One new bug came out of the verification: an on-demand mount serves a folder listing that
+  predates the sync engine's own uploads, so files the mirror pass created are invisible until
+  the folder is switched back (`docs/BUGS.md` B86). Nothing is lost — they are on the server and
+  come back on the next mirror pass — but they cannot be seen or named while the folder is
+  on-demand, which is also what blocks B42's live verification.
+
 ## [1.8.0] — 2026-08-16
 
 A queued write whose node had left the local tree was retried every five seconds forever and
@@ -279,6 +439,9 @@ First stable release: FUSE files-on-demand mount, sync daemon under `proton-driv
 - The outstanding FUSE defects tracked in `docs/BUGS.md`, plus a truncate defect, validated
   by a new POSIX compliance suite for the filesystem.
 
+[1.8.2]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.8.2
+[1.8.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.8.1
+[1.8.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.8.0
 [1.7.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.7.1
 [1.7.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.7.0
 [1.6.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.6.0

packaging/PKGBUILD

@@ -9,7 +9,7 @@
 # Build + install:   cd packaging && makepkg -fi
 
 pkgname=proton-drive-linux
-pkgver=1.8.0
+pkgver=1.8.2
 pkgrel=1
 pkgdesc="Proton Drive client for Linux: FUSE files-on-demand mount, CLI, GTK4 app + tray"
 arch=('x86_64')