release(1.8.2)

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(())

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, block_len};
+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,
@@ -715,6 +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
@@ -737,6 +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);
@@ -751,6 +772,7 @@ impl StateRegistry {
             state: Arc::downgrade(state),
             notifier,
             session_live,
+            size_upgrades,
         });
     }
 
@@ -777,9 +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()
@@ -789,6 +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)
@@ -908,9 +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.
     ///
@@ -3934,6 +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
@@ -5853,7 +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()),
@@ -5904,19 +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()),
@@ -5994,6 +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();
@@ -6007,13 +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()),
@@ -6041,7 +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()),
@@ -6104,7 +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()),
@@ -6372,7 +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()),
@@ -6426,7 +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/reads.rs

@@ -102,22 +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.
@@ -128,14 +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;
         }
@@ -176,8 +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>>);
@@ -417,11 +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,
@@ -500,7 +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
@@ -512,10 +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) {
@@ -547,12 +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.
@@ -578,11 +622,9 @@ impl Core {
         uid: &NodeUid,
         mtime: i64,
         fsize: u64,
-        bidx: u64,
+        span: BlockSpan,
         cache_blocks: bool,
     ) -> BlockResult {
-        let bstart = bidx * BLOCK_SIZE;
-        let blen = block_len(fsize, bidx);
         // 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
@@ -591,25 +633,33 @@ impl Core {
         // kind takes the file off the range path entirely (bugs.md B84).
         let bytes = if reader.size() != fsize {
             warn!(
-                %uid, bidx, fsize,
+                %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, bidx).await?
+            self.repair_block(uid, mtime, fsize, span).await?
         } else {
-            let bytes = reader.read_at(bstart, blen).await.map_err(|e| {
-                warn!(%uid, bstart, blen, error = %e, "block read failed");
+            // 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 == blen {
+            if bytes.len() as u64 == span.len {
                 bytes
             } else {
                 warn!(
-                    %uid, bidx, blen, got = bytes.len(), fsize,
-                    "block came back shorter than the file's size implies"
+                    %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, bidx).await?
+                self.repair_block(uid, mtime, fsize, span).await?
             }
         };
         let bytes = Arc::new(bytes);
@@ -618,16 +668,16 @@ impl Core {
             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 block `bidx` of a file whose per-block reads cannot be trusted,
+    /// 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.
     ///
@@ -648,10 +698,9 @@ impl Core {
         uid: &NodeUid,
         mtime: i64,
         fsize: u64,
-        bidx: u64,
+        span: BlockSpan,
     ) -> Result<Vec<u8>, Errno> {
-        let blen = block_len(fsize, bidx);
-        let bstart = bidx * BLOCK_SIZE;
+        let (bstart, blen) = (span.start, span.len);
         if fsize > REPAIR_MAX {
             warn!(
                 %uid, fsize,
@@ -683,11 +732,11 @@ impl Core {
         }
         match self.cache.read_range(uid, mtime, fsize, bstart, blen) {
             Some(bytes) if bytes.len() as u64 == blen => {
-                info!(%uid, bidx, "served a short block from a repaired whole-file download");
+                info!(%uid, bidx = span.idx, "served a short block from a repaired whole-file download");
                 Ok(bytes)
             }
             _ => {
-                error!(%uid, bidx, fsize, "repairing a short block did not produce the block");
+                error!(%uid, bidx = span.idx, fsize, "repairing a short block did not produce the block");
                 Err(Errno::EIO)
             }
         }
@@ -776,17 +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() {
@@ -803,7 +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
                 };
@@ -811,10 +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");
                     }
                 });
             }
@@ -844,8 +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
@@ -853,31 +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);
                 }
             }
         }
@@ -890,12 +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());
@@ -913,18 +969,20 @@ 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 file's size implies — 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 != block_len(fsize, bidx) {
-                error!(%uid, bidx, len = block.len(), fsize, "assembling a read hit a short block");
+            // 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;
@@ -1016,13 +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 {
@@ -1142,31 +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);
     }
 
@@ -1175,13 +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);
     }
 
@@ -1189,9 +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));
@@ -1201,11 +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/sync.rs

@@ -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,
@@ -777,6 +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,
@@ -832,6 +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,
@@ -857,70 +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,
-                        mtime_ns: None,
-                        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),
-                        mtime_ns: system_mtime_ns(&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
@@ -1186,6 +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))
             }
@@ -1295,6 +1246,10 @@ impl Core {
             )
             .await
             .map_err(|e| format!("upload {rel}: {e}"))?;
+        // 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
@@ -1546,6 +1501,95 @@ fn removed_locally(result: std::io::Result<()>) -> std::io::Result<()> {
     }
 }
 
+/// 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> {
@@ -1579,6 +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(),
@@ -1925,6 +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
@@ -2126,6 +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)