Merge pull request 'Export/import: keep archived items, spam samples and the spam model' (#31) from fix/export-all-subspaces into main

--export skipped three things, so a move from one database to another
(RocksDB to PostgreSQL, say) lost them without a word:

- archived items (subspace j), the records behind undelete;
- spam training samples (subspace w);
- the trained spam classifier and its trainer state, blobs stored under
  fixed names that no blob link points at, so the walk over links never
  reached them.

j and w now travel with the registry family, where their indexes and id
counters already were, so EXPORT_TYPES=registry keeps them consistent.
The two named blobs travel with the blob family. The file format is
unchanged and import reads any subspace it is given, so an export made
by an older binary still imports.

The full-text index (subspace z) stays out, on purpose. It belongs to one
search backend: PostgreSQL and MySQL index into their own tables and have
no z table at all, and external engines keep the index themselves. So
--import now returns the subspaces it wrote, and boot queues the
reindexAccounts and reindexTelemetry store maintenance tasks, the same
ones an administrator can queue by hand, to rebuild the index for
whichever search store the server runs with once it starts.

The round trip also turned up a loss in import itself: the SQL stores
add a negative amount with an UPDATE, which does nothing to a row that
isn't there yet, so every negative counter or quota vanished on import
into PostgreSQL, MySQL or SQLite. Import now creates the row first.

The in-memory subspaces (m, y) stay out: rate limits, locks, greylisting,
ACME challenge tokens and OAuth codes, all short-lived. Issued
certificates are registry objects and travel.

The store test now writes archived items, spam samples, directory
entries, the fork's own subspace and the named blobs, checks they come
back in place, then imports the same export into a fresh store of the
other local backend (RocksDB to SQLite, or SQLite to RocksDB), compares
it key for key and counter for counter, and checks the queued reindex.
It fails on the old export code ("Subspace j was not exported").
--help now says what an export holds.

.dockerignore

@@ -1,10 +1,16 @@
-// Ignore everything
+# Ignore everything
 *
 
-// Allow what is needed
+# Allow what is needed
 !crates
 !tests
 !resources
 
+# The patched dependency Cargo.toml's [patch.crates-io] points at. Without
+# it the build context has no vendor/, and `cargo chef cook` fails on
+# "failed to load source for dependency sieve-rs" -- which CI cannot see,
+# because CI builds from a checkout and only the image build has a context.
+!vendor
+
 !Cargo.lock
 !Cargo.toml

.gitea/workflows/ci.yml

@@ -35,6 +35,11 @@ jobs:
       - run: python3 tools/fork/name-check.py
       - if: always()
         run: python3 tools/fork/notice-check.py
+      # Cargo can patch a dependency to a directory in this repository, and
+      # the image builds from a context .dockerignore prunes to almost
+      # nothing. CI never sees the difference; a release does.
+      - if: always()
+        run: python3 tools/fork/context-check.py
 
   build:
     # Either runner (host1 or host2): the build needs no docker socket.

Dockerfile

@@ -19,6 +19,10 @@ RUN export DEBIAN_FRONTEND=noninteractive && \
     g++-x86-64-linux-gnu binutils-x86-64-linux-gnu
 RUN rustup target add "$(cat /target.txt)"
 COPY --from=planner /recipe.json /recipe.json
+# inbuxa: [patch.crates-io] points sieve-rs at vendor/, and the recipe only
+# carries the workspace's own manifests, so cooking the dependencies needs the
+# vendored crate itself (the context allows it since #27; this puts it here).
+COPY vendor/ vendor/
 RUN RUSTFLAGS="$(cat /flags.txt)" cargo chef cook --target "$(cat /target.txt)" --release --no-default-features --features "sqlite postgres mysql rocks s3 redis azure nats" --recipe-path /recipe.json
 COPY . .
 RUN RUSTFLAGS="$(cat /flags.txt)" cargo build --target "$(cat /target.txt)" --release -p inbuxa --no-default-features --features "sqlite postgres mysql rocks s3 redis azure nats"

crates/common/src/manager/backup.rs

@@ -23,6 +23,13 @@ use utils::{UnwrapFailure, codec::leb128::Leb128_};
 
 pub(super) const MAGIC_MARKER: u8 = 123;
 
+// inbuxa: blobs kept under a fixed name instead of a content hash. Nothing
+// links to them, so the export names them outright.
+const NAMED_BLOBS: &[&[u8]] = &[
+    crate::manager::SPAM_CLASSIFIER_KEY,
+    crate::manager::SPAM_TRAINER_KEY,
+];
+
 #[derive(Debug, Clone, Copy, Hash, PartialEq, Eq)]
 pub(super) enum Family {
     Data = 0,
@@ -143,15 +150,21 @@ impl Core {
                     .await
                     .failed("Failed to iterate over data store");
 
-                for hash in blobs {
+                // inbuxa: the trained spam classifier and its trainer state are
+                // blobs stored under fixed names with no blob link, so the walk
+                // over links above never reaches them.
+                let named = NAMED_BLOBS.iter().map(|key| key.to_vec());
+                for key in blobs
+                    .into_iter()
+                    .map(|hash| hash.as_slice().to_vec())
+                    .chain(named)
+                {
                     if let Some(blob) = blob_store
-                        .get_blob(hash.as_slice(), 0..usize::MAX)
+                        .get_blob(&key, 0..usize::MAX)
                         .await
                         .failed("Failed to get blob")
                     {
-                        writer
+                        writer.send((key, blob)).failed("Failed to send key");
-                            .send((hash.as_slice().to_vec(), blob))
-                            .failed("Failed to send key");
                     }
                 }
             }),
@@ -323,7 +336,13 @@ impl Family {
                 SUBSPACE_REGISTRY_IDX,
                 SUBSPACE_REGISTRY_PK,
                 SUBSPACE_DIRECTORY,
-                store::SUBSPACE_INBUXA, // inbuxa: masked email
+                // inbuxa: registry objects the upstream list left out, so an
+                // export dropped them: archived items (undelete) and spam
+                // training samples. Their indexes and id counters already
+                // travel in this family and in `data`, so they ride along.
+                SUBSPACE_DELETED_ITEMS,
+                SUBSPACE_SPAM_SAMPLES,
+                store::SUBSPACE_INBUXA, // inbuxa: the fork's own data (masked email, undelete, policies)
             ],
             Family::Changelog => &[SUBSPACE_LOGS],
             Family::Queue => &[SUBSPACE_QUEUE_MESSAGE, SUBSPACE_QUEUE_EVENT],

crates/common/src/manager/boot.rs

@@ -54,6 +54,13 @@ Options:
   -o, --console                    Open the store console
   -h, --help                       Print help
   -V, --version                    Print version
+
+An export holds everything in the data and blob stores except short-lived
+in-memory state (rate limits, locks, greylisting) and the full-text search
+index, which belongs to one search backend. An import into an empty store
+queues the index to be rebuilt when the server next starts. EXPORT_TYPES
+limits an export to some of: data, registry, blob, changelog, queue, report,
+telemetry, tasks.
 "#
 );
 
@@ -256,10 +263,10 @@ impl BootManager {
                 telemetry.enable();
 
                 // Parse settings and restore
-                Box::pin(Core::parse(&mut bootstrap, storage))
+                let core = Box::pin(Core::parse(&mut bootstrap, storage)).await;
-                    .await
+                let imported = core.restore(path).await;
-                    .restore(path)
+                // inbuxa: the search index isn't exported; rebuild it
-                    .await;
+                core.queue_reindex(&imported).await;
                 std::process::exit(0);
             }
             StoreOp::Console => {

crates/common/src/manager/restore.rs

@@ -9,15 +9,22 @@
 use super::backup::MAGIC_MARKER;
 use crate::{Core, DATABASE_SCHEMA_VERSION};
 use lz4_flex::frame::FrameDecoder;
-use registry::schema::enums::CompressionAlgo;
+use registry::{
+    schema::{
+        enums::{CompressionAlgo, TaskStoreMaintenanceType},
+        structs::{Task, TaskStatus, TaskStoreMaintenance},
+    },
+    types::EnumImpl,
+};
 use std::{
     fs::File,
     io::{BufReader, ErrorKind, Read},
     path::{Path, PathBuf},
 };
 use store::{
-    BlobStore, IterateParams, SUBSPACE_BLOBS, SUBSPACE_COUNTER, SUBSPACE_INDEXES, SUBSPACE_QUOTA,
+    BlobStore, IterateParams, SUBSPACE_BLOBS, SUBSPACE_COUNTER, SUBSPACE_INDEXES,
-    SUBSPACE_REGISTRY_PK, Store, U32_LEN,
+    SUBSPACE_PROPERTY, SUBSPACE_QUOTA, SUBSPACE_REGISTRY_PK, SUBSPACE_TELEMETRY_SPAN, Store,
+    U32_LEN,
     write::{
         AnyClass, AnyKey, BatchBuilder, ValueClass,
         key::{DeserializeBigEndian, is_node_id_key},
@@ -27,7 +34,9 @@ use types::{collection::Collection, field::Field};
 use utils::{UnwrapFailure, failed};
 
 impl Core {
-    pub async fn restore(&self, src: PathBuf) {
+    /// Imports an export into an empty store and returns the subspaces it
+    /// wrote. inbuxa: the caller hands them to [`Core::queue_reindex`].
+    pub async fn restore(&self, src: PathBuf) -> Vec<u8> {
         // Backup the core
         let paths = if src.is_dir() {
             let mut paths = Vec::new();
@@ -64,6 +73,13 @@ impl Core {
             std::process::exit(1);
         }
 
+        let mut imported = paths
+            .iter()
+            .map(|path| KeyValueReader::new(path).subspace)
+            .collect::<Vec<_>>();
+        imported.sort_unstable();
+        imported.dedup();
+
         let mut tasks = Vec::new();
         for path in paths {
             let storage = self.storage.clone();
@@ -76,6 +92,54 @@ impl Core {
         for task in tasks {
             task.await.failed("Failed to wait for task");
         }
+
+        imported
+    }
+
+    /// inbuxa: an export never carries the full-text index. It is built by
+    /// and for one search backend (the SQL stores index into their own
+    /// tables, the key-value stores into a subspace, external engines keep it
+    /// themselves), so it would be wrong or unreadable after a move to
+    /// another one. Instead, an import queues the same reindex tasks an
+    /// administrator can queue by hand (`reindexAccounts` and
+    /// `reindexTelemetry` store maintenance), and the server rebuilds the
+    /// index for whatever search store it is configured with once it starts.
+    pub async fn queue_reindex(&self, imported: &[u8]) -> Vec<TaskStoreMaintenanceType> {
+        let mut queued = Vec::new();
+        if imported.contains(&SUBSPACE_PROPERTY) {
+            queued.push(TaskStoreMaintenanceType::ReindexAccounts);
+        }
+        if imported.contains(&SUBSPACE_TELEMETRY_SPAN) {
+            queued.push(TaskStoreMaintenanceType::ReindexTelemetry);
+        }
+        if queued.is_empty() {
+            return queued;
+        }
+
+        let mut batch = BatchBuilder::new();
+        for maintenance_type in &queued {
+            batch.schedule_task(Task::StoreMaintenance(TaskStoreMaintenance {
+                maintenance_type: *maintenance_type,
+                status: TaskStatus::now(),
+                shard_index: None,
+            }));
+        }
+        self.storage
+            .data
+            .write(batch.build_all())
+            .await
+            .failed("Failed to queue the reindex tasks");
+
+        println!(
+            "Queued {} to rebuild the search index; it runs when the server starts.",
+            queued
+                .iter()
+                .map(|t| t.as_str())
+                .collect::<Vec<_>>()
+                .join(" and ")
+        );
+
+        queued
     }
 }
 
@@ -125,17 +189,22 @@ async fn restore_file(store: Store, blob_store: BlobStore, path: &Path) {
         }
         SUBSPACE_COUNTER | SUBSPACE_QUOTA => {
             while let Some((key, value)) = reader.next() {
-                batch.add(
+                let class = ValueClass::Any(AnyClass {
-                    ValueClass::Any(AnyClass {
                     subspace: reader.subspace,
                     key,
-                    }),
+                });
-                    u64::from_le_bytes(
+                let value = u64::from_le_bytes(
                     value
                         .try_into()
                         .expect("Failed to deserialize counter/quota"),
-                    ) as i64,
+                ) as i64;
-                );
+                // inbuxa: the SQL stores add a negative amount with an UPDATE,
+                // which does nothing to a row that isn't there yet, so a
+                // negative counter vanished on import. Create the row first.
+                if value < 0 {
+                    batch.add(class.clone(), 0);
+                }
+                batch.add(class, value);
                 if batch.is_large_batch() {
                     store
                         .write(batch.build_all())

crates/types/src/branding.rs

@@ -81,7 +81,7 @@ fn legacy_setting(name: &str, is_set: impl Fn(&str) -> bool) -> Option<String> {
 #[macro_export]
 macro_rules! brand_version {
     () => {
-        "2026.9.24"
+        "2026.9.24.3"
     };
 }
 

tests/src/store/import_export.rs

@@ -2,6 +2,8 @@
  * SPDX-FileCopyrightText: 2020 Stalwart Labs LLC <[email protected]>
  *
  * SPDX-License-Identifier: AGPL-3.0-only OR LicenseRef-SEL
+ *
+ * Modified by Coffey Labs in 2026 for INBUXA.
  */
 
 use crate::utils::{
@@ -9,14 +11,22 @@ use crate::utils::{
     server::TestServer,
     temp_dir::TempDir,
 };
-use ::registry::schema::enums::CompressionAlgo;
+use ::registry::schema::{
+    enums::{CompressionAlgo, TaskStoreMaintenanceType},
+    prelude::ObjectType,
+    structs::Task,
+};
 use ahash::AHashSet;
-use common::{DATABASE_SCHEMA_VERSION, manager::backup::BackupParams};
+use common::{
+    DATABASE_SCHEMA_VERSION,
+    manager::{SPAM_CLASSIFIER_KEY, SPAM_TRAINER_KEY, backup::BackupParams},
+};
 use store::{
     rand,
     write::{
         AnyClass, AnyKey, BatchBuilder, BlobLink, BlobOp, Operation, QueueClass, QueueEvent,
-        RegistryClass, ValueClass, key::KeySerializer,
+        RegistryClass, TaskQueueClass, ValueClass,
+        key::{DeserializeBigEndian, KeySerializer},
     },
     *,
 };
@@ -167,6 +177,50 @@ pub async fn test(test: &TestServer) {
     }
     db.write(batch.build_all()).await.unwrap();
 
+    // inbuxa: registry objects kept outside the registry subspace (archived
+    // items for undelete, spam training samples, directory entries) and the
+    // fork's own subspace. Exports used to leave the first two behind.
+    println!("Creating archived items, spam samples and fork data...");
+    let mut batch = BatchBuilder::new();
+    for item_id in [1u64, 2, 3] {
+        for object in [
+            ObjectType::ArchivedItem,
+            ObjectType::SpamTrainingSample,
+            ObjectType::Account,
+        ] {
+            batch.set(
+                ValueClass::Registry(RegistryClass::Item {
+                    object_id: object as u16,
+                    item_id,
+                }),
+                random_bytes(item_id as usize * 64),
+            );
+        }
+        batch.set(
+            ValueClass::Any(AnyClass {
+                subspace: SUBSPACE_INBUXA,
+                key: [b'U', b'x']
+                    .into_iter()
+                    .chain(item_id.to_be_bytes())
+                    .collect(),
+            }),
+            random_bytes(32),
+        );
+    }
+    db.write(batch.build_all()).await.unwrap();
+
+    // inbuxa: the trained spam classifier lives in blobs with fixed names
+    let mut named_blobs = Vec::new();
+    for key in [SPAM_CLASSIFIER_KEY, SPAM_TRAINER_KEY] {
+        let data = random_bytes(4096);
+        test.server
+            .blob_store()
+            .put_blob(key, &data, CompressionAlgo::Lz4)
+            .await
+            .unwrap();
+        named_blobs.push((key, data));
+    }
+
     // Create directory data
     println!("Creating directory data...");
     let mut batch = BatchBuilder::new();
@@ -185,6 +239,17 @@ pub async fn test(test: &TestServer) {
     println!("Calculating store hash...");
     let snapshot = Snapshot::new(&db).await;
     assert!(!snapshot.keys.is_empty(), "Store hash counts are empty",);
+    for subspace in [
+        SUBSPACE_DELETED_ITEMS,
+        SUBSPACE_SPAM_SAMPLES,
+        SUBSPACE_INBUXA,
+    ] {
+        assert!(
+            snapshot.keys.iter().any(|k| k.subspace == subspace),
+            "No test data in subspace {}",
+            char::from(subspace)
+        );
+    }
 
     // Export store
     println!("Exporting store...");
@@ -210,22 +275,188 @@ pub async fn test(test: &TestServer) {
             .finalize(),
     );
     db.write(batch.build_all()).await.unwrap();
-    test.server.core.restore(temp_dir.path.clone()).await;
+    for (key, _) in &named_blobs {
+        test.server.blob_store().delete_blob(key).await.unwrap();
+    }
+    let imported = test.server.core.restore(temp_dir.path.clone()).await;
     let mut batch = BatchBuilder::new();
     batch.clear(ValueClass::NodeId(0));
     db.write(batch.build_all()).await.unwrap();
+    for subspace in [
+        SUBSPACE_DELETED_ITEMS,
+        SUBSPACE_SPAM_SAMPLES,
+        SUBSPACE_INBUXA,
+    ] {
+        assert!(
+            imported.contains(&subspace),
+            "Subspace {} was not exported",
+            char::from(subspace)
+        );
+    }
 
     // Verify hash
     print!("Verifying store hash...");
     snapshot.assert_is_eq(&Snapshot::new(&db).await);
+    assert_named_blobs(test.server.blob_store(), &named_blobs).await;
     println!(" GREAT SUCCESS!");
 
+    // inbuxa: import the same export into a fresh store of another backend,
+    // the way a move from one database to another does it
+    #[cfg(all(feature = "rocks", feature = "sqlite"))]
+    cross_backend(test, &db, &temp_dir, &named_blobs).await;
+
     // Destroy store
+    for (key, _) in &named_blobs {
+        test.server.blob_store().delete_blob(key).await.unwrap();
+    }
     store_destroy(&db).await;
     store_assert_is_empty(&db, db.clone().into(), true).await;
     temp_dir.delete();
 }
 
+#[cfg(all(feature = "rocks", feature = "sqlite"))]
+async fn cross_backend(
+    test: &TestServer,
+    source: &Store,
+    export: &TempDir,
+    named_blobs: &[(&[u8], Vec<u8>)],
+) {
+    let source_type = std::env::var("STORE").unwrap();
+    let target_type = if source_type.eq_ignore_ascii_case("sqlite") {
+        "RocksDb"
+    } else {
+        "Sqlite"
+    };
+    println!("Importing the export into a fresh {target_type} store...");
+
+    let target_dir = TempDir::new("art_vandelay_cross_backend", true);
+    let target = Store::build(
+        crate::utils::storage::build_data_store(target_type, &target_dir.path.to_string_lossy())
+            .await,
+    )
+    .await
+    .unwrap();
+    target.create_tables().await.unwrap();
+    store_destroy(&target).await;
+
+    let mut core = test.server.core.as_ref().clone();
+    core.storage.data = target.clone();
+    core.storage.blob = target.clone().into();
+    let imported = core.restore(export.path.clone()).await;
+
+    // Counters are stored differently by the SQL and key-value backends, so
+    // compare their keys here and their values through the counter API.
+    print!("Verifying {target_type} store hash...");
+    Snapshot::new_portable(source)
+        .await
+        .assert_is_eq(&Snapshot::new_portable(&target).await);
+    for subspace in [SUBSPACE_COUNTER, SUBSPACE_QUOTA] {
+        let mut keys = Vec::new();
+        source
+            .iterate(
+                IterateParams::new(
+                    AnyKey {
+                        subspace,
+                        key: vec![0u8],
+                    },
+                    AnyKey {
+                        subspace,
+                        key: vec![u8::MAX; 10],
+                    },
+                )
+                .no_values(),
+                |key, _| {
+                    keys.push(key.to_vec());
+                    Ok(true)
+                },
+            )
+            .await
+            .unwrap();
+        for key in keys {
+            let class = || {
+                ValueClass::Any(AnyClass {
+                    subspace,
+                    key: key.clone(),
+                })
+            };
+            assert_eq!(
+                source.get_counter(class()).await.unwrap(),
+                target.get_counter(class()).await.unwrap(),
+                "Counter mismatch in {} for {key:?}",
+                char::from(subspace)
+            );
+        }
+    }
+    assert_named_blobs(&core.storage.blob, named_blobs).await;
+    println!(" GREAT SUCCESS!");
+
+    // The search index isn't exported; the import queues its rebuild
+    let queued = core.queue_reindex(&imported).await;
+    let expected = [
+        TaskStoreMaintenanceType::ReindexAccounts,
+        TaskStoreMaintenanceType::ReindexTelemetry,
+    ];
+    assert_eq!(queued, expected);
+    let mut task_ids = Vec::new();
+    target
+        .iterate(
+            IterateParams::new(
+                AnyKey {
+                    subspace: SUBSPACE_TASK_QUEUE,
+                    key: vec![0u8],
+                },
+                AnyKey {
+                    subspace: SUBSPACE_TASK_QUEUE,
+                    key: vec![u8::MAX; 20],
+                },
+            )
+            .no_values(),
+            |key, _| {
+                if key.deserialize_be_u64(0)? == 0 {
+                    task_ids.push(key.deserialize_be_u64(U64_LEN)?);
+                }
+                Ok(true)
+            },
+        )
+        .await
+        .unwrap();
+    let mut found = Vec::new();
+    for id in task_ids {
+        match target
+            .get_value::<Task>(ValueKey::from(ValueClass::TaskQueue(
+                TaskQueueClass::Task { id },
+            )))
+            .await
+            .unwrap()
+        {
+            Some(Task::StoreMaintenance(task)) => found.push(task.maintenance_type),
+            other => panic!("Unexpected task {other:?}"),
+        }
+    }
+    found.sort_by_key(|t| *t as u16);
+    assert_eq!(found, expected, "Queued tasks don't match");
+
+    store_destroy(&target).await;
+    drop(core);
+    drop(target);
+    target_dir.delete();
+}
+
+async fn assert_named_blobs(blob_store: &BlobStore, named_blobs: &[(&[u8], Vec<u8>)]) {
+    for (key, data) in named_blobs {
+        assert_eq!(
+            blob_store
+                .get_blob(key, 0..usize::MAX)
+                .await
+                .unwrap()
+                .as_ref(),
+            Some(data),
+            "Blob {} was not restored",
+            String::from_utf8_lossy(key)
+        );
+    }
+}
+
 #[derive(Debug, PartialEq, Eq)]
 struct Snapshot {
     keys: AHashSet<KeyValue>,
@@ -240,7 +471,19 @@ struct KeyValue {
 
 impl Snapshot {
     async fn new(db: &Store) -> Self {
-        let is_sql = db.is_sql();
+        Self::build(db, !db.is_sql(), true).await
+    }
+
+    /// Comparable across backends: no counter values, which the SQL and
+    /// key-value stores encode differently, and no blobs, which only live in
+    /// the data store when it doubles as the blob store.
+    #[cfg(all(feature = "rocks", feature = "sqlite"))]
+    async fn new_portable(db: &Store) -> Self {
+        Self::build(db, false, false).await
+    }
+
+    async fn build(db: &Store, counter_values: bool, with_blobs: bool) -> Self {
+        let is_sql = !counter_values;
 
         let mut keys = AHashSet::new();
 
@@ -265,7 +508,12 @@ impl Snapshot {
             (SUBSPACE_QUOTA, !is_sql),
             (SUBSPACE_REPORT_OUT, true),
             (SUBSPACE_REPORT_IN, true),
+            (SUBSPACE_DIRECTORY, true),
+            (SUBSPACE_INBUXA, true),
         ] {
+            if subspace == SUBSPACE_BLOBS && !with_blobs {
+                continue;
+            }
             let from_key = AnyKey {
                 subspace,
                 key: vec![0u8],

tools/fork/context-check.py

@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: 2026 Coffey Labs
+# SPDX-License-Identifier: AGPL-3.0-or-later
+"""Every path Cargo patches has to be in the image's build context.
+
+Cargo.toml's [patch.crates-io] can point at a directory in this repository,
+and the Dockerfile builds from a context that .dockerignore prunes to almost
+nothing. Those two facts met on 2026-09-23: a vendored, patched sieve-rs
+landed, CI stayed green -- it builds from a checkout, where the directory is
+simply there -- and the release build failed on
+
+    failed to load source for dependency `sieve-rs`
+    failed to read /build/vendor/sieve-rs/Cargo.toml
+
+after a tag had already been pushed. This is seconds, and it runs beside the
+other fork checks rather than waiting for a release to find out.
+
+Being in the context isn't enough on its own: the Dockerfile cooks the
+dependencies (`cargo chef cook`) before it copies the tree in, from a recipe
+that carries only the workspace's manifests. So each patched path must also be
+copied into that stage before the cook step, or the same error comes back
+there -- as it did for 2026.9.24.2, the first tag after the context fix.
+"""
+
+import re
+import sys
+from pathlib import Path
+
+root = Path(__file__).resolve().parents[2]
+
+
+def patched_paths(manifest: Path) -> list[str]:
+    """Directories named by a [patch...] section's `path = "..."` entries."""
+    out, in_patch = [], False
+    for line in manifest.read_text().splitlines():
+        stripped = line.strip()
+        if stripped.startswith("["):
+            in_patch = stripped.startswith("[patch")
+            continue
+        if not in_patch:
+            continue
+        m = re.search(r'path\s*=\s*"([^"]+)"', stripped)
+        if m:
+            out.append(m.group(1))
+    return out
+
+
+def allowed(dockerignore: Path) -> set[str]:
+    """The first path segment of every re-inclusion rule."""
+    keep = set()
+    for line in dockerignore.read_text().splitlines():
+        stripped = line.strip()
+        if stripped.startswith("!"):
+            keep.add(stripped[1:].strip("/").split("/")[0])
+    return keep
+
+
+def copied_before_cook(dockerfile: Path) -> list[str] | None:
+    """Sources COPY'd into the stage that runs `cargo chef cook`, before it.
+
+    None when no stage cooks. A `COPY . .` covers everything.
+    """
+    stage: list[str] = []
+    for line in dockerfile.read_text().splitlines():
+        stripped = line.strip()
+        if re.match(r"(?i)^FROM\s", stripped):
+            stage = []
+            continue
+        if "cargo chef cook" in stripped:
+            return stage
+        m = re.match(r"(?i)^COPY\s+(?!--from)(.+)$", stripped)
+        if m:
+            parts = m.group(1).split()
+            stage.extend(p.strip("./").split("/")[0] or "." for p in parts[:-1])
+    return None
+
+
+def main() -> int:
+    paths = patched_paths(root / "Cargo.toml")
+    if not paths:
+        print("no patched paths to check")
+        return 0
+    keep = allowed(root / ".dockerignore")
+    bad = []
+    for p in paths:
+        top = p.strip("/").split("/")[0]
+        if top not in keep:
+            bad.append((p, top))
+        elif not (root / p).is_dir():
+            bad.append((p, None))
+    for path, top in bad:
+        if top is None:
+            print(f"Cargo.toml patches {path}, which does not exist", file=sys.stderr)
+        else:
+            print(
+                f"Cargo.toml patches {path}, but .dockerignore does not re-include {top!r}:\n"
+                f"  the image build would not see it, and cargo would fail on it.\n"
+                f"  Add `!{top}` to .dockerignore.",
+                file=sys.stderr,
+            )
+    copied = copied_before_cook(root / "Dockerfile")
+    if copied is not None and "." not in copied:
+        for p in paths:
+            top = p.strip("/").split("/")[0]
+            if top not in copied:
+                print(
+                    f"Cargo.toml patches {p}, but the Dockerfile doesn't copy {top!r} into the\n"
+                    f"  stage that runs `cargo chef cook` before that step, so cooking the\n"
+                    f"  dependencies fails on it. Add `COPY {top}/ {top}/` before the cook.",
+                    file=sys.stderr,
+                )
+                bad.append((p, top))
+    if bad:
+        return 1
+    print(f"build context and cook stage include every patched path: {', '.join(paths)}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())