Merge branch 'duplicate-detection'

.env.prod-example

@@ -5,6 +5,15 @@ CHITAI_TOKEN_SECRET=secret
 CHITAI_DEFAULT_LIBRARY_NAME=Books
 CHITAI_DEFAULT_LIBRARY_PATH="libraries/books"
 
+# Duplicate detection when importing files (optional).
+# Scope: "library" compares against the library being uploaded to, "global" against
+# every library, "off" disables the check.
+CHITAI_DUPLICATE_SCOPE=library
+
+# Where the consume watcher parks files it refused as duplicates. Keep it outside
+# CHITAI_CONSUME_PATH, or the watcher picks them straight back up.
+CHITAI_DUPLICATE_PATH="duplicates"
+
 # You probably should not change these
 CHITAI_API_URL="http://backend:8000"
 CHITAI_API_DEBUG=false

backend/AGENTS.md

@@ -88,6 +88,14 @@ The backend owns files on disk, not just rows:
 - **Layout** — `services/filesystem_library.py` (`BookPathGenerator`) renders a Jinja2 template
   against book metadata to decide where a book lives under the library's `root_path`
   (default: `author/series/position - title/`).
+- **One directory per book, never shared.** The generated path is a pure function of the metadata,
+  so two books with the same author and title produce the same one — two editions, or an
+  `allow_duplicates` copy. `BookService._reserve_book_path` moves the later one to `title (2)`
+  before anything is written, and `update_book` reserves the same way so a rename cannot move a
+  book in on top of another. This matters because `book.path` is what deletes, moves and file
+  lookups act on: books sharing a directory means one overwrites the other's files, and deleting
+  either takes both. `_unused_path` does the same job for filenames within a directory.
+  A book that already has a `path` keeps it — `add_files` must follow the book, not the template.
 - **Metadata extraction** — `services/metadata_extractor.py` reads EPUB (ebooklib) and PDF
   (pypdfium2) files; extracted values fill only *empty* fields on the incoming payload.
 - **Covers** — converted to WebP with a UUID filename under `settings.book_cover_path`, served by a
@@ -100,6 +108,140 @@ The backend owns files on disk, not just rows:
   if it differs, moves the directory contents and prunes empty parents. Keep that in mind before
   changing metadata handling.
 
+## Duplicate detection
+
+Every ingest path screens incoming files against what is already stored, keyed on
+**`(hash, size)`** — never the hash alone, because it samples 12 KiB (see below) and
+EPUBs from one toolchain often share their first window. `FileMetadata.hash` carries a
+plain, deliberately **non-unique** index: a collision must not be able to fail an import,
+and older databases may already hold duplicates.
+
+Scope comes from `CHITAI_DUPLICATE_SCOPE` (`library`, the default | `global` | `off`).
+
+The policy differs by how deliberate the import is:
+
+| Path | Behaviour |
+| --- | --- |
+| `create_many_from_files` (browser bulk) | Skip per file, skip a whole group whose files are all known, report everything skipped in `ImportResult.duplicates`. Re-dropping a folder to pick up what is new is the case this serves. |
+| `create_book` (single, with metadata) | All-or-nothing: raises `DuplicateFilesError`, which `controllers/book.py` renders as a **409** carrying the refused files in `extra`. |
+| `add_files` | A file the book already carries is a no-op; one stored under another book raises `DuplicateFilesError`. |
+| `create_many_from_existing_files` (consume watcher) | Skips, and **moves the file to `CHITAI_DUPLICATE_PATH/<library slug>/`** — nothing is deleted, and it cannot stay put because `watchfiles` only reports additions. That path must stay outside `consume_path` or the watcher re-imports it and tries to read the directory name as a library slug. |
+
+`allow_duplicates=true` overrides all of it, on every endpoint. Keep that working — the
+hash is not proof of identity, so a wrong verdict has to be recoverable, and a scripted
+import needs a way through. The **web UI deliberately does not offer it**: storing the
+same bytes twice splits reading progress and shelf membership across two records that
+can never converge, which is nothing anyone wants on purpose.
+
+Three things to preserve when touching this code:
+
+- **A match only counts while the file is on disk.** `find_duplicate_files` stats each
+  candidate, and `add_files` writes a missing file back into the row that already
+  describes it (`_restore_file`) instead of adding a second row beside it. The hash
+  lives in the database and the file does not, so without this a file deleted behind
+  the app's back would go on refusing its own replacement.
+
+- **Screening runs before anything is written.** `fingerprint_upload` reads the spooled
+  upload and rewinds it; the resulting fingerprints are handed to `_save_book_files`,
+  which skips its own `StreamingHasher` when it already has the answer. Passing them
+  through is what keeps the file from being read twice.
+- **`_screen_for_duplicates` extends the `known` dict as it goes**, so the same bytes
+  submitted twice in one request are caught. Those duplicates report `book_id: None` —
+  there is no row to point at yet.
+
+`POST /books/duplicate-files` answers the same question from fingerprints alone, for
+clients that want to ask before uploading anything.
+
+### Book-level detection — a different question
+
+The file check answers "are these the same bytes?". `find_duplicate_books` answers "is
+this the same book?", which a re-scan, a re-zipped EPUB or another edition cannot be
+asked with a hash. Two signals, either sufficient: a **shared identifier**, or a
+**matching normalized title with at least one shared author**.
+
+It **never blocks**. A metadata match is a guess — a work shares title and author with
+its own translation, its own second edition and its own audiobook — so the book is
+created and the candidates are reported alongside it in
+`ImportResult.possible_duplicates`. File-level dedupe keeps its refuse/skip behaviour;
+that one is near-certain and this one is not. Do not "improve" this into a refusal.
+
+Two rules narrow it, both applied in Python over the small candidate set:
+
+- **A shared author is required for a title match.** Without it every book the
+  extractors gave up on and titled `Unknown` is a duplicate of every other one. A book
+  with no authors can therefore only match on an identifier.
+- **The same series at a different `series_position` disqualifies a match.** A trilogy
+  shares an author and often most of its title; the position is the library saying
+  outright that these are two books.
+
+A book is compared under **several title keys, not one** (`_title_keys`). Ebook files
+are overwhelmingly named `Title - Author.epub`, and wherever nothing inside the file
+overrode that name the author ended up in the title column — so one copy is stored as
+`Building Microservices` and another as `Building Microservices Sam Newman`. Both
+directions are generated, the author stripped off and the author added on, which is why
+the query is `normalized_title.in_(keys)` rather than `==`. This is still exact matching
+on an indexed column: no similarity score, nothing to tune. It does not weaken the
+shared-author requirement, which is a separate condition.
+
+`find_duplicate_book_groups` is the library-wide pass behind
+`GET /books/duplicate-books`, since the import-time check says nothing about a
+collection someone already has. It buckets books by every key they carry and merges the
+buckets with union-find, so A~B by ISBN and B~C by title land in one group. Pairs in
+`duplicate_dismissals` are never merged — a reader disagreeing with one pairing must not
+silently break a group that stands on other evidence.
+
+### Author names have one stored form
+
+`Author.name` is always the canonical form, produced by `format_author_name`. Extractors
+hand over whatever the file said — `Newman, Sam;` from a `DC:creator` list, `Sam Newman`
+from a PDF, `Sam Newman.epub` from a filename — and storing those verbatim is how one
+person becomes four rows in the sidebar, four entries in the author filter, and four
+books that never look like each other.
+
+This is **display** canonicalization, distinct from `normalize_author`, which throws
+away case, accents and spacing to build a comparison key nobody sees. Tidying only
+removes what an extractor added: a trailing separator, a file extension, and the
+`Surname, Given` ordering. It never touches case or accents — `Michał Płachta` and
+`Steve McConnell` are the author's own spelling, not something to correct.
+
+Three places have to agree, and `Author` keeps them together: the `@validates("name")`
+hook, `unique_hash`, and `unique_filter`. `as_unique_async` looks a row up with the
+filter and then constructs with the validator, so if the lookup used the raw name and
+the insert used the tidy one, every variant spelling would miss the existing row and
+then collide with it on the unique index.
+
+A form the rule does not recognise is **left exactly as it was found** — `Dave Thomas,
+Andy Hunt` is two people in one string, and flipping it would invent a third. Leaving a
+mess visible beats rewriting it wrongly.
+
+### The normalized columns are written by validators
+
+`Book.normalized_title`, `Author.normalized_name` and `Identifier.normalized_value` are
+derived from `services/matching.py` and kept current by **SQLAlchemy `@validates` hooks
+on the models**, not by any service. Assigning them directly is always wrong.
+
+This is deliberate and it is invisible at the call sites: `BookService` writes titles
+through at least three paths (`to_model_on_create`, `to_model_on_update`, and the
+`setattr` loop in `_populate_with_unique_relationships`), and a validator is the only
+thing a fourth cannot bypass. The columns carry plain, **non-unique** btree indexes —
+two spellings collapsing onto one value is the entire point.
+
+`services/matching.py` is imported *inside* those validators rather than at module
+scope: reaching it initialises the `chitai.services` package, which imports the
+services, which import the models. Keep the local import.
+
+The validators only fire on write, so a migration that adds one of these columns must
+backfill existing rows through the same helpers — see the `data_upgrades()` hook in
+`2026-08-15_add_book_matching_keys_and_duplicate__4358e7d4743a.py`.
+
+**Changing anything in `services/matching.py` needs a revision that recomputes them.**
+The keys are derived and already written, so a normalization change silently invalidates
+every stored row: a book written under the old rules just stops matching one written
+under the new rules, with nothing to show that anything is wrong. Copy
+`2026-08-15_recompute_book_matching_keys_ed41acf21270.py`, which exists because
+`normalize_title` learned to strip compact edition markers (`2E`, `5e`). It is
+idempotent and safe to re-run.
+
 ## KOReader hashing
 
 `services/utils.py` reimplements KOReader's partial-MD5 document identifier: 1 KiB samples at

backend/migrations/versions/2026-08-13_add_file_hash_index_e9c2c7e875ae.py

@@ -0,0 +1,80 @@
+"""add file hash index
+
+Revision ID: e9c2c7e875ae
+Revises: 6d72d1bbc0ee
+Create Date: 2026-08-13 14:52:09.341906
+
+"""
+
+import warnings
+from typing import TYPE_CHECKING
+
+import sqlalchemy as sa
+from alembic import op
+from advanced_alchemy.types import EncryptedString, EncryptedText, GUID, ORA_JSONB, DateTimeUTC, StoredObject, PasswordHash, FernetBackend
+from advanced_alchemy.types.encrypted_string import PGCryptoBackend
+from advanced_alchemy.types.password_hash.argon2 import Argon2Hasher
+from advanced_alchemy.types.password_hash.passlib import PasslibHasher
+from advanced_alchemy.types.password_hash.pwdlib import PwdlibHasher
+from sqlalchemy import Text  # noqa: F401
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+__all__ = ["downgrade", "upgrade", "schema_upgrades", "schema_downgrades", "data_upgrades", "data_downgrades"]
+
+sa.GUID = GUID
+sa.DateTimeUTC = DateTimeUTC
+sa.ORA_JSONB = ORA_JSONB
+sa.EncryptedString = EncryptedString
+sa.EncryptedText = EncryptedText
+sa.StoredObject = StoredObject
+sa.PasswordHash = PasswordHash
+sa.Argon2Hasher = Argon2Hasher
+sa.PasslibHasher = PasslibHasher
+sa.PwdlibHasher = PwdlibHasher
+sa.FernetBackend = FernetBackend
+sa.PGCryptoBackend = PGCryptoBackend
+
+# revision identifiers, used by Alembic.
+revision = 'e9c2c7e875ae'
+down_revision = '6d72d1bbc0ee'
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with op.get_context().autocommit_block():
+            schema_upgrades()
+            data_upgrades()
+
+def downgrade() -> None:
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with op.get_context().autocommit_block():
+            data_downgrades()
+            schema_downgrades()
+
+def schema_upgrades() -> None:
+    """schema upgrade migrations go here."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table('file_metadata', schema=None) as batch_op:
+        batch_op.create_index('ix_file_metadata_hash', ['hash'], unique=False)
+
+    # ### end Alembic commands ###
+
+def schema_downgrades() -> None:
+    """schema downgrade migrations go here."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table('file_metadata', schema=None) as batch_op:
+        batch_op.drop_index('ix_file_metadata_hash')
+
+    # ### end Alembic commands ###
+
+def data_upgrades() -> None:
+    """Add any optional data upgrade migrations here!"""
+
+def data_downgrades() -> None:
+    """Add any optional data downgrade migrations here!"""

backend/migrations/versions/2026-08-15_add_book_matching_keys_and_duplicate__4358e7d4743a.py

@@ -0,0 +1,163 @@
+"""add book matching keys and duplicate dismissals
+
+Revision ID: 4358e7d4743a
+Revises: e9c2c7e875ae
+Create Date: 2026-08-15 15:09:02.708914
+
+"""
+
+import warnings
+
+import sqlalchemy as sa
+from alembic import op
+from advanced_alchemy.types import EncryptedString, EncryptedText, GUID, ORA_JSONB, DateTimeUTC, StoredObject, PasswordHash, FernetBackend
+from advanced_alchemy.types.encrypted_string import PGCryptoBackend
+from advanced_alchemy.types.password_hash.argon2 import Argon2Hasher
+from advanced_alchemy.types.password_hash.passlib import PasslibHasher
+from advanced_alchemy.types.password_hash.pwdlib import PwdlibHasher
+from sqlalchemy import Text  # noqa: F401
+
+__all__ = ["downgrade", "upgrade", "schema_upgrades", "schema_downgrades", "data_upgrades", "data_downgrades"]
+
+sa.GUID = GUID
+sa.DateTimeUTC = DateTimeUTC
+sa.ORA_JSONB = ORA_JSONB
+sa.EncryptedString = EncryptedString
+sa.EncryptedText = EncryptedText
+sa.StoredObject = StoredObject
+sa.PasswordHash = PasswordHash
+sa.Argon2Hasher = Argon2Hasher
+sa.PasslibHasher = PasslibHasher
+sa.PwdlibHasher = PwdlibHasher
+sa.FernetBackend = FernetBackend
+sa.PGCryptoBackend = PGCryptoBackend
+
+# revision identifiers, used by Alembic.
+revision = '4358e7d4743a'
+down_revision = 'e9c2c7e875ae'
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with op.get_context().autocommit_block():
+            schema_upgrades()
+            data_upgrades()
+
+def downgrade() -> None:
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with op.get_context().autocommit_block():
+            data_downgrades()
+            schema_downgrades()
+
+def schema_upgrades() -> None:
+    """schema upgrade migrations go here."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table('duplicate_dismissals',
+    sa.Column('id', sa.BigInteger().with_variant(sa.Integer(), 'sqlite'), nullable=False),
+    sa.Column('book_a_id', sa.BigInteger().with_variant(sa.Integer(), 'sqlite'), nullable=False),
+    sa.Column('book_b_id', sa.BigInteger().with_variant(sa.Integer(), 'sqlite'), nullable=False),
+    sa.ForeignKeyConstraint(['book_a_id'], ['books.id'], name=op.f('fk_duplicate_dismissals_book_a_id_books'), ondelete='cascade'),
+    sa.ForeignKeyConstraint(['book_b_id'], ['books.id'], name=op.f('fk_duplicate_dismissals_book_b_id_books'), ondelete='cascade'),
+    sa.PrimaryKeyConstraint('id', name=op.f('pk_duplicate_dismissals')),
+    sa.UniqueConstraint('book_a_id', 'book_b_id', name=op.f('uq_duplicate_dismissals_book_a_id'))
+    )
+    with op.batch_alter_table('duplicate_dismissals', schema=None) as batch_op:
+        batch_op.create_index(batch_op.f('ix_duplicate_dismissals_book_a_id'), ['book_a_id'], unique=False)
+        batch_op.create_index(batch_op.f('ix_duplicate_dismissals_book_b_id'), ['book_b_id'], unique=False)
+
+    # `server_default` so the column can be added to a table that already has rows;
+    # `data_upgrades` fills in the real keys immediately afterwards.
+    with op.batch_alter_table('authors', schema=None) as batch_op:
+        batch_op.add_column(sa.Column('normalized_name', sa.String(), nullable=False, server_default=''))
+        batch_op.create_index(batch_op.f('ix_authors_normalized_name'), ['normalized_name'], unique=False)
+
+    with op.batch_alter_table('books', schema=None) as batch_op:
+        batch_op.add_column(sa.Column('normalized_title', sa.String(), nullable=False, server_default=''))
+        batch_op.create_index(batch_op.f('ix_books_normalized_title'), ['normalized_title'], unique=False)
+
+    with op.batch_alter_table('identifiers', schema=None) as batch_op:
+        batch_op.add_column(sa.Column('normalized_value', sa.String(), nullable=True))
+        batch_op.create_index(batch_op.f('ix_identifiers_normalized_value'), ['normalized_value'], unique=False)
+
+    # ### end Alembic commands ###
+
+def schema_downgrades() -> None:
+    """schema downgrade migrations go here."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table('identifiers', schema=None) as batch_op:
+        batch_op.drop_index(batch_op.f('ix_identifiers_normalized_value'))
+        batch_op.drop_column('normalized_value')
+
+    with op.batch_alter_table('books', schema=None) as batch_op:
+        batch_op.drop_index(batch_op.f('ix_books_normalized_title'))
+        batch_op.drop_column('normalized_title')
+
+    with op.batch_alter_table('authors', schema=None) as batch_op:
+        batch_op.drop_index(batch_op.f('ix_authors_normalized_name'))
+        batch_op.drop_column('normalized_name')
+
+    with op.batch_alter_table('duplicate_dismissals', schema=None) as batch_op:
+        batch_op.drop_index(batch_op.f('ix_duplicate_dismissals_book_b_id'))
+        batch_op.drop_index(batch_op.f('ix_duplicate_dismissals_book_a_id'))
+
+    op.drop_table('duplicate_dismissals')
+    # ### end Alembic commands ###
+
+def data_upgrades() -> None:
+    """
+    Fill the matching keys in for rows that already exist.
+
+    The validators on the models only fire when something is written, so without this
+    every book imported before today is invisible to duplicate detection. Run through
+    the same helpers the validators use, so a backfilled row and a freshly written one
+    are guaranteed to agree.
+    """
+    from chitai.services.matching import (
+        normalize_author,
+        normalize_identifier,
+        normalize_title,
+    )
+
+    connection = op.get_bind()
+
+    books = connection.execute(sa.text("SELECT id, title FROM books")).fetchall()
+    _apply(
+        connection,
+        "UPDATE books SET normalized_title = :key WHERE id = :id",
+        [{"id": id, "key": normalize_title(title)} for id, title in books],
+    )
+
+    authors = connection.execute(sa.text("SELECT id, name FROM authors")).fetchall()
+    _apply(
+        connection,
+        "UPDATE authors SET normalized_name = :key WHERE id = :id",
+        [{"id": id, "key": normalize_author(name)} for id, name in authors],
+    )
+
+    identifiers = connection.execute(
+        sa.text("SELECT id, name, value FROM identifiers")
+    ).fetchall()
+    _apply(
+        connection,
+        "UPDATE identifiers SET normalized_value = :key WHERE id = :id",
+        [
+            {"id": id, "key": normalize_identifier(name, value)}
+            for id, name, value in identifiers
+        ],
+    )
+
+
+def _apply(connection, statement: str, parameters: list[dict]) -> None:
+    """Run one update per row, in batches, skipping the work when there are none."""
+    batch_size = 1000
+
+    for start in range(0, len(parameters), batch_size):
+        connection.execute(sa.text(statement), parameters[start : start + batch_size])
+
+
+def data_downgrades() -> None:
+    """Add any optional data downgrade migrations here!"""

backend/migrations/versions/2026-08-15_canonicalize_author_names_49a9e85a0ffc.py

@@ -0,0 +1,130 @@
+"""canonicalize author names
+
+Revision ID: 49a9e85a0ffc
+Revises: ed41acf21270
+Create Date: 2026-08-15 15:59:47.331545
+
+"""
+
+import warnings
+
+import sqlalchemy as sa
+from alembic import op
+from advanced_alchemy.types import EncryptedString, EncryptedText, GUID, ORA_JSONB, DateTimeUTC, StoredObject, PasswordHash, FernetBackend
+from advanced_alchemy.types.encrypted_string import PGCryptoBackend
+from advanced_alchemy.types.password_hash.argon2 import Argon2Hasher
+from advanced_alchemy.types.password_hash.passlib import PasslibHasher
+from advanced_alchemy.types.password_hash.pwdlib import PwdlibHasher
+from sqlalchemy import Text  # noqa: F401
+
+__all__ = ["downgrade", "upgrade", "schema_upgrades", "schema_downgrades", "data_upgrades", "data_downgrades"]
+
+sa.GUID = GUID
+sa.DateTimeUTC = DateTimeUTC
+sa.ORA_JSONB = ORA_JSONB
+sa.EncryptedString = EncryptedString
+sa.EncryptedText = EncryptedText
+sa.StoredObject = StoredObject
+sa.PasswordHash = PasswordHash
+sa.Argon2Hasher = Argon2Hasher
+sa.PasslibHasher = PasslibHasher
+sa.PwdlibHasher = PwdlibHasher
+sa.FernetBackend = FernetBackend
+sa.PGCryptoBackend = PGCryptoBackend
+
+# revision identifiers, used by Alembic.
+revision = '49a9e85a0ffc'
+down_revision = 'ed41acf21270'
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with op.get_context().autocommit_block():
+            schema_upgrades()
+            data_upgrades()
+
+def downgrade() -> None:
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with op.get_context().autocommit_block():
+            data_downgrades()
+            schema_downgrades()
+
+def schema_upgrades() -> None:
+    """schema upgrade migrations go here."""
+    pass
+
+def schema_downgrades() -> None:
+    """schema downgrade migrations go here."""
+    pass
+
+def data_upgrades() -> None:
+    """
+    Rewrite every author into the canonical form, merging the rows that collide.
+
+    `Author.name` is only now guaranteed tidy — until this revision extractors wrote
+    whatever the file said, so one person could hold several rows: "Sam Newman" beside
+    "Newman, Sam;" beside "Sam Newman.epub" (the last from a filename whose extension
+    was never stripped). Each showed up as its own author in the sidebar and its own
+    filter, and no amount of fixing the extractors repairs a row already written.
+
+    Rows that canonicalize onto one name are merged into the lowest id, which keeps
+    whichever row the library has been referring to longest. `book.path` is stored, not
+    derived, so renaming an author moves nothing on disk.
+    """
+    from chitai.services.matching import format_author_name, normalize_author
+
+    connection = op.get_bind()
+    authors = connection.execute(sa.text("SELECT id, name FROM authors")).fetchall()
+
+    groups: dict[str, list[int]] = {}
+    for id, name in sorted(authors):
+        # A name with nothing left of it after tidying is left exactly as it was:
+        # merging those together would invent one author out of several unrelated
+        # broken rows, which is worse than leaving the mess visible.
+        if canonical := format_author_name(name):
+            groups.setdefault(canonical, []).append(id)
+
+    for canonical, ids in groups.items():
+        winner, losers = ids[0], ids[1:]
+
+        for loser in losers:
+            # A book credited to both rows would otherwise breach the
+            # (book_id, author_id) unique constraint the moment the link is repointed.
+            connection.execute(
+                sa.text(
+                    "DELETE FROM book_author_links WHERE author_id = :loser AND book_id IN"
+                    " (SELECT book_id FROM book_author_links WHERE author_id = :winner)"
+                ),
+                {"loser": loser, "winner": winner},
+            )
+            connection.execute(
+                sa.text(
+                    "UPDATE book_author_links SET author_id = :winner"
+                    " WHERE author_id = :loser"
+                ),
+                {"loser": loser, "winner": winner},
+            )
+            connection.execute(
+                sa.text("DELETE FROM authors WHERE id = :loser"), {"loser": loser}
+            )
+
+        # Only after the losers are gone, or this collides with the unique index.
+        connection.execute(
+            sa.text(
+                "UPDATE authors SET name = :name, normalized_name = :key WHERE id = :id"
+            ),
+            {"id": winner, "name": canonical, "key": normalize_author(canonical)},
+        )
+
+
+def data_downgrades() -> None:
+    """
+    Nothing to undo.
+
+    The rows a merge removed are gone, and the spellings it replaced were never
+    recorded anywhere else — there is nothing to restore them from.
+    """

backend/migrations/versions/2026-08-15_recompute_book_matching_keys_ed41acf21270.py

@@ -0,0 +1,122 @@
+"""recompute book matching keys
+
+Revision ID: ed41acf21270
+Revises: 4358e7d4743a
+Create Date: 2026-08-15 15:44:28.341020
+
+"""
+
+import warnings
+
+import sqlalchemy as sa
+from alembic import op
+from advanced_alchemy.types import EncryptedString, EncryptedText, GUID, ORA_JSONB, DateTimeUTC, StoredObject, PasswordHash, FernetBackend
+from advanced_alchemy.types.encrypted_string import PGCryptoBackend
+from advanced_alchemy.types.password_hash.argon2 import Argon2Hasher
+from advanced_alchemy.types.password_hash.passlib import PasslibHasher
+from advanced_alchemy.types.password_hash.pwdlib import PwdlibHasher
+from sqlalchemy import Text  # noqa: F401
+
+__all__ = ["downgrade", "upgrade", "schema_upgrades", "schema_downgrades", "data_upgrades", "data_downgrades"]
+
+sa.GUID = GUID
+sa.DateTimeUTC = DateTimeUTC
+sa.ORA_JSONB = ORA_JSONB
+sa.EncryptedString = EncryptedString
+sa.EncryptedText = EncryptedText
+sa.StoredObject = StoredObject
+sa.PasswordHash = PasswordHash
+sa.Argon2Hasher = Argon2Hasher
+sa.PasslibHasher = PasslibHasher
+sa.PwdlibHasher = PwdlibHasher
+sa.FernetBackend = FernetBackend
+sa.PGCryptoBackend = PGCryptoBackend
+
+# revision identifiers, used by Alembic.
+revision = 'ed41acf21270'
+down_revision = '4358e7d4743a'
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with op.get_context().autocommit_block():
+            schema_upgrades()
+            data_upgrades()
+
+def downgrade() -> None:
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with op.get_context().autocommit_block():
+            data_downgrades()
+            schema_downgrades()
+
+def schema_upgrades() -> None:
+    """schema upgrade migrations go here."""
+    pass
+
+def schema_downgrades() -> None:
+    """schema downgrade migrations go here."""
+    pass
+
+def data_upgrades() -> None:
+    """
+    Recompute every matching key against the current normalization.
+
+    The keys are derived, so changing a helper in `services/matching.py` silently
+    invalidates every row already written — a book stored under the old rules simply
+    stops matching one stored under the new ones, with nothing to show that anything
+    is wrong. `normalize_title` learned to strip the compact edition markers a cover
+    actually carries ("2E", "5e"), which moved `Building Microservices, 2E` onto the
+    same key as `Building Microservices`.
+
+    Any later change to those helpers wants a revision that looks exactly like this
+    one. It is idempotent and safe to re-run.
+    """
+    from chitai.services.matching import (
+        normalize_author,
+        normalize_identifier,
+        normalize_title,
+    )
+
+    connection = op.get_bind()
+
+    books = connection.execute(sa.text("SELECT id, title FROM books")).fetchall()
+    _apply(
+        connection,
+        "UPDATE books SET normalized_title = :key WHERE id = :id",
+        [{"id": id, "key": normalize_title(title)} for id, title in books],
+    )
+
+    authors = connection.execute(sa.text("SELECT id, name FROM authors")).fetchall()
+    _apply(
+        connection,
+        "UPDATE authors SET normalized_name = :key WHERE id = :id",
+        [{"id": id, "key": normalize_author(name)} for id, name in authors],
+    )
+
+    identifiers = connection.execute(
+        sa.text("SELECT id, name, value FROM identifiers")
+    ).fetchall()
+    _apply(
+        connection,
+        "UPDATE identifiers SET normalized_value = :key WHERE id = :id",
+        [
+            {"id": id, "key": normalize_identifier(name, value)}
+            for id, name, value in identifiers
+        ],
+    )
+
+
+def _apply(connection, statement: str, parameters: list[dict]) -> None:
+    """Run one update per row, in batches, skipping the work when there are none."""
+    batch_size = 1000
+
+    for start in range(0, len(parameters), batch_size):
+        connection.execute(sa.text(statement), parameters[start : start + batch_size])
+
+
+def data_downgrades() -> None:
+    """Add any optional data downgrade migrations here!"""

backend/src/chitai/config.py

@@ -1,9 +1,25 @@
+from enum import StrEnum
+
 from pydantic import Field, PostgresDsn, computed_field
 from pydantic_settings import BaseSettings, SettingsConfigDict
 from advanced_alchemy.extensions.litestar import (
     SQLAlchemyAsyncConfig,
 )
 
+
+class DuplicateScope(StrEnum):
+    """How widely an incoming file is compared against what is already stored."""
+
+    LIBRARY = "library"
+    """Only files in the library being uploaded to count as duplicates."""
+
+    GLOBAL = "global"
+    """A file already held by any library counts as a duplicate."""
+
+    OFF = "off"
+    """No duplicate detection at all."""
+
+
 class Settings(BaseSettings):
     version: str = Field("0.0.1")
     project_name: str = Field("chitai")
@@ -33,6 +49,14 @@ class Settings(BaseSettings):
     # Path to consume directory
     consume_path: str = Field("./consume")
 
+    # Duplicate detection
+    duplicate_scope: DuplicateScope = Field(DuplicateScope.LIBRARY)
+
+    # Where the consume watcher parks files it refused as duplicates. Must sit
+    # outside `consume_path`, or the watcher picks them straight back up and
+    # tries to resolve the directory name as a library slug.
+    duplicate_path: str = Field("./duplicates")
+
     @computed_field
     @property
     def postgres_uri(self) -> PostgresDsn:

backend/src/chitai/controllers/book.py

@@ -12,7 +12,12 @@ from litestar.params import Dependency, Body
 from litestar.enums import RequestEncodingType
 from litestar.response import File, Stream
 from litestar.exceptions import HTTPException
-from litestar.status_codes import HTTP_400_BAD_REQUEST
+from litestar.status_codes import (
+    HTTP_200_OK,
+    HTTP_204_NO_CONTENT,
+    HTTP_400_BAD_REQUEST,
+    HTTP_409_CONFLICT,
+)
 from litestar.datastructures import UploadFile
 from advanced_alchemy.service.pagination import OffsetPagination
 from advanced_alchemy.filters import CollectionFilter
@@ -23,6 +28,24 @@ from chitai.services import dependencies as deps
 from chitai import schemas as s
 from chitai.database import models as m
 from chitai.services import BookService, BookProgressService
+from chitai.services.book import DuplicateFilesError
+
+
+def _duplicate_conflict(exc: DuplicateFilesError) -> HTTPException:
+    """
+    Turn refused files into a 409 the caller can act on.
+
+    The files ride along in `extra` so the client can name them and offer to send them
+    again with `allow_duplicates`, rather than being told only that something clashed.
+    """
+    return HTTPException(
+        status_code=HTTP_409_CONFLICT,
+        detail="These files are already in the library",
+        extra=[
+            s.DuplicateFileRead.model_validate(duplicate).model_dump(mode="json")
+            for duplicate in exc.duplicates
+        ],
+    )
 
 
 class BookController(Controller):
@@ -63,6 +86,7 @@ class BookController(Controller):
         books_service: BookService,
         library: m.Library,
         data: Annotated[s.BookCreate, Body(media_type=RequestEncodingType.MULTI_PART)],
+        allow_duplicates: bool = False,
     ) -> s.BookRead:
         """
         Create a new book with metadata and files.
@@ -73,6 +97,10 @@ class BookController(Controller):
         Path Parameters:
             library_id: The ID of the library the book belongs to.
 
+        Query Parameters:
+            allow_duplicates: If True, store the files even if the library already
+                holds them.
+
         Request Body:
             data: Book creation data including metadata and files.
 
@@ -83,9 +111,17 @@ class BookController(Controller):
         Returns:
             The created book as a BookRead schema.
 
+        Raises:
+            HTTPException: 409 if any of the files is already in the library.
         """
 
-        result = await books_service.create_book(data, library)
+        try:
+            result = await books_service.create_book(
+                data, library, screen_duplicates=not allow_duplicates
+            )
+        except DuplicateFilesError as exc:
+            raise _duplicate_conflict(exc)
+
         book = await books_service.get(result.id)
         return books_service.to_schema(book, schema_type=s.BookRead)
 
@@ -97,13 +133,20 @@ class BookController(Controller):
         data: Annotated[
             s.BooksCreateFromFiles, Body(media_type=RequestEncodingType.MULTI_PART)
         ],
-    ) -> OffsetPagination[s.BookRead]:
+        allow_duplicates: bool = False,
+    ) -> s.BooksUploadResult:
         """
         Create multiple books from uploaded files.
 
         Groups files by directory and creates separate books for each group.
         Metadata is automatically extracted from the files.
 
+        Files the library already holds are skipped rather than refused, and reported
+        back so the caller can say which ones did not make it in and why.
+
+        Query Parameters:
+            allow_duplicates: If True, store every file, even one already held.
+
         Request Body:
             data: Container with list of uploaded files.
 
@@ -112,19 +155,200 @@ class BookController(Controller):
             library: The library the books belong to.
 
         Returns:
-            Paginated list of created books.
+            The books created, and the files skipped as duplicates.
         """
         try:
-            results = await books_service.create_many_from_files(data, library)
+            result = await books_service.create_many_from_files(
+                data, library, allow_duplicates=allow_duplicates
+            )
         except ValueError:
             raise HTTPException(
                 status_code=HTTP_400_BAD_REQUEST, detail="Must upload at least one file"
             )
 
-        books = await books_service.list(
-            CollectionFilter("id", [result.id for result in results])
+        books = (
+            await books_service.list(
+                CollectionFilter("id", [book.id for book in result.books])
+            )
+            if result.books
+            else []
         )
-        return books_service.to_schema(books, schema_type=s.BookRead)
+
+        return s.BooksUploadResult(
+            created=[
+                books_service.to_schema(book, schema_type=s.BookRead) for book in books
+            ],
+            skipped=[
+                s.DuplicateFileRead.model_validate(duplicate)
+                for duplicate in result.duplicates
+            ],
+            possible_duplicates=[
+                s.PossibleDuplicateRead.model_validate(possible)
+                for possible in result.possible_duplicates
+            ],
+        )
+
+    @get(path="duplicate-books")
+    async def list_duplicate_books(
+        self, books_service: BookService, library: m.Library
+    ) -> list[s.DuplicateBookGroupRead]:
+        """
+        Report books already in the library that look like copies of one another.
+
+        The import-time check only ever sees what is arriving, so this is what covers
+        a collection someone already has. Matching is on metadata and therefore a
+        guess: a group is a question for the reader, not a verdict.
+
+        Query Parameters:
+            library_id: The library to review.
+
+        Injected Dependencies:
+            books_service: The book service for database operations.
+            library: The library to review.
+
+        Returns:
+            One entry per group of two or more books. Empty when there is nothing to
+            review, or when duplicate detection is switched off.
+        """
+        groups = await books_service.find_duplicate_book_groups(library)
+
+        return [
+            s.DuplicateBookGroupRead(
+                books=[s.DuplicateBookRead.model_validate(book) for book in group]
+            )
+            for group in groups
+        ]
+
+    @post(path="merge")
+    async def merge_books(
+        self, books_service: BookService, library: m.Library, data: s.BookMerge
+    ) -> s.BookRead:
+        """
+        Fold several books into one and delete the records folded in.
+
+        The survivor keeps its id, so links and bookmarks still resolve. Files, reading
+        progress, shelves, tags and unheld identifiers move onto it; metadata is only
+        changed by what `metadata` names, because choosing between two titles is the
+        reader's judgement rather than this endpoint's.
+
+        Nothing is removed from disk — a wrong merge should cost metadata that can be
+        retyped, not a book.
+
+        Query Parameters:
+            library_id: The library the books belong to.
+
+        Request Body:
+            data: The survivor, the books to fold in, and the resolved metadata.
+
+        Injected Dependencies:
+            books_service: The book service for database operations.
+            library: The library the books belong to.
+
+        Returns:
+            The surviving book.
+
+        Raises:
+            HTTPException: 400 if fewer than two distinct books were named, one is
+                unknown, or they do not all belong to one library.
+        """
+        try:
+            book = await books_service.merge_books(
+                data.survivor_id,
+                data.merged_ids,
+                library,
+                metadata=data.metadata.model_dump(exclude_unset=True)
+                if data.metadata
+                else None,
+            )
+        except ValueError as exc:
+            raise HTTPException(status_code=HTTP_400_BAD_REQUEST, detail=str(exc))
+
+        return books_service.to_schema(book, schema_type=s.BookRead)
+
+    @post(path="duplicate-books/dismissals", status_code=HTTP_204_NO_CONTENT)
+    async def dismiss_duplicate_books(
+        self, books_service: BookService, data: s.DuplicateDismissal
+    ) -> None:
+        """
+        Record that two books are not the same book.
+
+        Without this the review screen proposes the same wrong pair forever, which is
+        how a reader learns to stop looking at it.
+
+        Request Body:
+            data: The two book IDs. Order does not matter.
+
+        Injected Dependencies:
+            books_service: The book service for database operations.
+
+        Raises:
+            HTTPException: 400 if the two IDs are the same or either book is unknown.
+        """
+        try:
+            await books_service.dismiss_duplicates(data.book_a_id, data.book_b_id)
+        except ValueError as exc:
+            raise HTTPException(status_code=HTTP_400_BAD_REQUEST, detail=str(exc))
+
+    @delete(path="duplicate-books/dismissals")
+    async def restore_duplicate_books(
+        self, books_service: BookService, book_a_id: int, book_b_id: int
+    ) -> None:
+        """
+        Undo a dismissal, so the pair is proposed again.
+
+        Query Parameters:
+            book_a_id: One of the two books.
+            book_b_id: The other. Order does not matter.
+
+        Injected Dependencies:
+            books_service: The book service for database operations.
+        """
+        await books_service.restore_duplicates(book_a_id, book_b_id)
+
+    # A question, not a change: 200 rather than the 201 a POST would default to.
+    @post(path="duplicate-files", status_code=HTTP_200_OK)
+    async def check_duplicate_files(
+        self,
+        books_service: BookService,
+        library: m.Library,
+        data: list[s.FileFingerprint],
+    ) -> list[s.DuplicateFileRead]:
+        """
+        Report which of the given files the library already holds.
+
+        Lets a client ask before it uploads anything, which is the difference between
+        re-sending a folder of books and re-sending twelve kilobytes of hashes.
+
+        Query Parameters:
+            library_id: The library to check against.
+
+        Request Body:
+            data: Hash and size for each file, optionally with the name to echo back.
+
+        Injected Dependencies:
+            books_service: The book service for database operations.
+            library: The library to check against.
+
+        Returns:
+            One entry per submitted file that is already stored. Files that are not
+            are absent.
+        """
+        matches = await books_service.find_duplicate_files(
+            ((item.hash, item.size) for item in data), library
+        )
+
+        return [
+            s.DuplicateFileRead(
+                filename=item.filename or match.filename,
+                hash=match.hash,
+                size=match.size,
+                library_id=match.library_id,
+                book_id=match.book_id,
+                book_title=match.book_title,
+            )
+            for item in data
+            if (match := matches.get((item.hash, item.size))) is not None
+        ]
 
     @get(path="/{book_id:int}")
     async def get_book_by_id(
@@ -303,13 +527,20 @@ class BookController(Controller):
         ],
         library: m.Library,
         books_service: BookService,
+        allow_duplicates: bool = False,
     ) -> s.BookRead:
         """
         Add files to an existing book.
 
+        A file the book already carries is ignored, so re-sending one is harmless.
+
         Path Parameters:
             book_id: The ID of the book to modify
 
+        Query Parameters:
+            allow_duplicates: If True, store the files even if the library already
+                holds them.
+
         Request Body:
             files: The files to add to the book
 
@@ -320,9 +551,17 @@ class BookController(Controller):
         Returns:
             The modified book
 
+        Raises:
+            HTTPException: 409 if a file is already stored under a different book.
         """
 
-        await books_service.add_files(book_id, data, library)
+        try:
+            await books_service.add_files(
+                book_id, data, library, allow_duplicates=allow_duplicates
+            )
+        except DuplicateFilesError as exc:
+            raise _duplicate_conflict(exc)
+
         book = await books_service.get(book_id)
         return books_service.to_schema(book, schema_type=s.BookRead)
 

backend/src/chitai/database/models/__init__.py

@@ -3,6 +3,7 @@ from .book import Book, Identifier, FileMetadata
 from .book_list import BookList, BookListLink
 from .book_progress import BookProgress
 from .book_series import BookSeries
+from .duplicate_dismissal import DuplicateDismissal
 from .kosync_device import KosyncDevice
 from .kosync_progress import KosyncProgress
 from .library import Library

backend/src/chitai/database/models/author.py

@@ -5,6 +5,7 @@ from sqlalchemy import ColumnElement, ForeignKey, UniqueConstraint
 from sqlalchemy.orm import Mapped
 from sqlalchemy.orm import mapped_column
 from sqlalchemy.orm import relationship
+from sqlalchemy.orm import validates
 
 from advanced_alchemy.base import BigIntAuditBase, BigIntBase
 from advanced_alchemy.mixins import UniqueMixin
@@ -16,18 +17,55 @@ if TYPE_CHECKING:
 class Author(BigIntAuditBase, UniqueMixin):
     __tablename__ = "authors"
 
+    # Always the canonical form — see `_canonicalize`. Extractors hand over whatever
+    # the file happened to say: "Newman, Sam;" from a `DC:creator` list, or
+    # "Sam Newman.epub" from a filename. Storing those verbatim is how one person ends
+    # up as several rows in the sidebar.
     name: Mapped[str] = mapped_column(unique=True, index=True)
+
+    # Kept current by `_canonicalize` too — never assign it directly. Not unique: two
+    # spellings that survive canonicalization, "Steve Mcconnell" and "Steve McConnell",
+    # are still one person to a reader, which is what this column exists to express.
+    normalized_name: Mapped[str] = mapped_column(default="", index=True)
+
     description: Mapped[Optional[str]]
 
+    @validates("name")
+    def _canonicalize(self, _key: str, name: str) -> str:
+        """
+        Store the tidied name, and derive the matching key from it.
+
+        A validator so no write can get around it, and `unique_hash` / `unique_filter`
+        below tidy the same way so `as_unique_async` looks the row up under the name it
+        would actually be stored as. All three have to agree: if the lookup used the
+        raw name and the insert used the tidy one, every variant spelling would miss
+        the existing row and then collide with it on the unique index.
+        """
+        # Imported here rather than at module scope: `chitai.services.matching` cannot
+        # be reached without initialising the `chitai.services` package, which imports
+        # the services, which import this module.
+        from chitai.services.matching import format_author_name, normalize_author
+
+        name = format_author_name(name)
+        self.normalized_name = normalize_author(name)
+        return name
+
+    @classmethod
+    def _tidy(cls, name: str) -> str:
+        """The name as `_canonicalize` would store it."""
+        from chitai.services.matching import format_author_name
+
+        return format_author_name(name)
+
     @classmethod
     def unique_hash(cls, name: str) -> Hashable:
         """Generate a unique hash for deduplication."""
-        return name
+        return cls._tidy(name)
 
     @classmethod
     def unique_filter(cls, name: str) -> ColumnElement[bool]:
         """SQL filter for finding existing records."""
-        return cls.name == name
+        return cls.name == cls._tidy(name)
 
     def __repr__(self) -> str:
         return f"Author({self.name!r})"

backend/src/chitai/database/models/book.py

@@ -2,13 +2,10 @@ from datetime import date
 from typing import TYPE_CHECKING, Any, Optional
 
 from sqlalchemy import Index, ForeignKey, UniqueConstraint
-from sqlalchemy.orm import Mapped, mapped_column, relationship
-from sqlalchemy.orm import mapped_column
-from sqlalchemy.orm import relationship
+from sqlalchemy.orm import Mapped, mapped_column, relationship, validates
 from sqlalchemy.ext.orderinglist import ordering_list
 from sqlalchemy.ext.associationproxy import association_proxy
 from sqlalchemy.ext.associationproxy import AssociationProxy
-from sqlalchemy.orm.collections import attribute_keyed_dict
 
 from advanced_alchemy.base import BigIntAuditBase, BigIntBase
 
@@ -44,6 +41,13 @@ class Book(BigIntAuditBase):
     library: Mapped["Library"] = relationship(back_populates="books")
 
     title: Mapped[str]
+
+    # Kept current by `_normalize_title` below — never assign it directly.
+    #
+    # Deliberately not unique: two spellings collapsing onto one value is the whole
+    # point of the column, and a second edition is allowed to exist.
+    normalized_title: Mapped[str] = mapped_column(default="", index=True)
+
     subtitle: Mapped[Optional[str]]
     description: Mapped[Optional[str]]
     published_date: Mapped[Optional[date]]
@@ -111,6 +115,24 @@ class Book(BigIntAuditBase):
     def progress(self) -> Optional["BookProgress"]:
         return self.progress_records[0] if self.progress_records else None
 
+    @validates("title")
+    def _normalize_title(self, _key: str, title: str) -> str:
+        """
+        Derive `normalized_title` from whatever writes the title.
+
+        A validator rather than a service call because `BookService` sets titles from
+        at least three places — `to_model_on_create`, `to_model_on_update` and the
+        `setattr` loop in `_populate_with_unique_relationships` — and a fourth would
+        otherwise leave the key silently stale.
+        """
+        # Imported here rather than at module scope: `chitai.services.matching` cannot
+        # be reached without initialising the `chitai.services` package, which imports
+        # the services, which import this module.
+        from chitai.services.matching import normalize_title
+
+        self.normalized_title = normalize_title(title)
+        return title
+
     def __repr__(self) -> str:
         return f"Book({self.title=!r})"
 
@@ -132,6 +154,24 @@ class Identifier(BigIntBase):
     book_id: Mapped[int] = mapped_column(ForeignKey("books.id", ondelete="cascade"))
     value: Mapped[str]
 
+    # Kept current by `_normalize` below — never assign it directly. Null for an
+    # identifier that cannot carry a match: a per-build UUID, or an ISBN that fails
+    # its own checksum.
+    normalized_value: Mapped[Optional[str]] = mapped_column(index=True)
+
+    @validates("name", "value")
+    def _normalize(self, key: str, value: str) -> str:
+        """Recompute `normalized_value` whenever either half of the pair changes."""
+        from chitai.services.matching import normalize_identifier
+
+        name = value if key == "name" else self.name
+        raw = value if key == "value" else self.value
+
+        self.normalized_value = (
+            normalize_identifier(name, raw) if name and raw else None
+        )
+        return value
+
     def __repr__(self):
         return f"Identifier({self.name!r} : {self.value!r})"
 
@@ -139,6 +179,15 @@ class Identifier(BigIntBase):
 class FileMetadata(BigIntBase):
     __tablename__ = "file_metadata"
 
+    __table_args__ = (
+        # Deliberately not unique. The hash is KOReader's partial MD5, which samples
+        # 12 KiB of the file, so two genuinely different files can collide — and an
+        # existing database may already hold duplicates, which a unique index would
+        # refuse to build over. Duplicate detection pairs it with `size` and treats a
+        # match as advisory, so this only has to make the lookup cheap.
+        Index("ix_file_metadata_hash", "hash"),
+    )
+
     book_id: Mapped[int] = mapped_column(ForeignKey("books.id", ondelete="cascade"))
     book: Mapped[Book] = relationship(back_populates="files")
     hash: Mapped[str]

backend/src/chitai/database/models/duplicate_dismissal.py

@@ -0,0 +1,37 @@
+from sqlalchemy import ForeignKey, UniqueConstraint
+from sqlalchemy.orm import Mapped, mapped_column
+
+from advanced_alchemy.base import BigIntBase
+
+
+class DuplicateDismissal(BigIntBase):
+    """
+    Two books a reader has said are not the same book.
+
+    Title and author matching is probabilistic, so it will keep proposing a second
+    edition, a translation and a sequel that shares its predecessor's name. A review
+    screen with no way to disagree with it nags forever, which is how people learn to
+    ignore a screen.
+
+    The pair is stored ordered — `book_a_id` is always the lower id — so "A and B" and
+    "B and A" are one row and the unique constraint can do its job. Use `pair()` rather
+    than assigning the columns directly.
+    """
+
+    __tablename__ = "duplicate_dismissals"
+    __table_args__ = (UniqueConstraint("book_a_id", "book_b_id"),)
+
+    book_a_id: Mapped[int] = mapped_column(
+        ForeignKey("books.id", ondelete="cascade"), index=True
+    )
+    book_b_id: Mapped[int] = mapped_column(
+        ForeignKey("books.id", ondelete="cascade"), index=True
+    )
+
+    @staticmethod
+    def pair(first: int, second: int) -> tuple[int, int]:
+        """The two book ids in the order this table stores them."""
+        return (first, second) if first <= second else (second, first)
+
+    def __repr__(self) -> str:
+        return f"DuplicateDismissal({self.book_a_id!r}, {self.book_b_id!r})"

backend/src/chitai/schemas/__init__.py

@@ -4,7 +4,15 @@ from .book import (
     BookProgressCreate,
     BookProgressRead,
     BooksCreateFromFiles,
+    BooksUploadResult,
+    BookMerge,
     BookMetadataUpdate,
+    DuplicateBookGroupRead,
+    DuplicateBookRead,
+    DuplicateDismissal,
+    DuplicateFileRead,
+    FileFingerprint,
+    PossibleDuplicateRead,
     FileMetadataRead,
     BookSeriesRead,
 )

backend/src/chitai/schemas/book.py

@@ -129,6 +129,83 @@ class BooksCreateFromFiles(BaseModel):
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
 
+class FileFingerprint(BaseModel):
+    """What a client can say about a file it has not uploaded yet."""
+
+    hash: str
+    size: int
+    filename: str = ""
+
+
+class DuplicateFileRead(BaseModel):
+    """A file that was not stored because the library already holds its bytes."""
+
+    model_config = ConfigDict(from_attributes=True)
+
+    filename: str
+    hash: str
+    size: int
+    library_id: int
+
+    # Null when the match was another file in the same upload, which has no row yet.
+    book_id: int | None = None
+    book_title: str | None = None
+
+
+class DuplicateBookRead(BaseModel):
+    """
+    A stored book that may be the same book as another one.
+
+    Unlike `DuplicateFileRead` this is a guess: the evidence is metadata two editions
+    of one work legitimately share. Nothing was refused on the strength of it.
+    """
+
+    model_config = ConfigDict(from_attributes=True)
+
+    book_id: int
+    title: str
+    authors: list[str]
+    library_id: int
+    cover_image: Path | None = None
+
+    # Why it matched: "identifier" and/or "title-author".
+    matched_on: list[str]
+
+
+class PossibleDuplicateRead(BaseModel):
+    """A book that was imported, together with what it might be a second copy of."""
+
+    model_config = ConfigDict(from_attributes=True)
+
+    book_id: int
+    title: str
+    candidates: list[DuplicateBookRead]
+
+
+class DuplicateBookGroupRead(BaseModel):
+    """Books the library holds that all look like copies of one book."""
+
+    books: list[DuplicateBookRead]
+
+
+class DuplicateDismissal(BaseModel):
+    """Two books a reader is saying are not the same book."""
+
+    book_a_id: int
+    book_b_id: int
+
+
+class BooksUploadResult(BaseModel):
+    """The outcome of a multi-file upload: what was created, and what was skipped."""
+
+    created: list["BookRead"]
+    skipped: list[DuplicateFileRead]
+
+    # Created, not skipped — these are books that went in and look like something the
+    # library already had. The reader decides what to do about it.
+    possible_duplicates: list[PossibleDuplicateRead] = Field(default_factory=list)
+
+
 class BookMetadataUpdate(BaseModel):
     title: str | None = None
     subtitle: str | None = None
@@ -170,6 +247,20 @@ class BookMetadataUpdate(BaseModel):
         return v
 
 
+class BookMerge(BaseModel):
+    """
+    Fold several books into one.
+
+    `metadata` is the reader's resolution of the fields the records disagreed on.
+    Anything it does not name keeps the survivor's value — merging metadata is a
+    judgement, so nothing is guessed on the caller's behalf.
+    """
+
+    survivor_id: int
+    merged_ids: list[int]
+    metadata: Optional["BookMetadataUpdate"] = None
+
+
 class BookProgressCreate(BaseModel):
     percentage: float
     epub_cfi: str | None = None

backend/src/chitai/services/consume.py

@@ -1,6 +1,7 @@
 import asyncio
 from pathlib import Path
 from collections import defaultdict
+from chitai.config import settings
 from chitai.database.models.library import Library
 from chitai.services import BookService, LibraryService
 from chitai.services.metadata_extractor import Extractor
@@ -111,13 +112,32 @@ class ConsumeDirectoryWatcher:
         """Process a batch of files."""
         try:
 
-            books = await self.book_service.create_many_from_existing_files(
+            result = await self.book_service.create_many_from_existing_files(
                 list(file_paths),
                 self.watch_path / Path(library_slug),
                 library=await self._get_library(library_slug),
             )
 
-            print(f"Created {len(books)} books!")
+            print(f"Created {len(result.books)} books!")
+
+            if result.duplicates:
+                print(
+                    f"Moved {len(result.duplicates)} already-stored file(s) "
+                    f"to {settings.duplicate_path}"
+                )
+
+            # Imported all the same — a metadata match is a guess, and there is nobody
+            # here to ask. The library's duplicates screen is where these get decided.
+            for possible in result.possible_duplicates:
+                names = ", ".join(
+                    f"{candidate.title} (#{candidate.book_id})"
+                    for candidate in possible.candidates
+                )
+                print(
+                    f"Imported {possible.title!r} (#{possible.book_id}), which may "
+                    f"already be in the library as: {names}"
+                )
+
 
         except Exception as e:
             print(f"Error processing batch: {e}")

backend/src/chitai/services/dependencies.py

@@ -127,6 +127,31 @@ def create_book_filter_dependencies(
     # Get base filters first
     filters = create_filter_dependencies(config, dep_defaults)
 
+    # OVERRIDE: id filter typed by the configured id type, not always `str`
+    #
+    # advanced_alchemy's `provide_id_filter` annotates `ids` as `list[str]` and
+    # ignores `config["id_filter"]` entirely, so `?ids=12` reaches the database as
+    # the string "12" and Postgres refuses `bigint = character varying`. Nothing
+    # called `?ids=` until the duplicates screen needed to fetch a handful of books
+    # by id, which is why it went unnoticed.
+    if id_type := config.get("id_filter"):
+        id_field = config.get("id_field", "id")
+
+        def provide_typed_id_filter(
+            ids=Parameter(query="ids", default=None, required=False),
+        ) -> CollectionFilter:
+            return CollectionFilter(field_name=id_field, values=ids)
+
+        # Attached as a type object rather than written as an annotation: this module
+        # has `from __future__ import annotations`, so a written one is stored as the
+        # string "Optional[list[id_type]]" and resolved against module globals, where
+        # a local named `id_type` does not exist.
+        provide_typed_id_filter.__annotations__["ids"] = Optional[list[id_type]]
+
+        filters[dep_defaults.ID_FILTER_DEPENDENCY_KEY] = Provide(
+            provide_typed_id_filter, sync_to_thread=False
+        )
+
     # OVERRIDE: Custom search filter with trigram search
     if config.get("search"):
         search_fields = config.get("search")

backend/src/chitai/services/matching.py

@@ -0,0 +1,207 @@
+# src/chitai/services/matching.py
+
+"""
+Normalisation for book-level duplicate detection.
+
+Two copies of one book rarely agree on how it is written down. One says
+`The Metamorphosis`, the other `Metamorphosis`; one credits `Kafka, Franz`, the other
+`Franz Kafka`; one carries the ISBN-10 and the other the ISBN-13 of the same edition.
+These functions reduce each of those to a single key, so the comparison is an equality
+test the database can index rather than a similarity score nobody can explain.
+
+Everything here is pure: the keys are computed once and stored on the row (see
+`Book.normalized_title`, `Author.normalized_name`, `Identifier.normalized_value`), so
+no Postgres extension is needed at query time.
+"""
+
+from __future__ import annotations
+
+import re
+import unicodedata
+
+from chitai.services.utils import is_valid_isbn, isbn10_to_isbn13
+
+
+# Asides a title carries that say nothing about which book it is:
+# "Frankenstein (Illustrated)", "Dune [Deluxe]".
+_BRACKETED = re.compile(r"[(\[{][^)\]}]*[)\]}]")
+
+# Edition and format qualifiers, matched only as a *trailing* run of words. Anchoring
+# to the end is what keeps "The Illustrated Man" a book and "Moby Dick Illustrated" a
+# format note — a qualifier trails the title, it is never the thing the title is about.
+_EDITION_NOISE = re.compile(
+    r"\s+(?:"
+    # "2nd edition", but also the compact forms publishers actually print on a
+    # cover: "2E", "3 Ed", "5e". The number alone is never enough — "Catch 22" is
+    # a title and must survive.
+    r"\d+(?:st|nd|rd|th)?\s*(?:edition|edn|ed|e)"
+    r"|(?:first|second|third|fourth|fifth|sixth|new|revised|expanded|updated|"
+    r"annotated|illustrated|unabridged|abridged|complete|definitive|deluxe|"
+    r"anniversary|collectors|international|kindle|paperback|hardcover|hardback|"
+    r"ebook|audiobook)"
+    r"(?:\s+(?:and|&)\s+\w+)*"
+    r"(?:\s+ed(?:ition|n)?)?"
+    r")$"
+)
+
+_LEADING_ARTICLE = re.compile(r"^(?:the|a|an)\s+")
+
+# Anything that is not a letter, a digit or a space, once accents are gone.
+_PUNCTUATION = re.compile(r"[^0-9a-z ]+")
+
+_WHITESPACE = re.compile(r"\s+")
+
+# Junk an extractor leaves on the end of a name: the `;` from a `DC:creator` list, the
+# `.epub` from a filename the author's name was read out of.
+_TRAILING_SEPARATORS = " ;,&/"
+_FILE_EXTENSION = re.compile(
+    r"\.(?:epub|pdf|mobi|azw3?|djvu|fb2|txt|cbz|cbr)$", re.IGNORECASE
+)
+
+# `J. R. R.` survives punctuation stripping as three one-letter words; `J.R.R.` as one.
+# Joining any run of them makes both `jrr`.
+_INITIAL_RUN = re.compile(r"\b(?:[a-z] )+[a-z]\b")
+
+# ISBNs are the same number under several names; everything else keeps its own.
+_ISBN_NAMES = {"isbn", "isbn-10", "isbn10", "isbn-13", "isbn13"}
+
+# Generated fresh for every build of a file, so two copies of one book never share one.
+# Matching on them would only re-find files the hash check already catches.
+_PER_BUILD_NAMES = {"uuid", "urn:uuid"}
+
+# Below this an identifier is not specific enough to be evidence: a Calibre `id` of
+# "42" would otherwise pair two unrelated books.
+_MIN_IDENTIFIER_LENGTH = 4
+
+
+def _fold(text: str) -> str:
+    """Casefolded, accent-free, punctuation-free, single-spaced."""
+    decomposed = unicodedata.normalize("NFKD", text.casefold())
+    unaccented = "".join(c for c in decomposed if not unicodedata.combining(c))
+
+    return _WHITESPACE.sub(" ", _PUNCTUATION.sub(" ", unaccented)).strip()
+
+
+def normalize_title(title: str | None) -> str:
+    """
+    Reduce a title to the key two copies of one book should share.
+
+    `Book.subtitle` is already split off by `Extractor.format_book_title`, so only what
+    is left in the title column is considered here.
+
+    Args:
+        title: The title as it was stored.
+
+    Returns:
+        The comparison key, or an empty string if nothing survives normalisation —
+        which is the signal not to match on the title at all.
+    """
+    if not title:
+        return ""
+
+    folded = _fold(_BRACKETED.sub(" ", title).replace("&", " and "))
+
+    # Repeated because qualifiers stack: "Dune Deluxe Edition Illustrated".
+    while (trimmed := _EDITION_NOISE.sub("", folded)) != folded:
+        folded = trimmed
+
+    # An article says nothing, but a title that is only an article is not improved by
+    # having none, and neither is one that noise removal emptied out.
+    return _LEADING_ARTICLE.sub("", folded, count=1) or folded
+
+
+def format_author_name(name: str | None) -> str:
+    """
+    Tidy an author's name into the one form the library writes them in.
+
+    Distinct from `normalize_author`, which throws away case, accents and spacing to
+    build a comparison key. This one is what a reader sees, so it keeps everything
+    that belongs to the name and only removes what an extractor added: a trailing
+    separator left over from a creator list, a file extension carried in from a
+    filename, and the `Surname, Given` ordering that EPUBs file names under.
+
+    Args:
+        name: The name as the file or filename gave it.
+
+    Returns:
+        The name to store, or an empty string if there is nothing left of it.
+    """
+    if not name:
+        return ""
+
+    tidied = _FILE_EXTENSION.sub("", name.strip().strip(_TRAILING_SEPARATORS).strip())
+
+    if tidied.count(",") == 1:
+        surname, given = (part.strip() for part in tidied.split(","))
+
+        # Only when the part before the comma is a single word. "Dave Thomas, Andy
+        # Hunt" is two people in one string, and flipping it would invent a third
+        # person who does not exist. Leaving an unrecognised form alone is the safe
+        # failure; rewriting it wrongly is not.
+        if surname and given and " " not in surname:
+            tidied = f"{given} {surname}"
+
+    return _WHITESPACE.sub(" ", tidied).strip()
+
+
+def normalize_author(name: str | None) -> str:
+    """
+    Reduce an author's name to the key their other books should share.
+
+    Deliberately not reduced to surname plus initial: that collides unrelated people,
+    and a wrong match here is a book pointed at a stranger's shelf.
+
+    Args:
+        name: The name as it was stored, in either `Franz Kafka` or `Kafka, Franz` form.
+
+    Returns:
+        The comparison key, or an empty string if nothing survives normalisation.
+    """
+    if not name:
+        return ""
+
+    # `Kafka, Franz` is one name written backwards. More than one comma is a list, or a
+    # suffix, and guessing at either does more harm than leaving it alone.
+    if name.count(",") == 1:
+        surname, forename = name.split(",")
+        name = f"{forename.strip()} {surname.strip()}"
+
+    folded = _fold(name)
+
+    return _INITIAL_RUN.sub(lambda run: run.group().replace(" ", ""), folded)
+
+
+def normalize_identifier(name: str, value: str) -> str | None:
+    """
+    Reduce one identifier to a `scheme:value` key, if it can carry a match at all.
+
+    Args:
+        name: What kind of identifier it is, as stored.
+        value: The identifier itself.
+
+    Returns:
+        The key, or None when the identifier is no use for matching: a per-build UUID,
+        something too short to be evidence, or an ISBN that fails its own checksum.
+    """
+    name = (name or "").strip().casefold()
+    value = (value or "").strip()
+
+    if not name or not value or name in _PER_BUILD_NAMES:
+        return None
+
+    if name in _ISBN_NAMES:
+        digits = re.sub(r"[^0-9Xx]", "", value).upper()
+
+        if not is_valid_isbn(digits):
+            return None
+
+        # One scheme for both forms: a publisher prints whichever it likes, and the
+        # ISBN-10 and ISBN-13 of an edition are the same number written twice.
+        isbn = digits if len(digits) == 13 else isbn10_to_isbn13(digits)
+        return f"isbn:{isbn}" if isbn else None
+
+    folded = _fold(value) or value.casefold()
+    if len(folded) < _MIN_IDENTIFIER_LENGTH:
+        return None
+
+    return f"{name}:{folded}"

backend/src/chitai/services/metadata_extractor.py

@@ -3,7 +3,6 @@
 # TODO: Code is a mess. Clean it up and add docstrings
 
 # Standard library
-from abc import ABC, abstractmethod
 import datetime
 from pathlib import Path
 from io import BytesIO
@@ -31,6 +30,147 @@ from chitai.services.utils import (
 logger = logging.getLogger(__name__)
 
 
+# Identifier schemes an EPUB can declare, mapped onto the names the rest of the app
+# uses. A scheme arrives either as an `opf:scheme` attribute or as a prefix on the
+# value itself (`urn:isbn:…`, `calibre:…`), and the two say the same thing.
+_IDENTIFIER_SCHEMES = {
+    "isbn": "isbn",
+    "isbn10": "isbn-10",
+    "isbn-10": "isbn-10",
+    "isbn13": "isbn-13",
+    "isbn-13": "isbn-13",
+    "uuid": "uuid",
+    "calibre": "calibre",
+    "doi": "doi",
+    "asin": "asin",
+    "amazon": "asin",
+    "mobi-asin": "asin",
+    "google": "google",
+    "goodreads": "goodreads",
+}
+
+# `scheme:rest`, with an optional `urn:` in front of it.
+_SCHEME_PREFIX = re.compile(r"^(?:urn:)?([A-Za-z][A-Za-z0-9.-]*):(.+)$")
+
+_UUID = re.compile(r"^[0-9a-f]{8}(?:-[0-9a-f]{4}){3}-[0-9a-f]{12}$", re.IGNORECASE)
+
+
+def parse_identifier(value: str, scheme: str | None = None) -> tuple[str, str] | None:
+    """
+    Work out what one raw identifier is, and what it is worth storing as.
+
+    EPUBs write the same ISBN as `9780486282114`, `978-0-486-28211-4` and
+    `urn:isbn:978-0-486-28211-4`, and carry plenty of identifiers that are not ISBNs
+    at all. Validating the string verbatim keeps only the first form and throws the
+    rest away, so normalise first and name whatever survives.
+
+    Args:
+        value: The identifier as the file wrote it.
+        scheme: What the file said it is, if it said anything — an `opf:scheme`
+            attribute. A prefix on the value takes precedence over this.
+
+    Returns:
+        The `(name, value)` to store, or `None` when there is nothing usable: an
+        empty value, or one declared to be an ISBN that fails its own checksum.
+    """
+    value = (value or "").strip()
+    if not value:
+        return None
+
+    name = _IDENTIFIER_SCHEMES.get((scheme or "").strip().casefold())
+
+    # An unrecognised prefix is part of the value rather than a scheme —
+    # "http://example.com/book" is not an identifier called "http".
+    if (match := _SCHEME_PREFIX.match(value)) and (
+        prefixed := _IDENTIFIER_SCHEMES.get(match.group(1).casefold())
+    ):
+        name = prefixed
+        value = match.group(2).strip()
+
+    if name is None or name.startswith("isbn"):
+        digits = re.sub(r"[^0-9Xx]", "", value).upper()
+        if is_valid_isbn(digits):
+            return f"isbn-{len(digits)}", digits
+
+        # Something that announced itself as an ISBN and is not one carries no
+        # information: storing it would link the reader to a page that does not exist.
+        if name is not None:
+            return None
+
+    return name or ("uuid" if _UUID.match(value) else "id"), value
+
+
+# Numbered editions, in the forms covers and catalogue records actually use:
+# "3rd Edition", "2E", "4e", "8_e", "/6e", "(2nd edition)", "Third International Edition".
+_ORDINAL_WORDS = {
+    "first": 1, "second": 2, "third": 3, "fourth": 4, "fifth": 5, "sixth": 6,
+    "seventh": 7, "eighth": 8, "ninth": 9, "tenth": 10, "eleventh": 11, "twelfth": 12,
+}
+
+# Words that sit between the number and "Edition" and belong to the same statement.
+_EDITION_QUALIFIER = (
+    r"(?:international|global|revised|updated|expanded|anniversary|deluxe|student|"
+    r"instructors?|annotated|illustrated|reprint)"
+)
+
+_EDITION = re.compile(
+    rf"""
+    [\s,;:/\-–—(\[]+                        # the separator the statement hangs off
+    (?:
+        (?P<num>\d{{1,2}})\s*(?:st|nd|rd|th)?[\s_]*
+            (?:{_EDITION_QUALIFIER}\s+)*(?:edition\b|edn\b|ed\b\.?|e\b)
+      | (?P<word>{"|".join(_ORDINAL_WORDS)})\s+
+            (?:{_EDITION_QUALIFIER}\s+)*(?:edition\b|edn\b|ed\b\.?)
+    )
+    [\s)\]]*                                # and its closing bracket, if it had one
+    """,
+    re.IGNORECASE | re.VERBOSE,
+)
+
+
+def split_edition(title: str | None) -> tuple[str | None, int | None]:
+    """
+    Separate a numbered edition statement from the title it is written into.
+
+    "Fluent Python, 2nd Edition" is one book with a field for the edition, not a
+    title. Left in place it also splits the library: the second edition never looks
+    like the first, and neither matches the copy whose file simply did not mention it.
+
+    The number is what makes this safe. Nothing is stripped without one, so
+    "Catch 22" and "Blade Runner 2049" keep their numbers and "Global Edition" —
+    which is a variant, not a numbered edition, and has nowhere to go in an
+    integer column — is left in the title where it can still be read.
+
+    Args:
+        title: The title as the file or filename gave it.
+
+    Returns:
+        The title without the edition statement, and the edition number. The title
+        unchanged and None when there is no numbered edition in it, or when removing
+        it would leave nothing behind.
+    """
+    if not title:
+        return title, None
+
+    if (match := _EDITION.search(title)) is None:
+        return title, None
+
+    edition = (
+        int(match["num"]) if match["num"] else _ORDINAL_WORDS[match["word"].casefold()]
+    )
+
+    stripped = _EDITION.sub(" ", title)
+    stripped = re.sub(r"\s{2,}", " ", stripped)
+    stripped = re.sub(r"\s+([,;:.!?])", r"\1", stripped)  # "Works : What" → "Works: What"
+    stripped = stripped.strip(" ,;:-–—/")
+
+    # A title that is only an edition statement is not improved by having none.
+    if not stripped:
+        return title, None
+
+    return stripped, edition
+
+
 class FileExtractor(Protocol):
     @classmethod
     async def extract_metadata(
@@ -54,15 +194,36 @@ class Extractor:
         # EPUB tends to give better metadata results over pdf
         sorted_files = sorted(files, key=lambda f: Extractor._get_file_priority(f))
 
+        # Identifiers accumulate across formats instead of replacing each other. Every
+        # other field is a single value where the later, better-trusted format simply
+        # wins, but identifiers are a *collection*: a book holding an EPUB and a PDF
+        # genuinely carries what both of them declare, and merging the dict wholesale
+        # threw away everything the earlier format found. An EPUB that declares an
+        # ASIN, a Google volume id and a Calibre id kept none of them once a PDF
+        # contributed a single ISBN.
+        identifiers: dict[str, str] = {}
+
         for file in sorted_files:
             match get_file_extension(file):
                 case "epub":
-                    metadata = metadata | await EpubExtractor.extract_metadata(file)
+                    extracted = await EpubExtractor.extract_metadata(file)
                 case "pdf":
-                    metadata = metadata | await PdfExtractor.extract_metadata(file)
+                    extracted = await PdfExtractor.extract_metadata(file)
                 case _:
                     break
 
+            # First writer wins per name, and the files are already ordered by how
+            # far their metadata can be trusted. A `dc:identifier` the publisher
+            # declared outranks an ISBN scraped out of a PDF's copyright page, which
+            # routinely prints the ISBNs of other formats and older editions too.
+            for name, value in (extracted.pop("identifiers", None) or {}).items():
+                identifiers.setdefault(name, value)
+
+            metadata = metadata | extracted
+
+        if identifiers:
+            metadata["identifiers"] = identifiers
+
         # Get metadata from file names
         for file in files:
             metadata = FilenameExtractor.extract_metadata(file) | metadata
@@ -75,7 +236,14 @@ class Extractor:
 
         # format the title
         if metadata.get('title', None):
-            title, subtitle = Extractor.format_book_title(metadata["title"])
+            # Before the subtitle split, so the edition cannot be mistaken for one:
+            # "How Linux Works, 3rd Edition: What Every Superuser Should Know" has to
+            # lose the edition first for the colon count to mean anything.
+            title, edition = split_edition(metadata["title"])
+            if edition is not None:
+                metadata.setdefault("edition", edition)
+
+            title, subtitle = Extractor.format_book_title(title)
             metadata["title"] = title
             metadata["subtitle"] = subtitle
 
@@ -236,7 +404,7 @@ class PdfExtractor(FileExtractor):
         try:
             return datetime.datetime.strptime(date_portion, "%Y%m%d").date()
 
-        except Exception as e:
+        except Exception:
             return None
 
     @classmethod
@@ -363,15 +531,23 @@ class EpubExtractor(FileExtractor):
 
     @classmethod
     def _extract_identifiers(cls, epub: epub.EpubBook) -> dict[str, str]:
+        """
+        Every `DC:identifier` the file carries, keyed by what kind of thing it is.
+
+        Non-ISBN identifiers are kept: `Identifier` is a free-form name/value pair, so
+        a Calibre id or an ASIN costs nothing to store and is one more thing two copies
+        of a book can be recognised by.
+        """
         identifiers = {}
 
-        for id in epub.get_metadata("DC", "identifier"):
-            if is_valid_isbn(id[0]):
-                if len(id[0]) == 13:
-                    identifiers.update({"isbn-13": id[0]})
+        for value, attributes in epub.get_metadata("DC", "identifier"):
+            scheme = None
+            if isinstance(attributes, dict):
+                scheme = attributes.get("opf:scheme") or attributes.get("scheme")
 
-                elif len(id[0]) == 10:
-                    identifiers.update({"isbn-10": id[0]})
+            if (parsed := parse_identifier(value, scheme)) is not None:
+                name, parsed_value = parsed
+                identifiers[name] = parsed_value
 
         return identifiers
 
@@ -380,7 +556,7 @@ class EpubExtractor(FileExtractor):
         try:
             return epub.get_metadata("DC", "description")[0][0]
 
-        except:
+        except Exception:
             return None
 
     @classmethod
@@ -389,15 +565,15 @@ class EpubExtractor(FileExtractor):
             date_str = epub.get_metadata("DC", "date")[0][0].split("T")[0]
             return datetime.date.fromisoformat(date_str)
 
-        except:
+        except Exception:
             return None
 
     @classmethod
     def _extract_publisher(cls, epub: epub.EpubBook) -> str | None:
         try:
-            epub.get_metadata("DC", "publisher")[0][0]
+            return epub.get_metadata("DC", "publisher")[0][0]
 
-        except:
+        except Exception:
             return None
 
     @classmethod
@@ -421,7 +597,7 @@ class EpubExtractor(FileExtractor):
                             cover_item = epub.get_item_with_id(cover_id)
                             if cover_item:
                                 return PIL.Image.open(BytesIO(cover_item.content))
-        except Exception as e:
+        except Exception:
             pass  # Fallback to next strategy
 
         # Strategy 2: Search image filenames for "cover" keyword
@@ -489,7 +665,11 @@ class FilenameExtractor(FileExtractor):
         elif isinstance(input, Path):
             filename = get_filename(input, ext=False)
         elif isinstance(input, UploadFile):
-            filename = Path(input.filename).name
+            # `.stem`, not `.name`: the extension is not part of the metadata, and
+            # this is the browser upload path, so keeping it is how a library fills
+            # up with authors called "Sam Newman.epub". The other two branches have
+            # always stripped it.
+            filename = Path(input.filename).stem
         else:
             raise ValueError("Input type not supported")
 

backend/src/chitai/services/utils.py

@@ -3,13 +3,11 @@
 # Standard library
 from __future__ import annotations
 
+import errno
 import hashlib
 from pathlib import Path
 import shutil
-from typing import TYPE_CHECKING, BinaryIO
-
-if TYPE_CHECKING:
-    from hashlib import _Hash
+from typing import BinaryIO
 
 # Third-party libraries
 import PIL
@@ -32,6 +30,9 @@ KO_STEP = 1024
 KO_SAMPLE_SIZE = 1024
 KO_INDICES = range(-1, 11)  # -1 to 10 inclusive
 
+# How much is read at a time while hashing.
+HASH_CHUNK_SIZE = 262144  # 256 KiB
+
 
 def _lshift32(val: int, shift: int) -> int:
     """
@@ -100,10 +101,9 @@ async def calculate_koreader_hash(file_path: Path) -> str:
     offsets = _get_koreader_offsets()
 
     file_pos = 0
-    chunk_size = 262144  # 256 KiB
 
     async with aiofiles.open(file_path, "rb") as f:
-        while chunk := await f.read(chunk_size):
+        while chunk := await f.read(HASH_CHUNK_SIZE):
             _partial_md5_from_chunk(chunk, hasher, offsets, file_pos)
             file_pos += len(chunk)
 
@@ -132,6 +132,49 @@ class StreamingHasher:
         """Return the final hash."""
         return self.hasher.hexdigest()
 
+    @property
+    def size(self) -> int:
+        """Total number of bytes fed in so far."""
+        return self.position
+
+
+async def fingerprint_upload(file: UploadFile) -> tuple[str, int]:
+    """
+    Calculate the hash and byte size of an uploaded file without storing it.
+
+    Duplicate detection has to answer before anything is written to the library, so
+    the file is read here and rewound for whoever writes it afterwards.
+
+    Args:
+        file: The uploaded file to read.
+
+    Returns:
+        The file's `(hash, size)` pair.
+    """
+    hasher = StreamingHasher()
+
+    await file.seek(0)
+    while chunk := await file.read(HASH_CHUNK_SIZE):
+        hasher.update(chunk)
+    await file.seek(0)
+
+    return hasher.hexdigest(), hasher.size
+
+
+async def fingerprint_file(file_path: Path) -> tuple[str, int]:
+    """
+    Calculate the hash and byte size of a file already on disk.
+
+    Args:
+        file_path: Path to the file to read.
+
+    Returns:
+        The file's `(hash, size)` pair.
+    """
+    stats = await aios.stat(file_path)
+    return await calculate_koreader_hash(file_path), stats.st_size
+
+
 ##################################
 #  Filesystem related utilities  #
 ##################################
@@ -184,7 +227,16 @@ async def move_file(src_path: Path, dest_path: Path, create_dirs=True) -> None:
         if dest_dir:  # Only create if there's a directory path
             await aios.makedirs(dest_dir, exist_ok=True)
 
-    await aios.rename(src_path, dest_path)
+    try:
+        await aios.rename(src_path, dest_path)
+    except OSError as exc:
+        if exc.errno != errno.EXDEV:
+            raise
+
+        # Source and destination are on different filesystems, which rename cannot
+        # cross. Libraries, the consume directory and the duplicates directory are
+        # all configured separately, so they can easily be separate mounts.
+        shutil.move(str(src_path), str(dest_path))
 
 async def move_dir_contents(source_dir: Path | str, target_dir: Path | str) -> None:
     """
@@ -432,7 +484,7 @@ def is_valid_isbn(isbn: str) -> bool:
             return is_valid_isbn13(isbn)
         else:
             return False
-    except:
+    except Exception:
         return False
 
 
@@ -457,6 +509,29 @@ def is_valid_isbn10(isbn: str) -> bool:
     return str(check_digit) == isbn[-1] or (check_digit == 10 and isbn[-1] in "Xx")
 
 
+def isbn10_to_isbn13(isbn: str) -> str | None:
+    """
+    Convert an ISBN-10 to the ISBN-13 naming the same edition.
+
+    The two are the same number written twice: prefix `978`, drop the ISBN-10 check
+    digit, recompute the check digit under the ISBN-13 rule. Matching only works if
+    both forms collapse onto one, since a publisher may print either.
+
+    Args:
+        isbn: A 10-character ISBN, digits and an optional trailing `X` only.
+
+    Returns:
+        The equivalent ISBN-13, or None if the input is not a valid ISBN-10.
+    """
+    if not is_valid_isbn(isbn) or len(isbn) != 10:
+        return None
+
+    digits = f"978{isbn[:9]}"
+    total = sum(int(digit) * (1 if i % 2 == 0 else 3) for i, digit in enumerate(digits))
+
+    return f"{digits}{(10 - total % 10) % 10}"
+
+
 def is_valid_isbn13(isbn: str) -> bool:
     """
     Validate an ISBN-13 number using its check digit.

backend/tests/conftest.py

@@ -40,7 +40,13 @@ pytest_plugins = [
 
 @pytest.fixture(autouse=True)
 def _patch_settings(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
-    monkeypatch.setattr(settings, "book_cover_path", f"{tmp_path}/covers")
+    # PIL will not create the directory it is asked to save into, so anything that
+    # imports a file carrying a cover needs it to exist first.
+    covers = tmp_path / "covers"
+    covers.mkdir()
+
+    monkeypatch.setattr(settings, "book_cover_path", str(covers))
+    monkeypatch.setattr(settings, "duplicate_path", str(tmp_path / "duplicates"))
 
 
 @pytest.fixture(name="engine")

backend/tests/integration/test_book.py

@@ -37,7 +37,8 @@ from pathlib import Path
         (
             Path("tests/data_files/Calculus Made Easy - Silvanus Thompson.pdf"),
             2,
-            "The Project Gutenberg eBook #33283: Calculus Made Easy, 2nd Edition",
+            # The ", 2nd Edition" is split off into `edition`, not kept in the title.
+            "The Project Gutenberg eBook #33283: Calculus Made Easy",
             ["Silvanus Phillips Thompson"],
         ),
     ],
@@ -208,6 +209,21 @@ async def test_get_book_file(
     assert downloaded_content == file_content
 
 
+async def test_list_books_by_id(populated_authenticated_client: AsyncClient) -> None:
+    """
+    `?ids=` has to reach the database as integers.
+
+    advanced_alchemy's stock id filter annotates the parameter as `list[str]` whatever
+    the configured id type, so the ids arrived as strings and Postgres refused to
+    compare a bigint primary key against them. Nothing called it until a screen needed
+    to fetch a handful of books by id.
+    """
+    response = await populated_authenticated_client.get("/books?ids=1&ids=2&pageSize=10")
+
+    assert response.status_code == 200
+    assert sorted(book["id"] for book in response.json()["items"]) == [1, 2]
+
+
 async def test_get_book_by_id(populated_authenticated_client: AsyncClient) -> None:
     """Test retrieving a specific book by ID."""
 
@@ -366,7 +382,8 @@ async def test_create_multiple_books_from_directory(
 
     assert response.status_code == 201
     data = response.json()
-    assert len(data.get("items") or data.get("data")) >= 1
+    assert len(data["created"]) == 2
+    assert data["skipped"] == []
 
 
 async def test_create_books_from_parent_directory_keeps_embedded_title(
@@ -396,7 +413,7 @@ async def test_create_books_from_parent_directory_keeps_embedded_title(
 
     assert response.status_code == 201
 
-    books = response.json()["items"]
+    books = response.json()["created"]
     assert len(books) == 1
     assert books[0]["title"] == "Metamorphosis"
 
@@ -419,11 +436,336 @@ async def test_create_books_groups_formats_within_one_folder(
 
     assert response.status_code == 201
 
-    books = response.json()["items"]
+    books = response.json()["created"]
     assert len(books) == 1
     assert len(books[0]["files"]) == 2
 
 
+class TestDuplicateHandling:
+    """A file the library already holds must not be stored a second time."""
+
+    epub_path = Path("tests/data_files/Metamorphosis - Franz Kafka.epub")
+
+    def upload(self, name: str | None = None) -> list[tuple[str, tuple]]:
+        return [
+            (
+                "files",
+                (
+                    name or self.epub_path.name,
+                    self.epub_path.read_bytes(),
+                    "application/epub+zip",
+                ),
+            )
+        ]
+
+    async def test_bulk_upload_reports_skipped_files(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        """Re-dropping a folder must import what is new and name what was not."""
+        first = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload()
+        )
+        assert first.status_code == 201
+        created = first.json()["created"][0]
+
+        second = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload()
+        )
+
+        assert second.status_code == 201
+        result = second.json()
+        assert result["created"] == []
+        assert len(result["skipped"]) == 1
+
+        skipped = result["skipped"][0]
+        assert skipped["filename"] == self.epub_path.name
+        assert skipped["book_id"] == created["id"]
+        assert skipped["book_title"] == created["title"]
+
+    async def test_bulk_upload_can_be_forced(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload()
+        )
+
+        response = await authenticated_client.post(
+            "/books/fromFiles?library_id=1&allow_duplicates=true", files=self.upload()
+        )
+
+        assert response.status_code == 201
+        assert len(response.json()["created"]) == 1
+        assert response.json()["skipped"] == []
+
+    async def test_single_book_create_conflicts(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        """Naming files deliberately earns a refusal rather than a silent drop."""
+        await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload()
+        )
+
+        response = await authenticated_client.post(
+            "/books?library_id=1", files=self.upload(), data={"library_id": 1}
+        )
+
+        assert response.status_code == 409
+        assert response.json()["extra"][0]["filename"] == self.epub_path.name
+
+        forced = await authenticated_client.post(
+            "/books?library_id=1&allow_duplicates=true",
+            files=self.upload(),
+            data={"library_id": 1},
+        )
+        assert forced.status_code == 201
+
+    async def test_adding_another_books_file_conflicts(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        created = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload()
+        )
+        book_id = created.json()["created"][0]["id"]
+
+        other = await authenticated_client.post(
+            "/books?library_id=1",
+            files=[
+                (
+                    "files",
+                    (
+                        "war.epub",
+                        Path("tests/data_files/The Art of War - Sun Tzu.epub").read_bytes(),
+                        "application/epub+zip",
+                    ),
+                )
+            ],
+            data={"library_id": 1},
+        )
+        other_id = other.json()["id"]
+
+        response = await authenticated_client.post(
+            f"/books/{book_id}/files",
+            files=[
+                (
+                    "files",
+                    (
+                        "war.epub",
+                        Path("tests/data_files/The Art of War - Sun Tzu.epub").read_bytes(),
+                        "application/epub+zip",
+                    ),
+                )
+            ],
+        )
+
+        assert response.status_code == 409
+        assert response.json()["extra"][0]["book_id"] == other_id
+
+    async def test_resending_a_books_own_file_changes_nothing(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        created = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload()
+        )
+        book_id = created.json()["created"][0]["id"]
+
+        response = await authenticated_client.post(
+            f"/books/{book_id}/files", files=self.upload()
+        )
+
+        assert response.status_code == 201
+        assert len(response.json()["files"]) == 1
+
+    async def test_duplicates_can_be_checked_before_uploading(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        """The pre-flight check answers from hashes alone, with no file sent."""
+        created = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload()
+        )
+        book = created.json()["created"][0]
+        stored = book["files"][0]
+
+        response = await authenticated_client.post(
+            "/books/duplicate-files?library_id=1",
+            json=[
+                {
+                    "hash": stored["hash"],
+                    "size": stored["size"],
+                    "filename": "local-copy.epub",
+                },
+                {"hash": stored["hash"], "size": stored["size"] + 1},
+                {"hash": "0" * 32, "size": 1234},
+            ],
+        )
+
+        assert response.status_code == 200
+
+        matches = response.json()
+        assert len(matches) == 1
+        assert matches[0]["filename"] == "local-copy.epub"
+        assert matches[0]["book_id"] == book["id"]
+
+
+@pytest.mark.asyncio
+class TestDuplicateBooks:
+    """A second copy of a book is imported and reported, never refused."""
+
+    epub_path = Path("tests/data_files/Metamorphosis - Franz Kafka.epub")
+
+    def upload(self, name: str, pad: bool = False) -> list[tuple[str, tuple]]:
+        """
+        The fixture, optionally padded so it is a different file and the same book.
+
+        Padding the archive changes its size and its sampled hash without disturbing
+        the metadata, which is exactly the case the file-level check cannot see.
+        """
+        data = self.epub_path.read_bytes() + (b"\0" * 64 if pad else b"")
+
+        return [("files", (name, data, "application/epub+zip"))]
+
+    async def test_a_second_edition_is_created_and_reported(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        first = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("first.epub")
+        )
+        original = first.json()["created"][0]
+
+        second = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("second.epub", pad=True)
+        )
+
+        assert second.status_code == 201
+        body = second.json()
+
+        # Created, not skipped: a metadata match is a guess, and refusing a legitimate
+        # second edition costs more than a note does.
+        assert len(body["created"]) == 1
+        assert body["skipped"] == []
+
+        assert len(body["possible_duplicates"]) == 1
+        possible = body["possible_duplicates"][0]
+        assert possible["book_id"] == body["created"][0]["id"]
+
+        candidate = possible["candidates"][0]
+        assert candidate["book_id"] == original["id"]
+        assert candidate["title"] == original["title"]
+        assert candidate["authors"] == ["Franz Kafka"]
+        assert "title-author" in candidate["matched_on"]
+
+    async def test_the_review_screen_groups_them(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        first = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("first.epub")
+        )
+        second = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("second.epub", pad=True)
+        )
+
+        book_ids = sorted(
+            [first.json()["created"][0]["id"], second.json()["created"][0]["id"]]
+        )
+
+        response = await authenticated_client.get("/books/duplicate-books?library_id=1")
+
+        assert response.status_code == 200
+        groups = response.json()
+        assert len(groups) == 1
+        assert [book["book_id"] for book in groups[0]["books"]] == book_ids
+
+    async def test_a_dismissed_group_stays_dismissed(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        first = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("first.epub")
+        )
+        second = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("second.epub", pad=True)
+        )
+
+        pair = {
+            "book_a_id": first.json()["created"][0]["id"],
+            "book_b_id": second.json()["created"][0]["id"],
+        }
+
+        dismissed = await authenticated_client.post(
+            "/books/duplicate-books/dismissals", json=pair
+        )
+        assert dismissed.status_code == 204
+
+        response = await authenticated_client.get("/books/duplicate-books?library_id=1")
+        assert response.json() == []
+
+        restored = await authenticated_client.delete(
+            "/books/duplicate-books/dismissals"
+            f"?book_a_id={pair['book_b_id']}&book_b_id={pair['book_a_id']}"
+        )
+        assert restored.status_code == 204
+
+        response = await authenticated_client.get("/books/duplicate-books?library_id=1")
+        assert len(response.json()) == 1
+
+    async def test_two_books_merge_into_one(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        """The survivor keeps its id and gains the other's file; the other is gone."""
+        first = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("first.epub")
+        )
+        second = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("second.epub", pad=True)
+        )
+        keep = first.json()["created"][0]
+        fold = second.json()["created"][0]
+
+        response = await authenticated_client.post(
+            "/books/merge?library_id=1",
+            json={
+                "survivor_id": keep["id"],
+                "merged_ids": [fold["id"]],
+                "metadata": {"title": "Metamorphosis", "edition": 2},
+            },
+        )
+
+        assert response.status_code == 201
+        merged = response.json()
+
+        assert merged["id"] == keep["id"]
+        assert merged["edition"] == 2
+        assert len(merged["files"]) == 2
+
+        # The folded record is gone, and the group it formed with it.
+        assert (await authenticated_client.get(f"/books/{fold['id']}")).status_code == 404
+        groups = await authenticated_client.get("/books/duplicate-books?library_id=1")
+        assert groups.json() == []
+
+    async def test_merging_an_unknown_book_is_refused(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        created = await authenticated_client.post(
+            "/books/fromFiles?library_id=1", files=self.upload("first.epub")
+        )
+        keep = created.json()["created"][0]["id"]
+
+        response = await authenticated_client.post(
+            "/books/merge?library_id=1",
+            json={"survivor_id": keep, "merged_ids": [9999]},
+        )
+
+        assert response.status_code == 400
+
+    async def test_dismissing_an_unknown_book_is_refused(
+        self, authenticated_client: AsyncClient
+    ) -> None:
+        response = await authenticated_client.post(
+            "/books/duplicate-books/dismissals",
+            json={"book_a_id": 1, "book_b_id": 9999},
+        )
+
+        assert response.status_code == 400
+
+
 # NOTE: the multi-book ZIP download is covered at the service level, in
 # tests/unit/test_services/test_book_service.py. Driving `/books/download` through
 # AsyncTestClient hangs in fixture teardown: it is the only `Stream` endpoint in the

backend/tests/unit/test_matching.py

@@ -0,0 +1,233 @@
+"""Tests for the normalization behind book-level duplicate detection."""
+
+import pytest
+
+from chitai.services.matching import (
+    format_author_name,
+    normalize_author,
+    normalize_identifier,
+    normalize_title,
+)
+from chitai.services.metadata_extractor import parse_identifier
+from chitai.services.utils import isbn10_to_isbn13
+
+
+class TestNormalizeTitle:
+    """Two copies of one book rarely agree on how the title is written."""
+
+    @pytest.mark.parametrize(
+        ("title", "expected"),
+        [
+            ("The Metamorphosis", "metamorphosis"),
+            ("Metamorphosis", "metamorphosis"),
+            ("METAMORPHOSIS", "metamorphosis"),
+            ("A Tale of Two Cities", "tale of two cities"),
+            ("An Enquiry", "enquiry"),
+            # Accents, punctuation and ampersands are spelling, not identity.
+            ("Les Misérables", "les miserables"),
+            ("Moby Dick; Or, The Whale", "moby dick or the whale"),
+            ("Sense & Sensibility", "sense and sensibility"),
+            # Bracketed asides and trailing edition noise say nothing about the book.
+            ("Frankenstein (Illustrated)", "frankenstein"),
+            ("Frankenstein [Kindle Edition]", "frankenstein"),
+            ("Frankenstein, 2nd Edition", "frankenstein"),
+            # The compact forms a cover actually carries.
+            ("Building Microservices, 2E", "building microservices"),
+            ("Building Microservices 2e", "building microservices"),
+            ("Frankenstein 3 Ed", "frankenstein"),
+            ("Dungeons & Dragons 5e", "dungeons and dragons"),
+            ("Frankenstein Revised Edition", "frankenstein"),
+            ("Dune Deluxe Edition Illustrated", "dune"),
+            ("", ""),
+            (None, ""),
+        ],
+    )
+    def test_titles_that_should_agree(self, title: str | None, expected: str) -> None:
+        assert normalize_title(title) == expected
+
+    def test_a_qualifier_that_is_the_title_survives(self) -> None:
+        """A trailing qualifier is noise; the same word at the front is the book."""
+        assert normalize_title("The Illustrated Man") == "illustrated man"
+
+    def test_normalization_never_empties_a_title(self) -> None:
+        """An article-only title is not improved by having no article left."""
+        assert normalize_title("The") == "the"
+
+    @pytest.mark.parametrize(
+        "title", ["Catch 22", "Fahrenheit 451", "Blade Runner 2049", "1984", "Apollo 13"]
+    )
+    def test_a_number_is_not_an_edition(self, title: str) -> None:
+        """Edition stripping keys on the `e`; a bare number is part of the title."""
+        assert normalize_title(title) == title.casefold()
+
+
+class TestNormalizeAuthor:
+    """One person, written down several ways."""
+
+    @pytest.mark.parametrize(
+        ("name", "expected"),
+        [
+            ("Franz Kafka", "franz kafka"),
+            ("Kafka, Franz", "franz kafka"),
+            ("KAFKA, FRANZ", "franz kafka"),
+            ("Émile Zola", "emile zola"),
+            ("Doyle, Arthur Conan", "arthur conan doyle"),
+            # Runs of initials are joined, so spacing them out changes nothing.
+            ("J.R.R. Tolkien", "jrr tolkien"),
+            ("J. R. R. Tolkien", "jrr tolkien"),
+            ("JRR Tolkien", "jrr tolkien"),
+            ("", ""),
+            (None, ""),
+        ],
+    )
+    def test_names_that_should_agree(self, name: str | None, expected: str) -> None:
+        assert normalize_author(name) == expected
+
+    def test_two_people_are_not_reduced_together(self) -> None:
+        """Surname plus initial would collide unrelated writers; it is not used."""
+        assert normalize_author("Charles Dickens") != normalize_author("Colin Dexter")
+
+    def test_a_list_is_left_alone(self) -> None:
+        """More than one comma is a list or a suffix, and guessing does more harm."""
+        assert normalize_author("Smith, John, Jr.") == "smith john jr"
+
+
+class TestFormatAuthorName:
+    """What gets stored and shown, as opposed to what gets compared."""
+
+    @pytest.mark.parametrize(
+        ("written", "expected"),
+        [
+            # A leftover separator from a `DC:creator` list.
+            ("Newman, Sam;", "Sam Newman"),
+            ("Sam Newman ", "Sam Newman"),
+            ("  Dan   Vanderkam  ", "Dan Vanderkam"),
+            # `Surname, Given` is how EPUBs file a name, not how anyone reads it.
+            ("Kleppmann, Martin", "Martin Kleppmann"),
+            ("Huxley, Aldous", "Aldous Huxley"),
+            ("Liu, Cixin", "Cixin Liu"),
+            # An extension carried in from the filename the name was read out of.
+            ("Sam Newman.epub", "Sam Newman"),
+            ("Franz Kafka.mobi", "Franz Kafka"),
+            ("Brian W. Kernighan.epub", "Brian W. Kernighan"),
+            ("", ""),
+            (None, ""),
+        ],
+    )
+    def test_names_are_tidied(self, written: str | None, expected: str) -> None:
+        assert format_author_name(written) == expected
+
+    @pytest.mark.parametrize(
+        "written",
+        [
+            # Two people in one string. Flipping it would invent a third.
+            "Dave Thomas, Andy Hunt",
+            "Mark Richards, Neal Ford",
+            # A compound surname is not recognised, and is left alone rather than
+            # rearranged wrongly.
+            "García Márquez, Gabriel",
+        ],
+    )
+    def test_an_unrecognised_form_is_left_alone(self, written: str) -> None:
+        assert format_author_name(written) == written
+
+    def test_case_and_accents_belong_to_the_author(self) -> None:
+        """Tidying removes what an extractor added; it does not correct spelling."""
+        assert format_author_name("Michał Płachta.epub") == "Michał Płachta"
+        assert format_author_name("Steve McConnell") == "Steve McConnell"
+
+    @pytest.mark.parametrize(
+        "written", ["Newman, Sam;", "Sam Newman.epub", "Kleppmann, Martin"]
+    )
+    def test_tidying_is_idempotent(self, written: str) -> None:
+        """`unique_filter` tidies a name that may already be tidy; it must not drift."""
+        once = format_author_name(written)
+        assert format_author_name(once) == once
+
+
+class TestNormalizeIdentifier:
+    """Identifiers only help if the same edition produces the same key."""
+
+    def test_isbn_10_and_isbn_13_are_one_key(self) -> None:
+        assert normalize_identifier("isbn-10", "0486282112") == "isbn:9780486282114"
+        assert normalize_identifier("isbn-13", "9780486282114") == "isbn:9780486282114"
+
+    @pytest.mark.parametrize(
+        "written", ["978-0-486-28211-4", "978 0 486 28211 4", "9780486282114"]
+    )
+    def test_formatting_is_not_part_of_an_isbn(self, written: str) -> None:
+        assert normalize_identifier("isbn", written) == "isbn:9780486282114"
+
+    def test_an_isbn_that_fails_its_checksum_is_no_evidence(self) -> None:
+        assert normalize_identifier("isbn-13", "9780486282115") is None
+
+    def test_uuids_are_refused(self) -> None:
+        """Generated per build, so they only re-find what the hash check catches."""
+        assert normalize_identifier("uuid", "3f2b1c4e-1111-2222-3333-444455556666") is None
+        assert normalize_identifier("urn:uuid", "3f2b1c4e-1111-2222-3333-444455556666") is None
+
+    def test_other_schemes_keep_their_own_key(self) -> None:
+        assert normalize_identifier("asin", "B000FC0PDA") == "asin:b000fc0pda"
+        assert normalize_identifier("ASIN", "b000fc0pda") == "asin:b000fc0pda"
+
+    def test_something_too_short_is_not_evidence(self) -> None:
+        """A Calibre id of "42" would otherwise pair two unrelated books."""
+        assert normalize_identifier("calibre", "42") is None
+
+    @pytest.mark.parametrize(("name", "value"), [("", "1234567"), ("asin", "")])
+    def test_half_an_identifier_is_no_identifier(self, name: str, value: str) -> None:
+        assert normalize_identifier(name, value) is None
+
+
+class TestIsbnConversion:
+    def test_isbn_10_converts_to_its_isbn_13(self) -> None:
+        assert isbn10_to_isbn13("0486282112") == "9780486282114"
+
+    def test_a_trailing_x_is_a_digit(self) -> None:
+        assert isbn10_to_isbn13("043942089X") == "9780439420891"
+
+    @pytest.mark.parametrize("isbn", ["0486282113", "9780486282114", "nonsense"])
+    def test_anything_that_is_not_an_isbn_10_converts_to_nothing(self, isbn: str) -> None:
+        assert isbn10_to_isbn13(isbn) is None
+
+
+class TestParseIdentifier:
+    """What an EPUB writes, and what is worth storing for it."""
+
+    @pytest.mark.parametrize(
+        "written",
+        [
+            "9780486282114",
+            "978-0-486-28211-4",
+            "urn:isbn:9780486282114",
+            "urn:isbn:978-0-486-28211-4",
+            "ISBN:978-0-486-28211-4",
+        ],
+    )
+    def test_isbns_survive_however_they_are_written(self, written: str) -> None:
+        assert parse_identifier(written) == ("isbn-13", "9780486282114")
+
+    def test_the_scheme_attribute_is_read_too(self) -> None:
+        assert parse_identifier("0-486-28211-2", "ISBN") == ("isbn-10", "0486282112")
+
+    def test_non_isbn_identifiers_are_kept(self) -> None:
+        assert parse_identifier("urn:uuid:3f2b1c4e-1111-2222-3333-444455556666") == (
+            "uuid",
+            "3f2b1c4e-1111-2222-3333-444455556666",
+        )
+        assert parse_identifier("calibre:1234") == ("calibre", "1234")
+        assert parse_identifier("B000FC0PDA", "mobi-asin") == ("asin", "B000FC0PDA")
+
+    def test_an_unrecognised_prefix_is_part_of_the_value(self) -> None:
+        """"http://example.com/book" is not an identifier called "http"."""
+        assert parse_identifier("http://www.gutenberg.org/5200") == (
+            "id",
+            "http://www.gutenberg.org/5200",
+        )
+
+    def test_a_declared_isbn_that_is_not_one_is_dropped(self) -> None:
+        assert parse_identifier("urn:isbn:not-an-isbn") is None
+
+    @pytest.mark.parametrize("written", ["", "   ", None])
+    def test_nothing_yields_nothing(self, written: str | None) -> None:
+        assert parse_identifier(written) is None

backend/tests/unit/test_metadata_extractor.py

@@ -1,7 +1,13 @@
 import pytest
+from ebooklib import epub
 from pathlib import Path
 from datetime import date
-from chitai.services.metadata_extractor import EpubExtractor
+from chitai.services.metadata_extractor import (
+    EpubExtractor,
+    Extractor,
+    PdfExtractor,
+    split_edition,
+)
 
 
 @pytest.mark.asyncio()
@@ -15,3 +21,172 @@ class TestEpubExtractor:
         assert metadata["authors"] == ["Herman Melville"]
         assert metadata["published_date"] == date(year=2001, month=7, day=1)
 
+
+
+EPUB = Path("tests/data_files/Metamorphosis - Franz Kafka.epub")
+PDF = Path("tests/data_files/Calculus Made Easy - Silvanus Thompson.pdf")
+
+
+@pytest.mark.asyncio()
+class TestIdentifierMerging:
+    """A book's formats each contribute identifiers; none of them replaces the rest."""
+
+    async def test_every_format_contributes(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        """
+        Identifiers are a collection, not a single value.
+
+        Merging the whole dict meant the last format to report won outright: an EPUB
+        declaring an ASIN, a Google volume id and a Calibre id kept none of them once
+        a PDF contributed one ISBN.
+        """
+
+        async def epub(_file):
+            return {
+                "title": "How Linux Works",
+                "identifiers": {"isbn-13": "9781718500419", "asin": "1718500408"},
+            }
+
+        async def pdf(_file):
+            return {"identifiers": {"isbn-13": "9781718500402", "isbn-10": "1593270356"}}
+
+        monkeypatch.setattr(EpubExtractor, "extract_metadata", epub)
+        monkeypatch.setattr(PdfExtractor, "extract_metadata", pdf)
+
+        metadata = await Extractor.extract_metadata(
+            [Path("How Linux Works.epub"), Path("How Linux Works.pdf")]
+        )
+
+        assert metadata["identifiers"] == {
+            # Declared by the publisher's toolchain, so it outranks the PDF's, which
+            # was scraped off a copyright page that also prints the print edition's.
+            "isbn-13": "9781718500419",
+            "asin": "1718500408",
+            "isbn-10": "1593270356",
+        }
+
+    async def test_a_second_format_that_finds_nothing_erases_nothing(self) -> None:
+        """The PDF fixture carries no ISBN, so it must leave the EPUB's alone."""
+        metadata = await Extractor.extract_metadata([EPUB, PDF])
+
+        assert metadata["identifiers"] == {"id": "http://www.gutenberg.org/5200"}
+
+    async def test_one_format_on_its_own_is_unaffected(self) -> None:
+        metadata = await Extractor.extract_metadata([EPUB])
+
+        assert metadata["identifiers"] == {"id": "http://www.gutenberg.org/5200"}
+
+    async def test_no_identifiers_anywhere_leaves_the_field_absent(self) -> None:
+        """An empty dict would count as extracted metadata and overwrite nothing."""
+        metadata = await Extractor.extract_metadata([PDF])
+
+        assert "identifiers" not in metadata
+
+
+class TestSplitEdition:
+    """An edition is a field on the book, not part of what the book is called."""
+
+    @pytest.mark.parametrize(
+        ("title", "stripped", "edition"),
+        [
+            # Every form below is one that turned up in a real library.
+            ("Fluent Python, 2nd Edition", "Fluent Python", 2),
+            ("Building Microservices, 2E", "Building Microservices", 2),
+            ("Digital Image Processing, 4e", "Digital Image Processing", 4),
+            (
+                "Network Security Essentials: Applications and Standards/6e",
+                "Network Security Essentials: Applications and Standards",
+                6,
+            ),
+            (
+                "Refactoring: Improving the Design of Existing Code (2nd edition)",
+                "Refactoring: Improving the Design of Existing Code",
+                2,
+            ),
+            # Ordinal words, including a qualifier sitting inside the statement.
+            (
+                "The Art of Computer Programming: Volume 1 / Fundamental Algorithms, Third Edition",
+                "The Art of Computer Programming: Volume 1 / Fundamental Algorithms",
+                3,
+            ),
+            (
+                "Introduction to the Theory of Computation, Third International Edition",
+                "Introduction to the Theory of Computation",
+                3,
+            ),
+            # Mid-title, before a subtitle and before a trailing author.
+            (
+                "How Linux Works, 3rd Edition: What Every Superuser Should Know",
+                "How Linux Works: What Every Superuser Should Know",
+                3,
+            ),
+            (
+                "Code Complete, 2nd Edition - Steve McConnell",
+                "Code Complete - Steve McConnell",
+                2,
+            ),
+            # An underscore between the number and the "e", beside an unnumbered
+            # qualifier that has nowhere to go in an integer column and so stays put.
+            (
+                "Cryptography and Network Security, Global Edition, 8_e - Stallings",
+                "Cryptography and Network Security, Global Edition - Stallings",
+                8,
+            ),
+        ],
+    )
+    def test_editions_are_split_out(self, title: str, stripped: str, edition: int) -> None:
+        assert split_edition(title) == (stripped, edition)
+
+    @pytest.mark.parametrize(
+        "title",
+        [
+            # A number alone is never an edition — these are titles.
+            "Catch 22",
+            "Fahrenheit 451",
+            "Blade Runner 2049",
+            "1984",
+            "Apollo 13",
+            "Slaughterhouse 5",
+            "The Art of Computer Programming: Volume 1",
+            # "Edition" with no number cannot be stored, so it stays where it can
+            # still be read.
+            "Cryptography and Network Security: Principles and Practice, Global Edition",
+            "Building Microservices",
+        ],
+    )
+    def test_titles_are_left_alone(self, title: str) -> None:
+        assert split_edition(title) == (title, None)
+
+    def test_a_title_that_is_only_an_edition_is_kept(self) -> None:
+        """Stripping must never leave a book with no title at all."""
+        assert split_edition("2nd Edition") == ("2nd Edition", None)
+
+    @pytest.mark.parametrize("title", ["", None])
+    def test_nothing_yields_nothing(self, title: str | None) -> None:
+        assert split_edition(title) == (title, None)
+
+
+@pytest.mark.asyncio()
+class TestEditionFromFiles:
+    async def test_extraction_moves_the_edition_off_the_title(self) -> None:
+        """The PDF fixture calls itself a 2nd edition in its own metadata title."""
+        metadata = await Extractor.extract_metadata([PDF])
+
+        assert metadata["title"] == "The Project Gutenberg eBook #33283: Calculus Made Easy"
+        assert metadata["edition"] == 2
+
+
+class TestEpubPublisher:
+    """The publisher was looked up and then dropped on the floor."""
+
+    def test_a_declared_publisher_is_returned(self) -> None:
+        """
+        The lookup discarded its own result and fell off the end of the function, so
+        every EPUB reported no publisher no matter what it said.
+        """
+        book = epub.EpubBook()
+        book.add_metadata("DC", "publisher", "No Starch Press")
+
+        assert EpubExtractor._extract_publisher(book) == "No Starch Press"
+
+    def test_no_publisher_is_none(self) -> None:
+        assert EpubExtractor._extract_publisher(epub.EpubBook()) is None

docs/duplicates-in-library-settings.md

@@ -0,0 +1,146 @@
+# Implementation brief: move Duplicates into library settings
+
+Written for an agent picking this up cold. Read the repo-root `AGENTS.md` and
+`frontend/AGENTS.md` first — this brief assumes both.
+
+**This is a frontend-only change.** The backend already scopes everything by library
+(`GET /books/duplicate-books?library_id=`), so no endpoint, schema or migration is
+involved.
+
+## Where this starts from
+
+The duplicates review screen exists and works. It currently lives at
+
+```
+frontend/src/routes/(root)/(library)/library/[libraryId]/duplicates/
+    +page.server.ts     loads the groups, plus the full Book records merge needs
+    +page.svelte        group cards, "Not duplicates", "Merge…"
+```
+
+and is reached from a **Duplicates entry in the main sidebar**
+(`frontend/src/lib/components/layout/nav-main.svelte`), which is what this change
+removes.
+
+Settings today is a flat, entirely global four-item nav
+(`frontend/src/routes/(root)/settings/+layout.svelte`): Account, Appearance, Libraries,
+Devices. `settings/libraries/+page.svelte` is a single table of every library whose rows
+link *out* to the library itself. **There is nowhere that means "settings for this
+library"** — that is the gap this change fills.
+
+## What to build — option B
+
+Libraries expands in the settings nav. Every library is a sub-item; selecting one swaps
+the pane; Duplicates is a section inside that pane. All libraries and all their sections
+end up one click apart.
+
+```
+/settings/libraries                        the existing table (leave it as the index)
+/settings/libraries/[libraryId]            redirects to the first section
+/settings/libraries/[libraryId]/duplicates the review screen, moved
+```
+
+Suggested files:
+
+| Path | What |
+| --- | --- |
+| `settings/libraries/[libraryId]/+layout.svelte` | Library name, and the section tabs |
+| `settings/libraries/[libraryId]/+page.ts` | `redirect(303, …/duplicates)` |
+| `settings/libraries/[libraryId]/duplicates/+page.server.ts` | Moved verbatim |
+| `settings/libraries/[libraryId]/duplicates/+page.svelte` | Moved verbatim |
+
+Duplicates is the **only** real section today. Build the tab strip so General and Danger
+zone have somewhere obvious to land, but do not invent them now — an empty tab is worse
+than no tab.
+
+## The nav
+
+In `settings/+layout.svelte`, `items` is a flat `as const` array matched on
+`page.route.id`. Libraries needs to render its children beneath it:
+
+```svelte
+{#each libraryState.libraries as library (library.id)}
+  <a href={resolve('/(root)/settings/libraries/[libraryId]/duplicates', {
+        libraryId: String(library.id) })}> … </a>
+{/each}
+```
+
+`getLibraryState()` **is** available under `/settings` — it is set in
+`(root)/+layout.svelte`, above the settings group, and `settings/libraries/+page.svelte`
+already uses it. No new load function is needed to list the libraries.
+
+**Active state is matched on route id, not pathname.** There is a comment in
+`settings/+layout.svelte` explaining why: `resolve()` returns an absolute path on the
+client and a relative one during SSR, so a pathname comparison is false on the server and
+true after hydration, and the highlight flashes in. A nested library item is active when
+the route id matches **and** `page.params.libraryId === String(library.id)` — both, or
+every library lights up at once.
+
+## Things that will bite
+
+1. **Remove the sidebar entry in the same change.** `nav-main.svelte` gained a
+   `Duplicates` item and a `CopyCheck` import when the screen was built. Delete both, and
+   delete the old route directory. Doing the removal and the move together is the point —
+   split across two commits the screen is unreachable in between.
+
+2. **Delete the old route, do not leave it.** Two live copies of a screen that both write
+   is how they drift.
+
+3. **`setBookSelectionState` is not available under `/settings`.** It is set in
+   `(root)/(library)/+layout.svelte`, which the settings group is not inside. This is
+   fine — the duplicates page uses `BookImage` directly, not `book-thumbnail.svelte`, and
+   `MergeBooks` takes its `libraryId` as a prop. **Verify this stays true** if you touch
+   either component; a `getBookSelectionState()` under settings returns `undefined` and
+   fails at the first access, not at import.
+
+4. **Keep `depends('app:duplicate-books')`.** Both the dismiss action and `MergeBooks`
+   call `invalidate('app:duplicate-books')` to make a resolved group leave the screen.
+   Drop it and the page silently stops refreshing. `MergeBooks` also invalidates
+   `app:books`, which is a no-op under settings and should stay that way.
+
+5. **The settings shell is height-constrained.** `settings/+layout.svelte` is
+   `h-[calc(100vh-var(--header-height)-2rem)]` with `overflow-auto` on the content pane.
+   The review screen is a long list of cards — it must scroll *inside* that pane. Its
+   current `mx-auto max-w-5xl` wrapper will want revisiting.
+
+6. **Three levels of nav is option B's known cost.** Nav → library → section, and the
+   pane is narrower than the full-width route the screen was designed against. The group
+   cards are `w-36` covers in a wrapping flex row, so they reflow, but check a group of
+   four at a narrow window before calling it done.
+
+7. **`resolve()` must be a direct call in markup** for `svelte/no-navigation-without-resolve`.
+   Where `nav-main.svelte` computes a url through a variable it carries an
+   `eslint-disable-next-line`; prefer the direct call over inheriting that.
+
+8. **The loader depends on the `?ids=` fix.** `+page.server.ts` fetches full `Book`
+   records with `listBooks({ ids, pageSize })` because the merge workbench needs
+   identifiers, description and publisher, which `DuplicateBookRead` does not carry.
+   advanced_alchemy's stock id filter types that parameter as `list[str]` regardless of
+   config, which made Postgres refuse `bigint = character varying`; the override lives in
+   `backend/src/chitai/services/dependencies.py` (`create_book_filter_dependencies`).
+   If `GET /books?ids=1&ids=2` 500s, that override is missing — do not work around it in
+   the loader.
+
+## Out of scope
+
+The **General** and **Danger zone** sections (rename, path template, read-only, consume
+directory, delete), and any change to the merge workbench itself. The toolbar entry point
+for merge — select 2+ books in the library view — is unrelated and stays where it is.
+
+## Verification
+
+```bash
+cd frontend
+pnpm check     # baseline: 30 errors, 1 warning, 8 files — none of them yours
+pnpm lint      # not clean either; check the files you touched, not the tree
+pnpm build
+```
+
+By hand, with a library that has a duplicate group:
+
+- Settings → Libraries lists every library beneath it; clicking one opens its pane.
+- Duplicates shows the same groups the old route did, and scrolls inside the settings pane.
+- **Not duplicates** removes the group and it stays gone after a reload.
+- **Merge…** opens the workbench, merges, and the group leaves the screen.
+- The main sidebar no longer has a Duplicates entry, and
+  `/library/<id>/duplicates` no longer resolves.
+- A library with no duplicates shows the empty state, not a blank pane.

frontend/AGENTS.md

@@ -157,10 +157,12 @@ but take a baseline first, because neither is clean (see below).
 
 Observed in the current tree — don't mistake these for intentional patterns to copy:
 
-- `src/lib/schema/openapi/schema.d.ts` is **stale** — it predates the `BookProgress` rework and has
-  no `BookProgressRead`, so `book.progress` types as `{}`. It is the source of most of the errors
-  `pnpm check` reports; regenerating it should clear them. **Baseline as of 2026-08-11: 37 errors,
-  1 warning, 11 files.** Get your own baseline before assuming an error is yours.
+- `pnpm check` is not clean. **Baseline as of 2026-08-13: 30 errors, 1 warning, 8 files**, most of
+  them in `src/routes/api/[...path]/+server.ts` (see below). Get your own baseline before assuming
+  an error is yours. `src/lib/schema/openapi/schema.d.ts` was regenerated on that date and is
+  current; regenerate it again after any backend API change, with
+  `pnpm exec openapi-typescript http://localhost:8000/schema/openapi.json -o src/lib/schema/openapi/schema.d.ts`
+  against a backend running **your** branch — a stale server silently writes a stale file.
 - `pnpm lint` does not pass either — 131 pre-existing ESLint errors, 35 of them
   `svelte/no-navigation-without-resolve` on plain `href`s, plus ~59 files Prettier would rewrite
   (mostly vendored shadcn components). Check the files you touched, not the whole tree.

frontend/src/lib/api/book.remote.ts

@@ -8,12 +8,31 @@ import {
 	deleteBooksSchema,
 	editBookMetadataSchema,
 	updateBookProgressSchema,
+	duplicateDismissalSchema,
+	bookMergeSchema,
 	type Book,
+	type BooksUploadResult,
+	type DuplicateBookGroup,
 	bookFilesUpload
 } from '$lib/schema/index';
 import { stringCoerce, type PaginatedResponse } from '$lib/schema/common';
 import { createQueryParams } from '$lib/utils';
 
+/**
+ * The backend's own message for a failed response, rather than its JSON envelope.
+ *
+ * A refused duplicate answers 409 with a `detail` worth reading and the offending
+ * files in `extra`; passing the body through whole puts JSON in front of the reader.
+ */
+function detailOf(body: string): string {
+	try {
+		const parsed = JSON.parse(body);
+		return typeof parsed?.detail === 'string' ? parsed.detail : body;
+	} catch {
+		return body;
+	}
+}
+
 export const getBook = query(stringCoerce, async (id): Promise<Book> => {
 	const { locals } = getRequestEvent();
 
@@ -68,26 +87,29 @@ export const updateBookCover = form(bookCoverUpload, async ({ book_id, file }) =
 	return await response.json();
 });
 
-export const uploadBooks = form(booksUpload, async ({ library_id, files }) => {
-	const { locals } = getRequestEvent();
+export const uploadBooks = form(
+	booksUpload,
+	async ({ library_id, files }): Promise<BooksUploadResult> => {
+		const { locals } = getRequestEvent();
 
-	const formData = new FormData();
-	files.forEach((file) => {
-		formData.append('files', file);
-	});
+		const formData = new FormData();
+		files.forEach((file) => {
+			formData.append('files', file);
+		});
 
-	const response = await locals.api.postMultipart(
-		`/books/fromFiles?library_id=${library_id}`,
-		formData
-	);
+		const response = await locals.api.postMultipart(
+			`/books/fromFiles?library_id=${library_id}`,
+			formData
+		);
 
-	if (!response.ok) {
-		const message = await response.text();
-		error(response.status, message);
+		if (!response.ok) {
+			const message = await response.text();
+			error(response.status, message);
+		}
+
+		return await response.json();
 	}
-
-	return await response.json();
-});
+);
 
 export const uploadBookFiles = form(bookFilesUpload, async ({ book_id, files }) => {
 	const { locals } = getRequestEvent();
@@ -100,8 +122,9 @@ export const uploadBookFiles = form(bookFilesUpload, async ({ book_id, files })
 	const response = await locals.api.postMultipart(`/books/${book_id}/files`, formData);
 
 	if (!response.ok) {
-		const message = await response.text();
-		error(response.status, message);
+		// 409 here means the file is already stored under a different book, which is
+		// something the reader can act on — so the message has to survive the trip.
+		error(response.status, detailOf(await response.text()));
 	}
 
 	return await response.json();
@@ -133,6 +156,65 @@ export const deleteBookFiles = command(deleteBookFilesSchema, async ({ book_id,
 	}
 });
 
+/**
+ * Books already in the library that look like copies of one another.
+ *
+ * Metadata only, so every group is a question rather than a verdict — which is why
+ * the screen it feeds offers a way to disagree.
+ */
+export const listDuplicateBooks = query(
+	stringCoerce,
+	async (libraryId): Promise<DuplicateBookGroup[]> => {
+		const { locals } = getRequestEvent();
+
+		const response = await locals.api.get(`/books/duplicate-books?library_id=${libraryId}`);
+
+		if (!response.ok) error(response.status, detailOf(await response.text()));
+
+		return await response.json();
+	}
+);
+
+/**
+ * Fold several books into one and delete the records folded in.
+ *
+ * Irreversible, so the caller is expected to have shown what is about to happen.
+ */
+export const mergeBooks = command(
+	bookMergeSchema,
+	async ({ library_id, ...data }): Promise<Book> => {
+		const { locals } = getRequestEvent();
+
+		const response = await locals.api.post(`/books/merge?library_id=${library_id}`, data);
+
+		if (!response.ok) error(response.status, detailOf(await response.text()));
+
+		return await response.json();
+	}
+);
+
+/** Record that two books are not the same book, so the pair stops being proposed. */
+export const dismissDuplicateBooks = command(duplicateDismissalSchema, async (data) => {
+	const { locals } = getRequestEvent();
+
+	const response = await locals.api.post('/books/duplicate-books/dismissals', data);
+
+	if (!response.ok) error(response.status, detailOf(await response.text()));
+});
+
+/** Undo a dismissal, so the pair is proposed again. */
+export const restoreDuplicateBooks = command(duplicateDismissalSchema, async (data) => {
+	const { locals } = getRequestEvent();
+
+	const params = createQueryParams(data);
+
+	const response = await locals.api.delete(
+		`/books/duplicate-books/dismissals?${params.toString()}`
+	);
+
+	if (!response.ok) error(response.status, detailOf(await response.text()));
+});
+
 export const updateBookProgress = command(
 	updateBookProgressSchema,
 	async ({ book_ids, ...data }) => {

frontend/src/lib/components/forms/edit-book/edit-files.svelte

@@ -50,6 +50,22 @@
 		toast.error(`${file.name} was not added`, { description: reason });
 	};
 
+	/**
+	 * The API's own words, when it has any.
+	 *
+	 * Adding a file the library already holds under another book is refused with a
+	 * 409 naming it — far more use than "failed to add files". SvelteKit hands an
+	 * `error()` back as an HttpError on the client, so the message sits on `body`.
+	 */
+	function apiMessage(error: unknown): string | undefined {
+		if (typeof error !== 'object' || error === null) return undefined;
+
+		const body = (error as { body?: { message?: string } }).body;
+		if (typeof body?.message === 'string') return body.message;
+
+		return error instanceof Error ? error.message : undefined;
+	}
+
 	function confirmDelete(file: BookFile) {
 		fileToDelete = file;
 		confirmOpen = true;
@@ -148,7 +164,8 @@
 				toast.success('Files added');
 			} catch (error) {
 				console.error('Failed to add files: ', error);
-				toast.error('Failed to add files');
+				toast.error(apiMessage(error) ?? 'Failed to add files');
+				uploadBookFiles.fields.files.set([]);
 			}
 		})}
 		enctype="multipart/form-data"

frontend/src/lib/components/forms/merge-books/field-spec.ts

@@ -0,0 +1,149 @@
+import type { Book } from '$lib/schema';
+
+/**
+ * What kind of thing a metadata field holds, which is the only thing that decides
+ * what you can do with it when two records disagree.
+ */
+export type FieldKind = 'text' | 'number' | 'date' | 'list' | 'keyed' | 'longtext';
+
+/** An action offered on a field, beyond replacing it outright. */
+export type FieldAction = 'replace' | 'merge' | 'append';
+
+export interface FieldSpec {
+	/** The key sent in `BookMetadataUpdate`. */
+	key: string;
+	label: string;
+	kind: FieldKind;
+	group: string;
+	/** Extra actions past `replace`, which every field has. */
+	extra: FieldAction[];
+}
+
+/**
+ * The fields a merge can resolve, in the order and grouping the edit form uses.
+ *
+ * Deliberately a spec rather than markup: the merge workbench and, later, the
+ * provider review screen both render from this, so a field cannot exist in one and
+ * not the other. `cover` and `files` are absent because they are not choices —
+ * files always come across and the cover has its own endpoint.
+ */
+export const MERGE_FIELDS: FieldSpec[] = [
+	{ key: 'title', label: 'Title', kind: 'text', group: 'Identity', extra: [] },
+	{ key: 'subtitle', label: 'Subtitle', kind: 'text', group: 'Identity', extra: [] },
+	{ key: 'edition', label: 'Edition', kind: 'number', group: 'Identity', extra: [] },
+	{ key: 'series', label: 'Series', kind: 'text', group: 'Identity', extra: [] },
+	{ key: 'series_position', label: 'No.', kind: 'text', group: 'Identity', extra: [] },
+	{ key: 'language', label: 'Language', kind: 'text', group: 'Identity', extra: [] },
+
+	// Order is meaningful for authors, so the second list is appended rather than
+	// interleaved; tags are a set, so they merge.
+	{
+		key: 'authors',
+		label: 'Authors',
+		kind: 'list',
+		group: 'People and subjects',
+		extra: ['append']
+	},
+	{ key: 'tags', label: 'Tags', kind: 'list', group: 'People and subjects', extra: ['merge'] },
+
+	{ key: 'publisher', label: 'Publisher', kind: 'text', group: 'Publication', extra: [] },
+	{ key: 'published_date', label: 'Published', kind: 'date', group: 'Publication', extra: [] },
+	{ key: 'pages', label: 'Pages', kind: 'number', group: 'Publication', extra: [] },
+	{
+		key: 'identifiers',
+		label: 'Identifiers',
+		kind: 'keyed',
+		group: 'Publication',
+		extra: ['merge']
+	},
+
+	{
+		key: 'description',
+		label: 'Description',
+		kind: 'longtext',
+		group: 'Description',
+		extra: ['append']
+	}
+];
+
+/** A field's value, in the shape `BookMetadataUpdate` expects to receive it. */
+export type FieldValue = string | number | string[] | Record<string, string> | null;
+
+/** Read one field off a book, flattening the relations the API returns as objects. */
+export function readField(book: Book, key: string): FieldValue {
+	switch (key) {
+		case 'authors':
+			return book.authors.map((author) => author.name);
+		case 'tags':
+			return book.tags.map((tag) => tag.name);
+		case 'publisher':
+			return book.publisher?.name ?? null;
+		case 'series':
+			return book.series?.title ?? null;
+		case 'identifiers':
+			return book.identifiers ?? {};
+		default:
+			return (book as unknown as Record<string, FieldValue>)[key] ?? null;
+	}
+}
+
+/** Whether a field holds nothing, and so has no decision attached to it. */
+export function isEmpty(value: FieldValue): boolean {
+	if (value === null || value === undefined || value === '') return true;
+	if (Array.isArray(value)) return value.length === 0;
+	if (typeof value === 'object') return Object.keys(value).length === 0;
+	return false;
+}
+
+/** Whether two field values say the same thing, order included for lists. */
+export function isSame(left: FieldValue, right: FieldValue): boolean {
+	if (isEmpty(left) && isEmpty(right)) return true;
+	return JSON.stringify(left ?? null) === JSON.stringify(right ?? null);
+}
+
+/**
+ * Apply an action to a pair of values and return what the target becomes.
+ *
+ * `merge` on a keyed collection is per name and the target wins a clash, because
+ * `Identifier` is unique on `(name, book_id)` — a book cannot hold both its print
+ * and its ebook ISBN, so the second one has nowhere to go.
+ */
+export function applyAction(
+	action: FieldAction,
+	kind: FieldKind,
+	target: FieldValue,
+	incoming: FieldValue
+): FieldValue {
+	if (action === 'replace') return incoming;
+
+	if (kind === 'list') {
+		const current = Array.isArray(target) ? target : [];
+		const extra = Array.isArray(incoming) ? incoming : [];
+		// Order preserved, duplicates dropped — works for both append and merge.
+		return [...new Set([...current, ...extra])];
+	}
+
+	if (kind === 'keyed') {
+		return { ...(incoming as Record<string, string>), ...(target as Record<string, string>) };
+	}
+
+	if (kind === 'longtext') {
+		const current = typeof target === 'string' ? target.trim() : '';
+		const extra = typeof incoming === 'string' ? incoming.trim() : '';
+		return [current, extra].filter(Boolean).join('\n\n');
+	}
+
+	return incoming;
+}
+
+/** How a value reads in the inert reference column. */
+export function displayValue(value: FieldValue): string {
+	if (isEmpty(value)) return '';
+	if (Array.isArray(value)) return value.join(', ');
+	if (typeof value === 'object') {
+		return Object.entries(value as Record<string, string>)
+			.map(([name, id]) => `${name}: ${id}`)
+			.join('  ·  ');
+	}
+	return String(value);
+}

frontend/src/lib/components/forms/merge-books/merge-books.svelte

@@ -0,0 +1,394 @@
+<script lang="ts">
+	import { untrack } from 'svelte';
+	import { ArrowRight, GitMerge, Plus, Undo2 } from '@lucide/svelte';
+
+	import * as Dialog from '$lib/components/ui/dialog/index.js';
+	import { Button } from '$lib/components/ui/button/index.js';
+	import { Input } from '$lib/components/ui/input/index.js';
+	import { Textarea } from '$lib/components/ui/textarea/index.js';
+	import { Badge } from '$lib/components/ui/badge/index.js';
+	import BookImage from '$lib/components/view/book-image.svelte';
+	import GeneratedCover from '$lib/components/view/generated-cover.svelte';
+	import type { Book } from '$lib/schema';
+
+	import { mergeInto } from './merge';
+
+	import {
+		MERGE_FIELDS,
+		applyAction,
+		displayValue,
+		isEmpty,
+		isSame,
+		readField,
+		type FieldAction,
+		type FieldSpec,
+		type FieldValue
+	} from './field-spec';
+
+	let {
+		books,
+		libraryId,
+		open = $bindable(),
+		onmerged
+	}: {
+		books: Book[];
+		libraryId: number | string;
+		open: boolean;
+		/** Given the record that survived and the ones folded into it and deleted. */
+		onmerged?: (survivor: Book, folded: Book[]) => void;
+	} = $props();
+
+	// Seeded once per mount. The dialog is keyed on the group upstream, so a
+	// different group gets a fresh workbench rather than the previous one's draft.
+	let survivorId = $state(untrack(() => books[0]?.id));
+	let candidateId = $state(untrack(() => books[1]?.id));
+	let draft = $state<Record<string, FieldValue>>({});
+	let busy = $state(false);
+
+	const survivor = $derived(books.find((book) => book.id === survivorId) ?? books[0]);
+	const candidates = $derived(books.filter((book) => book.id !== survivorId));
+	/** The records that will be deleted — the same set, named for what happens to them. */
+	const folded = $derived(candidates);
+	const candidate = $derived(candidates.find((book) => book.id === candidateId) ?? candidates[0]);
+
+	/** The survivor's stored value for a field, or the draft if it has been touched. */
+	function current(field: FieldSpec): FieldValue {
+		return field.key in draft ? draft[field.key] : readField(survivor, field.key);
+	}
+
+	function take(field: FieldSpec, action: FieldAction) {
+		draft[field.key] = applyAction(
+			action,
+			field.kind,
+			current(field),
+			readField(candidate, field.key)
+		);
+	}
+
+	function undo(field: FieldSpec) {
+		delete draft[field.key];
+		draft = { ...draft };
+	}
+
+	/**
+	 * Fill only the fields the survivor has nothing in.
+	 *
+	 * The safe bulk action, and the one worth reaching for: it cannot overwrite a
+	 * value, so it needs no per-field protection to be pressed without reading.
+	 */
+	function fillEmpty() {
+		for (const field of MERGE_FIELDS) {
+			const incoming = readField(candidate, field.key);
+			if (isEmpty(current(field)) && !isEmpty(incoming)) draft[field.key] = incoming;
+		}
+	}
+
+	const changed = $derived(Object.keys(draft));
+
+	const differing = $derived(
+		candidate
+			? MERGE_FIELDS.filter((field) => !isSame(current(field), readField(candidate, field.key)))
+			: []
+	);
+
+	/** Fields shown as a row: anything the two disagree on, plus anything edited. */
+	const shown = $derived(
+		MERGE_FIELDS.filter((field) => differing.includes(field) || field.key in draft)
+	);
+
+	const agreed = $derived(MERGE_FIELDS.filter((field) => !shown.includes(field)));
+
+	const groups = $derived([...new Set(shown.map((field) => field.group))]);
+
+	async function submit() {
+		if (busy) return;
+		busy = true;
+
+		const merged = await mergeInto(libraryId, survivor, folded, changed.length ? draft : undefined);
+
+		busy = false;
+
+		if (merged) {
+			open = false;
+			onmerged?.(survivor, folded);
+		}
+	}
+</script>
+
+{#snippet cover(book: Book, size: string)}
+	<span class="{size} shrink-0 overflow-hidden rounded-sm border bg-muted">
+		{#if book.cover_image}
+			<BookImage src="/api/{book.cover_image}" class="h-full w-full object-cover" />
+		{:else}
+			<!-- Drawn rather than left blank, so the rail tells two coverless books
+			     apart the same way the shelves do. -->
+			<GeneratedCover {book} />
+		{/if}
+	</span>
+{/snippet}
+
+<!-- The inert reference column: the same shape as the control opposite it, with
+     nothing that invites a click. -->
+{#snippet reference(field: FieldSpec, book: Book)}
+	{@const value = readField(book, field.key)}
+	<div class="flex min-w-0 flex-col gap-1">
+		<span class="text-[11px] text-muted-foreground">{field.label}</span>
+		{#if isEmpty(value)}
+			<span class="min-h-8 py-1 text-sm text-muted-foreground italic">empty</span>
+		{:else if field.kind === 'list'}
+			<span class="flex min-h-8 flex-wrap items-center gap-1 py-0.5">
+				{#each value as string[] as item (item)}
+					<Badge variant="secondary" class="font-normal">{item}</Badge>
+				{/each}
+			</span>
+		{:else}
+			<span class="min-h-8 py-1 text-sm break-words">{displayValue(value)}</span>
+		{/if}
+	</div>
+{/snippet}
+
+<Dialog.Root bind:open>
+	<Dialog.Content
+		class="flex h-[85vh] max-w-[min(72rem,95vw)] flex-col gap-0 overflow-hidden p-0 sm:max-w-[min(72rem,95vw)]"
+	>
+		<Dialog.Header class="shrink-0 space-y-0 border-b px-5 py-3 text-left">
+			<Dialog.Title class="font-serif text-base font-normal">
+				Merge {books.length} books
+			</Dialog.Title>
+			<Dialog.Description class="text-xs">
+				{candidates.length}
+				{candidates.length === 1 ? 'record is' : 'records are'} deleted. Their files move onto the book
+				you keep — nothing is removed from disk.
+			</Dialog.Description>
+		</Dialog.Header>
+
+		<div class="grid min-h-0 flex-1 grid-cols-1 md:grid-cols-[15rem_1fr]">
+			<!-- Rail: the books being folded in, one open at a time -->
+			<aside
+				class="flex min-w-0 flex-col overflow-y-auto border-b bg-sidebar md:border-r md:border-b-0"
+			>
+				<p
+					class="px-3 pt-3 pb-1 font-mono text-[10px] tracking-widest text-muted-foreground uppercase"
+				>
+					Taking from · {candidates.length}
+				</p>
+				{#each candidates as book (book.id)}
+					{@const count = MERGE_FIELDS.filter(
+						(field) => !isSame(readField(survivor, field.key), readField(book, field.key))
+					).length}
+					<button
+						type="button"
+						onclick={() => (candidateId = book.id)}
+						class="flex items-start gap-2 border-l-2 px-3 py-2 text-left transition-colors hover:bg-muted/50 {book.id ===
+						candidate?.id
+							? 'border-l-primary bg-background'
+							: 'border-l-transparent'}"
+					>
+						{@render cover(book, 'h-10 w-7')}
+						<span class="min-w-0 flex-1">
+							<span class="line-clamp-2 font-serif text-[13px]">{book.title}</span>
+							<span class="block text-[10px] text-muted-foreground">
+								#{book.id} · {book.files.length}
+								{book.files.length === 1 ? 'file' : 'files'}
+							</span>
+							{#if count === 0}
+								<Badge variant="secondary" class="mt-1 text-[10px] font-normal">
+									nothing to take
+								</Badge>
+							{/if}
+						</span>
+					</button>
+				{/each}
+			</aside>
+
+			<div class="flex min-h-0 flex-col">
+				<!-- Which record survives -->
+				<div class="flex shrink-0 flex-wrap items-center gap-2 border-b px-5 py-2.5">
+					<span class="text-[11px] text-muted-foreground">Keeping</span>
+					{#each books as book (book.id)}
+						<button
+							type="button"
+							onclick={() => (survivorId = book.id)}
+							class="flex items-center gap-1.5 rounded-md border px-2 py-1 text-xs transition-colors {book.id ===
+							survivorId
+								? 'border-primary bg-accent text-accent-foreground'
+								: 'text-muted-foreground hover:bg-muted'}"
+						>
+							{@render cover(book, 'h-6 w-4')}
+							<span class="max-w-32 truncate">#{book.id}</span>
+						</button>
+					{/each}
+
+					<div class="ml-auto flex gap-2">
+						<Button variant="outline" size="sm" onclick={fillEmpty} disabled={!candidate}>
+							Fill empty fields
+						</Button>
+					</div>
+				</div>
+
+				<div class="min-h-0 flex-1 overflow-y-auto px-5 py-4">
+					{#if !candidate}
+						<p class="text-sm text-muted-foreground">Nothing left to fold in.</p>
+					{:else if shown.length === 0}
+						<p class="text-sm text-muted-foreground">
+							These records agree on every field. Merging keeps
+							<span class="font-medium text-foreground">#{survivor.id}</span> and moves the others' files
+							onto it.
+						</p>
+					{:else}
+						{#each groups as group (group)}
+							<h3
+								class="mt-5 border-b pb-1 font-mono text-[10px] tracking-widest text-primary uppercase first:mt-0"
+							>
+								{group}
+							</h3>
+
+							{#each shown.filter((field) => field.group === group) as field (field.key)}
+								{@const value = current(field)}
+								{@const incoming = readField(candidate, field.key)}
+								{@const edited = field.key in draft}
+								<div class="grid grid-cols-[1fr_auto_1fr] items-center gap-3 border-b py-2">
+									{@render reference(field, candidate)}
+
+									<!-- Actions, on the row rather than stacked beside it -->
+									<div class="flex items-center gap-1">
+										{#if edited}
+											<Button
+												variant="ghost"
+												size="icon"
+												class="size-7"
+												title="Undo"
+												onclick={() => undo(field)}
+											>
+												<Undo2 class="size-3.5" />
+												<span class="sr-only">Undo {field.label}</span>
+											</Button>
+										{/if}
+
+										{#if !isSame(value, incoming)}
+											{#if isEmpty(value)}
+												<!-- Nothing to weigh, so it is an offer rather than a choice -->
+												<Button
+													size="sm"
+													class="h-7 gap-1 px-2 text-[11px]"
+													onclick={() => take(field, 'replace')}
+												>
+													<ArrowRight class="size-3.5" />
+													Take
+												</Button>
+											{:else}
+												<Button
+													variant="outline"
+													size="icon"
+													class="size-7"
+													title="Replace"
+													onclick={() => take(field, 'replace')}
+												>
+													<ArrowRight class="size-3.5" />
+													<span class="sr-only">Replace {field.label}</span>
+												</Button>
+												{#each field.extra as action (action)}
+													<Button
+														variant="outline"
+														size="icon"
+														class="size-7"
+														title={action === 'merge' ? 'Merge' : 'Append'}
+														onclick={() => take(field, action)}
+													>
+														{#if action === 'merge'}
+															<GitMerge class="size-3.5" />
+														{:else}
+															<Plus class="size-3.5" />
+														{/if}
+														<span class="sr-only">{action} {field.label}</span>
+													</Button>
+												{/each}
+											{/if}
+										{/if}
+									</div>
+
+									<!-- The survivor's side: the edit form, live -->
+									<div class="flex min-w-0 flex-col gap-1">
+										<span class="flex items-center gap-1.5 text-[11px] text-muted-foreground">
+											{field.label}
+											{#if edited}
+												<Badge class="h-4 px-1.5 text-[9px] font-semibold">taken</Badge>
+											{/if}
+										</span>
+
+										{#if field.kind === 'longtext'}
+											<Textarea
+												rows={4}
+												class="text-sm"
+												value={(value as string) ?? ''}
+												oninput={(event) => (draft[field.key] = event.currentTarget.value)}
+											/>
+										{:else if field.kind === 'list' || field.kind === 'keyed'}
+											<!-- Edited through the take actions; typing here would need the
+											     tags and key/value editors, which belong to the edit form. -->
+											<div
+												class="flex min-h-8 flex-wrap items-center gap-1 rounded-md border bg-background px-2 py-1"
+											>
+												{#if isEmpty(value)}
+													<span class="text-sm text-muted-foreground">—</span>
+												{:else if field.kind === 'list'}
+													{#each value as string[] as item (item)}
+														<Badge variant="secondary" class="font-normal">{item}</Badge>
+													{/each}
+												{:else}
+													{#each Object.entries(value as Record<string, string>) as [name, id] (name)}
+														<Badge variant="secondary" class="font-mono text-[10px] font-normal">
+															{name}: {id}
+														</Badge>
+													{/each}
+												{/if}
+											</div>
+										{:else}
+											<Input
+												type={field.kind === 'number'
+													? 'number'
+													: field.kind === 'date'
+														? 'date'
+														: 'text'}
+												class="h-8 text-sm"
+												value={(value as string | number) ?? ''}
+												oninput={(event) =>
+													(draft[field.key] =
+														field.kind === 'number'
+															? Number(event.currentTarget.value) || null
+															: event.currentTarget.value)}
+											/>
+										{/if}
+									</div>
+								</div>
+							{/each}
+						{/each}
+
+						{#if agreed.length > 0}
+							<p class="pt-3 text-center text-[11px] text-muted-foreground italic">
+								{agreed.map((field) => field.label).join(', ')} — identical in both
+							</p>
+						{/if}
+					{/if}
+				</div>
+			</div>
+		</div>
+
+		<Dialog.Footer class="shrink-0 items-center gap-2 border-t px-5 py-3 sm:justify-between">
+			<p class="text-xs text-muted-foreground">
+				{#if changed.length}
+					{changed.length}
+					{changed.length === 1 ? 'change' : 'changes'} pending · this cannot be undone
+				{:else}
+					Metadata is left as #{survivor.id} has it · this cannot be undone
+				{/if}
+			</p>
+			<div class="flex gap-2">
+				<Button variant="outline" onclick={() => (open = false)}>Cancel</Button>
+				<Button onclick={submit} disabled={busy || candidates.length === 0}>
+					Merge into “{survivor.title}”
+				</Button>
+			</div>
+		</Dialog.Footer>
+	</Dialog.Content>
+</Dialog.Root>

frontend/src/lib/components/forms/merge-books/merge.ts

@@ -0,0 +1,44 @@
+import { invalidate } from '$app/navigation';
+import { toast } from 'svelte-sonner';
+
+import { mergeBooks } from '$lib/api/book.remote';
+import type { Book } from '$lib/schema';
+import type { FieldValue } from './field-spec';
+
+/**
+ * Fold `folded` into `survivor`, and tell the reader how it went.
+ *
+ * Shared because a merge is reachable two ways — the workbench, where the reader
+ * resolved the metadata field by field, and the quick action beside it, which
+ * takes the survivor's metadata as it stands. Only the `metadata` argument
+ * differs, and the invalidation and the wording should not.
+ *
+ * @returns whether the merge went through; the caller decides what to close or
+ * clear, and has no toast of its own to write either way.
+ */
+export async function mergeInto(
+	libraryId: number | string,
+	survivor: Book,
+	folded: Book[],
+	metadata?: Record<string, FieldValue>
+): Promise<boolean> {
+	try {
+		await mergeBooks({
+			library_id: libraryId,
+			survivor_id: survivor.id,
+			merged_ids: folded.map((book) => book.id),
+			metadata
+		});
+
+		// Both, because a merge is started from the duplicates review and from the
+		// library's selection toolbar, and each page depends on a different one.
+		await Promise.all([invalidate('app:books'), invalidate('app:duplicate-books')]);
+
+		toast.success(`Merged into “${survivor.title}”`);
+		return true;
+	} catch (error) {
+		console.error('Failed to merge books', error);
+		toast.error(error instanceof Error ? error.message : 'Could not merge these books');
+		return false;
+	}
+}

frontend/src/lib/components/layout/nav-main.svelte

@@ -33,13 +33,15 @@
 			title: 'Home',
 			icon: House,
 			routeId: '/(root)/(library)/library/[libraryId]',
-			path: (id?: number) => resolve('/(root)/(library)/library/[libraryId]', { libraryId: String(id ?? '') })
+			path: (id?: number) =>
+				resolve('/(root)/(library)/library/[libraryId]', { libraryId: String(id ?? '') })
 		},
 		{
 			title: 'Library',
 			icon: LibraryBig,
 			routeId: '/(root)/(library)/library/[libraryId]/view',
-			path: (id?: number) => resolve('/(root)/(library)/library/[libraryId]/view', { libraryId: String(id ?? '') })
+			path: (id?: number) =>
+				resolve('/(root)/(library)/library/[libraryId]/view', { libraryId: String(id ?? '') })
 		},
 		{ title: 'Shelves', icon: Rows3, routeId: null, path: () => '#', shelves: [] }
 	];

frontend/src/lib/components/layout/upload-tray.svelte

@@ -1,11 +1,13 @@
 <script lang="ts">
 	import { fly } from 'svelte/transition';
 	import { prefersReducedMotion } from 'svelte/motion';
-	import { ChevronDown, CircleAlert, RotateCcw, X } from '@lucide/svelte';
+	import { resolve } from '$app/paths';
+	import { ChevronDown, CircleAlert, Copy, RotateCcw, X } from '@lucide/svelte';
 
 	import { Button } from '$lib/components/ui/button/index.js';
 	import { Spinner } from '$lib/components/ui/spinner/index';
 	import { getUploadQueueState } from '$lib/state/upload-queue.svelte';
+	import type { DuplicateBook, DuplicateFile } from '$lib/schema';
 	import { formatFileSize } from '$lib/utils';
 
 	const queue = getUploadQueueState();
@@ -19,10 +21,48 @@
 	const heading = $derived.by(() => {
 		const noun = queue.total === 1 ? 'book' : 'books';
 		if (queue.active) return `Adding ${queue.settled} of ${queue.total} ${noun}`;
-		if (queue.failed === 0) return `Added ${queue.done} ${noun}`;
-		if (queue.done === 0) return `Couldn't add ${queue.failed} ${noun}`;
-		return `Added ${queue.done}, ${queue.failed} failed`;
+		if (queue.failed === 0 && queue.skipped === 0) return `Added ${queue.done} ${noun}`;
+		if (queue.done === 0 && queue.skipped === 0) return `Couldn't add ${queue.failed} ${noun}`;
+
+		const parts = [];
+		if (queue.done > 0) parts.push(`Added ${queue.done}`);
+		if (queue.skipped > 0) parts.push(`${queue.skipped} already here`);
+		if (queue.failed > 0) parts.push(`${queue.failed} failed`);
+		return parts.join(', ');
 	});
+
+	/**
+	 * What the reader needs to know about files that were not stored.
+	 *
+	 * One duplicate can name where its bytes already live; several would be a list,
+	 * and the row has no room for one, so they collapse to a count.
+	 */
+	function duplicateNote(duplicates: DuplicateFile[]) {
+		if (duplicates.length > 1) return `${duplicates.length} files are already in your library`;
+
+		const [only] = duplicates;
+		// No book to name: it matched another file in this same upload.
+		return only.book_title ? `Already in ${only.book_title}` : 'Already added by this upload';
+	}
+
+	/** The book a note can link to, when a single duplicate points at exactly one. */
+	function noteTarget(duplicates: DuplicateFile[]) {
+		return duplicates.length === 1 ? duplicates[0].book_id : null;
+	}
+
+	/**
+	 * What the reader needs to know about a book that went in and may already be here.
+	 *
+	 * Worded weaker than the file-level "Already in …" on purpose. That one means the
+	 * library holds these exact bytes; this one means the metadata agrees, which a
+	 * second edition, a translation and a re-scan all do. Nothing was refused.
+	 */
+	function possibleNote(candidates: DuplicateBook[]) {
+		if (candidates.length > 1)
+			return `Might already be in your library, ${candidates.length} times`;
+
+		return `Might already be in your library — ${candidates[0].title}`;
+	}
 </script>
 
 {#if queue.total > 0}
@@ -44,6 +84,9 @@
 				<Spinner class="size-4 shrink-0" />
 			{:else if queue.failed > 0}
 				<CircleAlert class="size-4 shrink-0 text-destructive" />
+			{:else if queue.skipped > 0}
+				<!-- Muted, not alarming: nothing went wrong, the books were already here. -->
+				<Copy class="size-4 shrink-0 text-muted-foreground" />
 			{/if}
 
 			<span class="min-w-0 flex-1 truncate text-sm font-medium">{heading}</span>
@@ -78,12 +121,61 @@
 		{#if !queue.collapsed}
 			<ul class="flex max-h-64 flex-col divide-y overflow-y-auto">
 				{#each queue.jobs as job (job.id)}
+					{@const duplicates = job.duplicates ?? []}
+					{@const target = noteTarget(duplicates)}
+					{@const possible = job.possibleDuplicates ?? []}
 					<li class="flex min-w-0 items-center gap-2 px-3 py-2">
 						<span class="min-w-0 flex-1">
 							<span class="block truncate text-sm" title={job.label}>{job.label}</span>
-							<span class="block font-mono text-[10px] text-muted-foreground tabular-nums">
-								{job.error ?? formatFileSize(job.size)}
-							</span>
+
+							{#if job.error}
+								<span class="block font-mono text-[10px] text-muted-foreground tabular-nums">
+									{job.error}
+								</span>
+							{:else if duplicates.length > 0}
+								<!--
+									Reported, with no offer to override. Storing the same bytes twice
+									splits reading progress and shelves across two records that can
+									never converge, and there is no version of that the reader wants.
+									`allow_duplicates` stays on the API for when the match is wrong.
+								-->
+								<span class="block min-w-0 text-[10px] text-muted-foreground">
+									{#if target}
+										<a
+											href={resolve('/(root)/(library)/book/[bookId]', { bookId: String(target) })}
+											class="truncate underline underline-offset-2"
+										>
+											{duplicateNote(duplicates)}
+										</a>
+									{:else}
+										<span class="truncate">{duplicateNote(duplicates)}</span>
+									{/if}
+								</span>
+							{:else if possible.length > 0}
+								<!--
+									A guess, not a fact, so it says so and stops there: the book was
+									stored, and the library's duplicates screen is where a reader
+									decides what to do about it.
+								-->
+								<span class="block min-w-0 text-[10px] text-muted-foreground">
+									{#if possible.length === 1}
+										<a
+											href={resolve('/(root)/(library)/book/[bookId]', {
+												bookId: String(possible[0].book_id)
+											})}
+											class="truncate underline underline-offset-2"
+										>
+											{possibleNote(possible)}
+										</a>
+									{:else}
+										<span class="truncate">{possibleNote(possible)}</span>
+									{/if}
+								</span>
+							{:else}
+								<span class="block font-mono text-[10px] text-muted-foreground tabular-nums">
+									{formatFileSize(job.size)}
+								</span>
+							{/if}
 						</span>
 
 						<span
@@ -96,6 +188,8 @@
 								Adding
 							{:else if job.status === 'done'}
 								Done
+							{:else if job.status === 'skipped'}
+								Skipped
 							{:else if job.status === 'failed'}
 								Failed
 							{:else}

frontend/src/lib/components/view/batch-operations-toolbar.svelte

@@ -1,15 +1,27 @@
 <script lang="ts">
 	import { buttonVariants } from '$lib/components/ui/button/button.svelte';
 	import * as Tooltip from '$lib/components/ui/tooltip/index';
-	import { BookOpenCheck, Download, Trash2, SquareCheckBig, X, Album, PlusIcon } from '@lucide/svelte';
+	import {
+		BookOpenCheck,
+		Download,
+		GitMerge,
+		Trash2,
+		SquareCheckBig,
+		SlidersHorizontal,
+		X,
+		Album,
+		PlusIcon
+	} from '@lucide/svelte';
 	import * as DropdownMenu from '$lib/components/ui/dropdown-menu/index';
 	import { getBookOperationsState } from '$lib/state/bookOperations.svelte';
 	import { getBookSelectionState } from '$lib/state/bookSelection.svelte';
 	import { getBookCollectionState } from '$lib/state/bookCollection.svelte';
 	import { getLibraryState } from '$lib/state/library.svelte';
 	import { getBookshelfState } from '$lib/state/bookshelf.svelte';
-	import { invalidate } from '$app/navigation';
 	import ShelfCreateDialog from '../forms/shelf-create-dialog.svelte';
+	import MergeBooks from '../forms/merge-books/merge-books.svelte';
+	import { mergeInto } from '../forms/merge-books/merge';
+	import type { Book } from '$lib/schema';
 
 	const libraryState = getLibraryState();
 	const bookshelfState = getBookshelfState();
@@ -17,10 +29,50 @@
 	const bookOps = getBookOperationsState();
 	const collectionState = getBookCollectionState();
 
-	let selectedBooks = $derived(selectionState.getSelectedBooks())
+	let selectedBooks = $derived(selectionState.getSelectedBooks());
 
-	let createShelfDialogOpen = $state(false)
+	let createShelfDialogOpen = $state(false);
 
+	// The books the merge dialog opened on. Snapshotted rather than read live from
+	// the selection, so clearing the selection on success cannot empty the dialog
+	// underneath itself.
+	let mergingBooks = $state<Book[] | null>(null);
+
+	// A quick merge has no dialog to disable, so the flag is what stops a slow
+	// round trip being started twice.
+	let quickMerging = $state(false);
+
+	/** Put the library back in step with a merge that went through. */
+	function afterMerge(folded: Book[]) {
+		// Only the folded records are gone; the survivor is still in the library.
+		libraryState.activeLibrary!.total! -= folded.length;
+		bookshelfState.deletedBooks(folded);
+		selectionState.deselectAll();
+	}
+
+	/**
+	 * Merge without opening the workbench, keeping the book selected first.
+	 *
+	 * Selection is held in insertion order, so the first record is the one the
+	 * reader started from — the one they were looking at when they decided the
+	 * rest were copies of it.
+	 */
+	async function quickMerge() {
+		if (quickMerging) return;
+
+		// Read once: the merge clears the selection, and the bookkeeping afterwards
+		// still needs the records that went away.
+		const [survivor, ...folded] = selectedBooks;
+		if (!survivor || folded.length === 0) return;
+
+		quickMerging = true;
+
+		const merged = await mergeInto(libraryState.activeLibrary!.id, survivor, folded);
+
+		quickMerging = false;
+
+		if (merged) afterMerge(folded);
+	}
 </script>
 
 <!-- Mark selected as finished button -->
@@ -95,12 +147,13 @@
 							{/each}
 						</DropdownMenu.Group>
 						<DropdownMenu.Separator />
-							<DropdownMenu.Item 
-							onclick={() => createShelfDialogOpen = true}
-							class="text-muted-foreground ">
-								<PlusIcon class="size-4" />
-								New Shelf
-							</DropdownMenu.Item>
+						<DropdownMenu.Item
+							onclick={() => (createShelfDialogOpen = true)}
+							class="text-muted-foreground "
+						>
+							<PlusIcon class="size-4" />
+							New Shelf
+						</DropdownMenu.Item>
 					</DropdownMenu.Content>
 				</DropdownMenu.Root>
 			{/snippet}
@@ -128,6 +181,44 @@
 	</Tooltip.Root>
 </Tooltip.Provider>
 
+<!-- Merge selected button. Two books is the smallest thing a merge can mean, so
+     it appears only once there are two. -->
+{#if selectedBooks.length > 1}
+	<Tooltip.Provider>
+		<Tooltip.Root ignoreNonKeyboardFocus>
+			<Tooltip.Trigger>
+				{#snippet child({ props })}
+					<!-- A choice rather than a jump straight into the workbench: two
+					     copies of one book have nothing to resolve, and the dialog is a
+					     step in the way of saying so. -->
+					<DropdownMenu.Root>
+						<DropdownMenu.Trigger
+							{...props}
+							disabled={quickMerging}
+							class={buttonVariants({ variant: 'ghost', size: 'icon' })}
+						>
+							<GitMerge />
+						</DropdownMenu.Trigger>
+						<DropdownMenu.Content>
+							<DropdownMenu.Item onclick={quickMerge}>
+								<GitMerge class="size-4 shrink-0" />
+								Merge into first selected
+							</DropdownMenu.Item>
+							<DropdownMenu.Item onclick={() => (mergingBooks = selectedBooks)}>
+								<SlidersHorizontal class="size-4 shrink-0" />
+								Manual merge…
+							</DropdownMenu.Item>
+						</DropdownMenu.Content>
+					</DropdownMenu.Root>
+				{/snippet}
+			</Tooltip.Trigger>
+			<Tooltip.Content>
+				<p>Merge {selectedBooks.length} books</p>
+			</Tooltip.Content>
+		</Tooltip.Root>
+	</Tooltip.Provider>
+{/if}
+
 <!-- Delete selected button -->
 <Tooltip.Provider>
 	<Tooltip.Root ignoreNonKeyboardFocus>
@@ -136,8 +227,8 @@
 				bookOps.deleteDialogTitle = `Delete ${selectionState.getSelectedIds().length} books?`;
 				bookOps.deleteFn = async (deleteFiles: boolean) => {
 					await bookOps.deleteBooks(selectionState.getSelectedIds(), deleteFiles);
-					libraryState.activeLibrary!.total! -= selectedBooks.length
-					bookshelfState.deletedBooks(selectedBooks)
+					libraryState.activeLibrary!.total! -= selectedBooks.length;
+					bookshelfState.deletedBooks(selectedBooks);
 					selectionState.deselectAll();
 				};
 				bookOps.deleteDialogOpen = true;
@@ -186,14 +277,43 @@
 	</Tooltip.Root>
 </Tooltip.Provider>
 
-
 <ShelfCreateDialog
 	bind:open={createShelfDialogOpen}
 	onSubmit={async (name: string) => {
-
-		const bookshelf = await bookshelfState.addBookshelf(name, libraryState.activeLibrary!.id, selectionState.getSelectedIds())
-		selectedBooks.forEach(book => book.lists.push(bookshelf))
-		selectionState.deselectAll()
-		createShelfDialogOpen = false
+		const bookshelf = await bookshelfState.addBookshelf(
+			name,
+			libraryState.activeLibrary!.id,
+			selectionState.getSelectedIds()
+		);
+		selectedBooks.forEach((book) => book.lists.push(bookshelf));
+		selectionState.deselectAll();
+		createShelfDialogOpen = false;
 	}}
 />
+
+<!--
+	Keyed on the selection so opening merge on a different pair starts from those
+	records rather than the previous dialog's draft.
+-->
+{#if mergingBooks}
+	{#key mergingBooks.map((book) => book.id).join()}
+		<MergeBooks
+			books={mergingBooks}
+			libraryId={libraryState.activeLibrary!.id}
+			bind:open={
+				() => mergingBooks !== null,
+				(value) => {
+					// Bound, not passed as a bare `true`: the dialog closes itself on an
+					// outside click or Escape, and if that never reaches `mergingBooks`
+					// the snapshot stays set — the menu item then re-assigns the same
+					// selection, nothing changes, and the dialog never reopens.
+					if (!value) mergingBooks = null;
+				}
+			}
+			onmerged={(_survivor, folded) => {
+				afterMerge(folded);
+				mergingBooks = null;
+			}}
+		/>
+	{/key}
+{/if}

frontend/src/lib/components/view/generated-cover.svelte

@@ -1,7 +1,14 @@
 <script lang="ts">
-	import type { Book } from '$lib/schema';
-
-	let { book, class: className = '' }: { book: Book; class?: string } = $props();
+	// Typed by what is drawn rather than by `Book`: the duplicates review holds
+	// candidates that carry a title and author names without a whole record, and a
+	// `Book` satisfies this shape anyway.
+	let {
+		book,
+		class: className = ''
+	}: {
+		book: { title: string; authors?: { name: string }[] | null };
+		class?: string;
+	} = $props();
 
 	/**
 	 * Bookcloth tones rather than a point on the hue wheel.

frontend/src/lib/schema/book.ts

@@ -6,6 +6,49 @@ export type Book = components['schemas']['BookRead'];
 export type BookFile = components['schemas']['FileMetadataRead'];
 export type BookProgress = components['schemas']['BookProgressRead'];
 
+/**
+ * A file the library already held, so the upload did not store it again.
+ *
+ * `book_id` is null when the match was another file in the same upload — there is
+ * no book to point at yet.
+ */
+export type DuplicateFile = components['schemas']['DuplicateFileRead'];
+export type BooksUploadResult = components['schemas']['BooksUploadResult'];
+
+/**
+ * A stored book that may be the same book as another one.
+ *
+ * Weaker than `DuplicateFile`, and deliberately so: that one means the library holds
+ * these exact bytes, this one means the metadata agrees. A second edition and a
+ * translation both look like this, so nothing is ever refused on the strength of it.
+ */
+export type DuplicateBook = components['schemas']['DuplicateBookRead'];
+
+/** A book that was imported, together with what it might be a second copy of. */
+export type PossibleDuplicate = components['schemas']['PossibleDuplicateRead'];
+
+/** Books the library holds that all look like copies of one book. */
+export type DuplicateBookGroup = components['schemas']['DuplicateBookGroupRead'];
+
+/** The metadata a reader resolved while merging, or while reviewing a provider. */
+export type BookMetadataUpdate = components['schemas']['BookMetadataUpdate'];
+
+/** Mirrors BookMerge in backend/src/chitai/schemas/book.py */
+export const bookMergeSchema = z.object({
+	library_id: z.coerce.number(),
+	survivor_id: z.coerce.number(),
+	merged_ids: z.array(z.coerce.number()).min(1, 'Pick at least one book to fold in'),
+	// Passed through untouched — the backend validates it as BookMetadataUpdate, and
+	// duplicating that shape here would be two places to keep in step for no gain.
+	metadata: z.record(z.string(), z.unknown()).optional()
+});
+
+/** Mirrors DuplicateDismissal in backend/src/chitai/schemas/book.py */
+export const duplicateDismissalSchema = z.object({
+	book_a_id: z.coerce.number(),
+	book_b_id: z.coerce.number()
+});
+
 export const bookQuerySchema = commonQuerySchema.extend({
 	libraries: stringArrayCoerce,
 	authors: stringArrayCoerce,
@@ -97,4 +140,6 @@ export type BookQuery = z.infer<typeof bookQuerySchema>;
 export type BookEditMetadata = z.infer<typeof editBookMetadataSchema>;
 export type DeleteBook = z.infer<typeof deleteBooksSchema>;
 export type DeleteBookFiles = z.infer<typeof deleteBookFilesSchema>;
+export type DuplicateDismissal = z.infer<typeof duplicateDismissalSchema>;
+export type BookMerge = z.infer<typeof bookMergeSchema>;
 export type UpdateBookProgress = z.infer<typeof updateBookProgressSchema>;

frontend/src/lib/schema/openapi/schema.d.ts

@@ -22,6 +22,23 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/books/duplicate-files": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        get?: never;
+        put?: never;
+        /** CheckDuplicateFiles */
+        post: operations["BooksDuplicateFilesCheckDuplicateFiles"];
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/books": {
         parameters: {
             query?: never;
@@ -58,6 +75,24 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/books/duplicate-books/dismissals": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        get?: never;
+        put?: never;
+        /** DismissDuplicateBooks */
+        post: operations["BooksDuplicateBooksDismissalsDismissDuplicateBooks"];
+        /** RestoreDuplicateBooks */
+        delete: operations["BooksDuplicateBooksDismissalsRestoreDuplicateBooks"];
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/books/{book_id}": {
         parameters: {
             query?: never;
@@ -110,6 +145,40 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/books/duplicate-books": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        /** ListDuplicateBooks */
+        get: operations["BooksDuplicateBooksListDuplicateBooks"];
+        put?: never;
+        post?: never;
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
+    "/books/merge": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        get?: never;
+        put?: never;
+        /** MergeBooks */
+        post: operations["BooksMergeMergeBooks"];
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/books/progress": {
         parameters: {
             query?: never;
@@ -637,6 +706,12 @@ export interface components {
             cover_image?: string | null;
             files?: string[];
         };
+        /** BookMerge */
+        BookMerge: {
+            survivor_id: number;
+            merged_ids: number[];
+            metadata?: components["schemas"]["BookMetadataUpdate"] | null;
+        };
         /** BookMetadataUpdate */
         BookMetadataUpdate: {
             title?: string | null;
@@ -710,6 +785,46 @@ export interface components {
         BooksCreateFromFiles: {
             files?: string[];
         };
+        /** BooksUploadResult */
+        BooksUploadResult: {
+            created: components["schemas"]["BookRead"][];
+            skipped: components["schemas"]["DuplicateFileRead"][];
+            possible_duplicates?: components["schemas"]["PossibleDuplicateRead"][];
+        };
+        /** DuplicateBookGroupRead */
+        DuplicateBookGroupRead: {
+            books: components["schemas"]["DuplicateBookRead"][];
+        };
+        /** DuplicateBookRead */
+        DuplicateBookRead: {
+            book_id: number;
+            title: string;
+            authors: string[];
+            library_id: number;
+            cover_image?: string | null;
+            matched_on: string[];
+        };
+        /** DuplicateDismissal */
+        DuplicateDismissal: {
+            book_a_id: number;
+            book_b_id: number;
+        };
+        /** DuplicateFileRead */
+        DuplicateFileRead: {
+            filename: string;
+            hash: string;
+            size: number;
+            library_id: number;
+            book_id?: number | null;
+            book_title?: string | null;
+        };
+        /** FileFingerprint */
+        FileFingerprint: {
+            hash: string;
+            size: number;
+            /** @default  */
+            filename: string;
+        };
         /** FileMetadataRead */
         FileMetadataRead: {
             id: number;
@@ -777,6 +892,12 @@ export interface components {
             refresh_token?: string | null;
             expires_in?: number | null;
         };
+        /** PossibleDuplicateRead */
+        PossibleDuplicateRead: {
+            book_id: number;
+            title: string;
+            candidates: components["schemas"]["DuplicateBookRead"][];
+        };
         /** PublisherRead */
         PublisherRead: {
             id: number;
@@ -828,6 +949,7 @@ export interface operations {
         parameters: {
             query?: {
                 library_id?: number | null;
+                allow_duplicates?: boolean;
             };
             header?: never;
             path: {
@@ -908,6 +1030,47 @@ export interface operations {
             };
         };
     };
+    BooksDuplicateFilesCheckDuplicateFiles: {
+        parameters: {
+            query?: {
+                library_id?: number | null;
+            };
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody: {
+            content: {
+                "application/json": components["schemas"]["FileFingerprint"][];
+            };
+        };
+        responses: {
+            /** @description Request fulfilled, document follows */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["DuplicateFileRead"][];
+                };
+            };
+            /** @description Bad request syntax or unsupported method */
+            400: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": {
+                        status_code: number;
+                        detail: string;
+                        extra?: null | {
+                            [key: string]: unknown;
+                        } | unknown[];
+                    };
+                };
+            };
+        };
+    };
     BooksListBooks: {
         parameters: {
             query?: {
@@ -917,7 +1080,7 @@ export interface operations {
                 tags?: number[] | null;
                 shelves?: number[] | null;
                 progress?: string[] | null;
-                ids?: string[] | null;
+                ids?: number[] | null;
                 searchString?: string | null;
                 searchIgnoreCase?: boolean | null;
                 currentPage?: number;
@@ -969,6 +1132,7 @@ export interface operations {
         parameters: {
             query?: {
                 library_id?: number | null;
+                allow_duplicates?: boolean;
             };
             header?: never;
             path?: never;
@@ -1047,6 +1211,7 @@ export interface operations {
         parameters: {
             query?: {
                 library_id?: number | null;
+                allow_duplicates?: boolean;
             };
             header?: never;
             path?: never;
@@ -1060,21 +1225,86 @@ export interface operations {
         responses: {
             /** @description Document created, URL follows */
             201: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["BooksUploadResult"];
+                };
+            };
+            /** @description Bad request syntax or unsupported method */
+            400: {
                 headers: {
                     [name: string]: unknown;
                 };
                 content: {
                     "application/json": {
-                        items?: components["schemas"]["BookRead"][];
-                        /** @description Maximal number of items to send. */
-                        limit?: number;
-                        /** @description Offset from the beginning of the query. */
-                        offset?: number;
-                        /** @description Total number of items. */
-                        total?: number;
+                        status_code: number;
+                        detail: string;
+                        extra?: null | {
+                            [key: string]: unknown;
+                        } | unknown[];
                     };
                 };
             };
+        };
+    };
+    BooksDuplicateBooksDismissalsDismissDuplicateBooks: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody: {
+            content: {
+                "application/json": components["schemas"]["DuplicateDismissal"];
+            };
+        };
+        responses: {
+            /** @description Request fulfilled, nothing follows */
+            204: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content?: never;
+            };
+            /** @description Bad request syntax or unsupported method */
+            400: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": {
+                        status_code: number;
+                        detail: string;
+                        extra?: null | {
+                            [key: string]: unknown;
+                        } | unknown[];
+                    };
+                };
+            };
+        };
+    };
+    BooksDuplicateBooksDismissalsRestoreDuplicateBooks: {
+        parameters: {
+            query: {
+                book_a_id: number;
+                book_b_id: number;
+            };
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Request fulfilled, nothing follows */
+            204: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content?: never;
+            };
             /** @description Bad request syntax or unsupported method */
             400: {
                 headers: {
@@ -1254,6 +1484,84 @@ export interface operations {
             };
         };
     };
+    BooksDuplicateBooksListDuplicateBooks: {
+        parameters: {
+            query?: {
+                library_id?: number | null;
+            };
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Request fulfilled, document follows */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["DuplicateBookGroupRead"][];
+                };
+            };
+            /** @description Bad request syntax or unsupported method */
+            400: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": {
+                        status_code: number;
+                        detail: string;
+                        extra?: null | {
+                            [key: string]: unknown;
+                        } | unknown[];
+                    };
+                };
+            };
+        };
+    };
+    BooksMergeMergeBooks: {
+        parameters: {
+            query?: {
+                library_id?: number | null;
+            };
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody: {
+            content: {
+                "application/json": components["schemas"]["BookMerge"];
+            };
+        };
+        responses: {
+            /** @description Document created, URL follows */
+            201: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["BookRead"];
+                };
+            };
+            /** @description Bad request syntax or unsupported method */
+            400: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": {
+                        status_code: number;
+                        detail: string;
+                        extra?: null | {
+                            [key: string]: unknown;
+                        } | unknown[];
+                    };
+                };
+            };
+        };
+    };
     BooksProgressSetBookProgressBatch: {
         parameters: {
             query: {
@@ -1931,7 +2239,7 @@ export interface operations {
                 tags?: number[] | null;
                 shelves?: number[] | null;
                 progress?: string[] | null;
-                ids?: string[] | null;
+                ids?: number[] | null;
                 searchString?: string | null;
                 searchIgnoreCase?: boolean | null;
                 currentPage?: number;
@@ -2018,7 +2326,7 @@ export interface operations {
     OpdsLibraryLibraryIdCollectionTypeGetLibraryCollectionFeed: {
         parameters: {
             query?: {
-                ids?: string[] | null;
+                ids?: number[] | null;
                 searchString?: string | null;
                 searchIgnoreCase?: boolean | null;
                 currentPage?: number;
@@ -2147,7 +2455,7 @@ export interface operations {
                 tags?: number[] | null;
                 shelves?: number[] | null;
                 progress?: string[] | null;
-                ids?: string[] | null;
+                ids?: number[] | null;
                 searchString?: string | null;
                 searchIgnoreCase?: boolean | null;
                 currentPage?: number;

frontend/src/lib/state/upload-queue.svelte.ts

@@ -1,9 +1,9 @@
 import { getContext, setContext } from 'svelte';
 import { invalidate } from '$app/navigation';
 
-import type { Book, PaginatedResponse } from '$lib/schema';
+import type { Book, BooksUploadResult, DuplicateBook, DuplicateFile } from '$lib/schema';
 
-export type UploadStatus = 'queued' | 'uploading' | 'done' | 'failed';
+export type UploadStatus = 'queued' | 'uploading' | 'done' | 'skipped' | 'failed';
 
 export interface UploadJob {
 	id: string;
@@ -15,10 +15,23 @@ export interface UploadJob {
 	status: UploadStatus;
 	error?: string;
 	book?: Book;
+	/**
+	 * Files the library already held, which were not stored again. A job with
+	 * nothing left over lands as `skipped`; one that had something new is `done`
+	 * and still carries these, since the reader asked for those files too.
+	 */
+	duplicates?: DuplicateFile[];
+	/**
+	 * Books already in the library that the one this job created might be a second
+	 * copy of. Much weaker than `duplicates`: the bytes are new and only the metadata
+	 * agrees, which a second edition and a translation both do. The book was stored.
+	 */
+	possibleDuplicates?: DuplicateBook[];
 }
 
 export interface UploadSummary {
 	created: number;
+	skipped: number;
 	failed: number;
 	firstBook?: Book;
 }
@@ -70,12 +83,30 @@ export class UploadQueueState {
 
 	readonly total = $derived(this.jobs.length);
 	readonly done = $derived(this.jobs.filter((job) => job.status === 'done').length);
+	readonly skipped = $derived(this.jobs.filter((job) => job.status === 'skipped').length);
 	readonly failed = $derived(this.jobs.filter((job) => job.status === 'failed').length);
 	readonly active = $derived(
 		this.jobs.some((job) => job.status === 'queued' || job.status === 'uploading')
 	);
 	readonly current = $derived(this.jobs.find((job) => job.status === 'uploading'));
-	readonly settled = $derived(this.done + this.failed);
+	readonly settled = $derived(this.done + this.skipped + this.failed);
+
+	/**
+	 * Whether the run left something the reader still has to see.
+	 *
+	 * A skipped book is not a failure, but it is the only place that says the file
+	 * was already here — and the only place to override it from. A book that may
+	 * already be in the library counts too: it is a note the reader has to actually
+	 * read, and a tray that clears itself after six seconds is one they never will.
+	 */
+	readonly needsAttention = $derived(
+		this.jobs.some(
+			(job) =>
+				job.status === 'failed' ||
+				(job.duplicates?.length ?? 0) > 0 ||
+				(job.possibleDuplicates?.length ?? 0) > 0
+		)
+	);
 
 	#running = false;
 	#dismissTimer: ReturnType<typeof setTimeout> | undefined;
@@ -129,20 +160,20 @@ export class UploadQueueState {
 
 	release() {
 		this.#held = false;
-		if (!this.active && this.failed === 0 && this.total > 0) this.#scheduleDismiss();
+		if (!this.active && !this.needsAttention && this.total > 0) this.#scheduleDismiss();
 	}
 
 	/**
-	 * Clears itself only when everything worked. A run with failures stays until
-	 * dismissed — it is the only record of what did not make it in, and the only
-	 * place to retry from.
+	 * Clears itself only when everything went in cleanly. A run with failures or
+	 * skipped files stays until dismissed — it is the only record of what did not
+	 * make it in, and the only place to retry or override from.
 	 */
 	#scheduleDismiss() {
 		clearTimeout(this.#dismissTimer);
 		if (this.#held) return;
 
 		this.#dismissTimer = setTimeout(() => {
-			if (!this.active && this.failed === 0) this.jobs = [];
+			if (!this.active && !this.needsAttention) this.jobs = [];
 		}, DISMISS_AFTER_MS);
 	}
 
@@ -169,26 +200,35 @@ export class UploadQueueState {
 				const index = this.jobs.findIndex((job) => job.status === 'queued');
 				if (index === -1) break;
 
+				const job = this.jobs[index];
 				this.#patch(index, { status: 'uploading' });
 
 				try {
 					const body = new FormData();
-					for (const file of this.jobs[index].files) body.append('files', file);
+					for (const file of job.files) body.append('files', file);
 
 					const response = await fetch(
-						`/api/books/fromFiles?library_id=${encodeURIComponent(String(this.jobs[index].libraryId))}`,
+						`/api/books/fromFiles?library_id=${encodeURIComponent(String(job.libraryId))}`,
 						{ method: 'POST', body }
 					);
 
 					if (!response.ok) throw new Error(`The server returned ${response.status}`);
 
-					const result: PaginatedResponse<Book> = await response.json();
-					const book = result.items[0];
+					const result: BooksUploadResult = await response.json();
+					const book = result.created[0];
 
-					created += result.total ?? result.items.length;
+					created += result.created.length;
 					firstBook ??= book;
 
-					this.#patch(index, { status: 'done', book });
+					// Nothing created means every file in this folder was already here.
+					// That is not a failure, and it is not something to hide either.
+					this.#patch(index, {
+						status: result.created.length === 0 ? 'skipped' : 'done',
+						book,
+						duplicates: result.skipped,
+						// One job is one book, so it has at most one set of candidates.
+						possibleDuplicates: result.possible_duplicates?.[0]?.candidates
+					});
 				} catch (error) {
 					// One bad book must not take the rest of the queue with it.
 					console.error(`Failed to upload ${this.jobs[index].label}`, error);
@@ -205,9 +245,9 @@ export class UploadQueueState {
 
 		if (created > 0) await invalidate('app:books');
 
-		if (this.failed === 0) this.#scheduleDismiss();
+		if (!this.needsAttention) this.#scheduleDismiss();
 
-		onFinished?.({ created, failed: this.failed, firstBook });
+		onFinished?.({ created, skipped: this.skipped, failed: this.failed, firstBook });
 	}
 
 	#patch(index: number, changes: Partial<UploadJob>) {

frontend/src/routes/(root)/settings/+layout.svelte

@@ -1,9 +1,14 @@
 <script lang="ts">
 	import { page } from '$app/state';
 	import { resolve } from '$app/paths';
+	import { getLibraryState } from '$lib/state/library.svelte';
 
 	let { children } = $props();
 
+	// Set in (root)/+layout.svelte, above this group, so listing the libraries in
+	// the nav needs no load function.
+	const libraryState = getLibraryState();
+
 	// Route ids rather than paths: resolve() is called in the markup so it stays a
 	// direct call the lint rule can see, and the active check compares route ids.
 	// A pathname comparison would miss during SSR, where resolve() returns a
@@ -14,12 +19,28 @@
 		{ title: 'Libraries', routeId: '/(root)/settings/libraries' },
 		{ title: 'Devices', routeId: '/(root)/settings/devices' }
 	] as const;
+
+	/**
+	 * Whether a nested library item is the one being looked at.
+	 *
+	 * Both halves are needed: the route id alone is shared by every library, so
+	 * matching on it would light all of them up at once.
+	 */
+	function isActiveLibrary(id: number) {
+		return (
+			page.route.id?.startsWith('/(root)/settings/libraries/[libraryId]') === true &&
+			page.params.libraryId === String(id)
+		);
+	}
 </script>
 
 <div class="flex h-[calc(100vh-var(--header-height)-2rem)] flex-col">
 	<h1 class="font-serif text-xl font-medium tracking-tight">Settings</h1>
 	<div class="mt-6 flex flex-1 gap-8 overflow-hidden">
-		<nav class="flex w-48 shrink-0 flex-col gap-1">
+		<!-- Scrolls on its own now that it holds a list that grows with the number
+		     of libraries; the row above it is overflow-hidden, so without this a
+		     long list is clipped rather than reachable. -->
+		<nav class="flex w-48 shrink-0 flex-col gap-1 overflow-y-auto">
 			{#each items as item}
 				<a
 					href={resolve(item.routeId)}
@@ -30,6 +51,26 @@
 				>
 					{item.title}
 				</a>
+
+				<!-- Libraries is the one item with children: a library's own settings
+				     hang off it, so every library and every section it has stay one
+				     click from here. -->
+				{#if item.routeId === '/(root)/settings/libraries'}
+					{#each libraryState.libraries as library (library.id)}
+						<a
+							href={resolve('/(root)/settings/libraries/[libraryId]/duplicates', {
+								libraryId: String(library.id)
+							})}
+							class="ml-3 truncate rounded-md border-l px-3 py-1.5 text-sm transition-colors hover:bg-muted {isActiveLibrary(
+								library.id
+							)
+								? 'bg-muted font-medium'
+								: 'text-muted-foreground'}"
+						>
+							{library.name}
+						</a>
+					{/each}
+				{/if}
 			{/each}
 		</nav>
 		<div class="flex-1 overflow-auto">{@render children()}</div>

frontend/src/routes/(root)/settings/libraries/[libraryId]/+layout.svelte

@@ -0,0 +1,48 @@
+<script lang="ts">
+	import { page } from '$app/state';
+	import { resolve } from '$app/paths';
+	import { getLibraryState } from '$lib/state/library.svelte';
+
+	let { children } = $props();
+
+	// Set in (root)/+layout.svelte, above the settings group, so the libraries are
+	// already here — this pane needs no load function of its own.
+	const libraryState = getLibraryState();
+
+	const library = $derived(
+		libraryState.libraries.find((lib) => String(lib.id) === page.params.libraryId)
+	);
+
+	// Duplicates is the only section today. The strip exists so General and a
+	// danger zone have an obvious place to land; neither is built yet, and an
+	// empty tab is worse than no tab.
+	//
+	// Active state is matched on route id, not pathname, for the reason spelled
+	// out in settings/+layout.svelte.
+	const sections = [
+		{ title: 'Duplicates', routeId: '/(root)/settings/libraries/[libraryId]/duplicates' }
+	] as const;
+</script>
+
+<div class="flex flex-col gap-4">
+	<div>
+		<h2 class="text-lg font-semibold">{library?.name ?? 'Library'}</h2>
+		<p class="text-sm text-muted-foreground">Settings for this library</p>
+	</div>
+
+	<nav class="flex gap-1 border-b">
+		{#each sections as section (section.routeId)}
+			<a
+				href={resolve(section.routeId, { libraryId: page.params.libraryId! })}
+				class="-mb-px border-b-2 px-3 py-2 text-sm font-medium transition-colors hover:text-foreground {page
+					.route.id === section.routeId
+					? 'border-foreground'
+					: 'border-transparent text-muted-foreground'}"
+			>
+				{section.title}
+			</a>
+		{/each}
+	</nav>
+
+	{@render children()}
+</div>

frontend/src/routes/(root)/settings/libraries/[libraryId]/+page.ts

@@ -0,0 +1,10 @@
+import { resolve } from '$app/paths';
+import { redirect } from '@sveltejs/kit';
+
+export async function load({ params }) {
+	// The library pane is its sections; land on the first one.
+	redirect(
+		303,
+		resolve('/(root)/settings/libraries/[libraryId]/duplicates', { libraryId: params.libraryId })
+	);
+}

frontend/src/routes/(root)/settings/libraries/[libraryId]/duplicates/+page.server.ts

@@ -0,0 +1,18 @@
+import { listBooks, listDuplicateBooks } from '$lib/api/book.remote.js';
+
+export async function load({ params, depends }) {
+	// Dismissing or merging a group re-runs this, so the card leaves the screen.
+	depends('app:duplicate-books');
+
+	const groups = await listDuplicateBooks(params.libraryId);
+
+	// The groups carry only enough to render a card. Merging needs the whole record —
+	// identifiers, description, publisher — so fetch them in one go rather than per
+	// card, and let the dialog pick out the books for its own group.
+	const ids = [...new Set(groups.flatMap((group) => group.books.map((book) => book.book_id)))];
+	const books = ids.length
+		? await listBooks({ ids, pageSize: ids.length })
+		: { items: [] };
+
+	return { groups, books: books.items };
+}

frontend/src/routes/(root)/settings/libraries/[libraryId]/duplicates/+page.svelte

@@ -0,0 +1,249 @@
+<script lang="ts">
+	import { page } from '$app/state';
+	import { resolve } from '$app/paths';
+	import { invalidate } from '$app/navigation';
+	import { toast } from 'svelte-sonner';
+	import { CopyCheck, Fingerprint, GitMerge, SlidersHorizontal, Type } from '@lucide/svelte';
+
+	import * as Card from '$lib/components/ui/card/index.js';
+	import * as Empty from '$lib/components/ui/empty/index.js';
+	import * as DropdownMenu from '$lib/components/ui/dropdown-menu/index.js';
+	import { Badge } from '$lib/components/ui/badge/index.js';
+	import { Button, buttonVariants } from '$lib/components/ui/button/index.js';
+	import BookImage from '$lib/components/view/book-image.svelte';
+	import GeneratedCover from '$lib/components/view/generated-cover.svelte';
+	import { dismissDuplicateBooks } from '$lib/api/book.remote';
+	import MergeBooks from '$lib/components/forms/merge-books/merge-books.svelte';
+	import { mergeInto } from '$lib/components/forms/merge-books/merge';
+	import type { Book, DuplicateBookGroup } from '$lib/schema';
+
+	let { data }: { data: { groups: DuplicateBookGroup[]; books: Book[] } } = $props();
+
+	// The group open in the workbench, or null when it is closed.
+	let merging = $state<DuplicateBookGroup | null>(null);
+
+	/** The full records behind a group, in the order the group lists them. */
+	function recordsFor(group: DuplicateBookGroup): Book[] {
+		return group.books
+			.map(({ book_id }) => data.books.find((book) => book.id === book_id))
+			.filter((book): book is Book => book !== undefined);
+	}
+
+	// Which groups have a request in flight — a dismissal or a quick merge — so a
+	// slow round trip cannot be started twice.
+	let dismissing = $state<number[]>([]);
+	let quickMerging = $state<number[]>([]);
+
+	/**
+	 * Merge without opening the workbench, keeping the first book as it stands.
+	 *
+	 * The common case by far: the group is the same book twice, one record is as
+	 * good as the other, and the point is to end up with one. Resolving metadata
+	 * field by field is the other button.
+	 */
+	async function quickMerge(group: DuplicateBookGroup) {
+		const key = keyOf(group);
+		if (quickMerging.includes(key)) return;
+
+		const [survivor, ...folded] = recordsFor(group);
+		if (!survivor || folded.length === 0) return;
+
+		quickMerging = [...quickMerging, key];
+
+		try {
+			await mergeInto(page.params.libraryId!, survivor, folded);
+		} finally {
+			quickMerging = quickMerging.filter((id) => id !== key);
+		}
+	}
+
+	/** A group is named by its lowest book id, which the backend orders it by. */
+	function keyOf(group: DuplicateBookGroup) {
+		return group.books[0].book_id;
+	}
+
+	function reasonLabel(reason: string) {
+		if (reason === 'identifier') return 'Same identifier';
+		if (reason === 'title-author') return 'Same title and author';
+		return reason;
+	}
+
+	/**
+	 * Dismiss every pairing in a group at once.
+	 *
+	 * A group is held together pair by pair, so saying "these are not duplicates"
+	 * about three books means saying it about all three pairs — dismissing only the
+	 * first would leave the rest of the group standing and the screen unchanged.
+	 */
+	async function notDuplicates(group: DuplicateBookGroup) {
+		const key = keyOf(group);
+		if (dismissing.includes(key)) return;
+
+		dismissing = [...dismissing, key];
+
+		const ids = group.books.map((book) => book.book_id);
+		const pairs = ids.flatMap((a, index) =>
+			ids.slice(index + 1).map((b) => ({ book_a_id: a, book_b_id: b }))
+		);
+
+		try {
+			await Promise.all(pairs.map((pair) => dismissDuplicateBooks(pair)));
+			await invalidate('app:duplicate-books');
+			toast.success('Marked as different books');
+		} catch {
+			toast.error('Could not mark these as different books');
+		} finally {
+			dismissing = dismissing.filter((id) => id !== key);
+		}
+	}
+</script>
+
+<!-- No max width and no padding of its own: the settings pane sets the width and
+     is the thing that scrolls, so a wrapper that centres inside it only makes the
+     cards narrower than they need to be. -->
+<div class="flex flex-col gap-6 pb-4">
+	<p class="text-sm text-muted-foreground">
+		Books whose metadata matches. A second edition, a translation and a different scan of one book
+		all look like this, so nothing here has been changed or removed — this is a list to read, not a
+		problem to fix.
+	</p>
+
+	{#if data.groups.length === 0}
+		<Empty.Root>
+			<Empty.Header>
+				<Empty.Media variant="icon">
+					<CopyCheck />
+				</Empty.Media>
+				<Empty.Title>Nothing looks duplicated</Empty.Title>
+				<Empty.Description>
+					No two books in this library share an identifier, or a title and an author.
+				</Empty.Description>
+			</Empty.Header>
+		</Empty.Root>
+	{:else}
+		{#each data.groups as group (keyOf(group))}
+			{@const busy = dismissing.includes(keyOf(group)) || quickMerging.includes(keyOf(group))}
+			{@const records = recordsFor(group)}
+			<Card.Root>
+				<Card.Header>
+					<Card.Title>{group.books.length} books look like the same book</Card.Title>
+					<Card.Action class="flex gap-2">
+						<Button
+							variant="outline"
+							size="sm"
+							disabled={busy}
+							onclick={() => notDuplicates(group)}
+						>
+							Not duplicates
+						</Button>
+						<!-- A choice rather than a jump straight into the workbench:
+						     most groups are one book twice, where there is nothing to
+						     resolve and the dialog is a step in the way.
+
+						     Disabled until every record in the group came back, so
+						     neither action can run on a partial group. -->
+						<DropdownMenu.Root>
+							<DropdownMenu.Trigger
+								class={buttonVariants({ size: 'sm' })}
+								disabled={busy || records.length !== group.books.length}
+							>
+								Merge…
+							</DropdownMenu.Trigger>
+							<DropdownMenu.Content align="end">
+								<DropdownMenu.Item onclick={() => quickMerge(group)}>
+									<GitMerge class="size-4 shrink-0" />
+									Merge into first selected
+								</DropdownMenu.Item>
+								<DropdownMenu.Item onclick={() => (merging = group)}>
+									<SlidersHorizontal class="size-4 shrink-0" />
+									Manual merge…
+								</DropdownMenu.Item>
+							</DropdownMenu.Content>
+						</DropdownMenu.Root>
+					</Card.Action>
+				</Card.Header>
+
+				<Card.Content>
+					<ul class="flex flex-wrap gap-4">
+						{#each group.books as book (book.book_id)}
+							<li class="w-36">
+								<a
+									href={resolve('/(root)/(library)/book/[bookId]', {
+										bookId: String(book.book_id)
+									})}
+									class="group flex flex-col gap-2"
+								>
+									<span
+										class="block aspect-9/12 w-full overflow-hidden rounded-sm bg-muted shadow-md transition-all group-hover:brightness-75"
+									>
+										{#if book.cover_image}
+											<BookImage src="/api/{book.cover_image}" class="h-full w-full object-cover" />
+										{:else}
+											<!-- A blank swatch here is worse than anywhere else: this screen
+											     is a side-by-side comparison, and two of them are impossible
+											     to tell apart. The candidate carries author names rather than
+											     records, which is all the cover draws. -->
+											<GeneratedCover
+												book={{
+													title: book.title,
+													authors: book.authors.map((name) => ({ name }))
+												}}
+											/>
+										{/if}
+									</span>
+
+									<span class="line-clamp-2 font-serif text-sm group-hover:underline">
+										{book.title}
+									</span>
+								</a>
+
+								<p class="line-clamp-2 text-xs text-muted-foreground">
+									{book.authors.join(', ')}
+								</p>
+
+								<!-- Why this book is in the group, so the reader can judge the
+								     evidence rather than take the grouping on trust. -->
+								<p class="mt-1 flex flex-wrap gap-1">
+									{#each book.matched_on as reason (reason)}
+										<Badge variant="secondary" class="gap-1 text-[10px] font-normal">
+											{#if reason === 'identifier'}
+												<Fingerprint class="size-3" />
+											{:else}
+												<Type class="size-3" />
+											{/if}
+											{reasonLabel(reason)}
+										</Badge>
+									{/each}
+								</p>
+							</li>
+						{/each}
+					</ul>
+				</Card.Content>
+			</Card.Root>
+		{/each}
+	{/if}
+</div>
+
+<!--
+	Keyed on the group so a second merge starts from that group's records rather
+	than the previous one's draft, the same way the edit dialog keys on the book.
+-->
+{#if merging}
+	{#key merging.books[0].book_id}
+		<MergeBooks
+			books={recordsFor(merging)}
+			libraryId={page.params.libraryId!}
+			bind:open={
+				() => merging !== null,
+				(value) => {
+					// Bound, not passed as a bare `true`: the dialog closes itself on an
+					// outside click or Escape, and if that never reaches `merging` the
+					// group stays set — the menu item then re-assigns the same group,
+					// nothing changes, and the workbench never reopens.
+					if (!value) merging = null;
+				}
+			}
+			onmerged={() => (merging = null)}
+		/>
+	{/key}
+{/if}