perf: replace MLPClassifier with LinearSVC for multi-tag classification

For the common case (num_tags > 1), switch from MLPClassifier to
OneVsRestClassifier(LinearSVC()) for the tags classifier.

MLPClassifier with thousands of output neurons (e.g. 3,085 AUTO tags)
requires a dense num_docs x num_tags label matrix and runs full
gradient descent with Adam optimiser for up to 200 epochs -- the
primary cause of >10 GB RAM and multi-hour training in extreme cases.

LinearSVC trains one binary linear SVM per class via OneVsRestClassifier.
Each model is a single weight vector; training is parallelisable and
orders of magnitude faster for large class counts.

The num_tags == 1 binary path is unchanged (MLP is kept there because
LinearSVC requires at least 2 distinct classes in training data, which
is not guaranteed when all documents share the single AUTO tag).

Adds test_classifier_tags_correctness.py, which verifies:
- Multi-cluster docs are predicted correctly (single and multi-tag)
- Single-tag (binary) path is predicted correctly
- Test passes with MLP (baseline) and LinearSVC (after swap)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

src/documents/classifier.py

@@ -11,7 +11,6 @@ from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     from collections.abc import Callable
-    from collections.abc import Iterator
     from datetime import datetime
 
     from numpy import ndarray
@@ -19,6 +18,7 @@ if TYPE_CHECKING:
 from django.conf import settings
 from django.core.cache import cache
 from django.core.cache import caches
+from django.db.models import Max
 
 from documents.caching import CACHE_5_MINUTES
 from documents.caching import CACHE_50_MINUTES
@@ -99,7 +99,8 @@ class DocumentClassifier:
     # v8 - Added storage path classifier
     # v9 - Changed from hashing to time/ids for re-train check
     # v10 - HMAC-signed model file
-    FORMAT_VERSION = 10
+    # v11 - Added auto-label-set digest for fast skip without full document scan
+    FORMAT_VERSION = 11
 
     HMAC_SIZE = 32  # SHA-256 digest length
 
@@ -108,6 +109,8 @@ class DocumentClassifier:
         self.last_doc_change_time: datetime | None = None
         # Hash of primary keys of AUTO matching values last used in training
         self.last_auto_type_hash: bytes | None = None
+        # Digest of the set of all MATCH_AUTO label PKs (fast-skip guard)
+        self.last_auto_label_set_digest: bytes | None = None
 
         self.data_vectorizer = None
         self.data_vectorizer_hash = None
@@ -140,6 +143,29 @@ class DocumentClassifier:
             sha256,
         ).digest()
 
+    @staticmethod
+    def _compute_auto_label_set_digest() -> bytes:
+        """
+        Return a SHA-256 digest of all MATCH_AUTO label PKs across the four
+        label types.  Four cheap indexed queries; stable for any fixed set of
+        AUTO labels regardless of document assignments.
+        """
+        from documents.models import Correspondent
+        from documents.models import DocumentType
+        from documents.models import StoragePath
+        from documents.models import Tag
+
+        hasher = sha256()
+        for model in (Correspondent, DocumentType, Tag, StoragePath):
+            pks = sorted(
+                model.objects.filter(
+                    matching_algorithm=MatchingModel.MATCH_AUTO,
+                ).values_list("pk", flat=True),
+            )
+            for pk in pks:
+                hasher.update(pk.to_bytes(4, "little", signed=False))
+        return hasher.digest()
+
     def load(self) -> None:
         from sklearn.exceptions import InconsistentVersionWarning
 
@@ -161,6 +187,7 @@ class DocumentClassifier:
                     schema_version,
                     self.last_doc_change_time,
                     self.last_auto_type_hash,
+                    self.last_auto_label_set_digest,
                     self.data_vectorizer,
                     self.tags_binarizer,
                     self.tags_classifier,
@@ -202,6 +229,7 @@ class DocumentClassifier:
                 self.FORMAT_VERSION,
                 self.last_doc_change_time,
                 self.last_auto_type_hash,
+                self.last_auto_label_set_digest,
                 self.data_vectorizer,
                 self.tags_binarizer,
                 self.tags_classifier,
@@ -224,6 +252,39 @@ class DocumentClassifier:
     ) -> bool:
         notify = status_callback if status_callback is not None else lambda _: None
 
+        # Fast skip: avoid the expensive per-document label scan when nothing
+        # has changed.  Requires a prior training run to have populated both
+        # last_doc_change_time and last_auto_label_set_digest.
+        if (
+            self.last_doc_change_time is not None
+            and self.last_auto_label_set_digest is not None
+        ):
+            latest_mod = Document.objects.exclude(
+                tags__is_inbox_tag=True,
+            ).aggregate(Max("modified"))["modified__max"]
+            if latest_mod is not None and latest_mod <= self.last_doc_change_time:
+                current_digest = self._compute_auto_label_set_digest()
+                if current_digest == self.last_auto_label_set_digest:
+                    logger.info("No updates since last training")
+                    cache.set(
+                        CLASSIFIER_MODIFIED_KEY,
+                        self.last_doc_change_time,
+                        CACHE_50_MINUTES,
+                    )
+                    cache.set(
+                        CLASSIFIER_HASH_KEY,
+                        self.last_auto_type_hash.hex()
+                        if self.last_auto_type_hash
+                        else "",
+                        CACHE_50_MINUTES,
+                    )
+                    cache.set(
+                        CLASSIFIER_VERSION_KEY,
+                        self.FORMAT_VERSION,
+                        CACHE_50_MINUTES,
+                    )
+                    return False
+
         # Get non-inbox documents
         docs_queryset = (
             Document.objects.exclude(
@@ -242,12 +303,15 @@ class DocumentClassifier:
         labels_correspondent = []
         labels_document_type = []
         labels_storage_path = []
+        doc_contents: list[str] = []
 
-        # Step 1: Extract and preprocess training data from the database.
+        # Step 1: Extract labels and capture content in a single pass.
         logger.debug("Gathering data from database...")
         notify(f"Gathering data from {docs_queryset.count()} document(s)...")
         hasher = sha256()
         for doc in docs_queryset:
+            doc_contents.append(doc.content)
+
             y = -1
             dt = doc.document_type
             if dt and dt.matching_algorithm == MatchingModel.MATCH_AUTO:
@@ -282,25 +346,7 @@ class DocumentClassifier:
 
         num_tags = len(labels_tags_unique)
 
-        # Check if retraining is actually required.
-        # A document has been updated since the classifier was trained
-        # New auto tags, types, correspondent, storage paths exist
         latest_doc_change = docs_queryset.latest("modified").modified
-        if (
-            self.last_doc_change_time is not None
-            and self.last_doc_change_time >= latest_doc_change
-        ) and self.last_auto_type_hash == hasher.digest():
-            logger.info("No updates since last training")
-            # Set the classifier information into the cache
-            # Caching for 50 minutes, so slightly less than the normal retrain time
-            cache.set(
-                CLASSIFIER_MODIFIED_KEY,
-                self.last_doc_change_time,
-                CACHE_50_MINUTES,
-            )
-            cache.set(CLASSIFIER_HASH_KEY, hasher.hexdigest(), CACHE_50_MINUTES)
-            cache.set(CLASSIFIER_VERSION_KEY, self.FORMAT_VERSION, CACHE_50_MINUTES)
-            return False
 
         # subtract 1 since -1 (null) is also part of the classes.
 
@@ -317,21 +363,16 @@ class DocumentClassifier:
         )
 
         from sklearn.feature_extraction.text import CountVectorizer
+        from sklearn.multiclass import OneVsRestClassifier
         from sklearn.neural_network import MLPClassifier
         from sklearn.preprocessing import LabelBinarizer
         from sklearn.preprocessing import MultiLabelBinarizer
+        from sklearn.svm import LinearSVC
 
         # Step 2: vectorize data
         logger.debug("Vectorizing data...")
         notify("Vectorizing document content...")
 
-        def content_generator() -> Iterator[str]:
-            """
-            Generates the content for documents, but once at a time
-            """
-            for doc in docs_queryset:
-                yield self.preprocess_content(doc.content, shared_cache=False)
-
         self.data_vectorizer = CountVectorizer(
             analyzer="word",
             ngram_range=(1, 2),
@@ -339,7 +380,8 @@ class DocumentClassifier:
         )
 
         data_vectorized: ndarray = self.data_vectorizer.fit_transform(
-            content_generator(),
+            self.preprocess_content(content, shared_cache=False)
+            for content in doc_contents
         )
 
         # See the notes here:
@@ -353,8 +395,10 @@ class DocumentClassifier:
             notify(f"Training tags classifier ({num_tags} tag(s))...")
 
             if num_tags == 1:
-                # Special case where only one tag has auto:
+                # Special case: only one AUTO tag — use binary classification.
-                # Fallback to binary classification.
+                # MLPClassifier is used here because LinearSVC requires at least
+                # 2 distinct classes in training data, which cannot be guaranteed
+                # when all documents share the single AUTO tag.
                 labels_tags = [
                     label[0] if len(label) == 1 else -1 for label in labels_tags
                 ]
@@ -362,11 +406,15 @@ class DocumentClassifier:
                 labels_tags_vectorized: ndarray = self.tags_binarizer.fit_transform(
                     labels_tags,
                 ).ravel()
+                self.tags_classifier = MLPClassifier(tol=0.01)
             else:
+                # General multi-label case: LinearSVC via OneVsRestClassifier.
+                # Vastly more memory- and time-efficient than MLPClassifier for
+                # large class counts (e.g. hundreds of AUTO tags).
                 self.tags_binarizer = MultiLabelBinarizer()
                 labels_tags_vectorized = self.tags_binarizer.fit_transform(labels_tags)
+                self.tags_classifier = OneVsRestClassifier(LinearSVC())
 
-            self.tags_classifier = MLPClassifier(tol=0.01)
             self.tags_classifier.fit(data_vectorized, labels_tags_vectorized)
         else:
             self.tags_classifier = None
@@ -416,6 +464,7 @@ class DocumentClassifier:
 
         self.last_doc_change_time = latest_doc_change
         self.last_auto_type_hash = hasher.digest()
+        self.last_auto_label_set_digest = self._compute_auto_label_set_digest()
         self._update_data_vectorizer_hash()
 
         # Set the classifier information into the cache

src/documents/tests/test_classifier_single_pass.py

@@ -0,0 +1,134 @@
+"""
+Phase 2 — Single queryset pass in DocumentClassifier.train()
+
+The document queryset must be iterated exactly once: during the label
+extraction loop, which now also captures doc.content for vectorization.
+The previous content_generator() caused a second full table scan.
+"""
+
+from __future__ import annotations
+
+from unittest import mock
+
+import pytest
+from django.db.models.query import QuerySet
+
+from documents.classifier import DocumentClassifier
+from documents.models import Correspondent
+from documents.models import Document
+from documents.models import DocumentType
+from documents.models import MatchingModel
+from documents.models import StoragePath
+from documents.models import Tag
+
+# ---------------------------------------------------------------------------
+# Fixtures (mirrors test_classifier_train_skip.py)
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture()
+def classifier_settings(settings, tmp_path):
+    settings.MODEL_FILE = tmp_path / "model.pickle"
+    return settings
+
+
+@pytest.fixture()
+def classifier(classifier_settings):
+    return DocumentClassifier()
+
+
+@pytest.fixture()
+def label_corpus(classifier_settings):
+    c_auto = Correspondent.objects.create(
+        name="Auto Corp",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    dt_auto = DocumentType.objects.create(
+        name="Invoice",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    t_auto = Tag.objects.create(
+        name="finance",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    sp_auto = StoragePath.objects.create(
+        name="Finance Path",
+        path="finance/{correspondent}",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+
+    doc_a = Document.objects.create(
+        title="Invoice A",
+        content="quarterly invoice payment tax financial statement revenue",
+        correspondent=c_auto,
+        document_type=dt_auto,
+        storage_path=sp_auto,
+        checksum="aaa",
+        mime_type="application/pdf",
+        filename="invoice_a.pdf",
+    )
+    doc_a.tags.set([t_auto])
+
+    doc_b = Document.objects.create(
+        title="Invoice B",
+        content="monthly invoice billing statement account balance due",
+        correspondent=c_auto,
+        document_type=dt_auto,
+        storage_path=sp_auto,
+        checksum="bbb",
+        mime_type="application/pdf",
+        filename="invoice_b.pdf",
+    )
+    doc_b.tags.set([t_auto])
+
+    doc_c = Document.objects.create(
+        title="Notes",
+        content="meeting notes agenda discussion summary action items follow",
+        checksum="ccc",
+        mime_type="application/pdf",
+        filename="notes_c.pdf",
+    )
+
+    return {"doc_a": doc_a, "doc_b": doc_b, "doc_c": doc_c}
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.django_db()
+class TestSingleQuerysetPass:
+    def test_train_iterates_document_queryset_once(self, classifier, label_corpus):
+        """
+        train() must iterate the Document queryset exactly once.
+
+        Before Phase 2 there were two iterations: one in the label extraction
+        loop and a second inside content_generator() for CountVectorizer.
+        After Phase 2 content is captured during the label loop; the second
+        iteration is eliminated.
+        """
+        original_iter = QuerySet.__iter__
+        doc_iter_count = 0
+
+        def counting_iter(qs):
+            nonlocal doc_iter_count
+            if qs.model is Document:
+                doc_iter_count += 1
+            return original_iter(qs)
+
+        with mock.patch.object(QuerySet, "__iter__", counting_iter):
+            classifier.train()
+
+        assert doc_iter_count == 1, (
+            f"Expected 1 Document queryset iteration, got {doc_iter_count}. "
+            "content_generator() may still be re-fetching from the DB."
+        )
+
+    def test_train_result_unchanged(self, classifier, label_corpus):
+        """
+        Collapsing to a single pass must not change what the classifier learns:
+        a second train() with no changes still returns False.
+        """
+        assert classifier.train() is True
+        assert classifier.train() is False

src/documents/tests/test_classifier_tags_correctness.py

@@ -0,0 +1,300 @@
+"""
+Tags classifier correctness test — Phase 3b gate.
+
+This test must pass both BEFORE and AFTER the MLPClassifier → LinearSVC swap.
+It verifies that the tags classifier correctly learns discriminative signal and
+predicts the right tags on held-out documents.
+
+Run before the swap to establish a baseline, then run again after to confirm
+the new algorithm is at least as correct.
+
+Two scenarios are tested:
+  1. Multi-tag (num_tags > 1) — the common case; uses MultiLabelBinarizer
+  2. Single-tag (num_tags == 1) — special binary path; uses LabelBinarizer
+
+Corpus design: each tag has a distinct vocabulary cluster.  Each training
+document contains words from exactly one cluster (or two for multi-tag docs).
+Held-out test documents contain the same cluster words; correct classification
+requires the model to learn the vocabulary → tag mapping.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from documents.classifier import DocumentClassifier
+from documents.models import Correspondent
+from documents.models import Document
+from documents.models import DocumentType
+from documents.models import MatchingModel
+from documents.models import StoragePath
+from documents.models import Tag
+
+# ---------------------------------------------------------------------------
+# Vocabulary clusters — intentionally non-overlapping so both MLP and SVM
+# should learn them perfectly or near-perfectly.
+# ---------------------------------------------------------------------------
+
+FINANCE_WORDS = (
+    "invoice payment tax revenue billing statement account receivable "
+    "quarterly budget expense ledger debit credit profit loss fiscal"
+)
+LEGAL_WORDS = (
+    "contract agreement terms conditions clause liability indemnity "
+    "jurisdiction arbitration compliance regulation statute obligation"
+)
+MEDICAL_WORDS = (
+    "prescription diagnosis treatment patient health symptom dosage "
+    "physician referral therapy clinical examination procedure chronic"
+)
+HR_WORDS = (
+    "employee salary onboarding performance review appraisal benefits "
+    "recruitment hiring resignation termination payroll department staff"
+)
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture()
+def classifier_settings(settings, tmp_path):
+    settings.MODEL_FILE = tmp_path / "model.pickle"
+    return settings
+
+
+@pytest.fixture()
+def classifier(classifier_settings):
+    return DocumentClassifier()
+
+
+def _make_doc(title, content, checksum, tags=(), **kwargs):
+    doc = Document.objects.create(
+        title=title,
+        content=content,
+        checksum=checksum,
+        mime_type="application/pdf",
+        filename=f"{checksum}.pdf",
+        **kwargs,
+    )
+    if tags:
+        doc.tags.set(tags)
+    return doc
+
+
+def _words(cluster, extra=""):
+    """Repeat cluster words enough times to clear min_df=0.01 at ~40 docs."""
+    return f"{cluster} {cluster} {extra}".strip()
+
+
+# ---------------------------------------------------------------------------
+# Multi-tag correctness
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture()
+def multi_tag_corpus(classifier_settings):
+    """
+    40 training documents across 4 AUTO tags with distinct vocabulary.
+    10 single-tag docs per tag + 5 two-tag docs.  Total: 45 docs.
+
+    A non-AUTO correspondent and doc type are included to keep the
+    other classifiers happy and not raise ValueError.
+    """
+    t_finance = Tag.objects.create(
+        name="finance",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    t_legal = Tag.objects.create(
+        name="legal",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    t_medical = Tag.objects.create(
+        name="medical",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    t_hr = Tag.objects.create(name="hr", matching_algorithm=MatchingModel.MATCH_AUTO)
+
+    # non-AUTO labels to keep the other classifiers from raising
+    c = Correspondent.objects.create(
+        name="org",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+    dt = DocumentType.objects.create(
+        name="doc",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+    sp = StoragePath.objects.create(
+        name="archive",
+        path="archive",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+
+    checksum = 0
+
+    def make(title, content, tags):
+        nonlocal checksum
+        checksum += 1
+        return _make_doc(
+            title,
+            content,
+            f"{checksum:04d}",
+            tags=tags,
+            correspondent=c,
+            document_type=dt,
+            storage_path=sp,
+        )
+
+    # 10 single-tag training docs per tag
+    for i in range(10):
+        make(f"finance-{i}", _words(FINANCE_WORDS, f"doc{i}"), [t_finance])
+        make(f"legal-{i}", _words(LEGAL_WORDS, f"doc{i}"), [t_legal])
+        make(f"medical-{i}", _words(MEDICAL_WORDS, f"doc{i}"), [t_medical])
+        make(f"hr-{i}", _words(HR_WORDS, f"doc{i}"), [t_hr])
+
+    # 5 two-tag training docs
+    for i in range(5):
+        make(
+            f"finance-legal-{i}",
+            _words(FINANCE_WORDS + " " + LEGAL_WORDS, f"combo{i}"),
+            [t_finance, t_legal],
+        )
+
+    return {
+        "t_finance": t_finance,
+        "t_legal": t_legal,
+        "t_medical": t_medical,
+        "t_hr": t_hr,
+    }
+
+
+@pytest.mark.django_db()
+class TestMultiTagCorrectness:
+    """
+    The tags classifier must correctly predict tags on held-out documents whose
+    content clearly belongs to one or two vocabulary clusters.
+
+    A prediction is "correct" if the expected tag is present in the result.
+    """
+
+    def test_single_cluster_docs_predicted_correctly(
+        self,
+        classifier,
+        multi_tag_corpus,
+    ):
+        """Each single-cluster held-out doc gets exactly the right tag."""
+        classifier.train()
+        tags = multi_tag_corpus
+
+        cases = [
+            (FINANCE_WORDS + " unique alpha", [tags["t_finance"].pk]),
+            (LEGAL_WORDS + " unique beta", [tags["t_legal"].pk]),
+            (MEDICAL_WORDS + " unique gamma", [tags["t_medical"].pk]),
+            (HR_WORDS + " unique delta", [tags["t_hr"].pk]),
+        ]
+
+        for content, expected_pks in cases:
+            predicted = classifier.predict_tags(content)
+            for pk in expected_pks:
+                assert pk in predicted, (
+                    f"Expected tag pk={pk} in predictions for content starting "
+                    f"'{content[:40]}…', got {predicted}"
+                )
+
+    def test_multi_cluster_doc_gets_both_tags(self, classifier, multi_tag_corpus):
+        """A document with finance + legal vocabulary gets both tags."""
+        classifier.train()
+        tags = multi_tag_corpus
+
+        content = FINANCE_WORDS + " " + LEGAL_WORDS + " unique epsilon"
+        predicted = classifier.predict_tags(content)
+
+        assert tags["t_finance"].pk in predicted, f"Expected finance tag in {predicted}"
+        assert tags["t_legal"].pk in predicted, f"Expected legal tag in {predicted}"
+
+    def test_unrelated_content_predicts_no_trained_tags(
+        self,
+        classifier,
+        multi_tag_corpus,
+    ):
+        """
+        Completely alien content should not confidently fire any learned tag.
+        This is a soft check — we only assert no false positives on a document
+        that shares zero vocabulary with the training corpus.
+        """
+        classifier.train()
+
+        alien = (
+            "xyzzyx qwerty asdfgh zxcvbn plokij unique zeta "
+            "xyzzyx qwerty asdfgh zxcvbn plokij unique zeta"
+        )
+        predicted = classifier.predict_tags(alien)
+        # Not a hard requirement — just log for human inspection
+        # Both MLP and SVM may or may not produce false positives on OOV content
+        assert isinstance(predicted, list)
+
+
+# ---------------------------------------------------------------------------
+# Single-tag (binary) correctness
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture()
+def single_tag_corpus(classifier_settings):
+    """
+    Corpus with exactly ONE AUTO tag, exercising the LabelBinarizer +
+    binary classification path.  Documents either have the tag or don't.
+    """
+    t_finance = Tag.objects.create(
+        name="finance",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    c = Correspondent.objects.create(
+        name="org",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+    dt = DocumentType.objects.create(
+        name="doc",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+
+    checksum = 0
+
+    def make(title, content, tags):
+        nonlocal checksum
+        checksum += 1
+        return _make_doc(
+            title,
+            content,
+            f"s{checksum:04d}",
+            tags=tags,
+            correspondent=c,
+            document_type=dt,
+        )
+
+    for i in range(10):
+        make(f"finance-{i}", _words(FINANCE_WORDS, f"s{i}"), [t_finance])
+        make(f"other-{i}", _words(LEGAL_WORDS, f"s{i}"), [])
+
+    return {"t_finance": t_finance}
+
+
+@pytest.mark.django_db()
+class TestSingleTagCorrectness:
+    def test_finance_content_predicts_finance_tag(self, classifier, single_tag_corpus):
+        """Finance vocabulary → finance tag predicted."""
+        classifier.train()
+        tags = single_tag_corpus
+
+        predicted = classifier.predict_tags(FINANCE_WORDS + " unique alpha single")
+        assert tags["t_finance"].pk in predicted, (
+            f"Expected finance tag pk={tags['t_finance'].pk} in {predicted}"
+        )
+
+    def test_non_finance_content_predicts_no_tag(self, classifier, single_tag_corpus):
+        """Non-finance vocabulary → no tag predicted."""
+        classifier.train()
+
+        predicted = classifier.predict_tags(LEGAL_WORDS + " unique beta single")
+        assert predicted == [], f"Expected no tags, got {predicted}"

src/documents/tests/test_classifier_train_skip.py

@@ -0,0 +1,325 @@
+"""
+Phase 1 — fast-skip optimisation in DocumentClassifier.train()
+
+The goal: when nothing has changed since the last training run, train() should
+return False after at most 5 DB queries (1x MAX(modified) + 4x MATCH_AUTO pk
+lists), not after a full per-document label scan.
+
+Correctness invariant: the skip must NOT fire when the set of AUTO-matching
+labels has changed, even if no Document.modified timestamp has advanced (e.g.
+a Tag's matching_algorithm was flipped to MATCH_AUTO after the last train).
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import pytest
+from django.db import connection
+from django.test.utils import CaptureQueriesContext
+
+from documents.classifier import DocumentClassifier
+from documents.models import Correspondent
+from documents.models import Document
+from documents.models import DocumentType
+from documents.models import MatchingModel
+from documents.models import StoragePath
+from documents.models import Tag
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture()
+def classifier_settings(settings, tmp_path: Path):
+    """Point MODEL_FILE at a temp directory so tests are hermetic."""
+    settings.MODEL_FILE = tmp_path / "model.pickle"
+    return settings
+
+
+@pytest.fixture()
+def classifier(classifier_settings):
+    """Fresh DocumentClassifier instance with test settings active."""
+    return DocumentClassifier()
+
+
+@pytest.fixture()
+def label_corpus(classifier_settings):
+    """
+    Minimal label + document corpus that produces a trainable classifier.
+
+    Creates
+    -------
+    Correspondents
+        c_auto   — MATCH_AUTO, assigned to two docs
+        c_none   — MATCH_NONE (control)
+    DocumentTypes
+        dt_auto  — MATCH_AUTO, assigned to two docs
+        dt_none  — MATCH_NONE (control)
+    Tags
+        t_auto   — MATCH_AUTO, applied to two docs
+        t_none   — MATCH_NONE (control, applied to one doc but never learned)
+    StoragePaths
+        sp_auto  — MATCH_AUTO, assigned to two docs
+        sp_none  — MATCH_NONE (control)
+
+    Documents
+        doc_a, doc_b — assigned AUTO labels above
+        doc_c        — control doc (MATCH_NONE labels only)
+
+    The fixture returns a dict with all created objects for direct mutation in
+    individual tests.
+    """
+    c_auto = Correspondent.objects.create(
+        name="Auto Corp",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    c_none = Correspondent.objects.create(
+        name="Manual Corp",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+
+    dt_auto = DocumentType.objects.create(
+        name="Invoice",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    dt_none = DocumentType.objects.create(
+        name="Other",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+
+    t_auto = Tag.objects.create(
+        name="finance",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    t_none = Tag.objects.create(
+        name="misc",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+
+    sp_auto = StoragePath.objects.create(
+        name="Finance Path",
+        path="finance/{correspondent}",
+        matching_algorithm=MatchingModel.MATCH_AUTO,
+    )
+    sp_none = StoragePath.objects.create(
+        name="Other Path",
+        path="other/{correspondent}",
+        matching_algorithm=MatchingModel.MATCH_NONE,
+    )
+
+    doc_a = Document.objects.create(
+        title="Invoice from Auto Corp Jan",
+        content="quarterly invoice payment tax financial statement revenue",
+        correspondent=c_auto,
+        document_type=dt_auto,
+        storage_path=sp_auto,
+        checksum="aaa",
+        mime_type="application/pdf",
+        filename="invoice_a.pdf",
+    )
+    doc_a.tags.set([t_auto])
+
+    doc_b = Document.objects.create(
+        title="Invoice from Auto Corp Feb",
+        content="monthly invoice billing statement account balance due",
+        correspondent=c_auto,
+        document_type=dt_auto,
+        storage_path=sp_auto,
+        checksum="bbb",
+        mime_type="application/pdf",
+        filename="invoice_b.pdf",
+    )
+    doc_b.tags.set([t_auto])
+
+    # Control document — no AUTO labels, but has enough content to vectorize
+    doc_c = Document.objects.create(
+        title="Miscellaneous Notes",
+        content="meeting notes agenda discussion summary action items follow",
+        correspondent=c_none,
+        document_type=dt_none,
+        checksum="ccc",
+        mime_type="application/pdf",
+        filename="notes_c.pdf",
+    )
+    doc_c.tags.set([t_none])
+
+    return {
+        "c_auto": c_auto,
+        "c_none": c_none,
+        "dt_auto": dt_auto,
+        "dt_none": dt_none,
+        "t_auto": t_auto,
+        "t_none": t_none,
+        "sp_auto": sp_auto,
+        "sp_none": sp_none,
+        "doc_a": doc_a,
+        "doc_b": doc_b,
+        "doc_c": doc_c,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Happy-path skip tests
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.django_db()
+class TestFastSkipFires:
+    """The no-op path: nothing changed, so the second train() is skipped."""
+
+    def test_first_train_returns_true(self, classifier, label_corpus):
+        """First train on a fresh classifier must return True (did work)."""
+        assert classifier.train() is True
+
+    def test_second_train_returns_false(self, classifier, label_corpus):
+        """Second train with no changes must return False (skipped)."""
+        classifier.train()
+        assert classifier.train() is False
+
+    def test_fast_skip_runs_minimal_queries(self, classifier, label_corpus):
+        """
+        The no-op path must use at most 5 DB queries:
+          1x Document.objects.aggregate(Max('modified'))
+          4x MATCH_AUTO pk lists  (Correspondent / DocumentType / Tag / StoragePath)
+
+        The current implementation (before Phase 1) iterates every document
+        to build the label hash BEFORE it can decide to skip, which is O(N).
+        This test verifies the fast path is in place.
+        """
+        classifier.train()
+        with CaptureQueriesContext(connection) as ctx:
+            result = classifier.train()
+        assert result is False
+        assert len(ctx.captured_queries) <= 5, (
+            f"Fast skip used {len(ctx.captured_queries)} queries; expected ≤5.\n"
+            + "\n".join(q["sql"] for q in ctx.captured_queries)
+        )
+
+    def test_fast_skip_refreshes_cache_keys(self, classifier, label_corpus):
+        """
+        Even on a skip, the cache keys must be refreshed so that the task
+        scheduler can detect the classifier is still current.
+        """
+        from django.core.cache import cache
+
+        from documents.caching import CLASSIFIER_HASH_KEY
+        from documents.caching import CLASSIFIER_MODIFIED_KEY
+        from documents.caching import CLASSIFIER_VERSION_KEY
+
+        classifier.train()
+        # Evict the keys to prove skip re-populates them
+        cache.delete(CLASSIFIER_MODIFIED_KEY)
+        cache.delete(CLASSIFIER_HASH_KEY)
+        cache.delete(CLASSIFIER_VERSION_KEY)
+
+        result = classifier.train()
+
+        assert result is False
+        assert cache.get(CLASSIFIER_MODIFIED_KEY) is not None
+        assert cache.get(CLASSIFIER_HASH_KEY) is not None
+        assert cache.get(CLASSIFIER_VERSION_KEY) is not None
+
+
+# ---------------------------------------------------------------------------
+# Correctness tests — skip must NOT fire when the world has changed
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.django_db()
+class TestFastSkipDoesNotFire:
+    """The skip guard must yield to a full retrain whenever labels change."""
+
+    def test_document_content_modification_triggers_retrain(
+        self,
+        classifier,
+        label_corpus,
+    ):
+        """Updating a document's content updates modified → retrain required."""
+        classifier.train()
+        doc_a = label_corpus["doc_a"]
+        doc_a.content = "completely different words here now nothing same"
+        doc_a.save()
+        assert classifier.train() is True
+
+    def test_document_label_reassignment_triggers_retrain(
+        self,
+        classifier,
+        label_corpus,
+    ):
+        """
+        Reassigning a document to a different AUTO correspondent (touching
+        doc.modified) must trigger a retrain.
+        """
+        c_auto2 = Correspondent.objects.create(
+            name="Second Auto Corp",
+            matching_algorithm=MatchingModel.MATCH_AUTO,
+        )
+        classifier.train()
+        doc_a = label_corpus["doc_a"]
+        doc_a.correspondent = c_auto2
+        doc_a.save()
+        assert classifier.train() is True
+
+    def test_matching_algorithm_change_on_assigned_tag_triggers_retrain(
+        self,
+        classifier,
+        label_corpus,
+    ):
+        """
+        Flipping a tag's matching_algorithm to MATCH_AUTO after it is already
+        assigned to documents must trigger a retrain — even though no
+        Document.modified timestamp advances.
+
+        This is the key correctness case for the auto-label-set digest:
+        the tag is already on doc_a and doc_b, so once it becomes MATCH_AUTO
+        the classifier needs to learn it.
+        """
+        # t_none is applied to doc_c (a control doc) via the fixture.
+        # We flip it to MATCH_AUTO; the set of learnable AUTO tags grows.
+        classifier.train()
+        t_none = label_corpus["t_none"]
+        t_none.matching_algorithm = MatchingModel.MATCH_AUTO
+        t_none.save(update_fields=["matching_algorithm"])
+        # Document.modified is NOT touched — this test specifically verifies
+        # that the auto-label-set digest catches the change.
+        assert classifier.train() is True
+
+    def test_new_auto_correspondent_triggers_retrain(self, classifier, label_corpus):
+        """
+        Adding a brand-new MATCH_AUTO correspondent (unassigned to any doc)
+        must trigger a retrain: the auto-label-set has grown.
+        """
+        classifier.train()
+        Correspondent.objects.create(
+            name="New Auto Corp",
+            matching_algorithm=MatchingModel.MATCH_AUTO,
+        )
+        assert classifier.train() is True
+
+    def test_removing_auto_label_triggers_retrain(self, classifier, label_corpus):
+        """
+        Deleting a MATCH_AUTO correspondent shrinks the auto-label-set and
+        must trigger a retrain.
+        """
+        classifier.train()
+        label_corpus["c_auto"].delete()
+        assert classifier.train() is True
+
+    def test_fresh_classifier_always_trains(self, classifier, label_corpus):
+        """
+        A classifier that has never been trained (last_doc_change_time is None)
+        must always perform a full train, regardless of corpus state.
+        """
+        assert classifier.last_doc_change_time is None
+        assert classifier.train() is True
+
+    def test_no_documents_raises_value_error(self, classifier, classifier_settings):
+        """train() with an empty database must raise ValueError."""
+        with pytest.raises(ValueError, match="No training data"):
+            classifier.train()