feat(embeddings): every vector records the model whose space it lives in (#4132)

The four embedding tables stamped chunker_version but not the model, and
vector(384) is a width, not an identity: a same-width model swap would
write a second geometry beside the first with no error.

- embedding_model on note/rule/milestone/system embeddings (0109; existing
  rows stamped with the only model any install has ever run).
- Every write stamps EMBEDDING_MODEL; every backfill's "current" test is
  is_current_stamp(), both halves of calibration_stamp().
- migrate_floor refuses while any row its surface searches is off the live
  model, before sampling: re-embed, then migrate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/test_chunking.py

@@ -228,6 +228,7 @@ async def test_upsert_stores_one_versioned_row_per_chunk():
     assert [r.chunk_index for r in rows] == list(range(len(chunks)))
     assert [r.chunk_text for r in rows] == chunks
     assert {r.chunker_version for r in rows} == {emb.CHUNKER_VERSION}
+    assert {r.embedding_model for r in rows} == {emb.EMBEDDING_MODEL}
     assert {r.user_id for r in rows} == {42}
     session.execute.assert_awaited()  # the delete that makes replacement atomic
 
@@ -314,3 +315,17 @@ async def test_backfill_reembeds_notes_with_a_stale_chunker_version():
 
     embedded = [call.args[0] for call in upsert.call_args_list]
     assert embedded == [2], "only the stale note is re-embedded"
+
+
+def test_a_row_is_current_only_in_the_live_models_space():
+    """#4132: the column is a width, not an identity, so a same-width model
+    swap writes a second geometry with no error. The backfill's "current"
+    predicate has to name BOTH halves of the stamp, on every embedding table."""
+    from scribe.models.embedding import (
+        MilestoneEmbedding, NoteEmbedding, RuleEmbedding, SystemEmbedding,
+    )
+    from scribe.services import embeddings as emb
+
+    for table in (NoteEmbedding, RuleEmbedding, MilestoneEmbedding, SystemEmbedding):
+        clause = str(emb.is_current_stamp(table))
+        assert "chunker_version" in clause and "embedding_model" in clause, table

tests/test_integration_lesson_reach.py

@@ -23,7 +23,7 @@ from scribe.models.embedding import EMBEDDING_DIM, NoteEmbedding
 from scribe.models.note import Note
 from scribe.models.project import Project
 from scribe.services import lessons as lessons_svc
-from scribe.services.embeddings import CHUNKER_VERSION, semantic_search_notes
+from scribe.services.embeddings import CHUNKER_VERSION, EMBEDDING_MODEL, semantic_search_notes
 from tests.helpers import ensure_user
 
 pytestmark = [pytest.mark.integration, pytest.mark.usefixtures("_dispose_engine")]
@@ -71,6 +71,7 @@ async def corpus():
                 note_id=note.id, chunk_index=0, user_id=owner.id,
                 embedding=QUERY_VEC, chunk_text=note.title,
                 chunker_version=CHUNKER_VERSION,
+                embedding_model=EMBEDDING_MODEL,
             ))
         ids = {k: n.id for k, n in rows.items()}
         ids["owner"], ids["a"], ids["b"] = owner.id, a.id, b.id

tests/test_integration_milestone_search.py

@@ -17,7 +17,7 @@ from scribe.models.embedding import EMBEDDING_DIM, MilestoneEmbedding
 from scribe.models.milestone import Milestone
 from scribe.models.project import Project
 from scribe.services import dedup as dedup_svc
-from scribe.services.embeddings import CHUNKER_VERSION, semantic_search_milestones
+from scribe.services.embeddings import CHUNKER_VERSION, EMBEDDING_MODEL, semantic_search_milestones
 from tests.helpers import ensure_user
 
 pytestmark = [pytest.mark.integration, pytest.mark.usefixtures("_dispose_engine", "_no_embedding")]
@@ -48,7 +48,8 @@ async def roadmap():
         await s.flush()
         for ms, vec in ((m3, NEAR), (done, NEAR), (unrelated, FAR), (foreign, NEAR)):
             s.add(MilestoneEmbedding(milestone_id=ms.id, chunk_index=0, embedding=vec,
-                                     chunk_text=ms.title, chunker_version=CHUNKER_VERSION))
+                                     chunk_text=ms.title, chunker_version=CHUNKER_VERSION,
+                                     embedding_model=EMBEDDING_MODEL))
         ids = {"owner": owner.id, "stranger": stranger.id, "mine": mine.id,
                "m3": m3.id, "done": done.id, "unrelated": unrelated.id, "foreign": foreign.id}
         await s.commit()

tests/test_integration_pgvector_search.py

@@ -33,7 +33,7 @@ def _vec(*nonzero_first):
 
 def _emb(note_id, user_id, chunk_index, vec):
     """A chunk row at the current chunker version (#280, migration 0077)."""
-    from scribe.services.embeddings import CHUNKER_VERSION
+    from scribe.services.embeddings import CHUNKER_VERSION, EMBEDDING_MODEL
 
     return NoteEmbedding(
         note_id=note_id,
@@ -42,6 +42,7 @@ def _emb(note_id, user_id, chunk_index, vec):
         embedding=vec,
         chunk_text=f"chunk {chunk_index} of note {note_id}",
         chunker_version=CHUNKER_VERSION,
+        embedding_model=EMBEDDING_MODEL,
     )
 
 

tests/test_integration_rule_scope.py

@@ -18,7 +18,7 @@ from scribe.models.embedding import EMBEDDING_DIM, RuleEmbedding
 from scribe.models.project import Project
 from scribe.models.share import ProjectShare
 from scribe.services import rulebooks as rulebooks_svc
-from scribe.services.embeddings import CHUNKER_VERSION, semantic_search_rules
+from scribe.services.embeddings import CHUNKER_VERSION, EMBEDDING_MODEL, semantic_search_rules
 from tests.helpers import ensure_user
 
 pytestmark = [pytest.mark.integration, pytest.mark.usefixtures("_dispose_engine")]
@@ -66,6 +66,7 @@ async def homes():
             s.add(RuleEmbedding(
                 rule_id=rule.id, chunk_index=0, embedding=QUERY_VEC,
                 chunk_text=rule.title, chunker_version=CHUNKER_VERSION,
+                embedding_model=EMBEDDING_MODEL,
             ))
         await s.commit()
     ids.update(glob=glob.id, on_a=on_a.id, on_b=on_b.id)

tests/test_retrieval_migration.py

@@ -15,6 +15,9 @@ WHAT THIS PINS
      mid-backfill would end up with every bar at zero.
   4. **Every registry surface can be migrated.** A seventh arm that nobody adds
      a re-scorer for is one whose floor silently cannot survive a model change.
+  5. **A half-migrated corpus is a refusal (#4132).** While any row the surface
+     searches is stamped with another model, a re-score measures a blend of two
+     geometries — so nothing is sampled until the backfill has finished.
 """
 from unittest.mock import AsyncMock, MagicMock, patch
 
@@ -43,11 +46,42 @@ def _session_with(rows):
     return session
 
 
+@pytest.fixture(autouse=True)
+def _corpus_on_the_live_model():
+    """Every test below starts from a wholly re-embedded corpus; the one that
+    is about a half-migrated corpus patches this again."""
+    with patch.object(rm, "rows_off_the_live_model", AsyncMock(return_value=0)):
+        yield
+
+
 def test_every_surface_has_a_rescorer():
     """Otherwise a surface's floor cannot cross a model change at all."""
     assert set(rm._RESCORERS) == set(SURFACES)
 
 
+def test_every_surface_names_the_corpus_it_searches():
+    """Otherwise the half-migrated check has no table to count for it."""
+    assert set(rm._CORPUS) == set(SURFACES)
+
+
+@pytest.mark.asyncio
+async def test_a_corpus_part_way_through_a_model_change_refuses_before_sampling():
+    rows = _logs([("q1", None, 0.9), ("q2", None, 0.1)])
+    rescore = AsyncMock(return_value=0.4)
+    with patch.object(rm, "rows_off_the_live_model", AsyncMock(return_value=17)) as off, \
+         patch.object(rm, "async_session", MagicMock(return_value=_session_with(rows))), \
+         patch.object(rm, "floor_for", AsyncMock(return_value=0.5)), \
+         patch.object(rm, "set_dial", AsyncMock()) as set_dial, \
+         patch.dict(rm._RESCORERS, {"prompt_rule": rescore}):
+        out = await rm.migrate_floor(1, "prompt_rule", apply=True)
+
+    assert out["migrated"] is False
+    assert out["rows_off_model"] == 17
+    assert off.await_args.args[0] is rm.RuleEmbedding
+    rescore.assert_not_called()
+    set_dial.assert_not_called()
+
+
 def test_the_floor_that_admits_a_fraction_is_an_observed_score():
     scores = [0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1, 0.05]
     # 30% of ten is three; the third-best score is the bar that admits exactly