feat(ml): auto-applied tags don't train a head unless confirmed (milestone 139)

Makes auto-apply truly "soft" for heads: _ids_with_tag (head positives) and
_eligible_tag_ids (graduation count) now count human-applied + operator-confirmed
tags only, via a shared _AUTO_SOURCES (head_auto/ccip_auto/ml_auto) exclusion.
Unconfirmed auto-applied tags no longer train the head that judges them, so a
misfire can't reinforce itself and the retraction sweep can actually drop it.
Confirming a tag (TagPositiveConfirmation) promotes it to a positive AND protects
it from retraction. sklearn-free tests. CCIP reference exclusion is the companion
piece, next.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CDgx8bQS5YrGRK76v8HUnM

backend/app/services/ml/training_data.py

@@ -17,9 +17,20 @@ from typing import Any
 from sqlalchemy import func, select
 from sqlalchemy.orm import Session
 
-from ...models import ImageRecord, Tag, TagSuggestionRejection
+from ...models import (
+    ImageRecord,
+    Tag,
+    TagPositiveConfirmation,
+    TagSuggestionRejection,
+)
 from ...models.tag import image_tag
 
+# Auto-apply sources whose tags are PROVISIONAL: they never train a head (or seed
+# a CCIP reference) unless the operator confirms them (milestone 139). Keeping
+# auto-applied predictions out of training is what makes them "soft" — a misfire
+# can't reinforce itself, so the retraction sweep can actually drop it.
+_AUTO_SOURCES = ("head_auto", "ccip_auto", "ml_auto")
+
 
 def _hygiene_excluded_ids(session: Session) -> set[int]:
     """Ids of images carrying ANY system tag (wip / banner / editor
@@ -45,9 +56,23 @@ def _hygiene_excluded_ids(session: Session) -> set[int]:
 
 
 def _ids_with_tag(session: Session, tag_id: int) -> list[int]:
+    """Image ids that count as POSITIVES for this tag's head: human-applied
+    (manual / accepted) tags PLUS any auto-applied tag the operator explicitly
+    confirmed (TagPositiveConfirmation). Unconfirmed auto-applied tags are
+    EXCLUDED — they are provisional and must not train the head that judges
+    them (milestone 139)."""
+    confirmed = (
+        select(TagPositiveConfirmation.image_record_id)
+        .where(TagPositiveConfirmation.tag_id == tag_id)
+    )
     return [
         r[0] for r in session.execute(
-            select(image_tag.c.image_record_id).where(image_tag.c.tag_id == tag_id)
+            select(image_tag.c.image_record_id)
+            .where(image_tag.c.tag_id == tag_id)
+            .where(
+                image_tag.c.source.not_in(_AUTO_SOURCES)
+                | image_tag.c.image_record_id.in_(confirmed)
+            )
         ).all()
     ]