feat(ml): training hygiene — system-tagged images are absent from other concepts training

Step 2 of milestone #128. _hygiene_excluded_ids (training_data.py) is the
one shared predicate: images carrying any system tag are dropped from
every OTHER concepts head training — not positives (a rough wip tagged
as a character drags the head toward generic-sketch) and not rejection
or sampled negatives (a wip OF character X is not evidence against X).
A system tags own head trains on them unfiltered; that is what makes
auto-flagging banners work. Selection is split out of train_head as the
sklearn-free head_training_ids so CI (no sklearn) can pin the behavior.

CCIP: reference prototypes skip hygiene-tagged images — a faceless wip
figure region must never become an identity reference — and the ref
cache signature now counts hygiene applications, since tagging an image
wip changes the reference set without touching character/region counts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CDgx8bQS5YrGRK76v8HUnM

backend/app/services/ml/training_data.py

@@ -17,10 +17,33 @@ from typing import Any
 from sqlalchemy import func, select
 from sqlalchemy.orm import Session
 
-from ...models import ImageRecord, TagSuggestionRejection
+from ...models import ImageRecord, Tag, TagSuggestionRejection
 from ...models.tag import image_tag
 
 
+def _hygiene_excluded_ids(session: Session) -> set[int]:
+    """Ids of images carrying ANY system tag (wip / banner / editor
+    screenshot — milestone #128). These images are excluded from OTHER
+    concepts' head training entirely: not positives (a rough wip tagged as a
+    character drags that head toward 'generic sketch') and not sampled or
+    rejection negatives (a wip OF character X is not evidence against X) —
+    simply absent. A system tag's OWN head trains on them unchanged; that is
+    what makes auto-flagging banners/editor screenshots work.
+
+    Item-level by design: a wip-tagged process video contributes (or
+    withholds) ALL its sampled frames, though some may show the finished
+    piece. Operator call 2026-07-03: with enough clean data this washes out —
+    no per-frame handling.
+    """
+    return set(
+        session.execute(
+            select(image_tag.c.image_record_id)
+            .join(Tag, Tag.id == image_tag.c.tag_id)
+            .where(Tag.is_system.is_(True))
+        ).scalars().all()
+    )
+
+
 def _ids_with_tag(session: Session, tag_id: int) -> list[int]:
     return [
         r[0] for r in session.execute(