feat(ml): presentation-chrome auto-hide sweep + hard-skip + conflict flagging (#141 step 4)

presentation_auto_apply_sweep fires banner/editor-screenshot heads at the FLAT
presentation threshold (source=presentation_auto). Two guards: (1) hard-skip any
image already carrying a human/confirmed content tag — you valued it, so the model
can't bury it; (2) if an auto-hide ALSO scores >= presentation_conflict_threshold
on a content head, hide it but record a PresentationReview row (conflict tag +
score) for the Hidden view.

_auto_apply_heads now excludes system tags, so a graduated wip/banner can't fire
via the content path (and wip never auto-applies at all). presentation_auto added
to _AUTO_SOURCES so auto-hidden chrome never self-trains. Tests: applies,
hard-skip valued, conflict-flag, disabled no-op, ignores wip, content-path
excludes system. Settings UI + scheduling land 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/heads.py

@@ -32,13 +32,14 @@ from ...models import (
     ImageRecord,
     ImageRegion,
     MLSettings,
+    PresentationReview,
     Tag,
     TagHead,
     TagKind,
     TagPositiveConfirmation,
     TagSuggestionRejection,
 )
-from ...models.tag import image_tag
+from ...models.tag import PRESENTATION_SYSTEM_TAGS, image_tag
 from .training_data import (
     _AUTO_SOURCES,
     _auto_apply_point,
@@ -649,8 +650,11 @@ def start_head_auto_apply_run(session: Session, params: dict[str, Any]) -> int:
 
 
 def _auto_apply_heads(session: Session, embedding_version: str, min_pos: int):
-    """Eligible heads to fire: graduated (auto_apply_threshold set), enough
-    support, current embedding. Returns the row list (tag_id/name/weights/...)."""
+    """Eligible CONTENT heads to fire: graduated (auto_apply_threshold set),
+    enough support, current embedding, NON-system. System tags never auto-apply
+    via this path — `wip` never auto-applies at all, and banner/editor screenshot
+    go through the presentation path at their own flat threshold (#141). Returns
+    the row list (tag_id/name/weights/...)."""
     return session.execute(
         select(
             TagHead.tag_id, Tag.name, TagHead.weights, TagHead.bias,
@@ -660,6 +664,7 @@ def _auto_apply_heads(session: Session, embedding_version: str, min_pos: int):
         .where(TagHead.embedding_version == embedding_version)
         .where(TagHead.auto_apply_threshold.is_not(None))
         .where(TagHead.n_pos >= min_pos)
+        .where(~Tag.is_system)
     ).all()
 
 
@@ -743,6 +748,162 @@ def auto_apply_sweep(
     return {"n_applied": sum(applied), "concepts": concepts}
 
 
+_PRESENTATION_SOURCE = "presentation_auto"
+
+
+def _presentation_heads(session: Session, embedding_version: str):
+    """Trained heads for the presentation chrome tags (banner / editor screenshot).
+    They fire at the FLAT presentation threshold regardless of graduation — a head
+    exists once the operator has labelled enough chrome (head_min_positives)."""
+    return session.execute(
+        select(TagHead.tag_id, Tag.name, TagHead.weights, TagHead.bias)
+        .join(Tag, Tag.id == TagHead.tag_id)
+        .where(TagHead.embedding_version == embedding_version)
+        .where(Tag.is_system.is_(True))
+        .where(Tag.name.in_(PRESENTATION_SYSTEM_TAGS))
+    ).all()
+
+
+def _conflict_heads(session: Session, embedding_version: str):
+    """ALL content (non-system) heads — the "does this ALSO look like real
+    content" signal for the presentation conflict guard (#141)."""
+    return session.execute(
+        select(TagHead.tag_id, TagHead.weights, TagHead.bias)
+        .join(Tag, Tag.id == TagHead.tag_id)
+        .where(TagHead.embedding_version == embedding_version)
+        .where(~Tag.is_system)
+    ).all()
+
+
+def _valued_image_ids(session: Session) -> set[int]:
+    """Images the operator has shown they value: carrying a HUMAN or CONFIRMED
+    content (non-system) tag. The presentation sweep never auto-hides these
+    (guard 1) — you tagged it, so the model doesn't get to bury it (#141)."""
+    confirmed = exists().where(
+        TagPositiveConfirmation.image_record_id == image_tag.c.image_record_id,
+        TagPositiveConfirmation.tag_id == image_tag.c.tag_id,
+    )
+    rows = session.execute(
+        select(image_tag.c.image_record_id)
+        .join(Tag, Tag.id == image_tag.c.tag_id)
+        .where(~Tag.is_system)
+        .where(image_tag.c.source.not_in(_AUTO_SOURCES) | confirmed)
+    ).all()
+    return {r[0] for r in rows}
+
+
+def presentation_auto_apply_sweep(session: Session, dry_run: bool = False) -> dict:
+    """Auto-hide presentation chrome (banner / editor screenshot) at the FLAT
+    presentation threshold (#141) — NOT the per-head graduated threshold. Two
+    guards keep it safe: (1) never hide an image carrying a human/confirmed content
+    tag; (2) if an image about to be hidden ALSO scores >= the conflict threshold
+    on a content head, still hide it but flag it (PresentationReview) so the Hidden
+    view surfaces "also looks like <X>" for review. No-op unless
+    presentation_auto_apply_enabled. numpy-only (no sklearn). Returns
+    {n_applied, n_flagged, concepts}."""
+    import numpy as np
+    from sqlalchemy.dialects.postgresql import insert as pg_insert
+
+    settings = _settings(session)
+    if not dry_run and not settings.presentation_auto_apply_enabled:
+        return {"n_applied": 0, "n_flagged": 0, "concepts": []}
+    ver = settings.embedder_model_version
+    pres = _presentation_heads(session, ver)
+    if not pres:
+        return {"n_applied": 0, "n_flagged": 0, "concepts": []}
+    thr = float(settings.presentation_auto_apply_threshold)
+    conflict_thr = float(settings.presentation_conflict_threshold)
+
+    Wp = np.vstack([np.asarray(r.weights, dtype=np.float32) for r in pres])
+    bp = np.asarray([r.bias for r in pres], dtype=np.float32)
+    pres_tag_ids = [r.tag_id for r in pres]
+    pres_names = [r.name for r in pres]
+
+    conf = _conflict_heads(session, ver)
+    Wc = bc = conf_tag_ids = None
+    if conf:
+        Wc = np.vstack([np.asarray(r.weights, dtype=np.float32) for r in conf])
+        bc = np.asarray([r.bias for r in conf], dtype=np.float32)
+        conf_tag_ids = [r.tag_id for r in conf]
+
+    valued = _valued_image_ids(session)
+
+    # Skip images that already carry, or have rejected, each presentation tag.
+    skip = {tid: set() for tid in pres_tag_ids}
+    for tid in pres_tag_ids:
+        for (iid,) in session.execute(
+            select(image_tag.c.image_record_id).where(image_tag.c.tag_id == tid)
+        ):
+            skip[tid].add(iid)
+        for (iid,) in session.execute(
+            select(TagSuggestionRejection.image_record_id).where(
+                TagSuggestionRejection.tag_id == tid
+            )
+        ):
+            skip[tid].add(iid)
+
+    applied = [0] * len(pres)
+    n_flagged = 0
+    scanned = 0
+    all_ids = list(session.execute(
+        select(ImageRecord.id).where(ImageRecord.siglip_embedding.is_not(None))
+    ).scalars())
+    for start in range(0, len(all_ids), _AUTO_APPLY_CHUNK):
+        chunk = all_ids[start:start + _AUTO_APPLY_CHUNK]
+        emb = _load_embeddings(session, chunk)
+        cids = [i for i in chunk if i in emb]
+        if not cids:
+            continue
+        Xn = _l2norm(np.vstack([emb[i] for i in cids]).astype(np.float32), np)
+        probs = 1.0 / (1.0 + np.exp(-(Xn @ Wp.T + bp)))       # (N, P)
+        if Wc is not None:
+            cprobs = 1.0 / (1.0 + np.exp(-(Xn @ Wc.T + bc)))  # (N, C)
+            max_c = cprobs.max(axis=1)
+            arg_c = cprobs.argmax(axis=1)
+        scanned += len(cids)
+        for p in range(len(pres)):
+            tid = pres_tag_ids[p]
+            for idx in np.where(probs[:, p] >= thr)[0]:
+                iid = cids[int(idx)]
+                if iid in skip[tid] or iid in valued:
+                    continue
+                skip[tid].add(iid)
+                applied[p] += 1
+                if not dry_run:
+                    session.execute(
+                        pg_insert(image_tag)
+                        .values(
+                            image_record_id=iid, tag_id=tid,
+                            source=_PRESENTATION_SOURCE,
+                        )
+                        .on_conflict_do_nothing()
+                    )
+                # Guard 2: also looks like content → hide but flag for review.
+                if Wc is not None and float(max_c[idx]) >= conflict_thr:
+                    n_flagged += 1
+                    if not dry_run:
+                        session.execute(
+                            pg_insert(PresentationReview)
+                            .values(
+                                image_record_id=iid, tag_id=tid,
+                                conflict_tag_id=conf_tag_ids[int(arg_c[idx])],
+                                conflict_score=float(max_c[idx]),
+                            )
+                            .on_conflict_do_nothing()
+                        )
+        if not dry_run:
+            session.commit()
+
+    concepts = [
+        {"tag_id": pres_tag_ids[p], "name": pres_names[p],
+         "applied": applied[p], "scanned": scanned, "threshold": thr}
+        for p in range(len(pres))
+    ]
+    return {
+        "n_applied": sum(applied), "n_flagged": n_flagged, "concepts": concepts,
+    }
+
+
 def retract_auto_applied_heads(session: Session) -> int:
     """Soft auto-apply (milestone 139): re-score every standing source='head_auto'
     tag against its CURRENT head and REMOVE the ones now BELOW the head's

backend/app/services/ml/training_data.py

@@ -29,7 +29,7 @@ from ...models.tag import image_tag
 # 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")
+_AUTO_SOURCES = ("head_auto", "ccip_auto", "ml_auto", "presentation_auto")
 
 
 def _hygiene_excluded_ids(session: Session) -> set[int]: