feat(heads): incremental retraining — refit only changed tags (#1317 phase 2, m138)

train_all_heads is now incremental by default: a per-tag training-data
fingerprint (positive + rejection count/latest-timestamp, stored on
tag_head.train_fingerprint) means a manual Retrain refits ONLY the tags whose
data changed — O(what you touched), not O(all heads). The nightly
scheduled_train_heads passes full=True to reconcile sampled-negative + hygiene
drift across every head. First incremental run after deploy still refits
everyone (NULL fingerprints), stamping them, then it's incremental.

The refit decision + fingerprint are split into sklearn-free helpers
(_head_fingerprints, _heads_needing_retrain) so the incremental logic is
unit-tested directly (train_head itself needs scikit-learn). Migration 0080.

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

alembic/versions/0080_tag_head_train_fingerprint.py

@@ -0,0 +1,31 @@
+"""tag_head.train_fingerprint (#1317 phase 2) — incremental head retraining
+
+A per-head training-data fingerprint (positive + rejection count/latest-timestamp)
+so a manual Retrain refits only the tags whose data changed; the nightly run
+ignores it (full reconcile). Nullable — a NULL fingerprint (existing heads) forces
+a refit on the first incremental run, then it's stamped.
+
+Revision ID: 0080
+Revises: 0079
+Create Date: 2026-07-06
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0080"
+down_revision: Union[str, None] = "0079"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "tag_head",
+        sa.Column("train_fingerprint", sa.String(128), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("tag_head", "train_fingerprint")