fix(compare): census-sourced FSM/EAL benchmarks; never fall back across measure definitions

The FSM chip anchored against disadvantaged_pct (a different measure,
FSM6+CLA) whenever fsm_pct was null — which it always was, since the
performance df has no fsm_pct. New fact_census_benchmarks mart supplies
pupil-weighted FSM/EAL means per phase; the KS2-column medians that
produced a bogus 50% 'secondary disadvantaged' anchor are gone.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0146VHeLAWjDVE2B5uU67jCB

backend/models.py

@@ -234,6 +234,22 @@ class FactFinance(Base):
     premises_cost_pct = Column(Float)
 
 
+class CensusBenchmark(Base):
+    """State-school context benchmarks from the pupil census — one row per phase.
+
+    fsm_pct / eal_pct are pupil-weighted means. Computed at import time;
+    consumers label them "state-school average (computed from our dataset)".
+    """
+    __tablename__ = "fact_census_benchmarks"
+    __table_args__ = MARTS
+
+    phase = Column(String(20), primary_key=True)
+    year = Column(Integer)
+    fsm_pct = Column(Float)
+    eal_pct = Column(Float)
+    median_pupils = Column(Integer)
+
+
 class Ks4NationalAverage(Base):
     """Computed national KS4 averages (from our dataset) — one row per year."""
     __tablename__ = "fact_ks4_national_averages"