EFU divergence census: recompiler vs interpreter, 13 sites, no oracle
====================================================================
The EFU has no hardware-verified truth model, so unlike the FMAC and
Q censuses this quantifies DISAGREEMENT between the two engines, not
error against hardware. 400 cases per site, executed-effect
confirmed on every row (P pipeline readback normalized alongside Q).
"max dist" is the crude same-sign magnitude distance in raw-bit
space (a mantissa is 0x7FFFFF; 1e9-scale means sign or exponent
disagreement).

site      differ /400   max dist        tier
ESQRT          83              1        near-agreement (one ulp)
EEXP          337             20        near-agreement
ERSQRT        211        8388607        mantissa-scale
ERCPR         210        8388607        mantissa-scale
ESIN          346        8388608        mantissa-scale (boundary)
ESADD         391     2087018409        structural
ERSADD        387     1911999125        structural
ELENG         391     1602224127        structural
ERLENG        392     1257070360        structural
ESUM          374     1137448281        structural
EATANxy       400     1070141401        structural
EATANxz       400     1070141401        structural
EATAN         400     1068953638        structural

Reading:
- The structural tier (the sum-of-squares reductions and the whole
  atan family) disagrees at sign/exponent scale on nearly every
  case: at least one engine is far from hardware there, and these
  are the sites where a hardware truth model would pay first.
- The near-agreement tier (ESQRT one ulp; EEXP within ~20) says
  both implementations converge on the same approximation for
  those; hardware verification is low priority.
- No verdicts: without hardware captures neither engine is the
  oracle for the EFU. The campaign's truth-model acquisition for
  these thirteen sites is the standing prerequisite for turning
  this table into a correctness census.
