Before the whole-union experiment.
The phase/marked tests address inputs to the local reduction but do not match
Post65's whole-union H6. Test the actual deduplicated union F_b directly.
At record slope, height2, b8,10,12, enumerate all actual phase cells and
deduplicate the exact words. Use common binary period Q<=2^23, the same
two original windows floor(sqrt(Q)) andQ//64+1, and normalized translation.
Calculate the same exact predictable-variation diagnostics. No GPU; one
process120seconds; int64 guards remain mandatory. If the guard fails,
record that computational limit instead of quietly changing arithmetic.

An additional elementary Doob/Hölder argument gives
 E(T-mu)^6 <=32768 E V^3.
Thus a bound E V^3<=900mu^3 would give H6(C) with exactC1966080<2^21.
This is below Wren's REPORTED NUMERICAL budget2^21.4, which has not been
independently certified here. Do not claim a cycle exclusion from that
numerical budget or from the toy experiments.
