noise_floor.py

noise_floor — how precise is each observable, before anyone interprets a difference in it?

THE MISTAKE THIS PREVENTS, WHICH WAS ALREADY MADE ONCE
------------------------------------------------------
§5.6b of the whitepaper reported that H differs between GC and MGC on matched dates, concluded flow
persistence tracks participant mix, and was published. Two more matched windows flipped the sign. The
"effect" was 0.027 against a window-to-window spread that turned out to be ~0.05 within a single
contract — i.e. it was never resolvable, and nobody had measured the noise floor to notice.

So: measure each observable's WITHIN-CONTRACT, ACROSS-WINDOW dispersion first. That number is the
resolution limit of the instrument. Any between-group difference smaller than it is uninterpretable
no matter how clean the arithmetic looks, and any difference much larger than it is real.

Reported per observable:
    mean, SD across windows, and the DETECTION FLOOR (2 SD) -- the smallest difference worth a
    sentence.

Different observables computed from the SAME data can have wildly different precision. Establishing
that separately is the point: it tells you which of your findings you are entitled to believe.