Auditing: Widen tranche-escalation comparand to 16 bits

[philoniare]

Issue: The later-tranche audit rule (auditing.tex, eq. latertranches) compares a single byte of VRF output against the no-show count. The bias factor defined right under it is F = 2 — "the expected number of validators … required to issue a judgment … given a single no-show." A single byte can't actually deliver 2: at one no-show only the byte value 0 passes the test, so the mean number summoned is 1023/256 ≈ 3.996, about double. It's not a rounding thing either — with one byte every reachable mean is a multiple of 1023/256 ≈ 3.996, and 2 isn't on that grid.

Fix: decode bytes 0–1 of the same VRF output as a little-endian u16.

Explanation:
correction.pdf

POC — exact, by enumeration:

from fractions import Fraction
V, F = 1023, 2
def expected_summons(bits, m):
    space = 2 ** bits
    q = sum(1 for r in range(space) if Fraction(V * r, space * F) < m)
    return V * Fraction(q, space)
assert expected_summons(8, 1)  == Fraction(1023, 256)       # 3.996
assert expected_summons(16, 1) == Fraction(131967, 65536)   # 2.014