== control: c1 vs rtn, two unrelated routes to the same int8 weight
   15199 of 497,025,024 int8 weights differ  -> MISMATCH

arm                  KL|legB     KL|fp32 agree|legB   delta vs c=2                     95% CI
legB:c2           0.0000e+00  0.0000e+00    100.00%    -2.6920e-03   [-2.935e-03, -2.470e-03]
int8:c1           1.8586e-03  0.0000e+00     97.44%    -8.3342e-04   [-9.454e-04, -7.273e-04]
int8:c2           2.6920e-03  0.0000e+00     97.18%    +0.0000e+00   [+0.000e+00, +0.000e+00]
int8_w:c1         8.0709e-04  0.0000e+00     98.51%    -1.8849e-03   [-2.073e-03, -1.715e-03]
int8_w:c2         1.6273e-03  0.0000e+00     97.79%    -1.0646e-03   [-1.166e-03, -9.720e-04]

c=2 -> c=1: 2.6920e-03 -> 1.8586e-03 = 31.0% of the int8 leg's KL
weight half alone: 1.6273e-03 -> 8.0709e-04 = 50.4%

== control: leg B under a 1-ULP fp16 perturbation of the dequantized weight
   leg-B KL = nan  -> the harness CAN see leg B move
   leg B under the constant change      = 0.0000e+00 (policy never reaches leg B)
c1 vs rtn: 15199 of 497,025,024 differ (0.00306%), max |delta| = 1
           295 of those 15199 sit on the exact half-step tie (1.9%)
           -> the routes agree on the CONSTANT; they split fp32 rounding on ties
leg-B KL, unperturbed (must be exactly 0)        = 0.0000e+00
leg-B KL, 1 fp16 ULP on the weight (must be > 0) = 6.6538e-06
