[punithbm]

Part of #1274: the first two points.

The flow meter behind estimatesmartfee, Electrum and Esplora counted a steady stream at about half its rate, so near targets could quote low. note_admit decayed to now, which already applies 1 - alpha to every bucket, then applied 1 - alpha to the updated bucket again. 10,000 WU every second read 5,012 WU/s after 20 half-lives and now reads 10,000.

The gap a sample is divided by was measured from the last decay, and fullness and admit_rates_wu_s decay, so a fee refresh between admits changed the next sample. The gap is now measured from the previous admit.

two_admits_pin_the_bucket_ema moves from 91 to 92: 46.0 after the first admit decays to 45.58, plus 46.0 from the second.

CPFP members at the chunk rate is not in this PR. #1333 reworks the package admit path that change goes in, so it follows once that lands.

[rearden-grok]

APPROVE. The updated bucket was decayed twice, and the sample gap followed the decay clock, so a fee refresh shortened it. note_admit now decays once, then adds sample * alpha, and takes the gap from last_admit.

Float64 recurrence of the same update: two admits move from 91.151 (rounds to 91) to 91.571 (rounds to 92). A one-second stream of 10,000 WU settles at 5011.55 under the old double decay (rounds to 5,012) and at 9999.990 after 3,000 steps of this update (rounds to 10,000). The new steady-stream and snapshot tests both fail on the old formula. The added tests are a few thousand arithmetic steps with no sleeps.