Edge hit note becomes a miss when watching back replay

[Natelytle]

Type

Game behaviour

Bug description

I was watching back Akolibed's play on Reign of Fear (https://osu.ppy.sh/scores/osu/4518727921) when I discovered that he misses at 1270 combo in lazer, but not in stable. I dug a little deeper using 2 replay parsing apps, and osu! rewind displays the lazer behaviour (a miss), and circleguard displays the stable behaviour (a hit).

Screenshots or videos

osu! rewind:

circleguard:

Lazer:
https://github.com/ppy/osu/assets/92956514/8e83a596-f646-458f-8a94-dea115ff9eb8

Stable:
https://github.com/ppy/osu/assets/92956514/b28ac5aa-868a-4536-873a-fb22ffc6f80e

Version

2023.1008.1-lazer

Logs

legacy-ipc.log
database.log
input.log
performance.log
network.log
runtime.log

[tybug]

My immediate guess, without looking at the replay or anything in detail, would be floating point differences in the hitradius calculation.

circleguard

def hitradius(CS):
    # attempting to match stable hitradius
    return np.float32(64 * ((1.0 - np.float32(0.7) * (float(np.float32(CS)) - 5) / 5)) / 2) * np.float32(1.00041)

rewind

export function circleSizeToScale(CS: number) {
  return float32((1.0 - (0.7 * (CS - 5)) / 5) / 2);
}

lazer

            Scale = (1.0f - 0.7f * (difficulty.CircleSize - 5) / 5) / 2

stable (I think?)

            float circleRadius = (PLAYFIELD_WIDTH / 16.0f) * (1.0f - 0.7f * (float)HitObjectManager.AdjustDifficulty(Beatmap.DifficultyCircleSize));

I'll be honest with you...I have no idea anymore where some of the casts or constants in the circleguard code came from (what is np.float32(1.00041) doing there??). But maybe this will be useful for anyone looking into this. Circleguard has matched stable in all cases I know of, fwiw.

[peppy]

Path forward would be to reimplement to match stable calculation-wise, I believe. Should try and simplify the equation down as much as possible.

Will add this to the project as it seems important.

[peppy]

I'll look into this one.