From ac1250989cda4b34f49249aa5d28bca8602ce174 Mon Sep 17 00:00:00 2001 From: Roman Inflianskas Date: Tue, 6 Oct 2026 13:48:10 +0000 Subject: [PATCH] docs(disk-space): state the uniform arm's catch rate as measured The top of the arm's exponent range is TiB, where the scaling loop has nothing left to promote to, so a too-early promoter renders exactly as the real function there. The arm catches it on two fifths of its draws, not half. Signed-off-by: Roman Inflianskas --- crates/rocm-core/src/disk_space.rs | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/crates/rocm-core/src/disk_space.rs b/crates/rocm-core/src/disk_space.rs index 5f43b1bbc..c60d1cc8c 100644 --- a/crates/rocm-core/src/disk_space.rs +++ b/crates/rocm-core/src/disk_space.rs @@ -727,8 +727,10 @@ mod tests { /// * Uniform within one unit's range, `[1024^k, 1024^(k+1))`, except /// that the `B` range starts at 0 so zero is drawn too. About half /// of each range lies below half of the next unit, so a loop that - /// promotes too early is caught on about half of this arm's draws, - /// which is within a handful of draws across the whole strategy. + /// promotes too early is caught on about half of the draws in every + /// range but the largest, where there is nothing left to promote to + /// — two fifths of this arm's draws, which is within a handful of + /// draws across the whole strategy. /// * A window just below each `1024^k` boundary. The window has to SCALE /// with the boundary, because the band where `:.1` rounding pushes the /// mantissa up to 1024.0 is itself proportional: it spans the top