Struct KdaRowsStash 

pub struct KdaRowsStash {
    pub ring_snap: CudaSlice<f32>,
    pub raws: [CudaSlice<f32>; 3],
    pub scan: KdaScanInputs,
    pub rows: usize,
}

The rollback stash of one BATCHED verify-rows KDA call (lane/glm5-verify-batch): the per-layer t=K+1 twin of the per-row KdaScanInputs steal. Everything here is either stolen from buffers the call allocated anyway (raws, scan — zero copies) or one small clone per layer per round (ring_snap, 3*qkv*(kernel-1) floats ~ 96 KiB).

Rollback to keep rows rebuilds both state planes EXACTLY:

• conv ring: restore ring_snap, then re-issue kda_conv_ring_roll per plane over raws at T=keep — the roll is pure placement (no arithmetic), so the rebuilt ring is the sequential chain’s ring after row keep-1 byte-for-byte.
• ssm state: ONE kda_scan replay at T=keep from the caller’s pre-round snapshot over the batched scan inputs (the kernel walks rows 0..keep of the [t, ..] buffers) — the in-kernel T-loop IS the chained t=1 program (register-resident state, identical per-step order), held by the scan-chain bit-gate.

Fields

ring_snap: CudaSlice<f32>

The fused conv ring BEFORE this call’s rolls (one clone per layer per round).

raws: [CudaSlice<f32>; 3]

RAW (pre-conv) q/k/v projection rows [t, qkv], stolen post-roll (plane order).

scan: KdaScanInputs

Batched scan inputs [t, ..], stolen post-scan.

rows: usize

Row count of the call that filled this stash; rollback validates keep against it.