Struct KhlSlot 

#[repr(C, align(64))]
pub struct KhlSlot {
    pub packed_sequence: AtomicI64,
    pub _reserved: u64,
    pub items: [LineItem; 3],
}

Ring slot: Vyukov sequence (with n_items bit-packed into the low 2 bits) + 3 LineItem payloads. Fixed shape, 64 bytes, process-portable.

Cross-axis fusion: n_items packed into sequence

n_items is always in 1..=KHL_ITEMS_PER_SLOT = 3, which fits in 2 bits. Instead of paying a separate store to publish n_items alongside the sequence number, we encode it in the low 2 bits of packed_sequence. The producer’s ONE Release-store on packed_sequence publishes BOTH the protocol state AND the payload count - saving one store per slot, which fuses the K_inner axis (items per slot) with the K_gating axis (per-slot atomic) at the slot’s cache line.

Encoding:

• idx_value = high 62 bits of packed_sequence
• n_items = low 2 bits of packed_sequence
• State: idx_value == idx (empty) -> idx_value == idx + 1 (published, n_items valid) -> idx_value == idx + capacity (consumed)

Fields

packed_sequence: AtomicI64

Bit-packed Vyukov sequence: (idx_value << 2) | n_items.

_reserved: u64

Reserved 8 bytes for cache-line alignment of the items array (items start at offset 16, slot is 64 bytes total).

items: [LineItem; 3]

Caller’s byte-oriented payloads. Only the first unpack_n_items(packed_sequence.load()) are guaranteed valid.