fix(sprint-12/wave-F): codex P2 — AttentionMaskBackend impl for AttentionMaskSoA + canonical MailboxId import

Codex P2 review on PR #388 flagged that `AttentionMaskActor` is public
but `AttentionMaskSoA` (the only production backend in the crate)
doesn't implement `AttentionMaskBackend`. Downstream consumers can't
add the impl themselves because they own neither the trait nor the
SoA (Rust orphan rules), so `AttentionMaskActor::new(AttentionMaskSoA::new(...))`
would force them to wrap in a local newtype.

Fix #1 — production-backend impl in attention_mask_actor.rs
Added `impl AttentionMaskBackend for crate::attention_mask::AttentionMaskSoA`
with the four trait methods delegating to the existing inherent
methods on AttentionMaskSoA. Now downstream consumers can wire the
two directly: `AttentionMaskActor::new(AttentionMaskSoA::new(4))`
works out of the box.

Fix #2 — collapses CSI-10 from the W-Meta-Opus honest review
The W-F2 worker had defined a local `pub type MailboxId = u32` in
attention_mask.rs (Opus CSI-10 entry: "W-F2 used local MailboxId
shadow alias"). Switched to `pub use lance_graph_contract::collapse_gate::MailboxId;`
— same underlying u32 so all method signatures stay compatible,
but now both attention_mask.rs and attention_mask_actor.rs reference
the SAME canonical type and the trait impl in fix #1 can be added
without type-mismatch concerns.

Test status: 14/14 attention_mask + attention_mask_actor tests pass.
Clean compile (modulo 2 pre-existing unused_mut warnings unrelated
to this change).

Branch rebased on main (post PR #385 merge) to pick up the ndarray
hpc-extras feature flag that W-C1 added; this resolves the blake3
unresolved-crate compile failure that hit cognitive-shader-driver
prior to rebase.

https://claude.ai/code/session_01UwJuKqP828qyX1VkLgGJFS

Author: claude

crates/cognitive-shader-driver/src/attention_mask.rs

@@ -8,13 +8,12 @@
 //! `AttentionMaskSoA` is `!Send` by design (non-atomic interior mutation).
 //! Callers that need cross-thread access must wrap in `tokio::sync::Mutex`.
 //!
-//! # Registration
-//! This file is intentionally NOT registered in `lib.rs`. The main thread
-//! adds `pub mod attention_mask;` after the sprint-12 fleet completes.
+//! # MailboxId
+//! Imports the canonical `MailboxId` alias from
+//! `lance_graph_contract::collapse_gate` (also re-exported here for ergonomic
+//! `use cognitive_shader_driver::attention_mask::MailboxId` access).
 
-/// Re-export the canonical `MailboxId` alias so this module compiles
-/// standalone without an explicit `--extern` flag.
-pub type MailboxId = u32;
+pub use lance_graph_contract::collapse_gate::MailboxId;
 
 // ── SoA row ──────────────────────────────────────────────────────────────────
 

crates/cognitive-shader-driver/src/attention_mask_actor.rs

@@ -33,6 +33,25 @@ pub trait AttentionMaskBackend {
     fn is_active(&self, mailbox_id: MailboxId) -> bool;
 }
 
+/// Production-backend impl: wires the sibling `attention_mask::AttentionMaskSoA`
+/// into the actor without forcing downstream consumers to write a newtype
+/// (Rust orphan rules would block them — neither the trait nor the SoA is
+/// theirs). Per codex P2 review on PR #388.
+impl AttentionMaskBackend for crate::attention_mask::AttentionMaskSoA {
+    fn touch(&mut self, mailbox_id: MailboxId, w_slot: u8) -> bool {
+        crate::attention_mask::AttentionMaskSoA::touch(self, mailbox_id, w_slot)
+    }
+    fn evict_lru(&mut self) -> Option<MailboxId> {
+        crate::attention_mask::AttentionMaskSoA::evict_lru(self)
+    }
+    fn tick(&mut self) {
+        crate::attention_mask::AttentionMaskSoA::tick(self)
+    }
+    fn is_active(&self, mailbox_id: MailboxId) -> bool {
+        crate::attention_mask::AttentionMaskSoA::is_active(self, mailbox_id)
+    }
+}
+
 pub struct AttentionMaskActor<B: AttentionMaskBackend> {
     inner: B,
     pending_evictions: Vec<MailboxId>,