feat(codec): PR-X12 A2 mode bit-pack + A3-intra prediction kernel

A2 — mode.rs (~270 lines, 11 tests):
- 16-bit header: 2-bit mode (Skip/Merge/Delta/Escape) + 12-bit basin_idx
- MergeDir 2-bit pack/unpack (high bits masked)
- Whole-leaf compact pack/unpack: Skip=2B, Merge=3B, Delta=3B, Escape=6B
- packed_byte_len() const fn for buffer pre-sizing
- MAX_BASIN_IDX (4095) + BASIN_NONE sentinel
- Stream roundtrip test for mixed-mode leaves

A3-intra — predict.rs (~330 lines, 12 tests):
- IntraContext { basin_idx, delta_i32, NESW neighbours }
- IntraConfig { escape_next_idx: Option<u32> }
- predict_intra() decision tree: Skip → Merge → Delta → Escape
  (monotone wire cost 2 → 3 → 3 → 6 bytes; cheapest-fit policy)
- Merge match: same basin_idx + same δ as u8 (sign-tolerant wrapping cast)
- Escape fallback: lossy i8 clamp when allocator absent (never panics)
- End-to-end pack/unpack chain test through the decision

Deferred to follow-up: A3-inter (cross-tier neighbour scan from
BlockedGrid L2/L3), A4 transform, A6 RDO, A7 rANS, A8 stream framing.

Author: claude

src/hpc/codec/mod.rs

@@ -8,9 +8,13 @@
 //! # Module layout (per PR-X12 worker decomposition)
 //!
 //! - [`ctu`] — A1: `Ctu` carrier + `CtuPartition` enum + quad-tree
-//!   split / merge ops. **Shipped in this PR.**
-//! - `mode`, `predict`, `transform`, `quantize`, `rdo`, `ans`, `stream`
-//!   — A2-A8, queued as follow-up sprints.
+//!   split / merge ops.
+//! - [`mode`] — A2: bit-pack / unpack helpers for the on-wire 16-bit
+//!   header + per-mode tail (Skip/Merge/Delta/Escape).
+//! - [`predict`] — A3-intra: encoder-side mode-decision kernel that
+//!   picks the cheapest `LeafCu` from a cell + NESW neighbours.
+//! - `transform`, `quantize`, `rdo`, `ans`, `stream` — A4-A8, queued as
+//!   follow-up sprints.
 //!
 //! # Feature gate
 //!
@@ -22,6 +26,13 @@
 //! `.claude/knowledge/pr-x12-codec-x265-design.md` — master design doc.
 
 pub mod ctu;
+pub mod mode;
+pub mod predict;
 
 pub use ctu::{CellMode, MergeDir, MAX_QUAD_TREE_NODES, MAX_SPLIT_DEPTH};
 pub use ctu::{Ctu, CtuArena, CtuPartition, LeafCu, MaxSplitDepthReached, MergeError, NodeIdx};
+pub use mode::{
+    pack_header, pack_leaf, pack_merge_dir, packed_byte_len, unpack_header, unpack_leaf, unpack_merge_dir, BASIN_NONE,
+    MAX_BASIN_IDX,
+};
+pub use predict::{is_no_basin, predict_intra, IntraConfig, IntraContext};