docs(hhtl): PR-X4 pre-sprint prompt — 4×4 splat cascade (W4-W5)

Companion to GridLake (ade8edb) and PR-X14′ (56b2671) prompts, in the
same shape. Captures the W4-W5 splat-cascade sprint that promotes
splat3d from "bespoke 16×16 tile binner" to "typed multi-resolution
cognitive evolution operator" with the (4×4)×(4×4)×(4×4)×(4×4) tier
scheme as the load-bearing structural identity between Gaussian
splatting and the cognitive shader.

Verbatim citations:
- pr-x4-design.md:36-48 (graphics ≡ cognitive substrate identity table)
- pr-x4-design.md:57-72 (the tier scheme, area-uniform 16× branching,
  non-isotropic 4/16/4 per-dim branching rationale)
- pr-x4-design.md:108-118 (TileInstance v2 shape)
- pp13-brutally-honest-tester-verdict.md:32-37 (the P0-4 L4 Hilbert-3D
  bug — hilbert3d_encode([15,15,15], 4) = 2925, expected 4095)

Worker DAG: A1 TileInstance + BlockedGrid refactor (chain dep) →
A2 CascadeAddr [GATED on A12b L4 fix], A3 deg-3 SH inquiry-direction,
A4 INT4×32 packed dot (3 backends + parity), A5 NARS truth-revision
kernel + fast_exp_x16 precision audit.

Five forbidden constraints:
  1. No bespoke Hilbert-3D in PR-X4 — consume linalg::hilbert
  2. No crate::simd::* extension from inside PR-X4
  3. No write to lance-graph upstream
  4. No 4×4 tier-stride departure
  5. No NARS-revision activation by default (feature-gated)

Carry-overs back into other sprints:
- PR-X10 A12b L4 fix is a gate for A2 spawn
- PR-X11 Spd3 ops are deps for L5/L6
- PR-X9 (W6-W7) consumes CascadeAddr; PR-X4 must ship by end-W5
- W7 closure swap activates the splat4d-nars-compose feature

No new Q-marker; W4-W5 slot is fixed. PR-X4 starts late only if
GridLake + X14′ slip past W3 (per Q-NEW-1/Q-NEW-2 cell choice).

Author: claude