Feat/core layer architecture refactor

[Concode0]

Architectural Refactoring: Thin Layers & Polymorphic Runtime

• Thin Primitive Layers: All primitive layers (ProductLayer, RotorLayer, ReflectionLayer, MultiRotorLayer, CliffordLinear) are now ultra-thin wrappers over runtime behavior. They no longer branch explicitly on layouts.
• Centralized Dispatch: Dense/Compact dispatch is strictly handled by the core runtime (core/runtime/actions.py). CliffordAlgebra and AlgebraContext dynamically plan execution paths.

Benchmarks ( about planned core architecture and layer refactoring )

The "Grade Contract" (declared left_grades, right_grades, output_grades) allows the core to execute only the algebraically relevant lanes. This structurally avoids the $2^n$ Cayley-table materialization.

High-Dimensional Verification ((1)x(1)->(0,2), compact output):

• $n=16$ : 121 lanes (vs. Dense: 65,536) -> OK
• $n=32$ : 497 lanes (vs. Dense: 4.2 Billion) -> OK
• $n=63$ : 1,954 lanes (vs. Dense: $2^{63}$) -> OK

3. Performance Benchmarks: 75x Structural Speedup

The performance gain is driven by optimizing the computational tensor program itself, not just compiler magic.
Measured CPU results for vector GP (1)x(1) -> (0,2):

Dimension ($n$)	Dense Engine	Compact Planned	Speedup
$n=5$	0.477 ms	0.053 ms	8.9x
$n=6$	1.402 ms	0.065 ms	21.7x
$n=7$	2.147 ms	0.057 ms	37.8x
$n=8$	4.240 ms	0.057 ms	75.0x

Also checked:

• Pairwise wedge (2)x(1)->(3) works at n=12, n=16, n=24.
• At n=32, grade-3 output works when raised through PlanningLimits(max_lanes=8192).
• MultiRotorLayer works compactly at n=16, n=32, and n=63.
• aot_eager fullgraph captures compact GP at n=32 and n=63 with zero diff.
• Dense materialization at n=63 is intentionally blocked.