Element-typed parameters and index scaling (eliminate manual byte/cell math)

[tetsuo-cpp]

Summary

Make parameter (and shared) declarations carry element type and size so indexing is in elements by default, removing error-prone 4 * / CELLS patterns for narrow types.

Motivation

demo/attention.forth must use byte addressing for f32 globals:

2 PICK HEAD_DIM * I + 4 * Q + F32@

while f64 shared uses CELLS. Mixing F32@ (4-byte) with cell-oriented mental models is a footgun and obscures algorithms.

Proposed model

\! param Q f32[32768]
\! param SCALE f64
\! shared SCORES f64[1024]

Using Q pushes a typed base; indexing words scale by element size:

Word / form	Behavior
i Q []@ or Q I @	load element i of Q with correct width/address space
x i Q []!	store element
Or keep Q I CELLS + @ but define CELLS relative to the pointer’s element type

Minimum viable improvement:

1. Allow f32 / i32 / f16 / … in \! param / \! shared (not only i64/f64 storage with separate access words).
2. Document that addresses are byte addresses always, or introduce element-pointer types — pick one model and stick to it.
3. Provide []@ / []! (or typed @) that scale correctly.

Acceptance criteria

• Header accepts reduced-width element types for params (and ideally shared)
• At least one ergonomic indexing path that does not require manual 4 * for f32
• Attention-style kernel can be rewritten more clearly
• LIT + language doc updates
• Clarify relationship to existing F32@ / I32@ / CELLS words (compat policy)

Related

• Reduced-width memory words (Add load/store words for reduced-width memory types #52 closed)
• Shared split @ vs S@ (Proposal: split global vs shared memory words (@/! and S@/S!) #36 closed)
• Dynamic extents (separate issue)