DRAFT for writer prose pass before the PR opens.

Codifies the two-sorted semantics of pointer expressions: values are
either unbounded non-negative integers (no width) or bytes (with a
definite width). Arithmetic, JSON-number and odd-nibble-hex literals,
$wordsize, and lookups are integers; even-nibble hex, $read, and
$sizedN/$wordsized are bytes. An integer is accepted wherever an integer
is expected (a bytes value is read as its big-endian integer value), but
$concat and $keccak256 operands must be width-bearing bytes — a bare
integer there is an error and must be resized first. Fixes the keccak256
example's bare-integer operand to be word-sized, matching what the
mapping/string pointer examples already do.

The evaluator used a single width-carrying Data representation for
every expression, so a JSON number, $wordsize, or an arithmetic result
had an incidental width that leaked into $concat and $keccak256: for
instance `{ $keccak256: [{ $wordsized: k }, 0] }` hashed 32 bytes
rather than 64.

evaluate() now returns a Value of one of two sorts, following the
expression schema:

- integer (bigint, no width): JSON numbers, $wordsize, lookups,
  arithmetic results, and odd-digit hex literals
- bytes (Data, definite width): even-digit hex literals, $read, and
  the resize forms

Arithmetic takes its operands as integers (bytes coerce big-endian)
and yields an integer. $concat and $keccak256 require bytes operands
and throw, naming the offending operand and suggesting $wordsized or
$sizedN, when given a bare integer. Resize is the only integer-to-bytes
bridge. Variables carry the sort of their defining expression; list
counts, conditionals, and region slot/offset/length coerce to
integers. Cursor.Region is unchanged: its Data-valued properties now
encode integers minimally, with bytes keeping their width.

The previously skipped keccak256 test is rewritten against these
rules, along with tests for each sort rule, both coercion directions,
and the bare-integer errors.

Claude-Session: https://claude.ai/code/session_01RJFyifZxcSXZchLFNTNPuT

The expression schema's root-level `$keccak256` example still hashed a
bare JSON number and a bare lookup, which the two-sorted rules reject.
Wrap both operands in `$wordsized`, as the Keccak256 example already
does.

pointers: add a test that evaluates every `$keccak256` / `$concat`
expression found in the pointer schemas' examples against stubbed
regions and variables, so an example whose operand lacks a width fails
the suite rather than silently contradicting the spec.

Claude-Session: https://claude.ai/code/session_01RJFyifZxcSXZchLFNTNPuT

The guide extracts its code listings from src/evaluate.ts by symbol name,
and the two-sorted rewrite collapsed the five per-operation arithmetic
functions into one evaluateArithmetic, so the page failed to build.

Replace the five arithmetic listings with one, introduce the integer/bytes
sorts up front with listings of the exported Value type and namespace,
and add listings of the evaluateInteger and evaluateBytesOperands helpers
where coercion and width enforcement come up. Surrounding prose now
describes integer-sort arithmetic, literal sorting by hex digit count,
and the bytes requirement on $keccak256 / $concat operands.

Claude-Session: https://claude.ai/code/session_01RJFyifZxcSXZchLFNTNPuT