Numerals denote numbers

A bare word the numeral grammar accepts denotes that number in every position — argument, value, redirect target, list element — and a word that means bytes is quoted. Every number has exactly one printed spelling, so a canonical numeral crosses the shell unchanged and every other spelling normalises on the way out.

• No positional asymmetry. One question, asked the same way everywhere: f 007, let x = 007, [007], and > 007 all see the number seven. A position that read the digits instead would be an exception a user has to carry, and there is none.
• Lexical, never type-directed. Classification reads the token and nothing else — not an expected type, not the enclosing scope, not the installed handlers, not the printer. A word’s meaning is therefore visible at the prompt and stable under every refactoring that moves it, let-naming included.
• One number, one spelling. An Int prints its decimal digits. A Float prints the shortest decimal that reads back as the same number, always carrying a point (3.0, 1.0e300), through one printer that every float-rendering site shares.
• The printer’s image lies inside the grammar. Every spelling the printer emits is a numeral the grammar accepts, so printing then reading is the identity on numbers and a printed float never degrades into text. This is what fixes the point in 1.0e300: ryu’s own shortest form is 1e300, and the printer restores the point rather than the grammar admitting the exponent-only shape.
• The grammar is not widened to meet it. 1e6 is still a word. An exponent-only token is a short hex digest, a field name, or an identifier at least as often as it is a number, and a shell that silently turned 1e6 into a float would be the footgun the whole doctrine exists to avoid. The printer moved; the reader did not.
• Canonical spellings are fixed points. Print a number, hand the spelling back as a bare word, and the same number comes back; classifying then printing normalises (007 → 7, 1.50 → 1.5, +5 → 5, .5 → 0.5, -0 → 0). The property is checked by test and is never the definition: classification is a grammar, not a parse-then-print round trip.
• Bytes require quotes, and the emitter obeys the same rule. '007' for the padded digits, '3.10' for the version. Wherever ral renders a String back as source — a completion, a quoted argument — a numeral-shaped one comes back quoted: is_bare_word consults the grammar so that a bare emission always reads back as the text it came from. That direction is the permitted one. The ban is on the grammar consulting the printer, never on the printer’s side consulting the grammar.

This is a hard rule, not a stylistic preference. Do not let classification consult a type, a scope, or the printer; do not add a position in which a numeral means its digits; do not give a number a second printed spelling; and do not implement the grammar by round-tripping through the printer.

The consequence a user meets is named where it is decided: 3.10 is the numeral 3.1, so a version-like token is quoted (a-numeral-denotes-its-number). See types for the inference this keeps annotation-free and exec-argv-is-words for what a rendered number becomes at the operating-system call.