Question: should the surface have ONE canonical textual form (a formatter normalizes to it, gofmt/rustfmt style) or a lenient grammar that tolerates whitespace/style variation? And — the orthogonal axis — should the grammar be whitespace-insensitive?

Why it matters: it sets the human↔agent ergonomics trade. Inference + loose input is nice for humans; a single canonical, explicit form is nice for agents, diffs, and CI (less ambiguity, stable blame). The two concerns (type-annotation optionality, text-format canonicality) are SEPARATE axes and shouldn't be conflated — bidirectional typing already answers the first (optional, mandatory exactly where synthesis is stuck; ADR-0066).

Detail: the current surface is the worst quadrant on the format axis — whitespace-sensitive (operators are space-delimited: a + b parses, a+b does not — Bang/Frontend/Surface.lean tokenizer) AND non-canonical. The modern answer (Go, Rust, Elm) is both: a whitespace-insensitive grammar PLUS a canonical formatter. Humans write loose; the formatter normalizes to one true form; agents/diffs/CI see only canonical text. This is the same single-source-of-truth move as everywhere else (the formatter GENERATES the canonical form).

Options: (1) whitespace-insensitive grammar + canonical formatter (recommended — Go/Rust model; serves human, agent, machine at once); (2) lenient/whitespace-tolerant but no canonical form (compiles despite style drift — but diffs/agents see noise); (3) status quo (whitespace-sensitive operators — the trap; fix via #30).

Recommended: (1). Precondition is the Pratt parser (#30) which removes the space-delimited-operator hack → whitespace-insensitive; a bang fmt canonical formatter is then pure gravy (and pairs with the tree-sitter grammar, #111).

Blocked on: nothing hard; sequencing — do the Pratt parser (#30) first, then the formatter. Liquid until the surface stabilizes past the toy parser.

Revisit signal: building the Pratt parser (#30), or the first bang fmt / formatter; or agent-ergonomics friction from non-canonical diffs.