fix(formatter): preserve separated brace arguments

[vicmcorrea]

Problem and fix

Formatting a valid comma-optional call such as choose(id {42 : U32}) produces choose(id{42 : U32}). Bend reads the glued name and brace as a constructor, so the formatted program no longer parses.

Reuse the token's existing gap flag to preserve one space between a name and a separated brace. Adjacent constructors stay glued. This also preserves separated equality and reflexivity arguments without adding parser logic or helpers.

Blast radius

Two files in tools/bend-fmt-lsp: the formatter's name-before-brace spacing rule and its unit tests. The rule serves direct formatBend calls and LSP document formatting; other separated name/brace forms, including constructor declarations, also retain one space. Spacing after closing braces and angle brackets keeps its existing behavior. No language, compiler, bundled bend-lint, gate, dependency, or changelog changes.

This addresses a different defect from #1302's column-sensitive layouts, although both PRs touch nearby formatter code.

Regression evidence

The new seven-case table covers ordinary/dotted names, equality and reflexivity arguments, explicit commas, nested constructors, and case-lambda braces, including idempotence. With the original spacing expression, the final suite has 25 passing tests and the new regression fails; with the fix, all 26 pass.

Real Bend CLI checks cover nine programs before and after formatting. Four annotation/equality/reflexivity/constructor-field cases fail after baseline formatting and print 42 with the fix. Five constructor/pattern, comma, and closing-brace/angle controls print 42 throughout.

Checks

• npm test in tools/bend-fmt-lsp: TypeScript build and all 26 formatter/LSP tests pass, with no skips.
• bun gates/repo.ts with CI-pinned ttok 0.3: 54/54 pass.
• Real Bend CLI comparisons and formatter idempotence pass; git diff --check passes.

Validation limits

Local checks ran on macOS arm64 with Node 26.8.2 and Bun 1.3.11. Windows, the minimum supported Node 22, and the cluster test/performance/safety/ping gates were not run. These focused checks do not establish semantic preservation for every formatter input. #1302 may require a context rebase if it lands first; its complete suite was not rerun here.