nx_srcfmt.nx

nx_srcfmt.nx -- THE CANONICAL SOURCE FORMATTER for NishiLang (lang.plan rung LN9, watch symbol srcfmt_emit). One style, no options -- the gofmt / zig fmt / elm-format shape. The style was not chosen, it was MEASURED: every rule below is the majority convention of the whole runtime corpus (18,561 files, 2.97 M lines, census 2026-08-23; re-measure any time with `nx_srcfmt census`, which uses this same scanner so the ruler and the rule cannot drift apart): indent 4 spaces per brace depth (2,216,007 lines are multiples of 4 vs 9,644 that are not), tabs never (0 tab-indented lines); a continuation line inside an unclosed ( or [ takes base + 4 (20,695 vs 7,210 at +0 vs 4,536 at +8), except a line that opens with ) or ] (base) `{` stays on the line that opens it (179,575 vs 2 own-line); `} else {` on one line (33,968 vs 5,444 split) -- a lone `}` line followed by an `else` line is JOINED `, ` (1,098,119 vs 110,794) ; `; ` (455,500 vs 26,006) ; ` = ` and every binary operator spaced (955,786 vs 276,475 for `=`; 385,897 vs 134,264 for + == != <= >= && ||) ; ` -> ` (103,535 vs 1) `x: T` (781,324 vs 2,337; a space BEFORE the colon 10) ; `*u8` `&x` `-1` `!b` unary glued (700,588 vs 2 for the type star) ; `f(` (966,045 vs 19,129) but `if (` `return (` `while (` keyword-then-paren spaced (15,689 vs 1,315 for `if`) ; `{ stmt }` one-line blocks padded (338,077 vs 11,657) and `{}` empty blocks closed (457 vs 143) ; `a..b` `a.b` `a::b` glued (4,644 vs 91 for `..`) struct field TYPE COLUMNS are aligned (measured 2026-08-23: of 1,712 multi-field struct runs, 1,340 column-align their types, 193 single-space, 179 ragged) -- the type starts one space past the run's longest field name; a blank line or a comment breaks the run. This is the ONE place a token's spacing depends on other lines, and it is still whitespace between tokens so the oracle and the byte-identical-build proof are untouched. const / static / let / var declarations are NOT aligned -- single space after the colon. The corpus does NOT decisively align these the way it aligns struct fields: const runs split 4,234 aligned vs 4,760 single (single is the plurality), static 47 vs 174 single, and body-level let/var are single-space 437,751 vs 6,278. Where the corpus has no clear majority the simpler rule wins, so aligned const blocks reformat to single-space -- expected, measured churn. LF line endings (2,524 of 18,801 files carried CRLF), no trailing whitespace (1,914 lines had it), blank-line runs collapsed to one (348 runs of two or more), exactly one final newline (53 files had none) PRESERVED BYTE-FOR-BYTE, by construction: every string literal (including the multi-line literals this dialect permits -- 205 files carry one), every comment's text, every @-directive line (@macro bodies through their real extent, found by the tokenizer's own lexm_body_end), and the whitespace run that separates code from a trailing `//` comment (the corpus aligns those by column -- widths 2..6 are all common -- so canonicalising them would destroy deliberate alignment; this is a DECLARED floor, not an oversight). WHY THE TOKENIZER IS THE VERIFIER AND NOT THE EMITTER (measured, nx_tokenizer.nx): lex_source STRIPS comments (skip_ws_comments), PRE-EXPANDS @macro into a new buffer (so token positions no longer address the source), and CONSUMES @ifdef/@else/@endif branches at lex time (inactive code never