[nicolas-abril]

comp.ts goes from 63,652 to 63,386 ttok tokens on top of main. These are structural simplifications. The emitted program code is unchanged, and comments, if statements, types, section-prefixed names and every runtime cast stay as on main.

Compiler

• ty_srcs roots a type's datatypes directly, replacing type_adts.
• val_arms takes a box flag instead of a condition callback.
• emit_chain takes [cond, body] pairs.
• show_main reads main's layout from fun_of.
• emit_ctr's word and Array branches share their field emission.
• nat_add and nat_double share one tpl_ops entry.
• view_root folds into its one caller.
• emit_clo drops a binding copy.
• The lend edges come from split("<").reverse().

C runtime

• WL_CASE opens its own brace.
• WL_FN and OUTLINE reuse FAR.
• io_wait uses one fd-set buffer.
• f32_text NUL-terminates and prints its exponent with %+d.
• main starts thr at the CPU count.

JS runtime

• char_new, now in a Char section, is bound from the TS function, as f32_round is. The Chr builder reuses it.
• f32_show's digit loop covers nan, ±inf and -0.

Same output as main

• For all 1714 programs under tests/, demos/ and bench/runtime/:
  • the emitted C tables, spins and segments are byte-identical to main's, except for WL_CASE's brace;
  • the emitted JS program text is identical, except that non-scalar Char literals fold into string literals of the same value.
• Two review rounds reverted everything that was not a pure refactor:
  • term_aux's return type
  • the JS f32 bit views (Node keeps signaling-NaN payloads differently)
  • node_fill's name stems, wpoll and TAB_AT, which could flip INLINE/DFAR
  • task_deliver's loop bound
  • io_wait's running max fd
  • emit_jump's ref
  • WL_LOAD/WL_TAKE
  • emit_task's return
  • #undef WL_X
  • file_book's error order

Verification

• Gates: test 1634/1634, perf 124/124, ping and repo all pass. comp.ts typechecks clean.
• Both lanes: all 1710 non-network programs print the same as with main's comp.ts on the JS and C lanes, GPU builds included. The one mismatch was tree-radix's JS timeout under 8-way load. Its program JS is identical, and alone or under synthetic load it times the same: 9.0 s on both.
• Sanitizers: 745 test builds with -fsanitize=undefined print the same as with main's comp.ts, and report nothing.
• f32 text: C f32_text and JS f32_show were fuzzed against main's (2^24 and 2M+ bit patterns, plus edge values, under Bun and Node) with no differences.

🤖 Generated with Claude Code