[Lorenzobattistela]

comp.ts goes from 65,931 to 64,919 ttok, out of its 66,000 cap. That frees 1,012 tokens without changing what the compiler emits.

Comments (about −920 ttok)

32 long comment blocks are tightened. Each keeps its facts and drops narration, restatement of the code beneath it, and benchmark anecdotes:

• "56 to 369 ms"
• "3.05 s against 2.51 s"
• "15 MB -> 68 MB"
• "On the CPU, a call before each fork keeps a fork tree on a few threads: 256 leaves get 3.5x from 16 threads instead of 14x #1475: 3.2 s -> 1.1 s"
• "bitonic 1.35x without"

One term_drop comment line also comes back under 80 columns.

Code (about −94 ttok)

• f32_text builds the decimal form in one loop over the digit places, with a point before place −1. This replaces three memmove/memset cases (ex < 0, ex < p, else). The scientific branch now returns directly.
  • Old and new give the same text and length for all 2^32 f32 bit patterns, checked in a C harness.
  • p never reaches 9, so char d[9] cannot overflow.
• bend_dev: CUDA's early turn_groups return moves into the CUDA #else branch. The back-to-back #endif / #ifndef __METAL_VERSION__ pair goes. The text after preprocessing is the same on Metal, CUDA and the host.
• JS io_run requeues a step that has run 64 effects through io_push(op.kont, x, false, io), the helper io_wait already uses. Before, it mutated and re-pushed s.
• term_open and open_of return a named type Open, replacing the literal type written three times. emit_native reads f.h, which is always fun_of(k).h.

Two changes were left out because #1576's review restored them: dropping the trailing #undef WL_X, and emit_jump's seg_ref.

Same output as main

All 1,773 programs under tests/, bench/ and demos/ were built to .c, .js and .mjs with main's comp.ts and with this one:

• .mjs and build messages: byte-identical, including every expected build error.
• .c and .js: they differ only in comment lines and the three runtime edits above. The compiler's own output (tables, spins, segments, JS program text) is byte-identical.

Gates (on main 257f896)

• repo 54/54
• test 1688/1688
• ping 63/63
• perf 123/124 on two branch runs, against 123/124 for main run between them. The one failing cell, mandelbrot SEQ-CPU at about 1.33x, fails on main too. Geomeans per column match main within noise.

🤖 Generated with Claude Code