[nicolas-abril]

bend2/comp.ts drops from 63,975 to 63,532 ttok (−443) through simplifications, not minimizations: dead special cases, duplicated logic folded into one place, and helpers reused instead of copied. No identifiers are shortened and no comments are cut to save tokens.

What changed

• Compiler (A–C): a match's uses fold in the generic kids walk; lay_box is gone (only W32/W64/BOX have no arms, so it was x === BOX); seg_open renames a binding's words in place; compile_book opens defs and foreign defs in one loop and indents a segment with one seg_text; emit_fork picks frame bindings with one filter; a few dead guards confirmed unreachable by instrumenting the corpus.
• JS backend and runtime: one function emitter for foreign and filled defs; one tag test in js_match; a pure main's show descriptor is one array; show_chr takes the character.
• C runtime: a host WL_AGAIN is WL_JMP; blk_wcls is the one word-class rule; task_deal picks a ring in one expression; corpus_map loops its hints and falls back through pool_mmap; io_str drops a no-op io_seal; io_loop takes both of its every-64-steps checks under one test; main reads a missing flag value as "", so --threads/--gpu need no NULL checks; show_f32/show_val are simpler; gpu_hash lives in the CUDA branch, its only user.

Output changes

• The C and JS runtime text changes, so every emitted .c and .js differs in those runtime lines.
• js_def's shared emitter gives a foreign def the trailing blank line every other def has, so .js/.mjs files with a JS foreign def gain one blank line after it (this also fixes a }export default { run-on).
• Everything else emits byte-identical output, and every compile-error log is identical to main's.

Verification

• Output: all 1,074 tests, benches and demos were compiled to C, JS and .mjs and compared with main; the only differences are the ones listed above.
• Tests: every Base test on the JS and native (clang + Metal) lanes gives the same result as main.
• Gates: repo 50/50, ping 46/46, test 1587/1587. perf: every cell over 1.15x also fails on main (raytrace's COMPILER TIME sits at 1.15–1.18x on main; PAR-GPU mandelbrot/merkle up to 1.19x).
• Speed: the C runtime changes keep instruction counts; CPU cycles are within ±0.5% on all 17 benches, Metal within noise. Dropping term_const's fieldless-constructor shortcut slowed raytrace's compile by 3%, so it stays; raytrace now compiles in the same time as on main.
• Review fixes: ctr_build keeps its has check (a constructor and a def can share a name, and book_fam returns the constructor's own name when its type is no ADT); term_uses still skips DUMMY; io_loop reads the clock only when something is parked.
• Not tested here: the CUDA build (the CUDA change only moves gpu_hash into its branch and inlines gpu_shape).

Rejected for costing speed: a merged copy loop in ctr_take (5–10% faster on CPU, slower on Metal), WL_ARGS pushing N words (+1% queens), and removing WL_FRAME's wtw temporary (breaks an ldp/stp pairing).

🤖 Generated with Claude Code