[Lorenzobattistela]

comp.ts goes from 66,000 to 65,414 ttok (−586), leaving room under the 66k cap.

Trims (−120)

• term_peek, rfc_seal, term_keep and blk_read in the C runtime are one-line ternaries.
• ty_key(t) replaces five copies of Bend.term_key(Bend.term_lower(t)).
• A few comments are tighter, and three sentences that restated the code next to them are gone.

A pure main prints through Bend code (−466, from #1817)

• main_pure, taken from comp.ts: nothing reads emitted text, pure mains print through Bend #1817, writes a printer of main's type as Bend source and checks it in a copy of the book. It replaces the descriptor show_main built and the C and JS printers that read it.
• Base gains U32.hex, Char.quote, String.quote, F32.show.dot, Array.write and Array.show, and comp.ts: nothing reads emitted text, pure mains print through Bend #1817's four printer tests come along (tests/show/array_elements, imported_adt*).
• lay_eq walks the layout DAG, also from comp.ts: nothing reads emitted text, pure mains print through Bend #1817. The printer's defs take deep shared layouts, which comparing JSON expanded exponentially: tests/reg/show_shared_layout.bend took 93 s to build without it.
• comp.ts: nothing reads emitted text, pure mains print through Bend #1817's JS stdout buffering is not included: JS prints a pure main's text in one write.
• This overlaps comp.ts: nothing reads emitted text, pure mains print through Bend #1817's printer, so comp.ts: nothing reads emitted text, pure mains print through Bend #1817 will need to rebase over it.

Verification

• repo 54/54, test 1711/1711 (1707 + the 4 new tests), ping 73/73, all inside the 30 s cap.
• safe: identical to main's (1,114 agree, 0 false rejects).
• Before the port, JS and .mjs output were byte-identical to main's on all 1,308 programs that check, and C differed only by the four runtime ternaries. After the port, every runtime bench's C differs from main's only in the removed printer code, which MAIN_PURE (0 for them) already compiled out.
• cc -O3 host assembly is identical to main's for all 17 runtime benches, before and after the port. Metal device IR was identical too before the port.
• All test builds take 7.1 s in total, against main's 5.9 s.
• perf: 120–123 of 124 on this branch, and main also misses (122–123 of 124 in the same hour). The histogram SEQ-CPU cell reads ~1.25x on main and on this branch: built on its mini (cluster-0e), main's binary and this branch's have the same SHA and run 1.85–1.87 s there. The other misses (bfs, merkle, editdist, tree-bitonic) move between runs and also hit main.

🤖 Generated with Claude Code