comp.ts simplifications: ty_holds, CID_T, tele_fill, corpus_at and more

[nicolas-abril]

Behaviour-preserving simplifications of bend2/comp.ts. They free 183 ttok (63,905 → 63,722). Each commit covers one area:

• ty_holds, lay_el, task_deliver, gpu_load, startup:
  • ty_holds now looks inside every datatype except the words itself, so its predicate just returns true/false for what it is looking for. ty_value, ty_clo, lay_of's recursion check and js_marshal's Nat test no longer spell out the walk.
  • lay_el writes its type formers inline.
  • task_deliver computes its count cell once.
  • Metal gpu_load has one failure path.
  • corpus_setup and C main drop their one-use temporaries.
• term_const, lay_el, mat_ctrs:
  • Removes term_const's dead length === 0 guard ([].every is already true).
  • lay_el returns BOX directly for type formers, where lay_of always gave BOX on that branch.
  • Fixes a line break in mat_ctrs so its nested ternary reads the way it parses.
• node_fields, emit_open, show_main:
  • node_fields builds its scratch reads without patching strings.
  • emit_open names a parameter's words and marks the borrowed ones in one pass.
  • show_main's node checks its cache before computing anything.
• compile_book:
  • The shown constructor names are worked out once, for both the roots and SHOW_NAMES.
  • cid() stores each encoded arity where the constructor is met, so the ars array zipped by index with cids is gone.
  • js_def uses def_foreign.
• tele_fill: a wrapper next to tele_unbind for the four Bend.tele_fill(FL.book, T, xs, Bend.ctx_nil()) calls.
• corpus_at: one host-only helper for "map exactly at this address, give the mapping back otherwise", which corpus_map and corpus_grow both repeated.

Hot runtime paths and device code were deliberately left alone. During the drop work on #1438, an "obvious" merge of ctr_take's two loops cost about 6% on lexer's Metal run.

Validation

• Generated code: every #| test and all 17 runtime benches give byte-identical JS and compiler logs to main. Every .c file differs only in the task_deliver, gpu_load, corpus_setup, main and corpus_at lines. That was checked per commit, and again for the combined branch, where all 762 C diffs are identical.
• Gates on the cluster: repo 54/54, test 1597/1597, ping 47/47, perf 124/124.
• Perf flake: in 2 of 9 combined _run.ts runs on the first commit, perf came back 123/124. The full perf tables put the closest cells at merkle COMPILER and generics_3200 CHECKER, which are near the 1.15× limit on main too. Alternating rounds against main passed 124/124 every time.
• corpus_grow: exercised with an 8 GiB array, so the arena doubled twice through the new helper. Metal builds and --gpu on/off outputs match main.
• tsc: no errors in comp.ts.

#1438 carries the first commit too. Once this lands, it rebases onto it.

🤖 Generated with Claude Code