Prefix generated C locals: ts_32 is a host macro on macOS and the native build fails

[nood-co1]

bend prog.bend -o prog fails on macOS arm64 for a program that checks and runs on the JS
lane. name_local (bend2/comp.ts:648) builds every generated C local from the Bend-level
name plus a counter, with no prefix:

function name_local(fl: File, k: Bend.Name): string {
  const base = name_clean(k);          // <- "ts"
  const n = fl.fresh.get(base) ?? 0;
  fl.fresh.set(base, n + 1);
  return base + "_" + n;               // <- "ts_32"
}

so a generated local can land on a macro the platform headers already define. On macOS one
does: ts_32 is #define ts_32 uts.ts_32 in <mach/arm/thread_status.h>, and it is reachable
from Base itself — bend2/base.bend:1459 declares U32.divmod.go.shl(..., ts: Bool & Word(32n)), and compiling it mints ts_0 … ts_32.

Repro (12 lines)

import Base

def snd(p: Word(64n) & U32) -> U32:
  (q, r) = p
  r

def half(w: Word(64n), b: U32) -> U32:
  snd(U32.divmod.go(64n, w, b))

def main() -> IO(Unit):
  do IO<Unit>:
    IO.print(U32.show(half(Word.zero(64n), 10)))

$ bend repro.bend --check-only   # All terms check.
$ bend repro.bend                # the JS lane: 0
$ bend repro.bend -o repro       # native build
…/repro.c:1580:7: error: expected ';' at end of declaration
 1580 |   u32 ts_32 = r34;
/…/MacOSX.sdk/usr/include/mach/arm/thread_status.h:199:18: note: expanded from macro 'ts_32'
  199 | #define ts_32 uts.ts_32
…/repro.c:1582:655: error: member reference base type 'u32' is not a structure or union

The change

 function name_local(fl: File, k: Bend.Name): string {
-  const base = name_clean(k);
+  const base = name_clean(k).replace(/^bend_/, "");
   const n = fl.fresh.get(base) ?? 0;
   fl.fresh.set(base, n + 1);
-  return base + "_" + n;
+  return "bend_" + base + "_" + n;
 }

The strip is required for the case measured, not cosmetic: seg_open (comp.ts:1640) feeds
names it has already minted back through name_local, so a plain prefix accumulates — this
repro, without the strip, emits bend_b_0, bend_bend_b_0, bend_bend_bend_b_0 (12 doubled
occurrences, 5 distinct, longest identifier 18 characters; with the strip, 0 and 11). It closes
the observed path; it is not a general idempotence property.

name_local is shared with the JS emitter, so generated JS identifiers move as well as the C
ones. The JS output is unchanged here: 0 on this repro before and after, and no per-test
difference in the sampled differential below.

Scope: the names name_local mints — locals, parameters, continuations, and the fixed names
(sp, fb, fv, o, x, k, v). Top-level/def naming is untouched, and CID_/FID_ are
already prefixed.

Cost

bend2/comp.ts 64840 → 64852 ttok — a whole-file delta of 12 — against the repo gate's
65000 cap (headroom 160 → 148). I left out a comment line to keep it there; this message carries
the explanation instead. bun gates/repo.ts: PASS 46/46.

The emitted C is larger, because every generated local reference gains 5 characters: 140864 →
153104 bytes (+8.7%) on this 12-line repro. I have not measured that ratio on a large program.

What I checked

check	result
native build, before / after	rc=1, expected ';' at end of declaration / rc=0, binary prints 0
check (--check-only) and the JS lane, before / after	All terms check. and 0, both times
generated C, bare ts_N locals	33 distinct before, 0 after (33 bend_ts_N after)
interpreter differential, whole tests/ corpus	identical on all 1395 tests
cross-lane sample, 106 tests with a main	0 of 106 per-test records differ; 51 built on both compiled lanes; the same 2 pre-existing divergences on both sides (comptime/unsafe.bend, the interpreter-only unsafe banner; gfx/window.bend, which needs a display)
bun gates/repo.ts	PASS 46/46

gates/test.ts and gates/perf.ts need the SSH cluster and gates/ping.ts cannot run outside the
site project. None ran here, and the local differential above is not a substitute for them.

Not a duplicate

Searching the 920-item issue and PR snapshot, I found no direct duplicate about the C emitter's
generated locals meeting a host macro. The closest items use different mechanisms or layers:
#478 / PR #487 reserves __ for generated function names; #790, #799, #800 and #875 concern
different JS, FFI, or compiler-sentinel naming mechanisms; #904/#908 concern JS import-path
identifiers; and #905/#922 concern checker/parser name collisions.

What this does not claim

• Not that collisions are impossible after it: bend_ is an ordinary C identifier prefix and a
  foreign header could still define bend_x_0. A reserved prefix would be stronger; this is a
  small change that removes the reproduced failure.
• Not that the compiled lanes are regression-free: the interpreter differential is complete over
  tests/, the compiled lanes are sampled.
• Not that ts_32 is the only such name on any SDK: it is the one reachable from Base on this
  host, and the change addresses the class while claiming only the observed instance.

---

Found by an agent working for @blockbrain_labs

[VictorTaelin]

Merged into 2.0.25 under your authorship, thank you, with one trim: the prefix is a single _ (_ts_32), which closes the same class (every generated local is a block-scope identifier, where _x is not reserved) at +0.9 to +4.6 % of emitted C instead of +8.7 to +23 %. Your 12-line reproduction builds and prints 0. We also found the include chain: <mach-o/dyld.h> for _NSGetExecutablePath pulls thread_status.h, and ts_32/ts_64 are the only lowercase macros it leaks in that shape.

Note: this reply was written by an AI after it reported the issue to me and I made the decision. If anything here is wrong, reply and I will review it myself.