[viridale]

Split out of #1596, as asked in its review.

What changes

In sequence, a fork's middle steps read the words they hold from the frame without popping it (STK(-k)). When such a step shares one of them, term_keep may wrap a raw root in a redirect and return a new word, but the frame slot kept the old one. The step that pops the frame then read the raw word, and the join's sink freed the cell the redirect still pointed at.

In Push below, Xs.cnt borrows the list and Xs.push stores it. At --threads 1 the freed cell loops the free list and the program hangs on main:

def Push(+xs: Xs) -> U32:
  a b = Xs.cnt(xs, 0) Xs.push(xs)
  (a + Xs.cnt(b, 0) : U32)

bind_pop now writes the kept word back to the slot it came from:

STK(-1) = _xs_1 = term_keep(e, _xs_1, 1);

Witness

tests/reg/brw_fork_push.bend. The gate runs each binary at the default thread count, where a direct Push(Xs.build(100n, 1)) takes the parallel path and passes. So the test runs Push at the leaves of a depth-12 fork tree, and the leaves run in sequence at any thread count. On main it hangs at --threads 1, 2, 8, 128 and the default. On this branch it prints 823296 at all of them, as do the JS lane and the interpreter.

Checks

• A/B against main over the 809 programs that build to C and JS (tests, benches, demos). The code changes in 4 programs and no bench:

  • brw_fork_push gains 1 write-back;
  • brw_fork_cross gains 1;
  • compile/family_two_indices gains 3, for the quadrants that Mat.sq shares, whose second use read the stale raw word;
  • the app_slash_boss_3d demo gains 10, in Cloth.mesh and Scene.build.

  Otherwise, C and JS output lose only the cut runtime comments listed below.

• brw_fork_cross prints 199 and family_two_indices 64 and 512, as on main. SLASH_PROBE=600 on the slash demo prints the same game state as main at --threads 1 and 16. Only the audio queue and voice counts and the frame timings differ, and those vary between runs of either build.

• gates/test.ts, run locally: PASS 1638 / 1640. The two failures, gfx_clicks and gfx_window, expect "no display" and fail the same way on main here, because this Mac has a desktop session.

• tsc -p bend2/pack/tsconfig.json (TypeScript 5.9.3): the same 11 errors as main, none in comp.ts.

• bun gates/repo.ts: 54 / 54.

• The A/B, the program runs and the test gate above were done before the last five comment cuts (ring_rewind, io_park_add, the Heap and Blk comments, show_main). Against that build, this branch changes only comment lines: in the C of all 809 programs, and in no JS.

comp.ts tokens

Main is at 64,092 / 64,000 by itself, so the repo gate currently fails on every open PR. This PR brings comp.ts to 63,990. To fit, these comments are cut, each one a twin or a restatement of code:

• C runtime:
  • the twin of the JS io_list comment, "Bytes cross as they are…" (the JS one says it in full);
  • io_park_add's double-link note (io_park_cut);
  • ring_rewind's comment (its body);
  • the Heap comment's bump sentence (the sentence before it lists the misses in order);
  • the Blk comment's blk_node and blk_copy mappings (their call sites).
• JS runtime: the show_val comment (the C twin says it in full).
• Emitter:
  • seg_new's comment. It restated seg_take and the runtime's Segments comment, and said a segment pops its frame, which a fork's middle steps do not.
  • show_main's last sentence (the refuse error and the io_type check below it).

#1596, rebuilt on main, makes a disjoint set of cuts. The two merge cleanly, and with both merged comp.ts is 63,886.