el-retired/docs/v1/experiments/evidence/cycles/18-async-half-expressible/CLAIMS.md

CLAIMS — cycle 18, async — half expressible, and the cycle that was dogma

Source doc: docs/v1/experiments/cycles/18-async-half-expressible.md

This cycle is unusual: the doc already carries a chain-of-custody note stating that its original artifacts were destroyed and that it was rerun, not reconstructed. The verdicts below respect that. Where the original artifact is gone the verdict is NOT-CAPTURED, and the replication is graded separately.

Pinned commits

role	sha	subject
replication landed (system under test)	511db25230c69b9484ece39ed4131d3920b16596	rerun cycle 18 rather than reconstruct it
original measurement	—	no commit. The C stub lived in /tmp and was destroyed with the session worktrees

511db252 is the only commit that has ever contained lang/tests/integration/async_future.sh or lang/tests/integration/fixtures/future.c.

Verdicts

#	claim	artifact	commit	verdict
A1	the first attempt was dogma: pthread_join immediately after pthread_create, and the test printed the word DEFERRED itself	—	—	NOT-CAPTURED (artifact destroyed; the doc says so)
A2	second run, 4/4: [18 us] wrap RETURNS, [29 us] body START, caller continues, got 0, [50176 us] body END (computed 42)	—	—	NOT-CAPTURED (artifact destroyed; the doc says so)
A3	the caller got 0, not 42 on the second run	—	—	NOT-CAPTURED
A4	the fixture now lives at lang/tests/integration/fixtures/future.c and the harness at lang/tests/integration/async_future.sh	0002, repo tree at 511db252	511db252	REPRODUCED
A5	6/6 on the replication harness	0002-async-future-harness-6-assertions.out	511db252	REPRODUCED (exact)
A6	follow-on P1: el_val_t already carries tagged heap objects — 5 magic tags exist	0003-P1-five-magic-tags.out	511db252	REPRODUCED (exact)
A7	follow-on P2/P3: a future is one more; the caller awaits and gets 42	0005-P2P3-bound-run-interleaving.out	511db252	REPRODUCED
A8	follow-on P4: ZERO compiler changes — runtime C plus one binding	0004, 0005	511db252	REPRODUCED
A9	P1/P2 of the second run: the caller proceeds while the body runs; interleaving is observable in timestamps	0005	511db252	REPRODUCED
A10	follow-on P5: the unbound path still works — FALSE, SIGSEGV	—	—	NOT-CAPTURED (see note)
A11	the compiler builds at 511db252	0001-build-gen1-511db252.out	511db252	REPRODUCED

A5 — the replication, 6/6

0002, harness run against the compiler built from 511db252:

  ok    unbound: no construct, synchronous, correct result
  ok    unbound: el_await on a non-future passes through, no crash
  ok    bound: does not crash
  ok    bound: the awaited result is correct
  ok    bound: the caller continues BEFORE the body finishes
  ok    bound: the wrap returns in under 10ms while the body takes 50ms

  6 assertions, 6 passed, 0 failed

Exit 0. The harness's own header states it is a replication and not a blind test — the outcomes were known before it was written — and that grading is carried forward here unchanged.

A8 / A9 — zero compiler changes, demonstrated

The same El source (fixtures/probe.el), compiled by the stock elc, run twice against the same binary.

Unbound — no construct file, synchronous (0004):

CALLER_CONTINUED
RESULT 42

Bound — EL_CONSTRUCTS=c.txt containing work async wrap defer (0005):

WRAP_RETURNED 12
CALLER_CONTINUED
BODY_START 20
BODY_END 55034
RESULT 42

Same compiler, same emitted C, same binary; deferral appears only because a construct was bound after the build. The wrap returns at 12 µs while the body finishes at 55 034 µs, and CALLER_CONTINUED precedes BODY_END — the caller demonstrably proceeds while the body runs. The record's own numbers were 18 µs / 29 µs / 50 176 µs; the shape matches and the magnitudes are consistent (the body sleeps 50 ms by construction).

RESULT 42 is the follow-on P3: with a future as one more tagged heap object, the caller awaits and gets the real value.

A6 — five tags, and a future would be the sixth

0003, over el_runtime.c and el_runtime.h at 511db252:

#define EL_MAGIC_BIN
#define EL_MAGIC_GEOM
#define EL_MAGIC_LIST
#define EL_MAGIC_MAP
#define EL_MAGIC_MFLD

count: 5

and fixtures/future.c adds #define EL_MAGIC_FUT 0xE1F07000u in the fixture, not in the compiler.

A10 — why the SIGSEGV is NOT-CAPTURED

Follow-on P5 records that the first el_await read ->magic off an unvalidated slot and segfaulted on the unbound path. That version was never committed. The committed future.c carries the repair and a comment naming the defect:

el_val_t el_await(el_val_t h){
    if (h < 0x10000) return h;                 /* small ints / low addresses */
    if (h & 0x7)     return h;                 /* malloc returns 8-aligned   */
    ElFuture* f = (ElFuture*)(intptr_t)h;
    if (f->magic != EL_MAGIC_FUT) return h;    /* safe to read now */

so the harness's second assertion ("unbound: el_await on a non-future passes through, no crash") passes today. The historical failure has no artifact and one was not manufactured. The class it belongs to is independently reproduced in cycle 19 (0005-P6-sha256-hex-50000-exit-code.out, exit 139), which is the cycle P5 opened.

Provenance notes

All artifacts from /tmp/rerun-v1b-18, a dedicated worktree at 511db252, with capture.sh invoked by absolute path from inside it. No DIRTY-TREE flags in this cycle.

Three of this cycle's claims (A1, A2, A3) rest on a file that no longer exists. They are recorded as NOT-CAPTURED. Producing a /tmp C stub today and dating it to 2026-08-17 would be fabrication with a fresh timestamp; the doc's own chain-of-custody note already says as much, and this re-run does not go behind it.