Archived
will
d3495476f4
El SDK Release / build-and-release (push) Failing after 13m0s
The generated C, amalgams, vendored runtime pins, and compiled binaries from the Claude Code era are removed from the worktree. The El sources survive; this tree is now source-only for the first-principles rebuild. Per Principal direction 2026-08-19.
103 lines
4.4 KiB
Markdown
103 lines
4.4 KiB
Markdown
# DEFECT — string literals containing control bytes compile to a heap address
|
||
|
||
Not a cycle claim. This is a **live miscompilation in the current compiler**, found
|
||
incidentally during the cycle 01–10 evidence re-run and confirmed at branch HEAD.
|
||
|
||
It is filed with the evidence because it invalidates a *method* that cycles 01–04 rely
|
||
on: "the emitted C is byte-identical" is only meaningful if the emitter is
|
||
deterministic, and for affected inputs it is not.
|
||
|
||
## Provenance
|
||
|
||
Captured from a clean detached worktree at `9540f2399` (`iteration-2` HEAD), compiler
|
||
built at that commit. All artifacts `tree=clean`.
|
||
|
||
Found by the cycle 04 agent while checking prediction 5 ("existing `@manager` output
|
||
byte-identical") across the whole tree: of 378 `.el` files, 368 emitted identically
|
||
under the parent and the commit compiler, 9 failed under both, and **1 differed — and
|
||
differed from itself between two runs of the same compiler.** I reproduced it
|
||
independently at HEAD before accepting it.
|
||
|
||
## The defect
|
||
|
||
A string literal containing a raw control byte with no escape mapping is emitted as a
|
||
**decimal heap address** instead of the string.
|
||
|
||
`0002-control-char-vs-empty-and-printable.out` isolates it exactly:
|
||
|
||
```
|
||
8: el_val_t a = EL_STR("");
|
||
9: el_val_t b = EL_STR("ok");
|
||
10: el_val_t c = EL_STR("4324455808"); <-- source literal is a raw 0x1F
|
||
11: el_val_t d = EL_STR("\t");
|
||
```
|
||
|
||
- empty literal `""` — correct
|
||
- printable literal `"ok"` — correct
|
||
- **raw `0x1F` (unit separator) — emitted as `4324455808`**
|
||
- raw `0x09` (tab) — correct, `"\t"`
|
||
|
||
Tab has an escape mapping and survives. `0x1F` has none, and instead of being escaped
|
||
it falls through to a path that stringifies a pointer.
|
||
|
||
## It is nondeterministic
|
||
|
||
`0001-minimal-reproducer-control-char-literal.out`, three consecutive runs of the
|
||
**same** compiler on the **same** source:
|
||
|
||
```
|
||
--- emitted run 1 --- el_val_t sep = EL_STR("4344967392");
|
||
--- emitted run 2 --- el_val_t sep = EL_STR("4347851136");
|
||
--- emitted run 3 --- el_val_t sep = EL_STR("4384031472");
|
||
```
|
||
|
||
The value tracks the process heap base, which moves under ASLR.
|
||
|
||
## It occurs in real code in this repository
|
||
|
||
`0003-real-file-nondeterminism-arbor-parse.out`. `arbor/vessels/arbor-parse/src/main.el`
|
||
uses `0x1F` as a field separator — El source line 259, shown through `cat -v`:
|
||
|
||
```
|
||
if i > 0 { let out = out + "^_" }
|
||
```
|
||
|
||
Two runs of the same compiler over that file produce different sha256, differing in
|
||
three places, each a separator literal rendered as a different address. Within a single
|
||
run the corrupted values differ by fixed offsets (`+2240`, `+32208`), consistent with
|
||
distinct literals in one heap arena; across runs the whole block shifts.
|
||
|
||
The file applies **no decorators at all**, so this is unrelated to the construct work
|
||
in cycles 01–10, and it reproduces at cycle 04's parent `4f7568b` as well as at HEAD.
|
||
It predates the cycle series.
|
||
|
||
## Why it matters
|
||
|
||
1. **Silent miscompilation.** `arbor-parse` builds and links. At runtime it joins and
|
||
splits records on a separator that is a decimal rendering of an address rather than
|
||
`0x1F` — and the encode and decode sides get *different* values whenever they are
|
||
compiled separately. No diagnostic is produced at any stage.
|
||
2. **Reproducible builds are broken** for any source containing such a literal.
|
||
3. **It weakens byte-identity as a control.** Cycles 01, 02, 03 and 04 all assert
|
||
"emitted C is byte-identical". Those assertions were checked on decorator-bearing
|
||
sources, which are unaffected — the cycle 04 agent verified all 3 `@manager`/
|
||
`@accessor` files emit identically — so no cycle verdict changes. But the control
|
||
itself is only sound on inputs that avoid this defect, and nothing in the record
|
||
says so.
|
||
|
||
## Verdicts
|
||
|
||
| Claim | Artifact | Verdict |
|
||
|---|---|---|
|
||
| Control-byte literals emit as an address, not the string | `0002` | **CONFIRMED** |
|
||
| The emitter is nondeterministic for such inputs | `0001` (3 runs) | **CONFIRMED** |
|
||
| A real file in this repo is affected | `0003` | **CONFIRMED** |
|
||
| The defect predates cycles 01–10 | cycle 04's `0049`-series (parent `4f7568b` shows it too) | **CONFIRMED** |
|
||
|
||
## Not fixed here
|
||
|
||
This evidence pass does not change code. The defect is recorded, reproduced, and
|
||
minimally isolated so it can be fixed and regression-tested separately. A fix should
|
||
escape all non-printable bytes in the literal emitter and add a test asserting that
|
||
compiling the same source twice yields byte-identical output.
|