Layout/SIMD probe, 2026-09-07 — four new features, no findings
Verdict: zero defects, zero doc gaps. Probed 2026-09-07 against packed, align(N), vector[N]T, and swizzling with a 57-line pixel kernel. Outcome: recorded because a probe that finds nothing is the one most likely to go unwritten.
The first probe aimed at the surface the freeze was broken for. packed, align(N), vector[N]T and groups/swizzling shipped 2026-09-04 and 2026-09-05; every probe record before this one is dated 2026-08-19 or 2026-08-20, so nothing outside this repo had driven any of them.
The agent wrote a 57-line pixel kernel using all four, and it worked on the first pass. Zero doc gaps, zero defects. That is the result.
What it exercised
packed struct over four u8 fields with a to_bytes / from_bytes round trip; align(8) struct over three f32; vector[4]f32 multiplied by a broadcast scalar; and bright.(r, b) = bright.(b, r), the swizzle form of simultaneous assignment. It read the emitted C and confirmed the four attributes landed — packed, aligned(8), vector_size, and a swizzle lowered as read-both-then-write-both.
The five things it reported, and why none is a finding
Each was checked against main before being dismissed, per docs/internals/probe-procedure.md's "check each finding against main" step.
| reported | verdict |
|---|---|
no ternary; if is an expression only in tail position | documented, and the agent quotes the section that says so |
vector[N]T elements limited to int/float/f32 | documented with its rationale; the agent wanted vector[4]u8 for byte work, which is a design objection |
to_f32(u8) "worked via implicit widening" against a doc that says float | the agent misread the doc. docs/reference/builtins.md:122 says numeric -> f32, and docs/spec/06-conversions.md:40 documents the conversion as total over any sized int/float. Re-run here: to_f32 on a u8 field gives 200.0. Nothing undocumented occurred |
no range in for | documented |
println is string-only | documented |
The third is worth keeping. An agent reporting a doc as wrong, when the doc is right and says so two lines from where it looked, is a failure mode a probe report will always be able to produce — so a report's own citations have to be opened, not trusted.
Run under the sanitizers, separately
docs/internals/probe-procedure.md says to build the agent's code yourself because the 2026-08-19 run's best finding was never in its log. Done here: the emitted C at -fsanitize=address,undefined -fno-sanitize-recover=all -O1 -fwrapv with detect_leaks=1 exits 0 and prints output byte-identical to the plain build. No hidden finding this time.
A defect in the procedure itself
docs/internals/probe-procedure.md states its goal as the agent being unable to read compiler source, then lists deletions that leave compiler/ in place — 19 files including compiler/parse/parse.ty and the typechecker, which is a syntax oracle for exactly the features under test. It was deleted by hand for this run.
Every round since compiler/ entered the tree had this leak. That weakens those records rather than voiding them — an agent that can read a parser is still not the author — but it is unmeasured, and the delete list names compiler.
What this does not establish
One program, one surface, 361 seconds. The agent's own "did not check" list is the boundary: no non-x86 host, the packed pad byte never read back, and align(8)'s trailing padding confirmed in the emitted C but never measured at run time. The program was thrown away; this record is the artifact.