cyb/honeycrisp/audit/2026-10-09-unimem-block-creation.md

Checked native Block creation

2026-10-09, Apple Silicon. Implementation 6715335a3283e22218ad5f762316f2c735acb56f on merged nebu-identity parent 6560974dcd1a77b4b2e040a2b2a4187b702ace0b. This records a local combined candidate, with intentional API breaks and red workspace gates. It establishes no release, rung, GPU, pinning or Kadek row-6 closure.

Source and review identity

source-freeze.sha256 retains the original nine source/test/spec and two task-document hashes; manifest SHA-256 e28991a40b49a908ff1882ba14079217f39570187d2374c7601943c59368db87. The implementation commit contains exactly those nine frozen product files. The original full-overlay input was local nebu commit 76b2f26432cca15607ab1dd18e488df68db51ca0, strata 56aedb2d12b3126c601eb333419136d403614dbb, and those nine files; no task docs entered compilation. Original source record and archive hashes remain unchanged. Equivalence receipt is git diff --exit-code 76b2f264 6715335^ -- Cargo.toml Cargo.lock src acpu unimem rane aruminium, exit 0: the merged parent has the measured dependency/build-source bytes.

Fresh different-vendor PLAN and CODE reviews both returned APPROVE; reviewers ran no commands. Brief SHA-256:

  • PLAN: 4cba3dcdb74d4da3dd689178d320930b9e237fc2bfbd2e8bab4d7d2761b66080.
  • CODE: 7ef783a031bce898ba0e657c25da136146318f55b96bf1d28d3de355eb871ac5.

Invocation provenance distinguishes coordinator session-recorded exits from captured process receipts.

The raw CODE review's phrase “full slice::from_raw_parts contract” is too broad. This unit checks numeric and mapping bounds. Initialization, safe mutable-view aliasing, raw-handle/import lifetime, publication, physical residency, portable support and opaque CF allocator fallibility remain unresolved by this unit. The review separately identifies these exclusions; its original text is preserved.

Measured gates

Each receipt links to an exact .command record, process .exit, and raw output. Isolated receipts use one combined .log; overlay/baseline preserve stdout/stderr. All numbers below derive from those commands and the frozen source above.

Receipt Exit Observation
Isolated debug, release 0 Each: 50 native tests + 2 stable doctests
Isolated Clippy, fmt, docs, build 0 Clippy --all-targets -- -D warnings; no warnings
Nightly doctests 0 2 tests; actual matching nightly RUSTC/RUSTDOC enforce E0451/E0599
Native alignment 0 API alignment agreement, row shape and exact extent; includes 256 MiB
Semver origin, tag 100 Each: 202 checks, 201 pass, 1 fail, 58 skip; five added exhaustive MemError variants
Workspace fmt 1 Baseline-identical aruminium formatting output after root normalization
Workspace Clippy 101 Baseline-identical 16 unique acpu diagnostic/location pairs
Workspace release 0 Four pre-existing output-name collision warnings retained
Workspace tests 0 336 pass / 0 fail / 5 ignore; baseline 320 / 0 / 5; all 25 source-warning/location pairs identical
unimem build, tests 0 50 native tests + 2 doctests
Pipeline 0 Both reference verified: true and unimem verified: true; CPU/AMX and ANE executed
Bench 0 All 16 Criterion cases reached analysis; smoke execution only

Pipeline checks ANE identity output and finite FFN values; it has no executed GPU stage or full transformer golden comparison. Bench timings remain raw observations in the log, with no optimization claim. Five workspace ignores are unchanged (one acpu test, four aruminium doctests). See computed results and baseline comparison.

The first isolated Clippy failure is preserved. One new test lint and two old roundtrip lints were corrected: is_ok, nonnull ptr::dangling, and the exact 3.14f32 bits 0x4048_f5c3. Final debug/release cover those corrections. Standalone manifests drop unrelated dev dependencies/examples/benches only inside the harness; exact runtime dependencies and final per-candidate locks are in isolated/manifests. Original root/tag revisions are 9593c7218e14e2b5b816d9008b45ab60b9580cf3 / 96fa4f4509d7a966cb22a4bb9e12ac753cff43ae.

Compatibility and contract extent

Compiler probes independently show four raw-FFI source breaks: CFNumber type constant/parameter i32 → isize, CFDictionary capacity i64 → isize, and IOSurfaceCreate mutable → const dictionary input. The baseline compiles; the candidate reports exactly four E0308 errors. These were not reported by semver-checks. CFIndex is ABI-equivalent to i64 on Apple LP64; no previous runtime miscompile is claimed. Private plan construction fails E0451, the absent unchecked constructor E0599. No version bump or compatibility aliases are included.

The two specs add 51 lines / 3150 bytes by recorded wc -lc: README 480/17636 → 502/19371; API sketch 307/7884 → 336/9299. This adds a checked-constructor contract; it is not a strict-refinement size-reduction claim. Seven source/test files are 225/205/333/83/49/143/450 lines in manifest order, all within 500.

Reproduce

Requires macOS Apple Silicon, Nu, Git and the retained Rust toolchains/dependencies. Measured versions: Homebrew Rust/Cargo 1.95.0, rustc 59807616e, semver-checks 0.51.0; nightly-2025-11-26 rustc 80d8f292d; macOS 26.4.1/25E253, Mac16,5, 16384-byte pages. Exact version commands are in isolated/logs/.

From this audit directory, with repositories containing the pinned commits:

shasum -a 256 -c ARTIFACTS.sha256
nu check-records.nu
nu prepare.nu /tmp/block-repro --honeycrisp /path/to/honeycrisp --strata /path/to/strata
nu replay.nu /tmp/block-repro isolated debug-final
nu replay.nu /tmp/block-repro isolated build-origin
nu replay.nu /tmp/block-repro isolated build-final
nu replay.nu /tmp/block-repro isolated ffi-origin
nu replay.nu /tmp/block-repro isolated ffi-current
nu replay.nu /tmp/block-repro overlay tests

prepare only archives exact commits, applies checked hashes and copies the preserved standalone manifests/locks; no build, working-tree source, dependency shim or shared target is used. It requires a fresh destination. replay accepts any preserved isolated/overlay receipt except checkout-only diff; use table receipt names to run the remaining gates, including semver and actual nightly. It records and returns each process's real exit, including expected red gates. Original absolute paths stay in historical receipts; reproduction relocates them to private prepared roots. original/ retains the original harness scripts.

Reproduction checks preserve prepare/record verification and source-equivalence receipts. Preparation succeeds, refuses an existing output, and replay preserves both toolchain exit 0 and formatting exit 1. The first record-check script accidentally included compiler summary lines as warnings; that exit 1 is retained. Filtering summaries reproduces the original 25 source-warning pairs, exit 0; no product source changed.

Folder

Graph