acpu normalized package and published API — proposed unit

Read-only preparation, 2026-10-17. Honeycrisp canonical/origin HEAD is c7c9fabdb1e9a9b0be379917540355e89c899dfe, clean. No checkout, code, manifest, version or dependency edits have been made for this unit.

Placement and ownership

After root approval, recreate only /Users/master/cyber/honeycrisp/.worktrees/acpu-delivery/honeycrisp, branch chore/acpu-package-evidence, from that committed origin/main. /.worktrees/ is already excluded. Reuse the retained wrapper; archive its current task.json under the new owning audit before updating task/branch/base/retention metadata. Keep its existing Strata symlink target /Users/master/cyber/strata and revision 56aedb2d12b3126c601eb333419136d403614dbb unchanged. Cache: ~/.cache/cyber/honeycrisp/acpu-package-evidence/. Own .claude/plans/acpu-package-evidence.md and audit/2026-10-17-acpu-package-evidence/ only; Nox and production inputs stay fixed.

Smallest executable scope

1. Acquire official acpu0.2.0 archive/index with raw argv, responses/checksums, and compare complete source inventory/bodies with published VCS 96fa4f4509d7a966cb22a4bb9e12ac753cff43ae. Archive SHA256 is 81089272e2a248781b1e2c38ad7247a98bea63a602c0a445dc63e955b2035375. Require current owning compiler/config/asset inventory equal committed HEAD before diagnostic cargo package -p acpu --locked --offline --allow-dirty; audit-only dirt is permitted. Bind all package files, normalized/original manifests, generated lock and VCS identity. No source overlay or manifest fix.
2. Compare the complete library API on its sole supported target, aarch64-apple-darwin. acpu declares no feature matrix. Produce old/current rustdoc JSON in separate source/target roots using pinned Rust1.95.0 and explicit identical extraction flags; retain warnings. Run installed cargo-semver-checks0.51.0 with explicit baseline/current JSON and --release-type patch. No API filtering, shared-target discovery or claimed behavior/performance proof. Package target inventory separately covers bins, examples and tests that a library API checker does not adjudicate.
3. On the untouched normalized current package, run locked/offline debug and release cargo test --lib --tests, strict docs, and strict all-target Clippy; retain initial failures. Run bounded real matmul correctness, not performance benchmarks. This exercises merged admission/scaled-coefficient tests with registry unimem rather than the workspace's local unimem. Verify inclusion of the coefficient test's independent.tsv include_str input.
4. Replay the exact existing four-test Nox contract probe against normalized committed Nox and acpu packages. The approved scratch consumer override selects only the untouched acpu archive; record its manifest/lock and call it a prepublication diagnostic. Require one registry unimem and one registry strata-nebu identity, fast off, with matching Nox/direct Goldilocks type source. Run debug/release; preserve original probe semantics and hash. No owning Nox edits; actual unpatched registry acceptance waits for owner publication.
5. Reuse prior whole-workspace gate evidence only after exact current closure verification, and run required owning fmt/strict Clippy/release build/workspace tests/real matmul before the audit commit as required by Honeycrisp rules. Preserve all source epochs, exact command streams and a portable capsule verifier. Root independent CODE review precedes the ordinary audit commit.

Concrete packaging concern and limits

Current root Cargo.toml:41 declares NeBu with path/package but no version. Previous normalized acpu manifest drops it; acpu/bench/{zk,summary}.rs use it and remain declared examples. Thus all-target packaged checking is expected to expose unresolved NeBu even if default cargo package verification passes. Existing fix/nebu-package-identity only corrected the package name and explicitly left its version unspecified; that merged repair does not solve normalization. Measure the failure first and return a separate minimal source/manifest proposal.

Fresh official index reads still list acpu0.2.0 and unimem0.2.0 only. Unimem SHA256 a91794fe345844d09b134c55fde134bc9d51593964b906dde2f667162e49918d; its published VCS metadata declares dirty=true, so its archive checksum is authoritative. Registry strata-nebu0.1.1 SHA256 f4a881cdb3622cd6b38333246f516c70e3e5c63d2bc0c0e48344d8e7b0392766 has empty defaults; source Strata and registry tier versions remain separately identified.

The occupied current version0.2.0 archive remains diagnostic. An acpu-only0.2.1 bump is only a later proposal after actual package/API outcomes. Owner alone merges a bump, tags or publishes. Nox's later consumer repair must raise its acpu minimum to0.2.1, not merely refresh a lock; none of those edits belongs here.

Accepted normalization addendum — 2026-10-17

Root approved the measured repair in audit/2026-10-17-acpu-package-evidence/normalization-repair.md: add only version = "0.1.1" to the existing root workspace NeBu path/package dependency. Keep its path, package, defaults, Strata target/revision and all crate versions. Inspect root Cargo.lock explicitly; no graph change is intended. The normalized package must retain the published NeBu dev dependency used by its examples. The initial package/all-target failures remain immutable evidence. The proposed 15-file example lint repair is recorded separately for root review before edits. Final package preparation requires exact owning inputs equal committed HEAD; an intermediate red source commit, if needed, is explicitly wip: with the actual first error quoted. Ordinary owning all-target repairs come first.

Root also approved hygiene-proposal.md in full: exactly the 15 listed example/bench files, preserved float bits/AMX operands/timed operation order, no lint allowances or target exclusion. The compiled literal proof succeeded before production edits. Any wider source correction requires separate review.

Root approved the measured two-file extension hygiene-addendum.md after owner-1 exposed the previously unresolved examples: summary.rs and zk.rs. Preserve quantization's round→max→min sequence/NaN result using a temporary; combine only identical formatting arms at the retained 10_000 and 1 thresholds. The original first-stage diff/manifest and its actual red remain retained.