acpu delivery prerequisites

Agent /root/phase2_resume_audit under /root; 2026-10-16. Owner checkout honeycrisp/.worktrees/acpu-delivery/honeycrisp, branch chore/acpu-delivery, base 70dfb1f55e4cb2570d5a78a497be5828a203087b. Wrapper context preserves Strata 56aedb2d12b3126c601eb333419136d403614dbb; initial tracked/untracked/ignored status is clean. Build cache: ~/.cache/cyber/honeycrisp/acpu-delivery.

Initial ownership is this plan and audit/2026-10-16-acpu-delivery/ only. Run the whole-upstream gates before proposing any source or version changes: workspace fmt check, strict Clippy, release build, workspace tests, and the real acpu matmul example. Root/acpu CLAUDE.md define these gates. Record frozen source bytes, commands, toolchain, dependency revision and complete output; retain reds. Do not infer performance from this busy machine or from matmul's timing output.

Then inspect newly shipped sparse APIs, field malformed-input contracts and published/source API compatibility. Existing bounded arithmetic diagnostics in Nox establish the missing APIs' useful contract, not whole-upstream readiness. Prefer the existing upstream implementation and package-only 0.2.1 proposal; no duplicate kernels, fallback, dependency substitution or version edits now.

Review findings with root before ordinary fixes. Keep source fixes separate from the later version-only PR. Preparing that proposal is authorized in principle, but only the owner merges a bump, tags or publishes. Preserve this checkout until the reviewed delivery work and evidence are integrated.

First source unit for review: public input admission

Current source has a release-blocking soundness hole: safe CSR entry points can reach unchecked aarch64 reads/writes with arbitrary column indices or short output slices. Root confirmed this blocks publication. No UB witness is executed.

Proposed contract, retaining the existing safe signatures and panic convention:

• CSR: require a nonempty row-pointer array starting at zero; nondecreasing row offsets; final offset equal to the common column/value length; output length equal to the row count; every column less than the input length. Validate the entire input before pointer arithmetic, prefetch, unchecked loads/stores or mutation. Invalid inputs panic in debug and release with a field-specific message; all output/scratch arrays remain unchanged. [0] with empty data and output is the empty CSR. Rectangular matrices, duplicate/unsorted columns and empty rows remain valid. No floating-point value restriction belongs to this memory-safety unit.
• Chebyshev matvec: require row count equal to x length and d_inv_sqrt/y/t0/t1 lengths equal to that count, then the same CSR admission before scratch/output mutation. Existing recurrence and valid-input floating-point order stay intact.
• Field: require positive power-of-two lengths for public bit reversal and multilinear evaluation in every build; multilinear point length must equal log2(evaluation length). ntt_forward already rejects invalid lengths. Preserve caller-supplied root semantics and the existing algorithms.

Approved edit scope: acpu/src/sparse/{mod,chebyshev}.rs, optional private sparse/admission.rs, acpu/src/field/{ntt,poly}.rs, focused public regression tests, and compact contracts in existing acpu/specs/README.md/rustdoc. No version, dependency, FFI, AMX instruction or kernel scheduling changes. One shared CSR validator is sufficient; use established kernels after admission.

Regression scope: empty and rectangular CSR; accumulate/set semantics; duplicate columns; empty rows; scalar and four-wide rows; malformed partition/terminal, column/value counts, output length and column indices; every Chebyshev shape; invalid field lengths and both short/extra points. Check panic-before-mutation and valid arithmetic equivalence in debug/release. Freeze the dangerous CSR witnesses but execute them only after source review establishes the new checks precede every unsafe operation. Their parent baseline is the static soundness proof, never an executed out-of-bounds call. Safe field baseline outputs are already retained in the audit. The existing unmerged 904dd336 length-check branch is useful prior work but lacks CSR and point-count validation; do not merge it blindly or duplicate the algorithms.

Root approved the source contract and complete guard diff before any malformed CSR test ran. Implementation uses a private immutable Csr view with checked partition/columns and O(1) vector-length checks at each kernel call. Chebyshev constructs it once; field wrappers validate once before private kernels/recursion. The numerical algorithms and sparse unsafe loop contents remain unchanged.

Verification questions resolved by source inspection and public tests:

1. Can a later malformed row mutate an earlier output? Complete admission scans the immutable partition and every column before apply; late-column tests retain y.
2. Can a reused view receive different vector lengths? Apply checks both lengths before zeroing/prefetch; Chebyshev scratch has its admitted fixed dimensions.
3. Can empty input or a mismatched point reach polynomial recursion? Public length and exact-coordinate assertions precede the sole private recursive call.
4. Are valid sparse representations narrowed? Rectangular, empty, duplicate and unsorted cases remain accepted; no floating-point restrictions were introduced.
5. Does a shape error allocate coefficients first? All shape checks precede that call; huge-order malformed fixtures require the specific shape panic.

Separate work, still before version-only PR

Hygiene: one formatter file, duplicate allow attribute, three sparse Clippy patterns, and unused trace64 example binding; preserve initial reds. Whole workspace tests must then run beyond compilation. Proposed numerical follow-up: Chebyshev coefficients overflow for finite positive tau=100 and become NaN at 104 although scaled coefficients are bounded. Review the supported numeric domain and stable scaled evaluation with independent reference cases; do not invent a silent clamp or arbitrary tau cap. This numerical issue remains open after memory admission and is not repaired by the four existing Bessel examples.

Admission is committed in c5177c5 with its historical red Clippy gate retained. The separately reviewed hygiene unit now passes the required workspace gates, strict acpu docs/tests, all-target compilation and actual CPU/GPU/ANE matmul controls; exact receipts are in audit/2026-10-16-acpu-delivery/delivery-v4/. Its authorized scope extensions correct benchmark vDSP stride ABI, unique example names and parent-package automatic bench discovery, with independent SDK/metadata proofs. Versions, dependencies and numerical algorithms remain unchanged. Numerical scaling is the next distinct source unit.

Only after reviewed source fixes and all owning gates: prepare the package-only acpu 0.2.1 version/lock/changelog proposal with owner approval for the explicit package-version exception. Full gate/source evidence follows that exact versioned commit; owner alone merges/tags/publishes. The present source is not release-ready.