Blake batch owner audit

The Blake benchmark now owns initialized input/output allocations in one private Batch<'a>. Exclusive CPU views and synchronous GPU dispatch preserve allocation, device and completion lifetimes; unsuccessful native completion closes access. This unit follows Honeycrisp PR14's raw-access repair and serves Kadek foundation row 6. RTable/mine ownership and the physical-page policy remain open.

Frozen inputs and review

Source freeze records every changed path and hash, SHA-256 297e4959c0f4257c7b985ee9aeaa0194677e367a6e0e7ec0748bd187298d2d98. No post-freeze source edits. Original working trees were not build inputs or edited. The sole lock change deletes obsolete libc from unimem 0.2.0's normal dependency list; no package version, checksum or manifest changes.

Fresh independent Claude Opus PLAN and CODE reviews both APPROVE. CODE found no blockers; optional cosmetic changes were deferred. Review identities and hashes bind both verdicts to their frozen inputs. The root agent owns final acceptance, the decision label, commit, PR and merge. Passing scoped tests does not make the red workspace or the deliberate source incompatibilities green.

Execution

evidence.tar.gz preserves 140 payloads: complete receipts, 21 command records, all 42 raw streams, source snapshots, tested intermediate overlays, original locks, source/archive inventories and review input indexes. capsule-inventory.json identifies every archived payload. Archive SHA-256: 169ebb6357b95fd8c4252b6ff108a6d5224135bb50071a61ccaf08b3427bdb55. Archive README.md contains the exact command table; commands.json and logs/<label>.json retain argv, environment, executable paths, times and hashes. Twelve commands exited 0 and nine exited nonzero, including the initial corrected Clippy finding and both parent/candidate formatting diagnostics.

Native tests run on Apple M4 Max with pinned rustup 1.95.0, LLVM 22.1.2 and SDK 26.4.1; error-code doctests use actual nightly-2025-11-26 compiler and rustdoc. Archive toolchains.json records executable paths and hashes. Jobs were capped at 2 with private targets. No benchmark, huge epoch buildbench, mining, share/network or powermetrics loop ran. Packaging was not run.

The first scoped Clippy command found manual_is_multiple_of; the equivalent arithmetic correction preceded the freeze and final debug/release checks. Workspace diagnostics now expose three E0133 and one E0596 in RTable. These are remaining failures induced by Honeycrisp PR14's raw-access boundary, newly visible after this Blake migration. Further mine-owner errors and the two earlier chime overflow failures remain masked. Full Erga/rung gates remain red.

Semver uses a compile-only diagnostic parent with historical Honeycrisp 7b148906263da5ea8aabc5ddb0e2e44ff2288ecc and the same one-line lock normalization; no legacy raw-owner runtime executed. Intentional breaks include removing ZeroCopyBuf, replacing alloc_buffers with lifetime-bound alloc_batch, private factory fields, exclusive batch dispatch/views and fallible fill/verify APIs. Archive README.md catalogs type, lifetime and result changes beyond the tool's findings, including its misleading doc_hidden name for newly private fields.

Evidence boundaries

Real tests cover all four shaders, counts 1, 2, 3, 4, 63, 64 and 65, two fills and reuse, all output bytes, poisoned full-allocation initialization, padding canaries and pre-submit preservation. The independent published vector comes from pinned BoringSSL/Go generation; the existing known_vector_abc test is a false witness and supplies no independent evidence here.

Failed completion uses a simulated status/error matrix against the production completion helper plus source inspection. Different-device rejection and relative Buffer-before-Block destruction rely on source evidence; this single-device run does not measure them. No fabricated handles or invalid allocations were used. VM-region compatibility does not settle physical contiguity or pinning. Generic provider soundness, Kadek row 6, performance, packaging and release remain open. Zeroing both complete backings changes allocation cost; old timings are not comparable and no performance claim is made.

Full Apple SDK/web/PDF snapshots and full review packs remain in acquisition caches; primary provenance retains their pins and hashes.

Replay

Run from the repository root:

nu audit/2026-10-10-blake-batch-owner/verify.nu audit/2026-10-10-blake-batch-owner

The wrapper checks the archive hash, extracts into its own temporary directory, verifies all capsule payloads/records/streams/overlays/locks and removes that extraction. It replays evidence integrity without rerunning native workloads. Recorded replay exited 0. Historical absolute paths inside records identify actual execution directories. The root agent creates the final audit checksum manifest after adding its acceptance and merge receipts.