acpu0.2.1 preparation tooling and evidence

Accepted plan, initially committed as 0995cf9e51b5858df13f8349b083d409e0b306f5, now re-anchored to reviewed PR25 merge 14373cc8a0e18302751844579d94d1f8f80bd8b1. Pushed preparation 9e6cffe699278059c8940f13940727b10569e13e and all prior epochs remain retained. The source-only bump is committed separately at 2c70b3b929c791ca489bba12d2f5a6daef7d43e7. RESULTS-2.md identifies the final capsule and independent replay. Owner-only bump merge, fresh default-branch delivery, tagging and publication remain pending.

This directory contains Nu-only orchestration. common.nu imports the previous audit's generic byte/receipt/archive helpers unchanged. replay.nu binds raw receipts to fixed tool/source identities and rederives the reported outcomes. export.nu exports exact committed tooling and that helper with relative paths preserved, so the source-only bump checkout needs no tooling or audit additions.

After independent tooling review and its preparation commit, use explicit paths:

All target/output paths must be explicit external cache/tmp paths while the bump branch is active. Successful evidence later belongs on the preparation branch; the owner-facing bump PR contains only acpu/Cargo.toml, root Cargo.lock and acpu/CHANGELOG.md. Any red remains attached to its original epoch.

Historical tooling validation below precedes the candidate gates retained in candidate-2. Those gates and root's independent portable replay passed; independent-replay-candidate-2 preserves the root result and metadata receipt.

review-epoch-1 preserves the first review's source bytes and manifest unchanged. validation-2 records parsed entrypoints and the real unbumped-source refusal. validation-3 preserves the first failed guard test and its exact source bodies: Nu each deferred an error value. The eager-loop repair passed validation-4: one real retained registry control, four synthetic corruption refusals and exact current owner bytes equal the previously green source epoch.

validation-5 records the repaired closure in sources.json; its compact JSON SHA-256 is a685d5b1df3f959933002de58508e1d0e59cc2555afee371ec71ccc2bf08abb3. The exact commands and revisions are in its receipts and the accepted plan. Its replay controls check installed API binary hashes and raw versions, compiler environment selection, raw historical test summaries, actual matmul output, normalized package targets, consumer manifest/source paths, direct NeBu version, clean VCS, and the independently fixed Nox archive checksum. Synthetic changes are explicitly labeled and rejected; real inputs remain unchanged.

The historical PR21 raw workspace evidence contains 365 passing and 5 ignored tests across 28 summaries; each acpu library/integration run contains 196 passing and 1 ignored test across 6 summaries. These numbers are replayed from the merged PR21 receipts, identified in replay.nu and replay-check.nu; they are not acpu0.2.1 results. No compiler gate was rerun for that tooling-only unit. Its tooling-review.json and root approval are preserved unchanged; they apply only to the old source epoch. review-epoch-2 preserves that approved preparation's active files.

The new reanchor-1 records fresh target metadata, the complete PR25 compiler source bridge, replay of its eleven gates, current raw totals, entrypoint parsing, unbumped-source refusal and the updated tooling controls. The metadata command is cargo metadata --no-deps --format-version 1 --locked --offline; its inputs equal base 14373cc8a0e18302751844579d94d1f8f80bd8b1. It yields 40 targets: 32 examples, six integrations, one library and one binary. The integrations include the new gemv_admission and gemv_independent targets.

Current raw totals come from audit/2026-10-17-gemv-tape-admission/tape-1 commands cargo test --workspace --locked --offline (388 passing, five ignored, 31 summaries) and cargo test -p acpu --lib --tests --release --locked --offline (212 passing, one ignored, eight summaries). Its complete source identity is 9de4e9df1ff33dfca56a83ba8074f698658124aae914f0c2a98707f76487ebc3 and its compiler inventory equals the explicit merged base. At that preparation epoch, new package totals were expectations; the final capsule now retains the fresh candidate outcomes. The old 38-target and test counts are retained only as historical PR21 replay controls.

tooling-review-reanchor-1.json freezes the new review inputs. New validation does not reuse the old root approval. Root reviews this closure before the three-file bump or frozen candidate gates; that review and execution are now complete. The normalized package keeps registry unimem0.2.0: although acpu reexports its Tape, the owning-source Tape repair is not delivered by this acpu bump. The registry identity proof remains intact.

validation-repair-1.md and validation-repair-2.md retain the candidate's tooling refusals and narrow repairs. Owning, acpu package and API commands keep their original production bundle; the corrected consumer and validator use a separate committed bundle. The final capsule retains three exact bundles: original producer, successful consumer, and the later validator with the reviewed macOS path-alias repair. Each is bound to its own Git revision. The old package runner's missing result summary is never reconstructed; raw successful commands and archives are independently replayed by the validator.