Lossless archival Cargo manifests

Status: draft; fresh different-vendor PLAN approval required before edits to existing artifacts. Serves Kadek foundation row 4 through Radio; discovered from PR40's committed Git consumer. Parent: ce0fb312faa8770e80efe9afa864106ae6b51552.

Result

Rename exactly seven archived harness manifests from Cargo.toml to Cargo.toml.txt, preserving every byte. Keep source/workspace manifests and all dependency pins, locks, Rust sources, tests and specifications unchanged. Cargo must discover only the actual package when a fresh consumer pins an independent committed candidate snapshot. The candidate is a local packaging witness, never an origin release input.

Preserve the three original replay scripts under verify.original.nu.txt or reproduce.original.nu.txt beside their existing names. Replace only those entry scripts with thin Nu wrappers. Add one small shared staging helper in this task's new audit directory. Preserve each CLI and its defaults explicitly:

Preserve original --live checks, including their historical HEAD equality. They must still fail against a later merged HEAD. Research additionally checks a clean tracked diff; execution does not. Do not add or infer another cleanliness check.

The helper copies only its selected capsule to a newly created disposable /tmp directory, restores the explicitly listed Cargo logical names, and restores the original replay script at its historical filename by moving the copied .txt files, so the staged layout does not retain duplicate archived names. Replacing the copied entry wrapper is the sole intentional collision, after checking its identity; an existing logical Cargo path or missing archive fails before replay. It must never create a Cargo.toml inside the real repository. The original script then executes unchanged. All historical files.json, source indices, contexts, reviews, fixture bytes, raw manifests and command receipts remain unchanged. The wrapper is new packaging code, not the historically reviewed verifier. Its own bytes and the original-to-archive mappings are bound by the new capsule's manifest. historical-index-bindings.json records every affected index entry: index path and JSON pointer, historical logical path/hash/length, satisfying archived path, and deliberate direct-worktree outcome (missing logical Cargo path or wrapper replacing original script). An already historical pre-CODE verifier row is labeled separately and points to its existing preserved original. Physical indices are deliberately historical; a direct unstaged check is not the replay. Successful disposable replays are cleaned; failed ones retain their isolated directory and propagate the original nonzero status.

Steps and decisive checks

1. Freeze the seven manifest blobs, three original scripts, consuming paths and ce0 warning receipts. → Verify: every byte equals the exact committed object; discoverable audit manifest inventory is exactly the seven listed below; both original 2026-10-12 replay commands pass before mutation. These baseline checks are already measured; preserve them unchanged.
2. Rename the seven manifests losslessly and install only the replay wrappers, their archived originals and the small staging helper. → Verify: no audit Cargo.toml remains; the seven .txt contents match the original hashes; all real source manifests retain their paths and bytes. A complete committed tree comparison outside the explicit allowed audit/task paths must be empty, including root and cyber-bao manifests and root Cargo.lock.
3. Check the replay boundary. → Verify: research and execution wrappers reproduce the original successful JSON results with the same counts; their staging restores the complete original indexed file hashes including the original verifier itself. Run the unchanged proof CODE input verifier. Exercise 2026-10-09 staging and the original prepare-crate function for its historical parent/implementation. A child nu -c receives generated code with a quoted literal source path: source "<stage>/reproduce.nu"; prepare-crate "<repo>" "<rev>" "<export>" "<stage>". This makes the source path constant at the child's parse time and exposes the private definition; it does not call main. Revisions are db1d62e2… and 590e69ea…, fully pinned in evidence. Compare exported Cargo.toml, Cargo.lock, src/main.rs and the precise existing Hemera-dependency rewrite against inputs. Stage every capsule file, including Cargo.lock, probe.rs, SOURCE-FREEZE.json, FIXTURE-SHA256SUMS and expected-probe.stdout; preserve their exact hashes. The private-function mechanism already passes on both original revisions. Do not rerun the reproducer's codec matrix. A deliberate missing archived input/collision in a disposable copy must fail before running the historical script, leaving the real repository unchanged.
4. Freeze an independent committed temporary candidate snapshot containing the exact candidate tree: committed ce0 blobs plus an explicit allowlist of owned candidate changes, never another dirty worktree's inputs. → Verify: it is derived from ce0 with only the explicit archive changes, and all production source, normal manifest and lock blobs are equal. One private local Git object store contains both original ce0 and that candidate commit. Two independent fresh consumers use the same file:// Git URL and differ only in Radio revision and explicitly recorded consumer/cache/target paths. Each has a fresh private Cargo checkout/cache; neither uses a path dependency. Keep Hemera d5 pinned and acquire it through the same bounded cached-object mechanism; no new network claim.
5. Run a same-session differential. → Verify: before, between and after the two arms, /tmp/radio-bao-proof-parent-b933/cyber-bao/Cargo.toml exists with the same length/hash. Do not delete or alter it. The fresh ce0 arm must reproduce the duplicate warning; absence invalidates the witness. Lock generation, metadata (--locked --offline) and eight-case reader debug/release run in both arms with matched Rust 1.95, jobs=2, incremental=0 and isolated targets. Candidate commands must emit zero Cargo/compiler warnings and resolve cyber-bao from the exact candidate Git checkout. This warning-free claim covers only those invocations. Require byte-equal reader stdout across both arms/profiles and the frozen historical control, and equal (source path, bytes, SHA256) tuples for all selected cyber-bao/Hemera source files. Within each arm freeze inputs before/after. Compare metadata after only the explicit normalization below; do not require equality of changed Git pins or physical paths. This is not a broad codec gate.
6. Freeze compact new audit and fresh CODE input. → Verify: lossless historical replay, exact raw receipts, input mappings and source equality independently replay; different-vendor CODE approval precedes root's final commit/PR/merge. Preserve all historical red outcomes without claiming a new full-gate pass.

Comparison boundary

The historical snapshots are labeled historical-ce0-reader-before-runs.json and historical-ce0-reader-after-runs.json; their original timestamps/bytes remain unchanged, and v1's frozen packet stays intact. They span the old reader runs, not a repair. New paired measurements have separate ce0/candidate arm labels.

Within the new pair the local Radio Git URL is identical. Normalize only its validated ce0/candidate revision tokens in package/source IDs, the two consumer IDs and their exact consumer directories, the admitted checkout-root prefixes in manifest/target paths, and recorded target/build/workspace directories. Every replacement has an explicit old/new mapping and path-boundary check; no blanket regex stripping. Resolve-node IDs/edges use those same exact ID substitutions. Everything else must match, including features, dependency graph, package versions, rust-version, target lists and dependency requirements. Historical HTTPS and new local file URLs are intentionally different and are not a metadata equality claim. Snapshot/receipt timestamps and Radio commit identities are retained as distinct measurements, never edited to force equality. Consumer manifests/locks contain their explicit arm pins; only production source/manifests/locks and reader stdout have cross-arm byte-equality requirements.

File scope

Seven manifest renames, each adding only .txt:

audit/2026-10-09-kadek-slice-traversal/Cargo.toml
audit/2026-10-12-bao-proof-research/probe/Cargo.toml
audit/2026-10-12-bao-proof-research/primary-probe/Cargo.toml
audit/2026-10-12-bao-bounded-proof/fixture-probe/Cargo.toml
audit/2026-10-12-bao-bounded-proof/nostd/Cargo.toml
audit/2026-10-12-bao-bounded-proof/nostd-parent/Cargo.toml
audit/2026-10-12-bao-bounded-proof/seals/Cargo.toml

Three entry wrappers replace their names while preserving exact original bytes: audit/2026-10-09-kadek-slice-traversal/reproduce.nu, audit/2026-10-12-bao-proof-research/verify.nu, and audit/2026-10-12-bao-bounded-proof/verify.nu. New helper, mappings, documentation, review inputs and raw evidence live only in audit/2026-10-13-audit-cargo-packaging/; task artifacts live only in .claude/tasks/audit-cargo-packaging/. No nested historical audit archive is copied.

Five likely failures

1. A forgotten archived Cargo manifest still participates in discovery: exhaustive tracked basename inventory and a fresh committed-snapshot Git consumer.
2. Renaming silently changes historical bytes or hashes: compare all seven blobs, keep every historical index/receipt unchanged, replay the original scripts.
3. A wrapper reads temporary paths as the real source repository: resolve/pass explicit repo and sibling paths, including original CLI defaults.
4. Staging overwrites evidence or leaves an executable manifest in the repository: unique external destination, explicit mappings/collision checks, failure probe in a disposable copy, and before/after repository identity verification.
5. Packaging masks production/dependency drift: exact tree comparison, Git-pinned consumer identity, two unchanged literal result streams and honest warning counts.

Radio's cached registry part_of = soft3 and Kadek's direct dependency remain explicit. This is a packaging repair with no API/version floor change. There is no release, tag, registry publication or Kadek dependency update in this unit. The caller owns subsequent adoption of the actual merged Radio revision. Previous broad workspace/rung and last-tag gate results remain unrefreshed; no broader warning-free compiler, source, product or release claim follows.