ARCADIA-Studio/harness/experiments/analysis-source-s2b-windows-20260913/archive-review.md

# Independent S2b evidence archive audit

Reviewed `.local/analysis-source-s2b-evidence-20260913` without executing its archive helper, gate/admission helpers, compiler, tests, GUI or Gitea. No archive, root source, HEAD or index edit was made. This report records file/JSON/log inspection and read-only Git object comparison; it does not claim publication or public source8 acceptance.

**No actionable archive finding.** The archive accurately preserves a private lowering checkpoint with a failed full library gate. The corrected source is `a2a28f8d23a46ba1696b6a8410d4d3b5c1808f49`; original failed source is `472fe7a4339a23a3339b2b91e8d53d76442f60c4`; comparison baseline is `b2c970b4673e062e47eadd12906d2316c4b1a7b3`.

## Bytes and provenance

- Manifest: 12,001 bytes, SHA256 `d5b2668207438e319c2ab98395e2483c016088c5446b6bc99d797a066384a62f`.
- All 40 raw archived files match manifest lengths/hashes and their still-available original local bytes. Both generated files (`audit.json`, README) match their manifest bindings. These 42 payload files total 1,254,441 bytes; including the manifest, the exact archive tree contains 43 files. There are no unlisted files or missing payloads.
- `audit.json`: 73,824 bytes, SHA256 `34d10fafb8046943d7d0a1bd0ffee9de334bdff37019b033d1b2bbbf057d78c4`. Independently recomputed runtime outcome maps match it. All 32 full-failure section hashes/starting lines and complete diagnostic excerpts match the original raw logs.
- The bound archive helper is preserved as evidence; it was not rerun. Original local paths identify provenance, while actual archived files live under `evidence/.local/`. Binaries/toolchains are explicitly omitted. Recorded historical executable identities are preserved, not claimed to have been reread from mutable target paths.

## Corrected actual gates and admissions

| Gate | Actual result | Result SHA256 |
| --- | --- | --- |
| list-a2a28f8 | exit 0; 142 unique names | `86db5467964d3712510f867379ae32e0797b4d9f46cbde762119ed9055fd3a62` |
| typed-a2a28f8 | exit 0; 34 passed, 108 filtered | `9d8b9e432e31ffc4b621c38b80650a20846588cf96d6fc0f5dba3996ebda8d3d` |
| tokens-a2a28f8 | exit 0; 1 passed, 141 filtered | `7baa9e8c8d35f4e57e08539de69e215dbbeedbeb049585196f5e7109c1596890` |
| full-a2a28f8 | exit 101; 126 passed, 16 failed, 0 ignored | `81e171692b5c782bd234888e94468853b47a0e54375106e4cfce831e7b58e6c1` |
| strict-a2a28f8 | exit 0; all-targets Clippy, no warning/error in log | `fbee35f759af77124e83a646d15b34129ffe9a3873b34ac1ae8a600e89cc150e` |

All five receipts have matching source/head, nonempty ordered start/end and unchanged inputs. Every corrected input manifest is the same 531-row manifest, SHA256 `08b6dabbffbd06f7abea717f9aa34cd910c41c4a8c650247006a0f6858561c84`; its membership, sizes and hashes were independently compared with the exact committed input scope and raw Git blobs. Original472 likewise has 531 exact inputs; baseline b2 has 524. No source count was inferred from a report.

All four corrected test runs record one identical executable: 36,685,824 bytes, SHA256 `ce9d77ed1e68a21e06836264bf50ba9dddbfa69507e1a0bc2c05a7b55db325a9`. Strict records no test artifact. Actual commands match list/focused/full `oa-core --lib` and strict `oa-core --all-targets`; this is not a full all-targets runtime or doc-test claim.

Exactly three corrected admissions exist: listing, typed and raw tokens. Their input/result/log/helper/baseline hashes and executable identities match the archived originals. Their admitted names equal the actual selected raw outcomes. No corrected full-pass admission exists, and the README does not imply one.

The corrected and original472 listings each contain the same 142 names. Independently subtracting b2's 129 names gives exactly the 13 test functions authored in `language/analysis_source/lower/tests.rs`; all 129 prior names remain. The 13 new names match `audit.new_members` exactly and each is `ok` in both corrected typed and full logs. Entire ordered runtime membership and summaries were checked, including no unexpected ignored outcomes.

## Preserved failures and cause scope

The original472 focused log retains its actual 33 passed / 1 failed / 0 ignored, exit101. The failing `all_thirteen_expression_variants_lower_to_real_typed_values_and_admit` reports `analysis_type` for AsKind at `lower/tests.rs:45:54`. The archived correction review identifies an invalid named-input positive oracle. The 472-to-a2 Git diff is only that test file, 19 additions/5 deletions; it does not change production lowering or numeric semantics. The corrected run is separate evidence, not relabeling of the original failure. The original test artifact remains distinct: 36,686,336 bytes, SHA256 `732018626fceff6c6a75ea51e114a32ca96a11999cf6f135c077df49278853b7`.

Actual b2 full is 113 passed / 16 failed, exit101, with artifact 35,920,896 bytes, SHA256 `67249b37f30b596eb3deb95a217bd77eb06905bdf7b6c27b74fec16f9f0362e5`. Independently compared all 16 a2 and b2 panic blocks: failed names, panic source locations and complete diagnostic lines are identical apart from surrounding thread IDs. Each rejects source version at line1/column9/span8..9. A matching-name assertion alone was not used to establish this result.

The archived failure-cause review is appropriately bounded. Source readback independently confirms `model.rs:9` has schema8; `language.rs:506` writes that version while `language.rs:1335-1337` accepts only3-7. Direct nesting round trips fail at `language_nesting_tests.rs:63,124`. For bundle cases, `source_project.rs:297-298` invokes upgrade, `source_project/source_migration.rs:83` writes the current version, and line115 reparses it through the unchanged public parser. The transaction fixture begins as valid source5 and fails at `source_files.rs:1560` before its later save/recovery operation. All seven files named by the cause review, including the checked-in source3 entry, are byte-identical between b2 and a2. The 16 failures therefore do not validate later nesting/save/recovery behavior and cannot be waived into a full pass.

The cause review explicitly predates tokens/strict and does not claim their completion; their actual later receipts are separately present. Historical b2 production-language 7/9 is mentioned only as a separate S2a limitation, not a corrected a2 run. Public source8 writing/migration, native workflows and release acceptance remain outside this archive's successful scope.

## Readback and limitations

All manifest payloads and originals were checked again after review; manifest bytes remained unchanged. No new source test or artifact execution occurred. The archive is internally consistent evidence for the stated private milestone, with full-library failure still explicit. Source bodies/toolchains/executables are not packaged here; later publication must retain access to the identified source commits and must not promote this archive into public8, native or full-suite acceptance.