ARCADIA-Studio/harness/experiments/peer-union-windows-20260923/README.md

Combined peer candidate — partial Windows verification

This checkpoint records completed gates through strict workspace lint at source f82e29d6b8ed2b5fd23ac806675d3eafe851d104, published and ref-verified on codex/peer-union-956atf-20260923. It is not full workspace, desktop/native, public source8 or release acceptance. The live integration task is Gitea issue27; the source provenance and reviewed seams are in f82e29d:harness/peer-union-20260923.md.

What ran

Windows Rust/Cargo 1.98.1 gnullvm, LLVM-MinGW 20260908 UCRT, offline locked dependencies, one compiler at a time. Each receipt records command, environment, source HEAD before/after, input hashes, original log and test executable hashes. The final strict gate uses CARGO_BUILD_JOBS=1 and CARGO_INCREMENTAL=0.

Source	Gate	Observed result	Captured inputs
0b61771	Core inventory / Scenario layout	129 listed; 20 pass / 7 fail	742
2a0579d	Core inventory / Scenario layout / source defaults	129 listed; 27 / 22 pass	742
2a0579d	Default omission integration target	11 listed / 11 pass	742
015273a	Combined core inventory / Scenario layout / source defaults	129 listed; 27 / 22 pass	758
015273a	Strict workspace all-target Clippy	Failed: example Diagram constructor missing layout	758
c489748	Strict workspace all-target Clippy	Failed: missing attachment compile-time fixture	758
f82e29d	Strict workspace all-target Clippy	Exit 0, no diagnostics	760

These are 13 completed gate receipts. Passing scoped runs are matched to their actual same-source, same-executable inventories, with exact selected membership and outcomes. No historical counts are added into a combined total. The first seven failures are all source-lifecycle cases: parser field admission omitted layout. Fix 2a0579d (combined cherry-pick 846cae8) admits the field without adding synthetic layout-null editor defaults. Original failing logs remain.

The example fix c4897485b7058570a563707c4f5840f500a6c265 adds only layout: None to examples/release1/model.rs. The fixture restoration f82e29d6b8ed2b5fd23ac806675d3eafe851d104 imports exact peer b8685416 bytes at harness/experiments/edge-attachments-20260922/native-fixture.arcadia (Git blob f6419892143c9c08c22780de6060a7315be17930). Neither changes crate implementation. An include-file audit also identified the existing textual-authoring alarm-response.arcadia fixture outside the previous manifest roots. The final 760-input scope explicitly includes both compile-time fixtures; historical 742/758-input receipts are not retrospectively expanded.

Reproduce the evidence audit

Fetch the source commits named in manifest.json, then run from any directory:

python path/to/this/folder/verify.py --repo path/to/ARCADIA-Studio

The standard-library verifier reads the ZIP in memory without extraction or executing archived code. It checks every payload hash, receipt/log/manifest binding, frozen source, unchanged inputs, recorded command/exit, exact test membership/outcomes (including the seven failures), helper hashes and 3,760 distinct source-commit/input pairs against Git blobs. Executable bytes are not distributed in this archive; their recorded identity binds each scoped run to its listing but is not a substitute for rerunning the build.

The positive audit passes. Four falsification trials reject a corrupt payload, wrong outcome count, removed selected test outcome (with renewed wrapper hashes), and forged source input hash (also with renewed wrappers). The original local driver and hash-bound result are retained beside this report. The driver uses the original workspace paths; it is a record of these trials, not a second portable entrypoint.

To repeat the final strict gate from the identified source with the documented Windows toolchain and cached dependencies:

cargo clippy --locked --offline --workspace --all-targets -- -D warnings

Focused test commands, filters and tool environment are preserved verbatim in the receipts. The archived launchers document capture behavior; they reference the original local layout and are evidence rather than portable build setup.

Capture limits and pending acceptance

• Earlier helper 596dd821... was reconstructed by reversing the known scope extension and package selection change; its exact hash matches the historical receipts. Helper ddb37f5c... was copied after its last run. Neither is claimed to be a launch-time copy. launchers/capture-provenance.json preserves this. Later gates retain immutable hash-keyed helper bytes before launch.
• The final strict receipt's raw UTC span is 08:29:41–11:35:25, while Cargo reports 5m42s. The clock discontinuity is retained; that span is not a duration claim.
• The input inventory is a declared source scope, not a complete compiler file access trace. The discovered fixture omission and subsequent repair remain explicit. All recorded included inputs match their identified Git blobs.
• Desktop discovery/execution, remaining integration targets, full workspace tests, matching GUI build and a combined native save/reopen workflow remain pending at this checkpoint. No running gate is included in the archive.
• Public format adoption and normative Scenario language/API/OpenAPI coverage remain unresolved. The private analysis parser is a separate branch and is not part of this combined candidate. Linux remains user-deferred.

The source reviews retain their original pins and pre-execution dispositions. Passing these gates does not retroactively make a source review a runtime or native observation. Existing peer native evidence remains tied to its own source.