Scenario semantic reorder: portable Windows evidence
Tested production source: 1c4099a9ef76dd28d17ef892e95edf89e860569f, branch codex/scenario-multiselect-9e8r6t-20260924. The delivery descendant adds records only. Implementation, exact gate history, native observations and limitations; policy. Live coordination belongs to issue 29. This is branch evidence, not integrated/release acceptance. Linux remains user-paused.
Offline verification
From the checkout, Python 3.11+ standard library:
python -B harness/scenario-reorder-app-evidence-20260924/checks/verify.py
python -B harness/scenario-reorder-app-evidence-20260924/native/verify.py harness/scenario-reorder-app-evidence-20260924/native
Neither verifier executes archived code/binaries, launches a GUI or contacts a network. A consistency PASS does not turn a failed historical run into a passing test. The native verifier reports the original run-1 snapshot failure explicitly, with that attempt unaccepted.
| Archive | Bytes | SHA-256 |
|---|---|---|
| checks/evidence.zip | 29,229,448 | 80406d81e0a5d47bea2ef284bc50a079713d828e8ea291cb96cfcedfeaa16235 |
| native/native-evidence.zip | 12,792,711 | 0686d691a6ecea80310c5ee31056de36a7389d89a41ca501e6db9427f1ba4a82 |
Checks retain 17 runs, five failures, 17 exact source proofs and nine source capsules. Final-source Scenario315/0/1, format, strict desktop all-target Clippy and Studio+CLI build pass; final inventories each contain818 inputs. Filters are scoped, not the full desktop suite. The checks README describes the generic capture format; manifest records the actual selected reorder runs.
Native retains535 members: three closed native attempts, two closed CLI preflights, every original capture/receipt, strict source snapshots, two separately pinned oracle revisions, helpers and reviews. Run1 at14df is historical findings-only; its original whitespace-oracle failure stays unchanged. Current-source run2 has78 capture records and six exact snapshots; fresh-process run3 has15 capture records and one exact reopened snapshot. Both accepted attempts exit0 with empty streams and unchanged immutables. CLI preflight3 matches their source and v2 fixture manifest. Excluded executable/DLL bytes are represented by measured build/runtime/readback pins, not bundled runnable release assets.
Inside the native ZIP, helpers/run-2-observations.md records parent actions and recovery; helpers/run-2-review.md and helpers/run-3-review.md independently inspect original screenshots and exact source/lifecycle bytes. review-config.json declares each attempt's scope, phases and reopen relationship. runs/run-2/ and runs/run-3/ contain originals; oracle-revisions/oracle-v2/README.md explains the strict corrected expectation. Use the development review for the sequence of findings/corrections, with historical pending statements interpreted at their stated source.
Reproduce the Windows checks
Use the project's documented Windows GNU/LLVM toolchain and Rust/Cargo1.98.1 with CARGO_BUILD_JOBS=1, CARGO_INCREMENTAL=0. Original commands/environment and measured compiler/binary inputs are in receipts. At the tested commit:
cargo test -p oa-desktop scenario --message-format=json -- --nocapture
cargo fmt --all -- --check
cargo clippy -p oa-desktop --all-targets --message-format=json -- -D warnings
cargo build -p oa-desktop -p oa-cli --message-format=json
The native launcher copies the build's Studio executable and measured libunwind.dll into a clean bin directory, uses a separate cwd and a Windows-system-only child PATH, and opens working.arcadia through the application File UI. It does not claim a GUI model-path command-line argument. Preparation, launch, source phase snapshots, and normal closes are captured by helpers in the archive. Required runtime SHA is f45604a12e8bcd839fd86f667c1be1cc7c4f9da2b3c69ff381ac708f0e1ae7d4.
Native selection used contextual Add; mounted tests separately cover modifier handling. Actual capture sizes1605×872 and1104×732 do not establish all sizes or screen-reader accessibility. Runtime coverage of a real App source-write failure remains open. Aggregate removal and full issue29 acceptance are separate work. The published issue37 prerelease is unchanged.
Final bounded audit independently repeats both offline verifiers, compares package members with original records, checks local omitted-binary pins and inspects selected original native images. It records reviewer authorship limits and makes no remote-publication claim.