Files
ARCADIA-Studio/harness/experiments/analysis-output-acceptance-20260925/README.md
T

108 lines
6.0 KiB
Markdown

# Private actual-output acceptance — Windows evidence
Final source **edf096cd33350865ecfc3c65ec266a6c0d3c4000**, branch
`codex/analysis-mixed-acceptance-956atf-20260925`, extends baseline
`29417130c71f70c4eaefe8ed4c879efba78eb72c`. Source and this packet remain
**local-only**: Gitea rejected Git publication for exhausted server storage and
subsequently rejected the checkpoint comment. No peer-accessible commit link is
claimed. The published Windows preview and its artifacts are unchanged; Linux
remains user-deferred.
The [selected contract](../../analysis-mixed-acceptance-contract-20260925.md)
connects real emitted PendingRename ownership to both parser producer modes,
actual complete resolution/admission and full Project equality. AcceptedRename
owns the actual bundle, actual resolved model, mapped warnings, inherited policy
and final O+W inventory. Equality reserves the actual parsed expression count;
warning storage is counted once, and destination checks precede temporary-owner
destruction. There is no borrowed desired model in the accepted result.
## Observed gates
Windows Rust/Cargo 1.98.1 with LLVM-MinGW 20260908 gnullvm; one compiler job,
incremental disabled, locked/offline commands. All final gates bind 559 unchanged
declared inputs, clean source before/after and one captured launcher.
| Gate | Result |
| --- | --- |
| New acceptance tests | 11 passed, 281 filtered |
| Complete core library | 292 names: 276 passed, 16 inherited failures |
| Core all-target strict Clippy | Passed |
| Workspace format | Inherited exit 1; raw output byte-identical |
| Regression readback | All 281 prior outcomes and 16 failure bodies unchanged, normalizing numeric thread identifiers only |
The final executable is 43,066,880 bytes, SHA256
`7bc17993f265bdf7dc79b7ea5191401dd853f1dd47517235663113e4571031bb`.
Input-manifest SHA256 is
`6be8ae8911bb3d0aee346b137930d2b27a7d0ed4e77c4c9502b60d07982bd028`.
The executable is identified in the receipts, not embedded in this archive.
Full core and format are **not** all-green gates: public schema8 dispatch remains
unavailable, and the known failures remain obligations for later integration.
Actual tests establish the independently specified unchanged-profile boundaries:
152 bytes accepts the minimal complete rename and 151 refuses during output UUID
decoding; 161 accepts the unnamed warning and 160 refuses its mapping at the full
header span. Whole returned models/bytes, exact warning records, all policy
fields, separate verification work, source/desired lifetime independence,
admissible wrong output, malformed/model-invalid output, actual expression-bound
provenance and signed-zero semantic equality are checked. Rich OA/SA Unicode/CRLF
swaps and a four-file untouched-file case retain both real architecture warnings.
Initial source `4872141c036c16825f1c098bd7cf44709897e814` compiled but its focused
run had nine passes and two failed empty-warning assumptions. The repair changes
only new tests and the implementation note; production is byte-identical.
[The independent correction](WARNING-FIX-ORACLE.md) derives both ordered warnings,
locations and final storage from source and authored literals. Initial logs,
artifact identity and review omissions are preserved rather than replaced.
## Stack and evaluator scope
All ten exact inherited stack controls pass. The new scoped test performs actual
verification and result destruction in five 128 KiB workers: deep/wide success,
deep/wide mismatch and the genuine 151-byte late refusal. Prerequisites are built
outside these workers. Deep name mismatch can short-circuit; deep success performs
the expression comparison. Older separately unrerun full-pipeline setup limits
remain distinct.
[Stack observations](STACK-OBSERVATIONS.md) bind raw initial/final prologues.
The 17 selected old stack/active reservations are unchanged. New verification
reserves 8,648 bytes in its prologue; the carried parser wrapper changes from
536 to 568. Parsing, resolution, equality and destination checks have distinct
phases; these observations are not a whole-stack or headroom bound.
The first strict readback refused a raw CARGO_TARGET_DIR slash-spelling difference.
[The checker note](CHECKER-NOTE.md) preserves that refusal and original checker.
The amended checker accepts only the exact recorded old/new spellings after both
resolve to the same explicit workspace target. Every other environment field is
exact; failure comparison and source/artifact checks are unchanged. Both raw
environments are retained in [the readback](ACCEPTANCE-RUNTIME-READBACK.json).
## Evidence and reproduction
[evidence.zip](evidence.zip): **529,018 bytes, 67 members**, SHA256
`2f5e0c5257eba11908b60e02b76d529fcd53f11edd057c859d3cb1b34f41d7c4`.
[manifest.json](manifest.json) pins every member. The archive includes initial
and final receipts/logs/manifests, captured launchers, actual Git diffs, baseline
records, both checker versions, selected prologues, contract/oracles and raw review
mirrors. Read the [source](SOURCE-REVIEW.md), [test](TEST-REVIEW.md) and
[runtime](RUNTIME-REVIEW.md) reviews with their exact source/hash bindings.
The separate [package audit](PACKAGE-READBACK.md) checks every member against
originals and Git blobs without executing archived code.
From the identified source on the documented Windows toolchain, use one compiler
job and incremental disabled:
```text
cargo test --locked --offline -p oa-core --lib acceptance -- --test-threads=1
cargo test --locked --offline -p oa-core --lib -- --test-threads=1
cargo test --locked --offline -p oa-core --lib -- --list
cargo clippy --locked --offline -p oa-core --all-targets -- -D warnings
cargo fmt --all -- --check
```
Private acceptance grants no public source8, installation/history, repeated
carried editing or GUI integration. Independently bounded parser scratch,
numeric/decoder internals, model/validator allocation, allocator overhead and
native stack retain their explicit separate domains. In particular, source/file
bounds do not tightly bound copied traversal-path bytes. No universal memory or
before-every-allocation guarantee follows from this evidence.