5.0 KiB
Private owned emission and parser stack repair
Source 8d52dab97c is pushed and ref-verified
on codex/analysis-mixed-emission-956atf-20260925. It implements the
private emission contract:
consume a validated rename plan, preserve its exact owner for unchanged output,
or reserve complete owned output storage before copying changed files once in
original edit coordinates. Changed output is opaque and unverified; the exact
policies and desired whole Project survive. Borrowed carried owners are refused
on the changed path because dropping their guard cannot free their backing
allocation. There is no public source8, save/history or accepted-output authority.
Windows gnullvm, rustc/Cargo 1.98.1: 11 new tests pass, strict core all-target
Clippy passes and the listing contains 269 names. The full library remains
FAILED: 253 passed / 16 failed. All 258 prior outcomes and all 16 failure
bodies match fa7a980 after normalizing only numeric thread identifiers.
Formatting remains FAILED on eight inherited blocks in five files, with
byte-identical output. Linux/WSL checks were not run.
Source review, test review and the independent oracle cover complete OA/SA bytes, Unicode/CRLF, scoped and global swaps, untouched files, freshness and branch refusal, policy/lifetime retention, exact storage caps and scoped drop behavior. Minimal and three-file fixtures reach real emission peaks of 17 entries / 114 text bytes and 29 / 216 respectively; their cap-minus-one controls pass predecessor phases and refuse emission. The no-op returns the original owner below output-copy overlap limits. Two fixture issues were corrected before execution and are disclosed in the test review; neither is presented as a runtime failure.
The stack diagnosis preserves source19a4082's initial full-run abort after 135 completed passes. All eleven new focused cases had passed, but an unchanged 128 KiB ordinary-index control overflowed during its upstream numeric parsing. Exact-artifact LLDB and prologue records are retained. The mechanically reviewed repair splits the iterative binding driver into prefix and continuation operations while preserving task order, containers, accounting, limits and tests. Final numeric-prefix prologue reservations total 14,240 bytes instead of 27,272, a 13,032-byte reduction. This is not a complete stack bound. The unchanged control passes the final full suite and three additional fresh processes with no stack override. The older separate full-pipeline setup overflow has not been rerun or relabeled.
Runtime review verifies 555 declared inputs against immutable
Git and raw disk bytes, exact launchers, baseline records, logs and the final
42,402,816-byte test artifact, SHA256
29d8a4761fd8f09a3c0eafeb11cca07f0d45c711bf1b969d97c8b29d2b3ab446.
The initial artifact is identified by its own receipt and retained locally;
neither executable is included in this evidence archive.
evidence.zip contains 65 members, 479,039 bytes, SHA256
8d46b31379d75567b195b639cb3bc289ad1c9d26df53d29dbfd99ba22e38c5b9.
manifest.json pins each member.
The archive preserves initial and final gates, exact launchers, three control rechecks, debugger/prologue observations, baseline records, source patches, contract snapshots and reviews. Seven readable review/checker files are exact archive mirrors. Toolchain version text is an earlier unchanged capture. The independent package readback is recorded separately.
Reproduce from the identified source with the project Windows toolchain, one
compiler job, CARGO_INCREMENTAL=0, and no lifecycle or RUST_MIN_STACK override:
cargo test --locked --offline -p oa-core --lib -- --list
cargo test --locked --offline -p oa-core --lib emission -- --test-threads=1
cargo test --locked --offline -p oa-core --lib -- --test-threads=1
cargo clippy --locked --offline -p oa-core --all-targets -- -D warnings
cargo fmt --all -- --check
Helpers retain the producing machine's layout and must be read before reuse. The readback checker validates evidence without launching Cargo or the binary. Receipts bind declared inputs and recorded environment, not exhaustive machine state or authenticated build provenance. Debugger exit zero after killing the stopped process is not a passing test. Archived diagnostic helpers are recipes; their corresponding receipts identify actual executions.
Actual-output parse/resolve/admission, complete Project equality under justified scratch bounds, stronger parser producer timing and an owning accepted result remain required. Original expression counts alone do not bound arbitrary faulty output. This branch-specific prerequisite is separate from the published preview; live coordination and broader acceptance remain in issue #7.