sorokin.oleg.edu
36db9356cc
Preserve loaded draft bytes and model history, reject stale workspace and input events, and keep captured Escape under modal priority. Verified 181 focused Windows tests and strict desktop Clippy on exact inputs4; earlier full737 scope and component/native preparation evidence are archived separately. Native navigation acceptance remains pending.
Desktop source-navigation gates — 2026-09-11
Codex-9e8r6t /root/behavior_audit. This directory preserves completed Windows desktop gate records from .local/text-navigation-desktop-20260911. The original producer ran the gates; this archival task only copied bytes, inspected receipts/source, and independently verified hashes and ZIP contents. No compiler, application, server or Linux process was run for this archive.
What passed
| Record | Actual terminal result | Input scope |
|---|---|---|
check1 |
Exit 0 at 17:36:52 UTC; 13 unused-code warnings | Historical initial check, before inputs1 was captured; no exact inputs1 binding claimed. |
test1 |
Exit 0 at 17:51:59 UTC; 737 passed, 0 failed, 1 ignored; test execution 179.86 s | inputs1; predates the queued-action replay correction. Two unused-code warnings remain in this historical log. |
performance1 |
Exit 0 at 17:53:16 UTC; the one ignored trial passed | Exact preserved test1.exe, independently hash-checked during archival. |
focused2 |
Exit 0 at 18:01:39 UTC; 177 passed, 0 failed, 1 ignored, 562 filtered out; test execution 19.40 s | inputs2, including the replay correction. |
clippy2 |
Exit 0 at 18:03:25 UTC; all-target desktop Clippy with -D warnings |
inputs2. |
The raw statuses and helpers retain commands, timestamps and one-job configuration. Gates used the existing Windows gnullvm environment sourced from .local/toolchain/env.ps1. The helper is referenced, not copied here; this directory is evidence, not a bundled toolchain. capture.py is preserved as the original receipt-generating helper and depends on its original private directory location; use verify.py for this archive's portable, read-only integrity check.
Both captured snapshots name branch codex/release1-9e8r6t and base HEAD fb24def2f55699942d6fe689c44424933dac0d9e. Navigation changes were uncommitted above that HEAD. Each ZIP contains 499 exact captured files: inputs1 is 9,418,355 source bytes; inputs2 is 9,429,450. Seven paths differ; manifest.json lists them. These are the source scopes enumerated by capture.py (crates, Cargo files, .cargo, docs, examples and harness fixtures), not a hermetic compiler/dependency capture. Both original post-gate readbacks report no changed, added or removed input files. This archive separately verifies every ZIP member against its captured manifest, including original newline/UTF-8 bytes.
The second snapshot tests emitting-session and active-path envelopes for Query, Back, Close and Choose. Actual App-path regressions replay old messages after reset at the same path and after switching files in the same session, both while busy and after completion. They preserve current request/reveal state, interactions, model, raw text and histories; fresh-context actions still work. Earlier tests cover invalid/unapplied drafts without Apply, ambiguous/detached choices, exact Unicode/CRLF Back selections, worker/reset ordering and the 64-entry Back cap. Async completions are delivered deterministically by tests; this does not exercise native event timing. The full 737-test suite was not repeated after the replay correction.
Performance scope
The separate trial constructs 100 module files plus an entry, 1,000 process declarations and 10,000 lexical uses in 179,156 source bytes. It measured admission 364.9 µs, snapshot capture 137.7013 ms and reference query 28.7229 ms. The result contains one definition and 199 reference rows, with 9,801 uses explicitly omitted by the row budget. It asserts the source is unchanged.
This is a current-draft lexical fixture with process-to-process placeholder steps, not a semantically validated engineering model. It is neither GUI latency nor native usability acceptance. Reported process resident memory was 10,186,752 → 16,093,184 bytes, with peak 16,842,752 bytes; this includes the process/fixture and is not an allocator-exact index measurement. Performance was not rerun on inputs2.
Provenance and exclusions
The full-test executable is intentionally omitted (109,325,824 bytes, SHA256 F6C6312FF0C726AEC6A30A90658808DAF4EA7ACE6DF4F1977DFFC54CA09AABD9). Its original artifact receipt remains byte-exact, and manifest.json records the independent size/hash readback. No other binary, GUI build/native output, review-final, profile or work data is included.
Strict parser UUID-header spelling was already constrained to canonical decoded 8-4-4-4-12 shape. This slice aligns tolerant outline/analyze identity recognition with that existing rule; it does not introduce a new strict grammar restriction. Core, primitive-query and adapter evidence is owned separately in the component archive and the outline integration report. Those component payloads are not duplicated here.
manifest.json hashes the 20 exact copied payloads and three archive control files. readback.json pins that manifest and records verification of the complete archive file set and both source ZIPs. The local .gitattributes disables text normalization only inside this directory. Run python verify.py here or by absolute path to check without modifying any file. Root owns subsequent GUI/native acceptance, integration and publication; these receipts do not establish a release or the complete OA/SA goal. Live coordination is in issue 6.