[lvj-repox]

v2.0.8 candidate validation — 2026-10-19

Implementation tested: 56fa851906c937a309c4d67790177d293732371a.
PR: #14

Results

Layer	Evidence	Result
Linux unit suite	Python 3.12.13 on nmz6, 169 tests	PASS
Native end-to-end	tests/pr_gate_self_test.py on nmz1; also hosted CI	PASS
Bootstrap	Existing/offline interpreter, missing dependencies/network failure, explicit/versioned interpreter preference	PASS in hosted CI; optional Python 3.8 rejection case skips when unavailable
Hosted CI	Python 3.9, 3.11, 3.12, run 39215002651	All three PASS
Real pinned images	tests/real_tools_integration.py on nmz4, all 30 cases	30/30 expected outcomes
Fleet smoke	nmz1 and nmz6, clean/OpenMDW/config bypass/attribute bypass, all four scanner images	4/4 expected outcomes on each node
Real incident replay	Megatron-LM-das and pytorch-das exact ranges below; 1 GiB address-space / 60s CPU cap	PASS
Shell syntax	bash -n scripts/prepare-python.sh	PASS

Hosted CI evidence: https://github.com/HYGON-AI/quality-gate/actions/runs/39215002651

“Expected outcomes” includes intentionally blocked findings (exit 2), operational failures (exit 1) and clean/advisory passes (exit 0). Tests do not redefine expected errors as successes in production.

Historical replays

Only native Code compliance was replayed against existing checkouts, without changing their files, refs or configuration. All writeable test outputs live in separate temporary directories; target code is never executed.

• Megatron-LM-das #117: base 463e2f4536d2fe829947b5246b26f972cf78a728, head a6ca2d3bcf93def95354eccf1eebd0a305edf001. Initial fixed replay on nmz4: 0.132s, peak RSS 29,200 KiB, exit 0. Original stable attempt 4 grew to about 538 GiB by four minutes before cancellation. The 49 changed files include only two non-deleted files; the large Python file triggered duplicate branch-value growth, not model execution.
• pytorch-das #20: base acf085724329a8c07556709324cb757ce7046fbd, head d1591e5c528a017e14f2a3bf66bd4a9d9acf90cf. Fixed replay on nmz6: 0.038s, peak RSS 19,532 KiB, exit 0. Gitlinks no longer attempt to read a foreign commit as a source blob.

Negative controls and additional defects

• Original implementation failed 11/13 initial targeted regressions. Safe bounded fixtures used; no new unbounded production run was started.
• 500 unchanged branches now retain one candidate. Loop joins deduplicate; distinct bool/int values and all changed-line origins are preserved.
• Candidate/string/work exhaustion and empty-message MemoryError explicitly fail. The old orchestrator could replace MemoryError with an empty status detail and accidentally return Passed; this is covered separately.
• ruff.toml per-file F823 ignores, inline noqa, and .gitattributes -diff cannot hide the synthetic new F823 defect. Textconv suppression is separately unit tested.
• Unknown licenses still block; replacing an original OpenMDW notice with MIT still blocks.
• Missing/malformed reports, ENOSPC, missing images and tool timeouts retain safe diagnostics without echoing arbitrary secret/source output. Timeout cleanup targets only the captured container ID.
• Semgrep's network version check was unnecessary with pinned offline images. On nmz4 it took ~45.5s; disabling the version check reduced it to ~1.5s with the same scan results.

Release boundary

At validation completion this remains a candidate: required independent approval, merge and post-merge checks precede the release/tag update. No immutable tag, stable pointer, runner-group permission, system Python or node configuration was changed by these tests. GitHub status and final tested commit must be rechecked immediately before publishing.