test(BACKEND-GATE-ROCM-SGLANG): publish lifecycle diagnostic aggregate once

[localai-org-maint-bot]

Row: BACKEND-GATE-ROCM-SGLANG

Owner: current Strix campaign operator. Parent: #3108.

Strix diagnostic e6dba009-a81b-4e0f-a74e-b6f39cb3b43d completed vLLM startup, generation and normal owned-child teardown. Engine evidence is PASS, but the scratch validate-hardware-194a.py driver writes run/result.json in its per-engine finally block and then publishes the terminal aggregate to the same immutable path. audit.publish_result correctly refuses replacement with EEXIST, leaving the initial aggregate status FAIL. The job is not an overall PASS.

Repair the diagnostic driver, not the immutable publication contract: publish per-engine records once and the terminal aggregate once; retain failure evidence and stop after a failed engine. Add a focused regression that executes orchestration with fake engine operations and real no-replace publication, covering success and failure. Do not weaken engine binding checks, teardown assertions, pins, or acceptance rules. Re-run bounded Strix validation after independent review. The source lifecycle change and compliant replay are owned by #3108; this issue owns the diagnostic publication repair.

[localai-org-maint-bot]

The local implementation candidate is 656ce703e6c48794c84d608733fbdba4c3e4461b; it is not pushed or landed. It adds the source-bound lifecycle diagnostic, one terminal aggregate publication, finalized engine records, and pre-launch refusal of native-first selection without canonical IDs.

The author recorded pre-edit, staged-before-commit, and frozen-head gate exits of 0. The operator independently checked the frozen log hash: 367d6a41da857b49f275612f72e30d613844469108ece5ab9d33390b158476ae. The combined focused run passed 53 tests. Exact-range checks passed; four artifact-dependent checks remain unmeasured.

Independent review is running on the immutable head. Operator verification and leased hardware validation remain pending. This is not a GPU hang repair, token-parity result, throughput result, or speed advantage. Evidence is retained under campaign strix-four-engine-3053.X94a3J/lifecycle-publication-3110-evidence.OmFOGK on the resolved NAS. The author report is strix3110-implementation-report.md in that package.

[localai-org-maint-bot]

Operator validation passed on Strix job f49b95cf-82b6-4217-90c1-98351df6d101, 2026-09-09 11:57:11–12:00:53 UTC, exit 0. Reviewed controller 656ce703e6c48794c84d608733fbdba4c3e4461b ran the bound vLLM-only six-request c4 lifecycle diagnostic. The final aggregate is PASS and equals the published engine record. Parent exit 0; adopted tracker PID 88344 exited 0; process group absent; lifecycle restored; before/after bindings equal. Evidence: NAS campaign strix-four-engine-3053.X94a3J/lifecycle3110-validation.erTNGl/run/result.json, SHA256 9865c118e6065d34cbf047aeb70ce1a6f72faa3717ef2ffc6631efead7d08a89. Root full gate session 55630 exited 0 before hardware execution; four artifact-dependent checks remain unmeasured after explicit range classification. This proves the selected lifecycle/publication path, not cross-engine token parity or throughput acceptance. Patched SGLang validation follows separately; original failed evidence is preserved.