Retained failures and limits

The shared recipe parser recognized --served-model-name but ignored Ignis --model-id. It therefore supplied the repository ID to a correctly fail-closed monitoring guard. Managed load exited1 while the candidate was running. Independent source review accepted closing this attempt; any repair requires fresh source gates and a new campaign. No source change is part of this packet.

The raw /ready and /v1/models responses and inner registration exception were not durably retained. Readiness/model-ID observations have coordinator-tool and independent-review provenance; they are not reconstructed raw evidence. Selected startup logs corroborate loaded/ready state. A separate nonfatal KV-disk permission warning says loading continued without that tier; it is not assigned as the registration cause.

The configured49GiB/no-swap/16GiB-reserve variant retains historical52GiB prose in frozen serve.note, explicitly corrected by the separately retained erratum. No frozen byte is rewritten.

Report capture first used default storage and failed bad_campaign_file; the retained second read-only command explicitly selected the pilot run root/job store and passed. Neither command submitted jobs.

The first incumbent restoration command omitted the original explicit GPU selection/order and used --gpus all. Although the engine loaded and returned health/model-list HTTP200, monitoring correctly refused the unbound GPU identity. This is a coordinator launch omission, separate from the Ignis served-identity parser defect. The incorrect incarnation was removed through managed unload (exit0); an exact-order reload is in progress, not yet accepted. See restoration-failed-attempt.json.

Final after-state: the exact GPU-order reload passed and restored service/monitoring/original router. Typed routed coding, JSON, approximately128k needle and20of20 tools passed. Pi attempt1 inherited launcher stdin; attempt2 used DEVNULL and the exact198-byte prompt. Both returned extra prose/duplicated digest, failing strict final text despite successful real read, terminal completion and unchanged declared provider/model. Validation R2 recognizes agent_settled after agent_end; the correction does not change strict failure. Service recovery is complete; required client acceptance is not. No more retries are recorded.