The recorded llama.cpp b9892 pin names an unavailable object

[localai-bot]

The llama.cpp oracle registry records local-fork object 237ad9b961f009ae19ac29dbce4cd0c1251f94b3 with label b9892. Fresh provisioning on dgx.casa fails: the declared remotes return not our ref, the GitHub commit API returns HTTP 422, and no local or campaign cache contains the object. Official tag b9892 resolves to ee445f93d8a0a5033a46d1960e901ef5caec9a41, so the label and object cannot both describe a reproducible oracle.

This blocks Qwen3.8-27B Q4_K_M text, image, video, embedded-MTP, correctness, and performance validation. It also makes every new same-file llama.cpp gate impossible to reproduce from the registry.

Acceptance:

• reconcile the registry to an obtainable exact object without silent substitution;
• prove the old object fails exact provisioning;
• build and run the replacement on dgx.casa;
• record measured identity, source anchors, recipe, gateability, and evidence;
• mutation-prove that the oracle-pin checker rejects the old/unobtainable value and a mismatched label/object;
• rerun the four Qwen3.8 GGUF smoke arms before resuming the model campaign.