Mixed 3.5-bpw GLM: repaired loader and C4 evaluation

The repaired model ran at C1 and C4. Functional checks and repeated bounded quality passed; long retrieval passed at 216,307 actual input tokens. Strict capacity was inconsistent. The baseline was restored; no promotion occurred.

• Identity, C4 recipe, workloads, and comparison plan
• Loader patch, CPU fixture, pinned build, and containment
• C1: preflight, one-repetition scout, restoration
• C4: preflight, repeated quality, valid capacity cell, failed repeat, and long needle
• Baseline: repeated quality and failed capacity cell
• Final recovery: C4 restoration receipt, direct check, and routed check
• Decision, failures, sources, and artifact manifest

Primary Node ran the managed trials; Mini was not involved. Public artifacts redact operator paths, endpoint identities, container IDs, and GPU UUIDs. Native benchmark artifacts retain their full schemas, visible outputs, per-request validation, timings, and failures. Only endpoint and operator identities are redacted. The recipe remaps its endpoint port to a synthetic example value while preserving the load-bearing model, runtime, memory, and scheduling settings.