way-embed: verify scaled LoRA adapters on encoder rerankers in llama.cpp

[aaronsb]

Finding

ADR-190 (proposed in #554) stores each user's personal delta as a LoRA adapter on the reranker. It applies the adapter at a scale α: θ = (1−α)·θ_base + α·θ_delta, with α rising on reheat and annealing back. That relies on llama.cpp applying a LoRA adapter at a runtime scale to an encoder model (BERT or ModernBERT) running with rank pooling. This is unverified on the pinned commit (ec2b787).

Proposal

1. Train a toy LoRA adapter on one of the ADR-189 gate candidates.
2. Load it with a scale in llama.cpp on the rank path, and check that scores move with α and match a merged reference at α = 0 and α = 1.
3. If scaled adapters do not work on this path, measure the fallback ADR-190 names: merging the adapter and re-exporting the GGUF. Record how long that takes at 17–32M parameters, so the α anneal can run as a re-export.

Record the outcome in ADR-190 section 3, under "Personal delta".