fix: stop the registry callbacks erasing context_length and local_model

[flyworker]

Found while verifying #129 on a live node, and it answers a question #128 asked directly: "models.json supports a manual context_length override. Is it applied?" — it was not.

The bug

loadModelMappings unmarshals models.json into modelMappings correctly. Then the registry's added/updated callbacks rebuild each entry from a hand-written field list that omitted ContextLength and LocalModel — and they fire after that load, and again on every hot reload.

So the override survived only until the registry reported the model back, which is immediately.

Observed on this node

models.json says:

"Qwen/Qwen3.8-27B": { "endpoint": "http://localhost:30001", "context_length": 65536 }

and the agent reported:

Qwen/Qwen3.8-27B      -  (not reported)

ModelContext reads the override from modelMappings alone, so for a backend that publishes no max_model_len — llama.cpp, Ollama, an OpenAI-compatible proxy — the agent declared nothing, and the marketplace fell back to the catalog value for a window the operator had stated explicitly. Exactly the failure #128 is concerned with, arriving from the opposite direction: not a guess sent as a measurement, but a measurement discarded.

local_model was dropped too, and does not bite

Every forwarding path asks registry.GetLocalModelName() first and only falls back to mapping.LocalModel, so the name rewrite kept working. It was one refactor away from mattering.

The fix

Both callbacks now go through a single constructor that copies every field the registry holds, so a field added to RegisteredModel cannot be silently lost in the mirror again.

Verified

Two tests: one asserts every field round-trips, the other reproduces the end-to-end failure — an override on a backend that detects nothing, put through the registry callback, must still be declared as override rather than erased to unknown.

8/8 packages pass.