fix: stop the registry callbacks erasing context_length and local_model

internal/computing/inference_service.go

@@ -44,6 +44,29 @@ var streamingHttpClient = &http.Client{
 	},
 }
 
+// mappingFor mirrors a registered model into the legacy mapping table.
+//
+// Every field the registry holds has to be copied. These callbacks fire after
+// models.json is first read and again on every hot reload, so anything omitted
+// here is not merely missing — it silently overwrites a value that was loaded
+// correctly moments earlier. context_length was dropped this way: an operator
+// setting it in models.json saw it read at startup and then erased, so the
+// window went out as undetermined and the marketplace fell back to the catalog
+// value for a backend that publishes no max_model_len of its own.
+func mappingFor(model *RegisteredModel) ModelMapping {
+	return ModelMapping{
+		Container:     model.Container,
+		Endpoint:      model.Endpoint,
+		GPUMemory:     model.GPUMemory,
+		Category:      model.Category,
+		LocalModel:    model.LocalModel,
+		Format:        model.Format,
+		Quantization:  model.Quantization,
+		APIKey:        model.APIKey,
+		ContextLength: model.ContextLength,
+	}
+}
+
 // ModelMapping represents a model-to-endpoint mapping from models.json
 type ModelMapping struct {
 	Container     string `json:"container"`
@@ -146,15 +169,7 @@ func NewInferenceService(nodeID, cpPath string) *InferenceService {
 	registry.SetCallbacks(
 		func(model *RegisteredModel) {
 			// On model added
-			s.modelMappings[model.ID] = ModelMapping{
-				Container:    model.Container,
-				Endpoint:     model.Endpoint,
-				GPUMemory:    model.GPUMemory,
-				Category:     model.Category,
-				Format:       model.Format,
-				Quantization: model.Quantization,
-				APIKey:       model.APIKey,
-			}
+			s.modelMappings[model.ID] = mappingFor(model)
 			s.updateClientModels()
 		},
 		func(modelID string) {
@@ -164,15 +179,7 @@ func NewInferenceService(nodeID, cpPath string) *InferenceService {
 		},
 		func(model *RegisteredModel) {
 			// On model updated
-			s.modelMappings[model.ID] = ModelMapping{
-				Container:    model.Container,
-				Endpoint:     model.Endpoint,
-				GPUMemory:    model.GPUMemory,
-				Category:     model.Category,
-				Format:       model.Format,
-				Quantization: model.Quantization,
-				APIKey:       model.APIKey,
-			}
+			s.modelMappings[model.ID] = mappingFor(model)
 		},
 	)
 

internal/computing/mapping_mirror_test.go

@@ -0,0 +1,74 @@
+package computing
+
+import "testing"
+
+// The registry callbacks overwrite whatever models.json loaded, so a field
+// omitted from the mirror does not merely go missing — it erases a value that
+// was read correctly moments earlier. context_length was lost exactly that way.
+func TestMappingForCopiesEveryField(t *testing.T) {
+	model := &RegisteredModel{
+		ID:            "org/model",
+		Container:     "c",
+		Endpoint:      "http://backend:8000",
+		GPUMemory:     16000,
+		Category:      "text-generation",
+		LocalModel:    "model-local-name",
+		Format:        "awq",
+		Quantization:  "w4a16",
+		APIKey:        "sk-local",
+		ContextLength: 65536,
+	}
+
+	got := mappingFor(model)
+
+	for _, tc := range []struct {
+		field string
+		got   any
+		want  any
+	}{
+		{"Container", got.Container, model.Container},
+		{"Endpoint", got.Endpoint, model.Endpoint},
+		{"GPUMemory", got.GPUMemory, model.GPUMemory},
+		{"Category", got.Category, model.Category},
+		{"LocalModel", got.LocalModel, model.LocalModel},
+		{"Format", got.Format, model.Format},
+		{"Quantization", got.Quantization, model.Quantization},
+		{"APIKey", got.APIKey, model.APIKey},
+		{"ContextLength", got.ContextLength, model.ContextLength},
+	} {
+		if tc.got != tc.want {
+			t.Errorf("%s = %v, want %v", tc.field, tc.got, tc.want)
+		}
+	}
+}
+
+// The end-to-end consequence: an explicit override in models.json must reach
+// the declaration as an override, not be erased into "unknown".
+func TestModelsJSONOverrideSurvivesRegistryCallback(t *testing.T) {
+	s := newContextService(map[string]ModelMapping{
+		"org/proxied": {Endpoint: "http://proxy", ContextLength: 65536},
+	})
+	// The backend publishes nothing, as llama.cpp and Ollama do not.
+	s.healthChecker.recordDetectedContext("org/proxied", nil)
+
+	// Simulate the registry reporting the same model back, which is what
+	// overwrote the mapping before.
+	s.modelMappings["org/proxied"] = mappingFor(&RegisteredModel{
+		ID:            "org/proxied",
+		Endpoint:      "http://proxy",
+		ContextLength: 65536,
+	})
+
+	info := s.ModelContext("org/proxied")
+	if info.Length != 65536 {
+		t.Errorf("length = %d, want the operator's 65536", info.Length)
+	}
+	if info.Source != ContextSourceOverride {
+		t.Errorf("source = %q, want %q", info.Source, ContextSourceOverride)
+	}
+
+	declared := s.resolveModelContexts()
+	if declared["org/proxied"].Length != 65536 {
+		t.Errorf("declaration = %+v, want the override to be declared", declared["org/proxied"])
+	}
+}