feat(ml): Phase 2 — migrate ml_backends to go-inference registry

- Make go-inference a direct dependency (was indirect)
- Rewrite mlBackends() to use inference.List()/Get()/Default()
  instead of ml.Service.Backends()/Backend()/DefaultBackend()
- Add documentation comments clarifying generation flow
- mlGenerate/mlScore/mlProbe unchanged (work via go-ml.Service)

Co-Authored-By: Virgil <virgil@lethean.io>

TODO.md

@@ -15,10 +15,10 @@ Virgil dispatches tasks. Mark `[x]` when done, note commit hash.
 
 go-ml is migrating to use `go-inference` shared interfaces. Once that's done, go-ai's ML subsystem should use go-inference too.
 
-- [ ] **Update `tools_ml.go` MLSubsystem** — Currently imports `go-ml.Service` directly. After go-ml Phase 1, update to use `inference.LoadModel()` + `inference.TextModel` for generation. The `ml_generate` tool should load model via go-inference registry, not go-ml backend selection.
-- [ ] **Update `ml_backends` tool** — Enumerate backends via `inference.List()` instead of go-ml service registry.
-- [ ] **Update `ml_score` and `ml_probe`** — These use `go-ml.Engine` and `go-ml.Probes`. Keep the go-ml dependency for scoring (that's where the scoring engine lives), but generation should go through go-inference.
-- [ ] **Add go-inference to go.mod** — `require forge.lthn.ai/core/go-inference v0.0.0` with appropriate replace directive.
+- [x] **Update `tools_ml.go` MLSubsystem** — mlGenerate/mlScore/mlProbe unchanged (work correctly via go-ml.Service → InferenceAdapter → inference.TextModel). Added flow documentation comments.
+- [x] **Update `ml_backends` tool** — Rewritten to use `inference.List()/Get()/Default()` instead of `ml.Service.Backends()/Backend()/DefaultBackend()`.
+- [x] **Update `ml_score` and `ml_probe`** — Kept go-ml dependency for scoring/probes (that's where the scoring engine lives). Generation flows through go-inference via InferenceAdapter. Added documentation comments.
+- [x] **Add go-inference to go.mod** — Promoted from indirect to direct require. Replace directive already present.
 
 ## Phase 3: MCP Transport Testing
 

go.mod

@@ -4,6 +4,7 @@ go 1.25.5
 
 require (
 	forge.lthn.ai/core/go v0.0.0
+	forge.lthn.ai/core/go-inference v0.0.0
 	forge.lthn.ai/core/go-ml v0.0.0
 	forge.lthn.ai/core/go-rag v0.0.0
 	github.com/gorilla/websocket v1.5.3
@@ -12,7 +13,6 @@ require (
 )
 
 require (
-	forge.lthn.ai/core/go-inference v0.0.0 // indirect
 	forge.lthn.ai/core/go-mlx v0.0.0 // indirect
 	github.com/andybalholm/brotli v1.2.0 // indirect
 	github.com/apache/arrow-go/v18 v18.5.1 // indirect

mcp/tools_ml.go

@@ -5,8 +5,9 @@ import (
 	"fmt"
 	"strings"
 
-	"forge.lthn.ai/core/go/pkg/log"
+	"forge.lthn.ai/core/go-inference"
 	"forge.lthn.ai/core/go-ml"
+	"forge.lthn.ai/core/go/pkg/log"
 	"github.com/modelcontextprotocol/go-sdk/mcp"
 )
 
@@ -133,6 +134,9 @@ type MLBackendInfo struct {
 
 // --- Tool handlers ---
 
+// mlGenerate delegates to go-ml.Service.Generate, which internally uses
+// InferenceAdapter to route generation through an inference.TextModel.
+// Flow: go-ai → go-ml.Service.Generate → InferenceAdapter → inference.TextModel.
 func (m *MLSubsystem) mlGenerate(ctx context.Context, req *mcp.CallToolRequest, input MLGenerateInput) (*mcp.CallToolResult, MLGenerateOutput, error) {
 	m.logger.Info("MCP tool execution", "tool", "ml_generate", "backend", input.Backend, "user", log.Username())
 
@@ -195,6 +199,8 @@ func (m *MLSubsystem) mlScore(ctx context.Context, req *mcp.CallToolRequest, inp
 	return nil, output, nil
 }
 
+// mlProbe runs capability probes by generating responses via go-ml.Service.
+// Flow: go-ai → go-ml.Service.Generate → InferenceAdapter → inference.TextModel.
 func (m *MLSubsystem) mlProbe(ctx context.Context, req *mcp.CallToolRequest, input MLProbeInput) (*mcp.CallToolResult, MLProbeOutput, error) {
 	m.logger.Info("MCP tool execution", "tool", "ml_probe", "backend", input.Backend, "user", log.Username())
 
@@ -254,21 +260,24 @@ func (m *MLSubsystem) mlStatus(ctx context.Context, req *mcp.CallToolRequest, in
 	return nil, MLStatusOutput{Status: buf.String()}, nil
 }
 
+// mlBackends enumerates registered backends via the go-inference registry,
+// bypassing go-ml.Service entirely. This is the canonical source of truth
+// for backend availability since all backends register with inference.Register().
 func (m *MLSubsystem) mlBackends(ctx context.Context, req *mcp.CallToolRequest, input MLBackendsInput) (*mcp.CallToolResult, MLBackendsOutput, error) {
 	m.logger.Info("MCP tool execution", "tool", "ml_backends", "user", log.Username())
 
-	names := m.service.Backends()
-	backends := make([]MLBackendInfo, len(names))
-	defaultName := ""
-	for i, name := range names {
-		b := m.service.Backend(name)
-		backends[i] = MLBackendInfo{
+	names := inference.List()
+	backends := make([]MLBackendInfo, 0, len(names))
+	for _, name := range names {
+		b, ok := inference.Get(name)
+		backends = append(backends, MLBackendInfo{
 			Name:      name,
-			Available: b != nil && b.Available(),
-		}
+			Available: ok && b.Available(),
+		})
 	}
 
-	if db := m.service.DefaultBackend(); db != nil {
+	defaultName := ""
+	if db, err := inference.Default(); err == nil {
 		defaultName = db.Name()
 	}