feat: multi-backend proxy with auto rate-limit jailing

Refactor from monolithic proxy to multi-backend architecture
(Backend interface: OpenCode Zen + Kilo gateway).

- Add auto rate-limit jailing: models returning HTTP 429 are
  hidden from /v1/models immediately and re-probed every 30 min
- Backend interface supports model aliases, custom headers, and
  per-backend routing
- Full Anthropic Messages API support with OpenAI format conversion
- 9 free models across two backends with Claude name aliases

internal/proxy/anthropic.go

@@ -652,15 +652,19 @@ func (p *Proxy) ForwardAnthropicMessages(w http.ResponseWriter, r *http.Request,
 		return fmt.Errorf("invalid Anthropic request: %w", err)
 	}
 
-	// Resolve model
-	resolved, _ := p.ResolveModel(ar.Model)
+	// Route to the backend that owns this model (default = pass-through).
+	backend, resolved, known := p.resolveModel(ar.Model)
+	if !known {
+		resolved = ar.Model
+	}
 	// Keep original client-requested model name for response rewriting.
 	// Claude Code validates that the response model matches the request model.
 	clientModel := ar.Model
 
-	// Auto-route to vision model if the request has images and resolved model can't handle them
+	// Auto-route to a vision model only inside the opencode backend (which owns
+	// a vision-capable free model). Kilo models are forwarded as-is.
 	hasImages := requestHasImages(&ar)
-	if hasImages {
+	if hasImages && backend.Name() == "opencode" {
 		mi := ModelByID(resolved)
 		if mi == nil || !mi.SupportsVision {
 			fmt.Printf("[anthropic] model=%s lacks vision, auto-routing to mimo-v2.5-free\n", resolved)
@@ -668,8 +672,8 @@ func (p *Proxy) ForwardAnthropicMessages(w http.ResponseWriter, r *http.Request,
 		}
 	}
 
-	fmt.Printf("[anthropic] model=%s → %s, max_tokens=%d, stream=%v, messages=%d, tools=%d\n",
-		clientModel, resolved, ar.MaxTokens, ar.Stream, len(ar.Messages), len(ar.Tools))
+	fmt.Printf("[anthropic] model=%s → %s via %s, max_tokens=%d, stream=%v, messages=%d, tools=%d\n",
+		clientModel, resolved, backend.Name(), ar.MaxTokens, ar.Stream, len(ar.Messages), len(ar.Tools))
 
 	// Convert Anthropic → OpenAI
 	oaReq := AnthropicToOpenAI(&ar)
@@ -686,13 +690,13 @@ func (p *Proxy) ForwardAnthropicMessages(w http.ResponseWriter, r *http.Request,
 	requestID := RequestID()
 	sessionID := p.SessionID("anthropic")
 
-	// Build upstream request
-	upstreamReq, err := http.NewRequestWithContext(r.Context(), "POST", ZenChatURL(), bytes.NewReader(oaBody))
+	// Build upstream request to the chosen backend
+	upstreamReq, err := http.NewRequestWithContext(r.Context(), "POST", backend.ChatURL(), bytes.NewReader(oaBody))
 	if err != nil {
 		return fmt.Errorf("failed to create upstream request: %w", err)
 	}
 
-	for k, v := range UpstreamHeaders(requestID, sessionID) {
+	for k, v := range backend.Headers(requestID, sessionID) {
 		upstreamReq.Header.Set(k, v)
 	}
 
@@ -709,6 +713,9 @@ func (p *Proxy) ForwardAnthropicMessages(w http.ResponseWriter, r *http.Request,
 	if resp.StatusCode != 200 {
 		bodyBytes, _ := io.ReadAll(resp.Body)
 		fmt.Printf("[anthropic] upstream error body (%d bytes): %s\n", len(bodyBytes), string(bodyBytes))
+		if resp.StatusCode == http.StatusTooManyRequests {
+			p.markRateLimited(resolved)
+		}
 		w.Header().Set("Content-Type", "application/json")
 		w.WriteHeader(resp.StatusCode)
 		json.NewEncoder(w).Encode(AnthropicError{