fix(provider): fail the candidate on error-only streams instead of serving empty replies

Some upstreams (zen free pool) answer HTTP 200 with a single-chunk stream
whose only payload is finish_reason:"network_error" and an empty delta.
ChatStream used to hand that channel up as a success, so clients received
a laundered empty reply (agents ended their turn mid-loop) and the broken
slot kept its scheduling preference. ChatStream now holds back the first
chunk: an error-only done chunk fails the candidate before any byte
reaches the gateway, letting the scheduler fall through to the next
source. Standard finish reasons (including instant-empty "stop") are never
classified as errors.

internal/provider/provider.go

@@ -777,7 +777,7 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 		rc <- respOrErr{resp, err}
 	}()
 
-	ch := make(chan types.UnifiedChunk, 64)
+	inner := make(chan types.UnifiedChunk, 64)
 	sel := <-rc
 	if sel.err != nil {
 		// client disconnect/cancel before the first byte: not a scheduling
@@ -797,11 +797,12 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 	}
 	go func() {
 		defer p.Release()
-		defer close(ch)
+		defer close(inner)
 		defer sel.resp.Body.Close()
 		scanner := bufio.NewScanner(sel.resp.Body)
 		scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)
 		var chunks int
+		var realChunks int
 		var doneSeen bool
 		var doneSent bool
 		for scanner.Scan() {
@@ -821,7 +822,7 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 				// finish_reason downstream.
 				if !doneSent {
 					select {
-					case ch <- types.UnifiedChunk{Done: true}:
+					case inner <- types.UnifiedChunk{Done: true}:
 					case <-ctx.Done():
 					}
 				}
@@ -844,9 +845,12 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 			if ck.Done {
 				doneSent = true
 			}
+			if !errorOnlyChunk(ck) {
+				realChunks++
+			}
 			chunks++
 			select {
-			case ch <- ck:
+			case inner <- ck:
 			case <-ctx.Done():
 				return
 			}
@@ -857,15 +861,71 @@ func (p *Provider) ChatStream(ctx context.Context, req *types.ChatRequest) (<-ch
 		// success nor failure for scheduling purposes. A 200 that produced
 		// zero chunks and no [DONE] is an empty stream, i.e. a failure
 		// before the first chunk — record it so the slot can fall back.
+		// A stream whose only payload was error-only done chunks (e.g.
+		// finish_reason:"network_error") is likewise a failure, not a success.
 		if ctx.Err() == nil && scanner.Err() == nil {
-			if doneSeen || chunks > 0 {
+			if realChunks > 0 || (doneSeen && chunks == 0) {
 				p.RecordSuccess(model)
 			} else {
 				p.RecordFailure(model, 0)
 			}
 		}
 	}()
-	return ch, nil
+
+	// Hold back the first chunk to validate the stream actually carries
+	// content: some upstreams answer HTTP 200 with a degenerate stream whose
+	// only payload is an error finish reason (zen free pool sends
+	// finish_reason:"network_error" with empty delta). Failing the candidate
+	// here — before any byte reaches the gateway — lets the scheduler fall
+	// through to the next source instead of serving the client an empty reply.
+	type heldChunk struct {
+		ck types.UnifiedChunk
+		ok bool
+	}
+	var first heldChunk
+	select {
+	case ck, ok := <-inner:
+		first = heldChunk{ck, ok}
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	}
+	if !first.ok {
+		return nil, fmt.Errorf("provider %s: empty stream", p.Name())
+	}
+	if errorOnlyChunk(first.ck) {
+		go func() {
+			for range inner {
+			}
+		}()
+		return nil, fmt.Errorf("provider %s: upstream returned %q stream",
+			p.Name(), first.ck.FinishReason)
+	}
+	out := make(chan types.UnifiedChunk, 64)
+	go func() {
+		defer close(out)
+		out <- first.ck
+		for ck := range inner {
+			out <- ck
+		}
+	}()
+	return out, nil
+}
+
+// errorOnlyChunk reports whether ck carries nothing but an upstream error
+// signal: a done chunk with a non-standard finish reason and zero content,
+// tool calls, reasoning text or usage. Standard OpenAI finish reasons are
+// never classified as errors, so legitimate instant-empty completions
+// (finish_reason:"stop", no output) still reach the client.
+func errorOnlyChunk(ck types.UnifiedChunk) bool {
+	if !ck.Done || ck.FinishReason == "" {
+		return false
+	}
+	switch ck.FinishReason {
+	case "stop", "length", "tool_calls", "function_call", "content_filter":
+		return false
+	}
+	return ck.Content == "" && len(ck.ToolCalls) == 0 &&
+		ck.ReasoningContent == "" && ck.Usage == nil
 }
 
 // Image generates images via /v1/images/generations. Same scheduling-state