fix(proxy): wire LLM latency metrics into the streaming response path (#63)

`NewLLMMetricsCapture`/`ResponseWriterWrapper` had **zero call sites**, so `llm_ttft_seconds`, `llm_itl_seconds` and `llm_tokens_total` were exported as HELP/TYPE stubs with no samples (verified against the live `/metrics`).

These are the only per-request LLM latency series covering the Ollama models (`ornith:35b`, `qwen2.5:3b-instruct`), which expose no `/metrics` of their own — so those models had zero latency visibility.

## Change

Wire the wrapper around the response writer in `ServeHTTP` for model-routed requests, and delete the unused `io.WriteCloser` duplicate that made the gap hard to spot (two near-identical implementations, neither constructed).

## Correctness details

- **Only SSE responses are measured.** A non-streaming completion is written in one `Write`, which would otherwise be recorded as a TTFT sample equal to the whole request latency, with no ITL samples — plausible-looking but meaningless numbers.
- **`requestStart` is captured at the top of `ServeHTTP`** so TTFT measures from request acceptance, not from when the upstream response is copied.
- **Token count skips the `data: [DONE]` terminator.**
- **`Unwrap()`** keeps `http.ResponseController` (Flush/Hijack/deadlines) working through the wrapper, so SSE unbuffering (#33) is unaffected.

## Tests

| Test | Guards |
|---|---|
| `TestResponseWriterWrapperRecordsSSEMetrics` | TTFT=1, ITL>0, tokens=2 (terminator excluded) |
| `TestResponseWriterWrapperIgnoresNonStreaming` | JSON response records nothing |
| `TestResponseWriterWrapperPreservesStreaming` | bytes + Flush pass through |
| `TestHandlerRecordsLLMMetricsForStreaming` | end-to-end through `ServeHTTP` |

The last one fails if the wrapper is ever wired nowhere again — verified by disabling the wiring locally and watching it fail.

## Verification needed post-merge

Confirm samples appear:
```
curl -s https://api.riotpiao.com/metrics | grep -E "llm_(ttft|itl)_seconds_count|llm_tokens_total"
```

internal/proxy/llm_metrics.go

@@ -3,7 +3,6 @@ package proxy
 import (
 	"bufio"
 	"fmt"
-	"io"
 	"net"
 	"net/http"
 	"strings"
@@ -12,87 +11,37 @@ import (
 	"forgejo.riotpiao.com/rock/homelab-frontend/internal/observability"
 )
 
-// LLMMetricsCapture wraps a response writer to capture TTFT and ITL metrics
-type LLMMetricsCapture struct {
-	writer              io.WriteCloser
-	model               string
-	metrics             *observability.Metrics
-	firstTokenTime      time.Time
-	lastTokenTime       time.Time
-	requestStartTime    time.Time
-	ttftRecorded        bool
-	tokenCount          int64
-	responseStartTime   time.Time
-}
-
-// NewLLMMetricsCapture creates a new metrics capture wrapper
-func NewLLMMetricsCapture(writer io.WriteCloser, model string, metrics *observability.Metrics, startTime time.Time) *LLMMetricsCapture {
-	return &LLMMetricsCapture{
-		writer:            writer,
-		model:             model,
-		metrics:           metrics,
-		requestStartTime:  startTime,
-		responseStartTime: time.Now(),
-	}
-}
-
-// Write intercepts writes to detect tokens and record metrics
-func (c *LLMMetricsCapture) Write(p []byte) (int, error) {
-	// Record first token time
-	if !c.ttftRecorded && len(p) > 0 {
-		now := time.Now()
-		ttft := now.Sub(c.requestStartTime).Milliseconds()
-		c.metrics.RecordTTFT(c.model, ttft)
-		c.ttftRecorded = true
-		c.firstTokenTime = now
-		c.lastTokenTime = now
-	}
-
-	// Count tokens in SSE stream (simple: count "data: " lines)
-	if c.ttftRecorded {
-		tokenCount := strings.Count(string(p), "data: ")
-		if tokenCount > 0 {
-			now := time.Now()
-			if !c.firstTokenTime.IsZero() && c.lastTokenTime != now {
-				itl := now.Sub(c.lastTokenTime).Milliseconds()
-				c.metrics.RecordITL(c.model, itl)
-			}
-			c.lastTokenTime = now
-			c.tokenCount += int64(tokenCount)
-		}
-	}
-
-	return c.writer.Write(p)
-}
-
-// Close records final metrics and closes writer
-func (c *LLMMetricsCapture) Close() error {
-	if c.tokenCount > 0 {
-		c.metrics.RecordTokenCount(c.model, c.tokenCount)
-	}
-	return c.writer.Close()
-}
-
-// ResponseWriterWrapper wraps http.ResponseWriter to capture metrics
+// ResponseWriterWrapper wraps an http.ResponseWriter to capture per-request LLM
+// latency metrics (TTFT, ITL, generated tokens) while a streaming response is
+// copied from the upstream by httputil.ReverseProxy.
+//
+// Only server-sent-event responses are measured. A non-streaming completion is
+// written in a single Write, so recording it would emit one "TTFT" sample equal
+// to the whole request latency and no ITL samples at all — numbers that look
+// plausible but mean nothing.
 type ResponseWriterWrapper struct {
-	writer          http.ResponseWriter
-	statusCode      int
-	metrics         *observability.Metrics
-	model           string
-	startTime       time.Time
-	firstByteTime   time.Time
-	lastWriteTime   time.Time
-	ttftRecorded    bool
+	writer        http.ResponseWriter
+	statusCode    int
+	metrics       *observability.Metrics
+	model         string
+	startTime     time.Time
+	firstByteTime time.Time
+	lastWriteTime time.Time
+	ttftRecorded  bool
+	streamChecked bool
+	streaming     bool
 }
 
-// NewResponseWriterWrapper creates a wrapper for response writer
+// NewResponseWriterWrapper creates a wrapper for response writer.
+// startTime must be the time the request entered the gateway, not the time this
+// wrapper was constructed, otherwise TTFT measures the wrong interval.
 func NewResponseWriterWrapper(w http.ResponseWriter, model string, metrics *observability.Metrics, startTime time.Time) *ResponseWriterWrapper {
 	return &ResponseWriterWrapper{
-		writer:      w,
-		model:       model,
-		metrics:     metrics,
-		startTime:   startTime,
-		statusCode:  200,
+		writer:     w,
+		model:      model,
+		metrics:    metrics,
+		startTime:  startTime,
+		statusCode: 200,
 	}
 }
 
@@ -101,29 +50,48 @@ func (w *ResponseWriterWrapper) Header() http.Header {
 	return w.writer.Header()
 }
 
+// isStreaming reports whether this is an SSE response. Resolved once from the
+// Content-Type header, which ReverseProxy has already set before the first write.
+func (w *ResponseWriterWrapper) isStreaming() bool {
+	if !w.streamChecked {
+		w.streamChecked = true
+		w.streaming = strings.Contains(w.writer.Header().Get("Content-Type"), "text/event-stream")
+	}
+	return w.streaming
+}
+
+// countSSETokens counts generated tokens in a chunk of an SSE stream. These
+// upstreams emit one `data:` event per token, so the event count is the token
+// count; the `[DONE]` terminator is not a token.
+func countSSETokens(b []byte) int64 {
+	var n int64
+	for _, line := range strings.Split(string(b), "\n") {
+		line = strings.TrimSpace(line)
+		if strings.HasPrefix(line, "data:") && !strings.HasPrefix(line, "data: [DONE]") {
+			n++
+		}
+	}
+	return n
+}
+
 // Write implements http.ResponseWriter
 func (w *ResponseWriterWrapper) Write(b []byte) (int, error) {
-	// Record TTFT on first write
-	if !w.ttftRecorded && len(b) > 0 {
+	if len(b) > 0 && w.isStreaming() {
 		now := time.Now()
-		ttft := now.Sub(w.startTime).Milliseconds()
-		w.metrics.RecordTTFT(w.model, ttft)
-		w.ttftRecorded = true
-		w.firstByteTime = now
-		w.lastWriteTime = now
-	}
-
-	// Record ITL for subsequent writes (for streaming)
-	if w.ttftRecorded && len(b) > 0 {
-		now := time.Now()
-		if !w.firstByteTime.IsZero() && w.lastWriteTime != now {
-			itl := now.Sub(w.lastWriteTime).Milliseconds()
-			// Only record if ITL > 0 (avoid recording same millisecond twice)
-			if itl > 0 {
-				w.metrics.RecordITL(w.model, itl)
-			}
+		if !w.ttftRecorded {
+			w.metrics.RecordTTFT(w.model, now.Sub(w.startTime).Milliseconds())
+			w.ttftRecorded = true
+			w.firstByteTime = now
+		} else if itl := now.Sub(w.lastWriteTime).Milliseconds(); itl > 0 {
+			// Only record positive ITL: a zero means two chunks landed in the
+			// same millisecond, which carries no timing information.
+			w.metrics.RecordITL(w.model, itl)
 		}
 		w.lastWriteTime = now
+
+		if n := countSSETokens(b); n > 0 {
+			w.metrics.RecordTokenCount(w.model, n)
+		}
 	}
 
 	return w.writer.Write(b)
@@ -150,4 +118,8 @@ func (w *ResponseWriterWrapper) Hijack() (net.Conn, *bufio.ReadWriter, error) {
 	return nil, nil, fmt.Errorf("response writer does not implement Hijacker")
 }
 
-
+// Unwrap exposes the underlying writer to http.ResponseController so optional
+// capabilities (Flush, Hijack, deadlines) keep working through the wrapper.
+func (w *ResponseWriterWrapper) Unwrap() http.ResponseWriter {
+	return w.writer
+}