gateway: count usage from multiline and final SSE events - #94

Thanks for this. I pulled the branch and ran it: the new test passes and the gateway suite is green. But the change breaks a case that works on main, so I can't merge it yet.

Regression: the sniffer now parses only when it sees a blank line (or EOF). A stream that separates events with a single \n gets all its data: lines joined into one blob, which isn't valid JSON, so every usage in it is lost. Main parses each line on its own and counts both. Relays and self-hosted proxies emit this shape in practice. This test passes on main and fails on the PR (usage {Input:0 Output:0 …}):

func TestSniffSingleNewlineEvents(t *testing.T) {
	payload := "data: {\"type\":\"message_start\",\"message\":{\"usage\":{\"input_tokens\":7}}}\ndata: {\"type\":\"message_delta\",\"usage\":{\"output_tokens\":3}}\n"
	s := newSniffer(provider.Anthropic, "text/event-stream")
	s.write([]byte(payload))
	if got := s.usage(); got.Input != 7 || got.Output != 3 {
		t.Errorf("usage %+v", got)
	}
}

Suggested approach: keep the per-line parse. On each data: line, if the event collected so far plus this line is valid JSON (json.Valid), parse it and reset. Otherwise keep collecting, and flush what's left at a blank line or at EOF, as you do now. That handles multi-line data, the unterminated final event, and single-newline streams. Please add the test above next to yours.

Thanks for the quick v2. I checked it again, and it's merged.

What I verified:

• The single-newline test now passes, as do the existing sniffer tests. go test -race ./internal/gateway and go vet are clean.
• I ran extra edge cases, each written 1 byte, 7 bytes and whole: Chat with a final [DONE], CRLF line endings, single-newline Chat events followed by [DONE], a full Anthropic stream with ping, and data: with no space. All were counted correctly.
• I also ran it end to end: I built your branch and sent real streaming requests through it to DeepSeek, one over Chat Completions and one over Anthropic Messages. Both landed in usage.jsonl with the right input and output tokens.