engines tab: engine card shows all-zero metrics for a busy llama-swap model; unreported reads as zero #140

With a llama-swap engine wired up, the engine card shows DECODE 0.0 t/s, PREFILL 0.0 t/s, "in 0 tok, out 0 tok, queue 0", and a model row with the name and no values — even while the model is actively serving.

The chain (main @ 050866f):

1. The llama-swap reader (internal/engines/engines.go, llamaSwap) reads llama-swap's /running, then each ready model's proxy /metrics with the llamacpp parser.
2. llama-swap starts llama-server without --metrics, so the proxy answers 501 "This server does not support metrics endpoint. Start it with --metrics" and out.Metrics stays zero.
3. rates() computes decode/prefill from those counters, so they read 0, and the card renders (eng.decode_tps || 0).toFixed(1) — an unreported field is drawn as a measured zero.

Live state while this was captured: slot s0 for qwen-3.8-27b processing with 65,950 tokens in its KV window, and the node's own request log computing decode traffic for the same engine (the basic tab's ticket counter showed 14.5 t/s) — while the engine card shows 0.0 t/s and zero token totals.

The model row has the same shape: /running carries no per-model VRAM and the reader sets no ReportedVRAM for llama-swap models, so the row shows name and state with no values.

Suggestion:

• the card should render n/a (the util.js na() helper already does this on the same tab, e.g. the leaderboard's decode median) for fields the engine does not report, instead of 0.0/0;
• when the engine's counters are absent, the node's own request log already has the tps — falling back to it would keep the card honest for engines whose /metrics is off;
• on the llama-swap side, starting its llama-server with --metrics would give the reader real counters (a llama-swap default, not a berth change).