[outofcoffee]

Show the rate the engine reports for itself — tokens per second — on every metrics surface.

Closes #271

Summary

• The daemon parses the instantaneous tg figure from llama.cpp's print_timing lines out of the engine's captured output, and reports it as an optional speed reading on /v1/metrics plus per-tick s points in the retained history. Engines that print no rate, stopped engines and engines quiet for a sampling interval simply omit the figure — absent, not zero.
• Bar, combined (serve view), gauge, table and dashboard surfaces draw a Speed row: the t/s figure and, where history is drawn, a sparkline whose glyph heights scale to the window's own high watermark and whose points all draw in the terminal's default colour — the 80/90% thresholds belong to utilisation percentages, and a rate has no ceiling to cross.
• The GPU utilisation series is relabelled: GPU util/GPU N util become GPU/GPU N (memory keeps GPU mem).
• Backward/forward compatible: older daemons and the cloud control plane omit the fields; older clients ignore them.

Implementation details

• A lineTee wraps the supervisor's log-file capture sink: raw bytes reach the log file untouched, and complete lines are additionally handed to an injected OnOutput hook. One tee covers both capture paths (stderr and the pty pump's normalised stdout), so the parser reads lines where they are already assembled rather than re-reading the log. The hook must be wired before Supervisor.Start, which builds the sink as it launches.
• The tracker answers two questions with two clocks: the sampler's tick folds a history point for figures its interval saw (strictly since the previous tick — a point is an observation, never a continuation), and a metrics read reports the current figure only until one interval's worth of quiet ages it out. Engine starts reset tracker, baseline and history together, so one engine's rate never rides the next engine's report.
• The parser is deliberately narrow: tg\s*= cannot match tg_3s =, so the three-second average and prompt-evaluation rates never read as the current rate.

Spec deltas live in the openspec change engine-generation-speed (to be archived after merge).