Prime engine baselines when a turn starts

[charlesnutter]

Prime a counter-difference engine's baseline when a turn starts on it, so the first turn after OpenCode starts can show engine figures instead of engine telemetry from the next turn. Ships as 0.3.1.

What changes

• Lazy priming. At session.execution.started, the turn's engine is taken from the session's last assistant message or model switch (or a session.model.selected event). If it is vllm, sglang, vllmmlx, aphrodite, lmdeploy, llamacpp, llamafile, splash or omlx and has no baseline yet, that one engine is read once. Nothing is read at startup, and no other engine is read.
• Fallback heading uses the engine's display name (vllm-mlx, not vllmmlx).

Verified live (vllm-mlx, OpenCode 2.0.12)

• First turn after two separate restarts: baseline read in 4ms, first step ~7s later; windows matched exactly (909 = 909, 459 = 459 tok).
• New session on the default model: no model known, so unprimed — from the next turn, as before.

Not covered

• A brand-new session's first turn: OpenCode exposes no selected model to plugins before the first reply.