Count tool-call argument streaming as generation time

[charlesnutter]

Count the time a model spends writing a tool call's arguments as generation time. Ships as 0.3.3.

Problem

Tool-call arguments (e.g. a file's contents for write) are generated tokens and OpenCode counts them in the step's output, but the stream marks only watched text and reasoning deltas. Tokens were divided by too short a window: a write step showed 3,672 tokens "at 162.9 tok/s" against the engine's 36.4 for the turn, and the Session average read 55.1. Affects OpenCode-derived figures (universal line, per-step rates, Session speed), not engine figures.

Fix

Mark session.tool.input.started and .ended (and .delta, kept in case it is forwarded later).

Measured (MTPLX, OpenCode 2.0.12)

• session.tool.input.delta reaches the plugin 0 times per step, so deltas alone did not fix it (37.7 vs 32.4).
• With start/end marks, each step's last mark lands 0–1 ms before step.streamed.
• 8-step turn that wrote a file: Session 36.4 tok/s against the engine's 36.1.

Debug-only: a per-step stream window line compares the window's end with step.streamed.