[ptescher]

Part of #95 (plan: plans/battlefield-latency.md). Follows a review of the earlier conclusions; details and numbers in plans/battlefield-latency-results.md.

What changes

• Lag figure (teleoperation_battlefield::lag): a paint now says how old its picture is (time since the app asked for the camera it shows). The old figure ramped from 0 at the start of each pan, which made a steady ~500 ms delay look like a queue that grows with the pan's length. The picture from before a motion is not counted; the wait for the motion's first picture is logged apart.
• Parked main camera (RenderScale::parked, set by towers_view --hidden): the view's own camera, which nobody sees when the app starts it, draws a 64×36 picture of the sky where it stands instead of the full scene at 1920×1080 among the trees. It can't be turned off (Bevy's atmosphere needs a live view).
• One frame in flight (pacing::IN_FLIGHT 2 → 1), and the pacer runs in PostCleanup, after Bevy's map_buffers, so its wait also completes that frame's readback.

Numbers

	Before	After
View alone, request to arrival	203 ms (2 in flight)	155 ms (0 in flight: 122 ms, 13 % fewer pictures)
In a game, the view's frame	112–153 ms (7–9/s)	56–64 ms (16–18/s)
In a game, pictures sent to the app	7/s	16/s

The battlefield camera has no sun shadows to cut: Bevy's shadows reach 150 m and it sits 450 m from the ground. No look changes in this PR.

Tests

• New unit tests: a steady delay reads the same on short and long pans; the picture from before a motion is not counted; a parked camera draws a tiny picture and looks up from where it was; --hidden parks it. Three old lag tests changed on purpose for the new definition.
• cargo clippy --workspace --all-targets, cargo fmt clean. cargo test --workspace --release: all pass except cognition_league's acceptance suite, which ran out of its 120 s budget before its tests started (unrelated to this change; the machine was busy).
• «The picture keeps up with the camera» stays Status: pending: it still fails, and the app-side figures from an automated run are unreliable (the window is off screen). Next: an on-screen run.

🤖 Generated with Claude Code