fix(tracking): prevent overshoot from stale camera frames

[stefanleyb]

Issue

Closes #1331

Description

Face tracking was projecting a detected face from the head pose at processing time, although the camera frame was captured earlier. While the head was moving, this made the target escalate and caused repeated overshoot/correction.

This records each producer-side camera frame timestamp by IPC buffer offset, retains a short measured head-pose history, and interpolates the pose at capture time before projecting the face into world coordinates. A detector configured for timestamps drops a frame when its timestamp is unavailable instead of silently using a later pose. No public API or daemon route changes.

Testing

• ruff check and mypy pass for the changed source files.
• 68 focused unit tests pass, covering timestamp mapping, face tracking, backend pose history, and media-server pose/watchdog/WebRTC paths.
• Tested on one Reachy Mini Wireless running the 1.10.0 release. In three stationary-face positions, measured target overshoot fell from 12.1–14.9° to 0.7–3.9° and settling time fell from 2.46–5.04 s to 1.14–1.76 s.
• With instrumentation disabled, a normal daemon tracking run reacquired the operator after natural movement and brief occlusion, then held direct gaze.

Tested on

• Reachy Mini Wireless
• Reachy Mini Lite
• MuJoCo simulation (--sim)
• Mockup simulation (--mockup-sim)
• Not applicable (e.g. docs, typo)

AI assistance

• No AI involvement
• AI helped with wording or boilerplate
• AI wrote code here, and I ran and reviewed it
• An agent produced this PR, and I have not run it myself

Assisted-by: OpenAI Codex: GPT-5

[stefanleyb]

Videos of Before/after, where the overreach of the initial movement is visible in the current version (including blocking the movement because Reachy goes to its limits), which makes it a bit weird and unnatural.
Similar video with the new version, where Reachy directly looks at the user, builds connection and shows curiosity.

https://github.com/user-attachments/assets/7d5fd8d3-0929-41fb-9b25-f8036fa52f24
https://github.com/user-attachments/assets/07bf23fe-d940-45d9-b543-2bdb3d09c494

[RemiFabre]

This is a very good input, thank you. We're reworking how the face tracking works. Would you be ok to test our new implementation for feedback?

[stefanleyb]

@RemiFabre yes, I would definitely be happy to test the new version! I am working on a version where the body moves also, and I think I finally managed to have a clean "natural-looking" movement for the body. At some point, body movement could be a parameter when launching face tracking instead of an app built on top of head-only movement. Feel free to ping me here or on discord (stephaneleyb_29208).