Depth observability parity: rerun live view + frame store

[jeqcho]

Closes #236

Plan: plans/0036-depth-observability.md (pre-critique draft).

Three-part change, all core:

1. Rollout resolves depth once per step and shares it with every consumer (policy, rerun, frame store) — fixes the callable-timing hazard where each consumer would capture a different moment; also strips bulk depth from the persisted record (plan 0021 follow-up).
2. RerunSink logs rr.DepthImage per camera at {trial}/camera/{cam}/depth (metric hover), watermark-aware, SDK-absence guarded.
3. FrameStore persists float32 depth under a frames/<run>/depth/ subdirectory with a dtype-preserving DepthFrameRef — subdir chosen so inspect-robots video stream discovery and other flat-glob readers stay blind to it.

Non-goals: transcript imagery (transcripts are image-free by design; wire capture is the exact-payload record), depth video rendering, depth-only cameras.

Status: plan committed; implementation not started (holding per operator).

🤖 Generated with Claude Code