Retain relative reconstruction after camera stop

[JROChub]

Root cause

The relative lane had valid Keyxym geometry in memory, but presentation was hard-coded to hide every non-metric native surface. STOP also detached the video stream, clearing the only visible scene. Guidance could then overwrite the stopped state with tracking-loss language. The resulting screen was blank despite thousands of retained surfels.

Correction

• freezes the final sensory frame before hardware tracks close
• renders native relative triangles or surfels over that retained frame after capture stops
• keeps the live camera preview during active monocular capture
• exposes explicit capture-active, retained-frame, reconstruction-visible, and surface-mode state
• replaces terminal TRACKING FROZEN errors with recoverable frame rejection or relocalization states
• prevents guidance from overwriting a stopped reconstruction
• changes reacquisition guidance to RELOCALIZING / MAP RETAINED
• extends live deployment verification to require the retained-frame route and relative-native reconstruction markers

Verified reproduction and result

The pre-fix regression failed with 5,097 surfels and 659 native triangles present but a blank stopped stage. The same exact official Middlebury TempleRing archive now ends with:

• 5,097 surfels
• 1,977 native surface vertices / 659 triangles
• relative-native-triangles presentation
• retained recognizable source frame
• RECONSTRUCTION READY
• 2 Moments and a sealed Rootprint

Qualification: 31/31 contract tests, production TypeScript/Vite build, 3/3 Chromium theater/navigation tests, 2/2 iPhone WebKit tests, and calibrated metric regression at 14,614 surfels / 50,588 triangles. power_house was not modified.