# Depth-aware glass improvement

Goal: retain Ref1 rectangular color planes while making transmission follow the existing displaced photo and preserving manual framing.

Architecture: keep the current mesh, render targets, atlas and history. Add bounded screen-depth correction to Mosaic only, adaptive depth evidence from a histogram, and a subtle view-dependent environment reflection. No semantic dog segmentation or hidden-surface reconstruction is claimed. Preserve current maximum-depth placement for penetration safety; do not move the face forward by guessing its identity.

Approved scope: user's “작업해봐” follows the in-chat review and proposed order. Work directly in this non-Git workspace; original main/shader saved in docs/qa-depth-coupling. No new agent-assignment instructions.

- [x] Preserve manual zoom/pan on media input; explicit reset still resets.
- [x] Test then add depth histogram evidence thresholds (flat input disabled, sparse near outlier robust), constants under src/constants.
- [x] Add three bounded surface-depth iterations shared by atlas/final. Reject missing, front-of-glass or nonconvergent samples; retain existing artistic path as fallback. Do not attenuate color simplification with distance.
- [x] Add restrained common environment reflection after color simplification, with gloss/strength gates; preserve other modes.
- [x] Run focused regression scripts/build and browser AI, No Glass, Final, framing reload checks. Record limits and comparison evidence.

Review correction: a collapsed P50–P85 interval did not always mean flat depth. Added failing bimodal-distribution cases and P5–P95 fallback; all tests then passed. Quantile trimming can still ignore foreground occupying under 5% of the image. This is a bounded heuristic, not segmentation.

Acceptance: no automatic framing change; actual depth influences the sample correction, invalid samples fall back, no raw eye/nose restoration, gloss zero has no reflection, build and relevant checks pass. Visual comparison is qualitative, not proof of mobile performance or physical accuracy.