# Coordinated body/feet and visibility — frozen before candidate search

Brief: relocate Lunte one metre camera-right over source frames 80–103. Preserve camera, authored timing, face, hair, backpack and environment. Allow body height in [-0.15, 0.25] m and up to 0.5 m of horizontal camera-depth adjustment in either direction to find a viable placement. Explicit source discontinuity at frame 83 divides temporal constraints into 80–82 and 83–103. This adds control capacity, not merely a new statistic.

Source interpretation (explicit adapter, not automatic semantics): RIG-lunte_proxy; existing IK-MSTR-Foot.L/R controls; GEO-suit foot weights DEF-Foot/DEF-Toe; terrain surfaces ground, ground.001 and ground.002. Vegetation and props are excluded from support but retained as occluders. The character's explicitly identified head and suit are checked separately for external occlusion.

Measurement: each evaluated foot contains vertices with total foot/toe deform weight >0.5. A sole proxy is the mean of its lowest 20% world-Z vertices. This is evaluated footwear-region geometry, not a certified contact patch. For each source segment and foot, provisional support means clearance <= segment minimum +2 cm. Other poses retain source foot-to-body translation; no ground snap during provisional swing. Preserve source support clearance, including original stylized penetration. Labels are provisional and must be reported.

Both methods use the same candidate set and the same rig controls: camera-depth offsets [-.5,-.25,0,.25,.5] m and body heights -.15 through .25 m in .025 m increments; existing foot IK translates toward geometry-derived targets. Up to eight iterations per candidate. The conventional scripted comparator independently chooses the least-changing feasible pose each frame. The coordinated method chooses a whole-segment path through exactly those candidates, penalizing correction changes and enforcing an added-offset velocity bound. Neither is a benchmark against an expert animator or an external commercial tool.

Hard spatial checks: target sole residual <=2 mm; provisional-support clearance change <=2 cm; added penetration beyond source <=2 cm; lower-leg length change <=5%; additional external occlusion <=5 percentage points independently for sampled head and suit vertices; no new image-boundary clipping in those samples; camera and external transforms unchanged. Face/hair/backpack rigid-translation residual <=1 mm. Root-depth/height and foot-controls are the only edited rig channels. Foot joint limits and full mesh collisions are not exhaustively certified.

Temporal policy: within each segment, adjacent added body-depth, body-height and each foot's world correction vector must change by no more than 0.5 m/s. Use actual 1/24-second intervals, including held frames. The conventional comparator is scored against this same bound after selection. Optimization cost is normalized squared root/foot changes, plus squared between-frame correction changes. Missing candidates, infeasible transitions and uncovered surfaces remain failures; do not change thresholds to obtain a pass. Publish the least-cost spatial-only path if the temporal gate fails, explicitly rejected.

Render source reference (reuse exact previous PNGs explicitly), framewise comparator and coordinated candidate at 640×360 / Eevee16 / subdivision1. Render every selected frame, replay measurements in fresh processes, use existing monitor-core for final gates, inspect all candidate PNGs and playback. Mechanical acceptance is distinct from artistic approval. The original source stays unsaved. Preserve prior failed transfer outputs.