How Carry Clearance works
The model
The item is its rigid bounding box. If the box can make a move, so can any item inside it, which is why results are cautious for real furniture. The route is a list of steps, and one carrying orientation (which dimension is vertical) must work for all of them. Re-orienting between steps isn't credited.
Moves that are credited
- Openings and straight hallways: carried straight along one horizontal dimension. The cross-section must fit the clear opening, upright or tilted within the opening plane (Carver's exact rectangle-in-rectangle condition). "Must stay upright" disables tilting.
- Right-angle corners: carried flat, either straight across the corner without turning, or turned with its outer edge sliding along both outer walls. The turn is checked at 2,000 angles, with the item enlarged by the most any point can move between samples. The vertical dimension must clear the ceiling or door head.
- Turning into a door off a hallway: treated as a corner whose second leg is as narrow as the door. That's stricter than a real room, so it's cautious.
- Straight stair runs: cross-section against the width and the headroom × cos(pitch). Tops, bottoms and turns aren't modelled, so stairs are capped below "probably fits".
- Spaces: vertical dimension under the height, footprint flat on the floor at any angle.
Verdicts
- Probably fits: every step passes with the item enlarged by the safety margin (default 2 in / 5.1 cm).
- Straight run clears, but the ends can't be confirmed: as above, but the route includes stairs or a landing.
- Too tight to call: passes only without the margin.
- Won't fit by the moves we check: fails even with no margin. Other moves might work. Don't count on them.
It never says "fits" without qualification. Spare clearance is how much bigger the item could be in every dimension and still pass.
How it was tested
- Textbook cases: the 9 ft / 6 ft hallway pipe (21.07 ft); equal hallways (2√2·a − 2w); a stick in a square (√2).
- Carver's condition agrees with a brute-force rotation sweep on 3,000 random cases.
- Corner turns match a separately written search of positions and angles: feasible at 97 % of the limit, infeasible at 103 % for thin and medium items, and at 95 % / 103 % for wide ones.
- Every accepted turn is re-checked pose by pose by separate collision code, and every accepted door passage is rebuilt as an explicitly rotated 3D box.
- Property tests confirm that larger spaces never make results worse and larger items or margins never make them better. Every "probably fits" route is re-verified using only the independent checker.
The test code and logs are in the project source. Passing tests show the maths does what's described here. They can't make up for wrong measurements or for a move that isn't described here.