Keep collision geometry in world/local simulation coordinates. Apply cameras only when rendering. Sprite art, visible pixels and hardware sprite pieces are not automatically the collision shape.
Specify overlap versus contact, touching while moving away, grazing corners, starting inside a wall, nearest-hit ties and what happens when a work limit is reached. Do not hide these choices behind a generic Boolean.
Choose endpoint overlap for slow or intentionally discrete interactions; use swept queries when a body can cross an obstacle between ticks. A ray is a finite zero-width segment, not a substitute for sweeping a thick body.
For moving targets, sweep relative motion from both bodies' old positions. Detecting that contact does not provide pushing, platform carrying or an impulse solver. Sliding in the example is against static boxes only.

Keep the work and arithmetic bounded

Read query contracts for slab tests, exact hit ordering, fixed-point range and API examples. Do not reinterpret Q4 units as SGDK fix32. Read sliding and candidate selection for static response, initial overlap policy, tile candidates and limits. Never silently drop solids when a candidate buffer fills; stop or use a documented conservative fallback.

Use the project's existing collision model where practical. Keep slopes, one-way platforms and gameplay-specific interaction rules separate; this skill does not replace those with a full rigid-body framework. Collision layers or category masks should filter candidates before narrow tests, according to the game's actual rules, not SGDK palette or sprite priority flags.

Integrate and prove it

Verify types, rendering calls and resource syntax in the selected toolchain; pull ewrogers/sgdk:latest before validation. Read SGDK integration and sources.
Test zero/negative motion, thin walls, moving targets, grazing, equal-time corners, initial overlap, nearest hit, finite endpoints and range rejection. Use an independent oracle or hand-computed cases; optimized versus debug agreement alone can preserve the same bug.
Record candidate/narrow-test counts separately from CPU/frame timing. Keep response iteration and collision data lifetimes explicit. Do not grow the example's arithmetic limits or candidate arrays without checking the cost.
Check release builds and target arithmetic, then emulator/hardware movement and visuals. PAL/NTSC scheduler tests do not prove VBlank performance.

Collision Lab is a standalone executable example with endpoint/swept shots, movable aiming rays and wall sliding. Its art and controls demonstrate the algorithms; they do not prescribe another game's style.

For deciding whether a valid contact causes damage, is blocked or triggers a reaction, use sgdk-combat when available. Those policies do not change the geometric meaning of this skill’s query results.

sgdk-collision

SGDK collision

Decide what the query means