A looping clip needs both pose and motion continuity at the wrap. A pose offset is a C0 discontinuity: the runtime jumps when it returns to frame 0. A matching pose with mismatched incoming and outgoing velocity is a C1 discontinuity: it reaches the right point but changes direction or speed abruptly, producing a hitch or pulse once per cycle. Unity's looping-clip guide shows the same artist-facing start/end match problem and the special treatment root-motion axes may need.

Animsmith separates four questions:

• loop-closure finds the largest last-to-first model-space position and shortest-path model-space rotation delta across all skeleton bones. The default caps are 0.01 m and 1 degree.
• loop-seam-vel finds the largest difference between the model-space linear velocity entering the last sample and leaving frame 0. Its default cap is 0.1 m/s. A closed triangle-wave trajectory demonstrates the problem: the first and last position are identical, but the bone reverses direction at the wrap.
• loop-seam-rot finds the largest difference between the shortest-path model-space angular velocities entering the last sample and leaving frame 0. Its default cap is 5 deg/s. This is rotational C1 continuity: the pose can close rotationally (C0) yet still snap in direction or turn rate at the wrap when the two seam-adjacent rotation steps disagree.
• loop-seam remains the locomotion-specific test. It compares feet relative to hips and normalizes by the neighbouring stride step, so it needs resolved hips/foot roles and deliberately has no result for a stationary clip.

There is a fourth, narrower warning for an export-mode problem rather than a pose-continuity judgment: duplicate-loop-endpoint. Many DCC workflows export an inclusive range or bake a cycle by copying frame 0 again at the final frame. For a declared loop, animsmith recognizes only the mechanically certain subset: every authored channel has one common finite, strictly increasing timeline and valid cardinality; first and final vector components match within 1e-5 and quaternions match within a sign-invariant 1e-4 radians; and the clip has real interior motion. A stationary hold is therefore not mistaken for a cycle. The warning is default-on, needs no rig roles, and is otherwise not applicable unless loop = true is declared.

That duplicate can matter to an engine that loops by advancing over the clip duration: it evaluates or holds the same pose twice before wrapping, which can look like a one-frame hitch. transform --drop-duplicate-loop-endpoint turns only an eligible candidate into an open cycle: it atomically removes the same complete terminal key count from every channel (and cubic-spline key triplets), preserves retained authored data, and re-pins duration. It does not repair a nonclosing/root-travel clip, mismatched timelines, a C1 tangent problem, retargeting damage, or a runtime loop blend. Blender's Cycles F-curve modifier and Unity's looping-clip guide are useful places to inspect the authored range and engine import behavior.

An open cycle intentionally does not keep loop-closure green: that existing inclusive check compares a repeated final sample with frame 0. The loop_endpoint_mode measurement distinguishes strict duplicate endpoints, non-duplicate closing cycles, and non-closing cycles for declared loops.

The first three checks are per-bone and role independent, so idle, guard, block, aim-offset, facial, and prop loops remain testable without a humanoid profile or detectable stride. Model-space is intentional: a parent mismatch can move or rotate many descendants even when their local keys match. JSON measurements retain a stable bone_index plus display bone_name for every row, while findings name only the maximum offending bone for each judged dimension.

Typical causes are an export range ending one frame early, a cycle modifier or procedural controller not being baked, copied endpoint keys whose Bezier tangents do not match, retargeting or resampling that changes one endpoint, and an engine importer using a different clip range from the DCC. A useful diagnostic split is:

• nonzero position/rotation delta: repair the endpoint pose or clip range;
• zero pose delta but high velocity delta: repair the endpoint tangents or the seam-adjacent keys;
• closed rotational pose but high angular-velocity delta: repair the endpoint rotation tangents or the seam-adjacent rotation keys using the intended shortest turn;
• many descendants reporting the same displacement: inspect their first mismatching ancestor;
• only loop-seam failing: inspect locomotion phase and foot/hips-relative motion rather than a stationary pose loop.

The normal fix is in the DCC: select the intended complete cycle, bake procedural motion, make the last pose match the first where appropriate, match the derivative on both sides, then re-export and rerun animsmith measure and animsmith lint. Blender's Cycles F-curve modifier can preview repeated curves, but the evaluated endpoint and tangents still need to survive baking and export. Engine-side loop-pose blending can hide a small mismatch at runtime; it does not make the source measurement close.

Root-motion locomotion is the important exception. Intentional horizontal root travel does not return to its starting model-space position, and every child inherits that travel. Tune the global max_position_delta_m, or the clip-level max_loop_position_delta_m, to the contract; alternatively disable loop-closure for such a pipeline. Keep using loop-seam for feet-relative locomotion. loop-seam-vel and loop-seam-rot can still validate constant inherited linear and angular motion respectively.

When one project contains both stationary idles and root-motion locomotion, keep the global [checks.loop-closure], [checks.loop-seam-vel], and [checks.loop-seam-rot] caps strict, then put a finite, non-negative max_loop_position_delta_m, max_loop_rotation_delta_deg, max_loop_velocity_delta_mps, or max_loop_angular_velocity_delta_degps on the relevant [clips."run_*"] family. A [clips.run_forward] entry wins over matching globs for only the fields it supplies, so a one-off authored clip need not copy every inherited value. This is a contract choice, not a repair: it does not make a discontinuous source clip smooth.

There is no general automatic repair. transform --gait-anchor can rotate an explicitly in-place locomotion cycle in time to choose a better stride cut; it refuses accumulating root translation or yaw. Every nonconstant selected Root/Hips trajectory channel must contain exactly one key at each declared whole-frame sample; sparse, differently framed, duplicate-time, or off-grid evidence refuses so the phase shift cannot synthesize values at omitted frames. Duplicate (bone, property) channels refuse too. The safety grid refuses before allocation when declared frames × skeleton bones, declared frames × tracks, or maximum authored keys × skeleton bones exceeds 1,000,000 samples. It does not rewrite arbitrary bone endpoint poses or tangents. Angular-velocity C1 continuity, acceleration/jerk continuity, root-motion extraction policy, and runtime blend settings remain out of scope.

All four checks are judged only on clips declared loop = true — whether a clip is intended to loop is project knowledge — while raw measurements are always available for measurable clips through animsmith measure. Workflow: a project contract config shows the same walk cycle passing clean and failing with a popped seam, and why an undeclared loop is reported clean.