[huxiaoyi-ovo]

Summary

The quintic planner samples with np.arange(0.0, T + dt, dt). When dt does not divide T, the final sample is past the polynomial's terminal time, so the returned trajectory can overshoot its specified final position even though the polynomial itself satisfies the endpoint conditions.

For example, with T=5 and dt=0.7, the last sample is t=5.6. A zero-velocity, zero-acceleration quintic from x=0 to x=10 evaluates to about x=10.2054 at that time instead of the intended x=10.

This patch samples only times below T and then includes T exactly once. It keeps the existing sampling interval, avoids extrapolation, and does not change polynomial coefficients or trajectory search logic.

Regression tests

Parameterized tests cover:

• dt=0.5: divides the candidate duration exactly.
• dt=0.7: previously sampled beyond the endpoint.
• dt=6.0: longer than the candidate duration.

Each checks strictly increasing timestamps, no samples after T, consistent output lengths, and the requested terminal position.

Scope: two files, no new dependencies. Full cross-platform checks were started on fork CI PR #4; they are not claimed as passed yet.

AI assistance: ChatGPT (OpenAI GPT-6) helped identify the edge case, prepare the fix, and add regression tests.