Hold every joint to its acceleration limit where the path bends, timing the bends the way MoveIt's TOTG does

[Jepson2k]

The player plays one row per tick. Wherever the path bends between rows, the joint velocity changes within a single tick, however smoothly a profile times the path coordinate. Every profile that follows a path let joints exceed their acceleration limit at those bends:

• _enforce_segment_limits sized its acceleration stretch from the velocity-limited segment time. That time is usually far shorter than a tick, so the check rarely stretched anything.
• TOPP-RA's rows were never checked at all.

Peak joint acceleration as a multiple of the hard limit, with accel=0.5, so the allowed peak is 0.5× (shown as before → after):

Move	TOPPRA	LINEAR	QUINTIC	TRAPEZOID
MoveL +40x +40z from standby	3.1 → 0.5	4.2 → 0.5	2.6 → 0.5	4.2 → 0.5
Blended MoveJ chain, r=5	4.3 → 0.5	4.9 → 0.5	4.1 → 0.4	4.9 → 0.5
MoveC, r=15 mm half circle	1.8 → 0.5	1.2 → 0.5	0.3 → 0.3	1.2 → 0.5

These figures come from the dry-run plan record, whose rows are two ticks apart. At tick resolution, which is what the new test checks, every move stays within the acceleration it asks for.

Changes

• Every profile times the path's bends, as MoveIt2's TOTG does. _BlendedPath (one degree-2 PPoly) turns each corner into a parabolic blend, sized by the tick the rows are played at. TOPP-RA times its curvature, and LINEAR, TRAPEZOID and QUINTIC sample their path coordinate on it rather than on the polyline. A MoveJ chain's path parameter is its joint-space distance, so the density of the Bézier samples doesn't create speed changes.
• Process moves (MoveP) hold one feasible tool speed under every profile. The path-speed cap includes the bends.
• Corner blends turn the tool at a continuous rate. A spherical quadratic through the corner orientation replaces the entry→exit geodesic. A leg that turns the tool no longer forces a step in J6's speed.
• Requested durations are handled in one place: build(). Each profile times its fastest; the post-pass runs once on the final rows, including the seam after a wrist-turn prefix; then a longer request re-times the result through a C2 spline, so acceleration scales down smoothly. Ruckig gets the request as its minimum_duration.
• _segment_rates is a safeguard. This forward/backward pass over the played rows guarantees the tick-level limits. Its "Extending duration" message is a WARNING when it adds more than 10%.
• Planning cost. Cartesian constraints are evaluated vectorized over fewer gridpoints, and _segment_rates is compiled at warmup.
• CI lint fix. The async client's robot setter takes any waldoctl Robot (ty 0.0.86), the same change as in Add the SEPTIC motion profile #54.

Timing

Every row below stays within the acceleration limit (worst 0.99× the requested acceleration).

Move	First version of this PR	Now
Blended MoveJ chain, TOPPRA, r = 0 / 5 / 20 / 50	1.43 / 1.50 / 1.57 / 1.62 s	1.43 / 1.43 / 1.23 / 1.11 s
Same chain, LINEAR, r = 0 / 5 / 20 / 50	1.43 / 3.53 / 2.45 / 1.89 s	1.43 / 1.43 / 1.23 / 1.11 s
Two-leg J1→J2 chain, LINEAR / TRAPEZOID / QUINTIC	3.63 / 3.63 / 2.27 s	1.00 / 1.00 / 1.02 s
MoveP from standby, TOPPRA / LINEAR / QUINTIC	1.83 / 5.09 / 8.11 s, speed varies by up to 4× in TOPPRA's corners	2.50 s each, speed held within 0.3%
MoveP whose leg turns the tool, TOPPRA	39 s (J6 speed step)	7.98 s (the same legs as a MoveL chain: 4.22 s)
MoveL from standby, TOPPRA	2.08 s, planned in 213 ms	2.07 s, planned in 34 ms

Jerk is not bounded here. TOPPRA, LINEAR, TRAPEZOID and QUINTIC have unbounded jerk by design; RUCKIG and (after #54) SEPTIC bound it on the straight joint runs they time.

Tests

All of these enter through DryRunRobotClient, and the first three fail on the first version of this PR:

• test_blended_chain_is_no_slower_than_stops_and_respects_acceleration, parametrized over TOPPRA, LINEAR, TRAPEZOID and QUINTIC.

• test_process_move_holds_a_feasible_tool_speed, parametrized over the same four profiles.

• test_a_chain_along_one_line_takes_as_long_as_the_one_move: replaces test_blended_chain_is_longer_than_one_move, whose three waypoints lie on one joint-space line.

• A test that a MoveP whose leg turns the tool plans within 2× the same legs as a MoveL chain. It fails without the rotation blend.

• TestDryRunTimingRegressions, added after the second review round. Each test fails on 149c2ca:

  • a requested duration scales acceleration down smoothly (QUINTIC, TOPPRA, LINEAR, TRAPEZOID, RUCKIG);
  • QUINTIC paths ease acceleration in and out at both ends;
  • process-move timing doesn't depend on the sampling pitch;
  • MoveS doesn't stop where the tool's turn changes, and still turns along the geodesic;
  • a timed RUCKIG move lands on its requested tick count;
  • TOPP-RA's own rows stay within the acceleration limit at a corner.

Locally: 450 passed, 8 skipped. The one failure, test_msg_mounts_on_the_flange_not_inside_the_wrist, comes from rtree missing in the local venv; the dev extra installs it.

🤖 Generated with Claude Code

https://claude.ai/code/session_016uVshJDKTz7nCqrMGVxVJU