Hyundai: cap EV9 OP-long steering targets at 140 degrees

[Intelli]

Description

Sharp turns with openpilot longitudinal configured on the Kia EV9 can end in “Steering Assist Temporarily Unavailable” and a steering dropout. Assistance can briefly return and drop out again, including when longitudinal is disengaged and only lateral assistance is active. This was reproduced on stock stable StarPilot; I have not observed it with stock ACC. Dom still permits autonomous angle targets well beyond the roughly 140° assisted-wheel plateau seen in the logs.

This change caps the EV9 autonomous target at ±140° before filtering on the OP-long steering path. Saturating the target does not itself deactivate steering. Stock ACC, other vehicles, inactive measured-angle commands, existing rate limits and Panda safety limits retain their current behavior. This is a target limit, not a hard clamp on the physical wheel or every transmitted angle during reentry.

Verification

The reproduction used stock StarPilot stable at c3e4ec630f41c4baa43254a90f718abd1bf764a1, installed through an exact-tree mirror at 272b7edcb30238297efe934f71b826dd2d810d1d (the commit recorded in the logs), with stock bundled Panda firmware and no custom changes.

Reproduction route — full logs uploaded. Useful windows from route start: 4:25–4:29 for repeated left-turn faults, and 4:51–4:58 for another cluster and the displayed warning.

• The EPS asserts LKA_ANGLE_FAULT while ordinary LKA_FAULT remains clear. Near 4:25.48, speed is 19.7 km/h and the measured wheel angle is 137.6° left. The preceding logged CAN batch requests 206.7° left; the batch containing the first fault includes a 208.1° left request. No driver steering override is detected at that first onset.
• StarPilot withdraws steering after the EPS fault and retries as it clears. Steering-message counters remain continuous, with no rejected steering frames or checksum failures around these fault clusters.
• The 140° target limit was tested on an EV9 with OP-long: there were no errors or steering cutouts, and no noticeable reduction in available steering compared with the uncapped configuration.
• Five controller/CAN regression cases passed, covering saturation/reversal, lateral-only operation, stock ACC/other-vehicle scope and inactive-angle/reentry behavior. The three EV9 regression cases fail on unchanged Dom.
• Python compilation and diff checks passed, with no new Ruff findings. The full Hyundai suite has the same 13 existing failures as unchanged Dom (253 passed, 2 skipped); none were introduced by this change.