Curve Speed Controller

This feature separates learned cornering comfort, future curve geometry, and runtime control ownership. Its numerical policy is derived from StarPilot 678af78347d9656bc2f5dacd4204b012db861fa7; the applicable notice is in LICENSE. The pure modules do not read Params, publish commands, or enable longitudinal control.

LearnedCurve keeps 24 logarithmic curvature buckets and a weighted monotone comfort fit. Historical sample weight is capped at 600 when adding observations; the stored count continues to track calibration. Driver feedback uses a separate 20-sample weight. Ordinary observations and feedback require runtime permission: calling these methods alone does not establish driver ownership of speed.

The loader accepts the legacy bucket mapping or a version-1 document. Loading does not write, mark data dirty, or silently rewrite a damaged document. Invalid entries reject the whole document; the caller must preserve the original bytes and require explicit reset before collecting replacement data. Counts must be positive integers within the exact JSON integer range. Missing data is distinct from a storage read failure. A successful asynchronous write acknowledges only the revision it actually saved, preserving any newer dirty samples.

CurveProfile validates an immutable, bounded forward profile. evaluate produces a target and binding distance using the learned comfort, approach deceleration and distance-dependent curvature correction. Its weather reduction argument requires an independently qualified source; absent weather supplies zero reduction. Runtime must establish freshness and authority before a target can become a planner ceiling. Invalid or stale geometry is unavailable, not a straight road with an invented speed target.

TargetFilter preserves the frozen 20 Hz seeding, filter, release and slew behavior. A gap longer than 0.2 seconds resets it; runtime must handle restart and clock discontinuity before reentry. Lowering driver cruise immediately caps the filtered target. Saved malformed data and malformed geometry receive stricter validation than the frozen implementation; these changes are deliberate.

The planner integration requires the default-off saved switch and an exact tagged Ioniq 6 system-longitudinal session; CURVE_REPLAY_RUNTIME=1 retains the explicit developer path. A candidate only lowers the cruise ceiling after SLC composition; native lead/MPC braking, experimental model braking, force-deceleration and driver overrides retain their authority. Feedback is attributed to Curve only after the actual acceleration composition selects it. Card supplies separately identified physical acceleration presses after excluding SLC-owned actions. A stale event cannot become new feedback.

Live input validation checks the original model event and camera EOF in their respective clock domains. Startup and suspend/resume require new source events. Explicit offline replay uses the recorded event timeline. Tiny negative model origins caused by numerical roundoff are normalized only within one micrometre; negative later points and decreasing paths remain invalid.

Native settings and Galaxy share one calibration owner. Their confirmed parked actions require the saved switch already Off, unchanged source bytes and no active learning session. Adoption writes validated legacy samples to the current format. Reset writes an empty current document so older saved samples cannot silently reappear. Both retain the legacy bytes and leave the switch Off. A conflicting editor or second learner prevents activation for that session; recovery requires starting a new session after the conflict ends. Persistence uses the ordinary Params lock and an atomic replacement, and distinguishes a completed rename from a verified, durable save.

Live diagnostics share the existing planner status publisher but have their own session, model timestamp and expiry. SLC availability does not imply Curve availability. The optional large-screen glow and compact curve cue consume these diagnostics and independent longitudinal authority; saved progress and comfort remain explicitly separate from live learning.

Source-bound tests cover numerical learning and target traces, native planner composition, driver events, persistence races, reset/adoption and both UI profiles. Runtime weather adjustment still requires a qualified producer and currently contributes zero reduction. Device timing, current-model calibration, road comfort and full vehicle qualification remain required; host tests and recorded geometry alone do not establish them.