Deprecated constructors

The old constructor names remain as functions, not types. Each emits a Julia deprecation message and returns native annotations or a ShootingLayer:

The returned layers use the new Lux parameter/state layout and return (trajectory, state). Initial-state matching residuals live in trajectory.residuals. These functions do not restore legacy fields or support using their names as types in dispatch or isa checks.

Controls and single shooting

control = ControlParameter([0.0, 1.0, 2.0];
    name = :p₁, controls = [0.2, 0.4, 0.6], bounds = (0.0, 1.0))
layer = SingleShootingLayer(problem, algorithm; controls = [1 => control])

The equivalent native definition uses only the boundaries between intervals:

control = Tunable(1; kind = Val(:parameter),
    default = [0.2, 0.4, 0.6], lowerbound = 0.0, upperbound = 1.0) => [1.0, 2.0]
layer = ShootingLayer(problem, algorithm, control; controls = [1])

The deprecated shooting constructor selects the parameter using the outer index, so the old control name need not match the underlying symbolic name. A returned control definition can also be passed directly to ShootingLayer when its name is an actual parameter symbol. A problem's final endpoint is excluded from the translated control switching grid. At a problem start between legacy grid entries, the control value active at that time is used; before the first grid entry, the first control value is used.

Legacy control callbacks (rng, times, bounds) -> values are adapted to native time-based defaults and evaluated during setup. Scalar bounds and uniform bound vectors translate to scalar annotation bounds. Nonuniform per-time bounds are rejected because the native annotations do not represent them; construct the optimizer's per-entry bounds explicitly when migrating.

tunable_ic selects variables tunable at the problem start. Unselected variables pass through the problem's initial state. Non-control parameters remain tunable. bounds_ic and bounds_p provide annotation bounds; default guesses for tunable entries are clamped to those bounds. bounds_p indexes non-control parameters in problem parameter order. Extra keywords, including tspan, tolerances, and saving settings, remake the problem before constructing the native layer.

Multiple shooting

layer = MultipleShootingLayer(problem, algorithm, [0.0, 1.0, 2.0, 3.0];
    controls = [1 => control], ensemble_alg = EnsembleThreads())

Interior shooting points become tunable resets for every state variable. Points at the problem start or end do not introduce additional interior resets. The result supports the native serial, threaded, or distributed block execution. Parameters and controls are shared across blocks. Original tunable reset values are retained at their original times; newly lifted values default to the bounded problem initial values.

quadrature_indices becomes the cache's quadrature metadata. Quadratures are lifted like other states and have continuity residuals; use their terminal values as accumulated costs. Legacy per-stage quadrature summation, parameter/control lifting, and stage-based optimizer metadata are not reconstructed by these shims. Fixed reset schedules cannot be combined with tunable shooting resets on the same variable using this conversion, since a native per-variable layer has one fixed/tunable status.

Legacy whole-vector state_initialization and parameter_initialization callbacks, and whole-layer multiple-shooting initialization callbacks, are rejected with migration guidance. Express guesses through the scalar, vector, count-based, or time-based annotation defaults.