RustedSciThe 0.4.20

//! Symbolic BVP assembly and runtime preparation.
//!
//! This module is the main bridge between symbolic BVP construction and the
//! solver-facing residual/Jacobian callback layer. It owns discretization,
//! symbolic Jacobian preparation, lambdify/AOT runtime wiring, and sparse
//! backend selection.
//!
//! Current practical sparse mainline:
//! 1. `Jacobian` receives symbolic residual equations, unknown names,
//!    optional parameters, discretization settings, and boundary conditions.
//! 2. `generate_BVP_with_params(...)` discretizes the system and fills the
//!    symbolic residual cache plus dense/sparse Jacobian caches.
//! 3. `prepare_sparse_aot_problem(...)` turns the sparse symbolic cache into an
//!    owned `BvpPreparedSparseAotProblem`.
//! 4. Backend selection combines requested policy (`Numeric`, `Lambdify`,
//!    `Aot`) with resolver state and yields a `BvpSparseExecutionPlan`.
//! 5. The solver then consumes either:
//!    - `BvpSparseSolverProvider<'_>` / `BvpSparseSolverBundle` for sparse
//!      lambdify or linked compiled AOT,
//!    - `BvpLegacySolverBundle` for the remaining compatibility paths.
//!
//! Compiled AOT scenario:
//! - `BvpPreparedSparseAotProblem` converts into a shared
//!   `PreparedSparseProblem`.
//! - lifecycle modules can emit/build/register/resolve a generated crate.
//! - once a linked runtime backend is registered in-process, the exact same
//!   solver bundle shape becomes callable through compiled callbacks.
//!
//! The production-oriented branch in this file is the sparse faer backend
//! (`SparseColMat` / `Col`). Dense and older sparse matrix backends are kept
//! for compatibility, validation, and gradual migration of the wider BVP
//! stack.
//!
//! ExprLegacy differentiation and bandwidth discovery now live in
//! `legacy_symbolic`; the discretization/generation orchestrator remains here
//! as the next migration boundary. Public `Jacobian` methods are retained as
//! stable adapters so downstream code does not have to know about that split.
//!
//! Historical reference below:
//! the remaining long-form notes describe the original symbolic BVP module
//! layout and older backends. They are kept as reference only; the concise
//! summary above reflects the current pipeline.
//!
//! ## Main Purpose
//! This module provides high-performance symbolic computation and automatic differentiation
//! capabilities specifically designed for solving Boundary Value Problems (BVPs) in differential
//! equations. It bridges symbolic mathematics with numerical computation by:
//! - Converting continuous ODEs into discretized algebraic systems
//! - Computing analytical Jacobians for Newton-Raphson and other iterative methods
//! - Providing multiple sparse and dense matrix backends for optimal performance
//! - Supporting parallel computation throughout the symbolic-to-numerical pipeline
//!
//! ## Main Structures and Methods
//!
//! ### `Jacobian` Struct
//! The core structure that manages the entire symbolic-to-numerical transformation pipeline:
//! - **Fields**:
//!   - `vector_of_functions`: Symbolic expressions representing the discretized BVP system
//!   - `symbolic_jacobian`: 2D matrix of symbolic partial derivatives
//!   - `jac_function`: Compiled numerical Jacobian function (trait object)
//!   - `residiual_function`: Compiled numerical residual function (trait object)
//!   - `bandwidth`: Optional sparse matrix bandwidth for banded systems
//!   - `bounds`/`rel_tolerance_vec`: Constraint and tolerance vectors for variables
//!
//! ### Key Methods
//!
//! #### Discretization
//! - `discretization_system_BVP_par()`: Converts continuous BVP to discrete algebraic system
//!   with parallel boundary condition processing and variable tracking
//!
//! #### Jacobian Computation
//! - `calc_jacobian_parallel_smart()`: Parallel symbolic differentiation with sparsity optimization
//! - `calc_jacobian_parallel()`: Standard parallel symbolic differentiation
//!
//! #### Function Compilation (Multiple Backends)
//! - **Dense (nalgebra)**: `lambdify_jacobian_DMatrix_par()`, `lambdify_residual_DVector()`
//! - **Sparse (faer)**: `lambdify_jacobian_SparseColMat_parallel2()`, `lambdify_residual_Col_parallel2()`
//! - **Sparse (sprs)**: `lambdify_jacobian_CsMat()`, `lambdify_residual_CsVec()`
//! - Historical `Sparse_2`/nalgebra `CsMatrix` BVP runtime was removed because
//!   its solver implementation was incomplete.
//!
//! #### Utility Functions
//! - `find_bandwidths()`: Automatic sparse matrix bandwidth detection
//! - `process_bounds_and_tolerances()`: Efficient constraint processing with string optimization
//!
//! ## Interesting Tips and Non-Obvious Code Features
//!
//! ### Performance Optimizations
//! 1. **Parallel Outer Loop Pattern**: Functions like `lambdify_jacobian_SparseColMat_parallel2()`
//!    use `(0..n).into_par_iter().flat_map()` instead of nested loops for better load balancing
//!
//! 2. **Pre-compilation Strategy**: Functions are compiled once during setup and stored as
//!    `Box<dyn Fn + Send + Sync>` to avoid repeated lambdification during evaluation
//!
//! 3. **Smart Sparsity Detection**: `calc_jacobian_parallel_smart()` only computes derivatives
//!    for variables actually present in each equation, dramatically reducing computation
//!
//! 4. **String Processing Optimization**: `process_bounds_and_tolerances()` uses `rfind('_')`
//!    instead of regex for 10-100x faster variable name processing
//!
//! 5. **Boundary Condition Caching**: `discretization_system_BVP_par()` pre-computes HashSets
//!    and HashMaps for O(1) boundary condition lookups instead of O(n²) nested loops
//!
//! ### Thread Safety Patterns
//! 1. **Mutex-Protected Collections**: Parallel triplet collection uses `Mutex<Vec<Triplet>>`
//!    for thread-safe sparse matrix assembly
//!
//! 2. **Lifetime Management**: Functions convert `Vec<&str>` to `Vec<String>` early to avoid
//!    lifetime issues in `'static` closures: `variable_str.iter().map(|s| s.to_string()).collect()`
//!
//! 3. **Send + Sync Bounds**: Thread-safe lambdification uses `Box<dyn Fn + Send + Sync>`
//!    instead of regular `Box<dyn Fn>` for parallel evaluation
//!
//! ### Memory Management Tricks
//! 1. **Pre-allocation**: Vectors are pre-allocated with known capacity to avoid reallocations:
//!    `Vec::with_capacity(len)`
//!
//! 2. **Zero-Copy Operations**: Uses `ColRef::from_slice().to_owned()` for efficient faer
//!    matrix construction without intermediate copies
//!
//! 3. **Selective Cloning**: Only clones `Expr` objects when necessary due to their complex
//!    recursive structure - uses references and moves where possible
//!
//! ### Discretization Schemes
//! - **Forward Euler**: `"forward"` - First-order explicit scheme
//! - **Trapezoidal**: `"trapezoid"` - Second-order implicit scheme with better stability
//!
//! ### Matrix Backend Strategy
//! The module supports several matrix backends, each optimized for different scenarios:
//! - **Dense (nalgebra)**: Best for small, dense systems
//! - **Sparse (faer)**: Modern, high-performance sparse matrices with excellent parallel support
//! - **Sparse (sprs)**: Mature Rust sparse matrix library
//! - **Banded**: Lapack-style banded storage/solve path for narrow-band BVP Jacobians
//!
//! This multi-backend approach allows users to choose the optimal performance characteristics
//! for their specific problem size and sparsity pattern.

#![allow(non_camel_case_types)]
use crate::global::THRESHOLD as T;
use crate::numerical::BVP_Damp::BVP_traits::{
    BandedMatrixType, Fun, FunEnum, Jac, JacEnum, MatrixType, VectorType, Vectors_type_casting,
    convert_to_fun, convert_to_jac, dense_slice_or_copy,
};
use crate::numerical::BVP_Damp::BVP_utils::{elapsed_time, record_callback_stage_time};
use crate::somelinalg::banded::{
    BandedError, LinearSolverConfig, NodeMajorLayout, banded_assembly::BandedAssembly,
};
use crate::symbolic::View::bvp::DiscretizedBvpAtomSystem;
use crate::symbolic::View::bvp::discretization_system_bvp_par_atom;
use crate::symbolic::View::bvp_codegen::{
    AtomBvpCodegenPrepBreakdown, PreparedSparseAtomBvpCodegen,
    prepare_sparse_bvp_codegen_from_discretized_system_with_breakdown,
};
use crate::symbolic::View::conversions::atom_to_expr;
use crate::symbolic::View::jacobian::PreparedSparseAtomSystem;
use crate::symbolic::View::jacobian::SparseAtomJacobianEntry;
use crate::symbolic::bvp::aot_adapters::{
    AtomViewAotAdapter, BvpAotAdapter, BvpAotAdapterError, ExprLegacyAotAdapter,
};
use crate::symbolic::bvp::aot_telemetry::{
    BvpAotColdStage, BvpAotLifecycleEvent, BvpAotTelemetryMode,
};
use crate::symbolic::bvp::atom_aot::{
    AtomAotMatrixLayout, AtomAotPreparedPlan, AtomAotRuntimeError,
};
use crate::symbolic::bvp::atom_lambdify;
use crate::symbolic::bvp::direct::{
    BandedJacobianChunking, BandedLambdifyConfig, BandedStructurePlan, DirectBandedJacobianRuntime,
    DirectBandedProblem,
};
use crate::symbolic::bvp::legacy_lambdify;
use crate::symbolic::bvp::legacy_lambdify::{
    LegacyLambdifyTelemetry, LegacyLambdifyTelemetrySnapshot,
};
use crate::symbolic::bvp::parameter_binding::BvpParameterBindingHandle;
use crate::symbolic::bvp::telemetry::{
    BvpDirectJacobianTelemetry, BvpDirectJacobianTelemetrySnapshot, BvpGenerationTelemetrySnapshot,
    BvpLambdifyExecutionPolicy, BvpLambdifyTelemetryMode, BvpSymbolicPreparationTelemetrySnapshot,
};
use crate::symbolic::codegen::CodegenIR::{AtomOptimizationProfile, CodegenModule};
use crate::symbolic::codegen::codegen_aot_driver::{AotCodegenBackend, GeneratedAotArtifact};
use crate::symbolic::codegen::codegen_aot_lifecycle::AotLifecycleError;
use crate::symbolic::codegen::codegen_aot_resolution::{AotResolver, ResolvedAotArtifact};
use crate::symbolic::codegen::codegen_aot_runtime_link::{
    LinkedAotCallbackError, LinkedJacobianLayout, invoke_linked_callback,
    try_resolve_linked_sparse_backend,
};
use crate::symbolic::codegen::codegen_backend_selection::{
    BackendSelectionPolicy, SelectedBackendKind, select_backend, select_backend_by_manifest,
};
use crate::symbolic::codegen::codegen_manifest::PreparedProblemManifest;
use crate::symbolic::codegen::codegen_orchestrator::{
    ParallelExecutorConfig, ParallelFallbackPolicy, SparseAutoParallelPlan,
    auto_parallel_recommendation, recommended_sparse_auto_parallel_plan,
};
use crate::symbolic::codegen::codegen_provider_api::{
    BackendKind, MatrixBackend, PreparedBandedProblem, PreparedProblem, PreparedSparseProblem,
};
use crate::symbolic::codegen::codegen_runtime_api::{
    BandedCompactJacobianStructure, BandedJacobianRuntimePlan, BandedJacobianStructure,
    ResidualChunkPlan, ResidualChunkingStrategy, ResidualRuntimePlan, SparseJacobianRuntimePlan,
    SparseJacobianStructure, SparseJacobianValuesChunkPlan,
    recommended_banded_chunking_for_parallelism, recommended_residual_chunking_for_parallelism,
    recommended_row_chunking_for_parallelism,
};
use crate::symbolic::codegen::codegen_tasks::{
    BandedChunkingStrategy, BandedExprEntry, CodegenOutputLayout, CodegenTaskKind, CodegenTaskPlan,
    PlannedOutput, SparseChunkingStrategy, SparseExprEntry,
};
use crate::symbolic::codegen::rust_backend::codegen_aot_crate::GeneratedAotCrate;
use crate::symbolic::symbolic_engine::Expr;
use faer::col::ColRef;

use faer::sparse::{SparseColMat, Triplet};
use log::{debug, error, info};
use nalgebra::sparse::CsMatrix;
use nalgebra::{DMatrix, DVector};

use rayon::prelude::*;

use sprs::{CsMat, CsVec};
use std::collections::{HashMap, HashSet};
use std::fmt;
use std::panic::{AssertUnwindSafe, catch_unwind};
use std::time::{Duration, Instant};
use tabled::{builder::Builder, settings::Style};

fn normalize_timer_percent(value: &mut f64, total: f64) {
    if total <= f64::EPSILON {
        *value = 0.0;
    } else {
        *value /= total / 100.0;
    }
}

fn flatten_runtime_args(parameter_values: Option<&[f64]>, unknowns: &[f64]) -> Vec<f64> {
    let mut flat_args =
        Vec::with_capacity(parameter_values.map_or(0, |params| params.len()) + unknowns.len());
    if let Some(params) = parameter_values {
        flat_args.extend_from_slice(params);
    }
    flat_args.extend_from_slice(unknowns);
    flat_args
}

fn panic_message(panic: Box<dyn std::any::Any + Send>) -> String {
    if let Some(message) = panic.downcast_ref::<&str>() {
        (*message).to_string()
    } else if let Some(message) = panic.downcast_ref::<String>() {
        message.clone()
    } else {
        "callback panicked with a non-string payload".to_string()
    }
}

fn map_linked_callback_error(
    stage: &'static str,
    error: LinkedAotCallbackError,
) -> BvpBackendIntegrationError {
    match error {
        LinkedAotCallbackError::OutputLength {
            expected, actual, ..
        } => BvpBackendIntegrationError::CallbackShapeMismatch {
            stage: stage.to_string(),
            expected_rows: expected,
            expected_columns: 1,
            actual_rows: actual,
            actual_columns: 1,
        },
        LinkedAotCallbackError::NonFiniteOutput { index, .. }
        | LinkedAotCallbackError::NonFiniteInput { index, .. } => {
            BvpBackendIntegrationError::NonFiniteCallbackValue {
                stage: stage.to_string(),
                index,
            }
        }
        other => BvpBackendIntegrationError::CallbackExecutionFailed {
            stage: stage.to_string(),
            message: other.to_string(),
        },
    }
}

#[inline]
fn record_aot_stage_ms(plan: &AtomAotPreparedPlan, stage: BvpAotColdStage, milliseconds: f64) {
    if milliseconds.is_finite() && milliseconds > 0.0 {
        plan.record_cold_stage_duration(stage, Duration::from_secs_f64(milliseconds / 1_000.0));
    }
}

fn panic_symbolic_numeric_backend_requested() -> ! {
    panic!(
        "BVP symbolic pipeline cannot materialize BackendSelectionPolicy::NumericOnly; \
         use NRBVP::set_numeric_rhs/with_numeric_rhs so the solver can build the pure numeric \
         finite-difference discretization path, or choose LambdifyOnly/AotOnly/PreferAotThenLambdify"
    )
}

/// High-level backend family used by the BVP symbolic pipeline.
///
/// This enum intentionally separates "how the numerical evaluator appears"
/// from the lower-level matrix storage choice.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum BvpBackendKind {
    /// Reserved marker for externally supplied numeric closures.
    ///
    /// The symbolic BVP pipeline must not synthesize this by lambdifying
    /// symbolic equations. Pure numeric BVP solving is owned by
    /// `BVP_Damp::numeric_discretization`, where the source of truth is a Rust
    /// closure and the Jacobian is built by finite differences.
    Numeric,
    Lambdify,
    Aot,
}

/// Matrix/vector backend used by the BVP Jacobian/residual pipeline.
///
/// `FaerSparseCol` is the main production-oriented branch in this module.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum BvpMatrixBackend {
    Dense,
    SprsCsMat,
    FaerSparseCol,
    Banded,
}

impl BvpMatrixBackend {
    pub const fn legacy_method(self) -> &'static str {
        match self {
            Self::Dense => "Dense",
            Self::SprsCsMat => "Sparse_1",
            Self::FaerSparseCol => "Sparse",
            Self::Banded => "Banded",
        }
    }

    pub fn from_legacy_method(method: &str) -> Option<Self> {
        match method {
            "Dense" => Some(Self::Dense),
            "Sparse_1" => Some(Self::SprsCsMat),
            "Sparse" => Some(Self::FaerSparseCol),
            "Banded" => Some(Self::Banded),
            _ => None,
        }
    }

    pub const fn from_provider_matrix_backend(matrix_backend: MatrixBackend) -> Self {
        match matrix_backend {
            MatrixBackend::Dense => Self::Dense,
            MatrixBackend::Banded => Self::Banded,
            MatrixBackend::SparseCol => Self::FaerSparseCol,
            MatrixBackend::CsMat => Self::SprsCsMat,
            MatrixBackend::CsMatrix => Self::FaerSparseCol,
            MatrixBackend::ValuesOnly => Self::FaerSparseCol,
        }
    }
}

/// User-facing backend selection for BVP symbolic compilation.
///
/// For now only the `Lambdify` branch is executed inside this module, while
/// `Numeric` and `Aot` are reserved for the staged integration layer above it.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct BvpBackendConfig {
    pub backend_kind: BvpBackendKind,
    pub matrix_backend: BvpMatrixBackend,
}

impl BvpBackendConfig {
    pub const fn new(backend_kind: BvpBackendKind, matrix_backend: BvpMatrixBackend) -> Self {
        Self {
            backend_kind,
            matrix_backend,
        }
    }

    pub const fn lambdify(matrix_backend: BvpMatrixBackend) -> Self {
        Self::new(BvpBackendKind::Lambdify, matrix_backend)
    }
}

impl Default for BvpBackendConfig {
    fn default() -> Self {
        Self::lambdify(BvpMatrixBackend::FaerSparseCol)
    }
}

pub(crate) use crate::symbolic::bvp::legacy_symbolic::bandwidth as bvp_symbolic_bandwidth;
pub(crate) use crate::symbolic::bvp::legacy_symbolic::jacobian_smart as bvp_symbolic_jacobian_smart;
pub(crate) use crate::symbolic::bvp::legacy_symbolic::parse_variables as bvp_symbolic_parse_variables;
pub(crate) use crate::symbolic::bvp::legacy_symbolic::variables_for_functions as bvp_symbolic_variables_for_functions;

/// Symbolic assembly backend used before lambdify/AOT runtime selection.
///
/// This controls how the discretized residual equations are assembled:
/// - `ExprLegacy` keeps the historical boxed-expression path,
/// - `AtomView` performs residual assembly on the packed View path and only
///   materializes back to `Expr` at the integration boundary.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum BvpSymbolicAssemblyBackend {
    ExprLegacy,
    AtomView,
}

impl Default for BvpSymbolicAssemblyBackend {
    fn default() -> Self {
        Self::ExprLegacy
    }
}
/// Core structure for BVP symbolic-to-numerical transformation pipeline.
///
/// This struct represents a Jacobian for BVPs, which is a matrix of partial derivatives of a vector function.
/// It contains the symbolic and numerical representations of the jacobian, as well as the functions used to evaluate it.
/// Manages the complete workflow from symbolic BVP discretization to high-performance numerical evaluation.
///
/// # Performance Features
/// - Supports multiple matrix backends (dense/sparse) for optimal performance
/// - Parallel symbolic differentiation and function compilation
/// - Smart sparsity detection and bandwidth optimization
/// - Thread-safe evaluation with Send+Sync closures
//#[derive(Clone)]
pub struct Jacobian {
    /// Vector of symbolic functions/expressions representing the discretized BVP system.
    /// After discretization, contains the residual equations F(y) = 0 that need to be solved.
    pub vector_of_functions: Vec<Expr>,

    /// Vector of lambdified functions (symbolic functions converted to rust functions).
    /// Legacy field - modern implementations use trait objects in `jac_function` and `residiual_function`.
    pub lambdified_functions: Vec<Box<dyn Fn(Vec<f64>) -> f64>>,

    /// String identifier for the matrix backend method being used.
    /// Values: "Dense", "Sparse" (faer), "Sparse_1" (sprs), or "Banded".
    /// Historical "Sparse_2" / nalgebra CsMatrix is no longer exposed as a solver backend.
    pub method: String,

    /// Explicit backend configuration used by the modernized BVP entrypoints.
    ///
    /// The legacy `method` string is still kept for compatibility with the
    /// wider BVP stack, but new code should prefer this field.
    pub backend_config: BvpBackendConfig,

    /// Native linear solver configuration used by the banded runtime path.
    pub banded_linear_solver_config: LinearSolverConfig,

    /// Runtime compilation/evaluation controls for the banded lambdify path.
    ///
    /// This carries Jacobian chunking strategy and structural threshold
    /// settings used when lowering symbolic sparse Jacobian entries into
    /// `BandedAssembly`.
    pub banded_lambdify_config: BandedLambdifyConfig,

    /// Symbolic assembly path used for residual discretization before Jacobian
    /// generation begins.
    pub symbolic_assembly_backend: BvpSymbolicAssemblyBackend,

    /// Vector of symbolic variables representing the unknowns in the BVP system.
    /// Contains Expr::Var objects for each discretized variable (e.g., y_0, y_1, ..., y_n)
    pub vector_of_variables: Vec<Expr>,

    /// String representations of the variable names for efficient lookups.
    /// Used during lambdification and variable processing (e.g., ["y_0", "y_1", "z_0", "z_1"])
    pub variable_string: Vec<String>,

    /// Optional symbolic parameter names that affect evaluation but are not
    /// Newton unknowns of the BVP system.
    pub parameters_string: Vec<String>,

    /// Current numeric values for `parameters_string`.
    ///
    /// At runtime the residual/Jacobian evaluators flatten arguments in the
    /// order `[params..., unknowns...]`, while the outer Newton solver still
    /// passes only the unknown vector.
    pub parameter_values: Option<Vec<f64>>,

    /// Shared numeric binding used by prepared Lambdify callbacks. Changing
    /// values can therefore rebind an existing evaluator without symbolic
    /// differentiation or closure recompilation.
    parameter_binding: Option<BvpParameterBindingHandle>,

    /// 2D matrix of symbolic partial derivatives ∂F_i/∂x_j.
    /// Core of the analytical Jacobian - computed once symbolically, then compiled to numerical functions
    pub symbolic_jacobian: Vec<Vec<Expr>>,

    /// Sparse symbolic Jacobian cache containing only explicit non-zero entries.
    ///
    /// This cache is maintained alongside the legacy dense matrix so the BVP
    /// path can migrate gradually toward sparse-first symbolic storage.
    pub symbolic_jacobian_sparse: Vec<(usize, usize, Expr)>,

    /// AtomView sparse Jacobian kept in packed form for Atom-native Lambdify.
    /// The Expr sparse cache remains alongside it solely for compatibility
    /// consumers and the existing generated-backend bridge.
    pub atom_symbolic_jacobian_sparse: Option<Vec<SparseAtomJacobianEntry>>,

    /// Optional single Jacobian element evaluator function.
    /// Legacy field for element-wise Jacobian evaluation - modern code uses `jac_function`
    pub lambdified_jac_element: Option<Box<dyn Fn(f64, usize, usize) -> f64>>,

    /// Compiled numerical Jacobian function as trait object.
    /// Supports multiple active backends through JacEnum variants (Dense, Sparse_1, Sparse_3)
    /// plus the separate banded runtime path.
    pub jac_function: Option<Box<dyn Jac>>,

    /// Compiled numerical residual function as trait object.
    /// Evaluates F(x,y) for the discretized BVP system through FunEnum variants
    pub residiual_function: Box<dyn Fun>,

    /// Optional variable bounds as (min, max) pairs for each discretized variable.
    /// Used by constrained solvers to enforce physical constraints during Newton iteration
    pub bounds: Option<Vec<(f64, f64)>>,

    /// Optional relative tolerance vector for adaptive error control.
    /// Per-variable tolerances for adaptive mesh refinement and convergence criteria
    pub rel_tolerance_vec: Option<Vec<f64>>,

    /// Optional sparse matrix bandwidth (kl, ku) for banded Jacobians.
    /// kl = number of subdiagonals, ku = number of superdiagonals. Enables banded matrix optimizations
    pub bandwidth: Option<(usize, usize)>,

    /// Variable tracking for smart sparsity detection.
    /// variables_for_all_disrete[i] contains variables actually used in equation i,
    /// enabling zero-derivative skipping in calc_jacobian_parallel_smart()
    pub variables_for_all_disrete: Vec<Vec<String>>,

    /// Boundary condition positions and values for full solution reconstruction.
    /// Contains (position_in_full_vector, boundary_value) pairs to insert BC values
    /// back into the solution vector of unknowns to create the complete solution.
    pub BC_pos_n_values: Vec<(usize, usize, f64)>,

    /// Raw stage timings from the last `generate_BVP*` symbolic build.
    ///
    /// Values are kept in wall-clock seconds before percentage normalization so
    /// diagnostics can compare backend stages directly.
    pub last_generate_timer_snapshot: Option<HashMap<String, f64>>,

    /// Detailed timings from the last symbolic-Jacobian construction pass.
    ///
    /// The generated solver handoff promotes these values to runtime
    /// diagnostics so large BVP runs can distinguish differentiation work
    /// from legacy dense-cache materialization.
    pub last_symbolic_jacobian_timer_snapshot: Option<HashMap<String, f64>>,

    /// Typed stage timings for the most recent symbolic Jacobian build.
    ///
    /// The string-keyed snapshot above remains for compatibility; this field
    /// is the preferred diagnostics surface for new Lambdify stories.
    pub last_symbolic_jacobian_telemetry_snapshot: Option<BvpSymbolicPreparationTelemetrySnapshot>,

    /// Detailed timings from the last Lambdify callback compilation pass.
    ///
    /// In generated handoff diagnostics this distinguishes expensive closure
    /// compilation from backend selection and callback installation.
    pub last_lambdify_binding_timer_snapshot: Option<HashMap<String, f64>>,

    /// Optional packed View-side discretized system kept for Atom-native
    /// Lambdify and AOT lowering. Compatibility fields are still materialized
    /// as `Expr`, but native Banded Lambdify consumes this snapshot directly.
    pub atom_discretized_system: Option<DiscretizedBvpAtomSystem>,

    /// Telemetry for the active direct Banded Jacobian callback.
    ///
    /// This is kept as a shared handle rather than a string map so the direct
    /// no-Mutex callback can report runtime work without coupling its hot path
    /// to the legacy diagnostics representation.
    pub direct_banded_jacobian_telemetry: Option<BvpDirectJacobianTelemetry>,

    /// Typed runtime telemetry for the currently installed legacy Expr-based
    /// dense/faer Lambdify callbacks. AtomView Lambdify has a separate stream;
    /// the direct Banded callback has its own telemetry as well.
    pub legacy_lambdify_telemetry: Option<LegacyLambdifyTelemetry>,

    /// Typed runtime telemetry for the currently installed AtomView
    /// Dense/faer Lambdify callbacks.
    pub atom_lambdify_telemetry: Option<LegacyLambdifyTelemetry>,

    /// Runtime telemetry collection mode for both Lambdify callback streams.
    /// `Off` is deliberately the zero-overhead solver default.
    pub lambdify_telemetry_mode: BvpLambdifyTelemetryMode,

    /// Typed telemetry policy carried by the AtomView AOT preparation route.
    ///
    /// This is separate from callback telemetry because compiler lifecycle
    /// stages and warm generated callbacks have different cost semantics.
    pub aot_telemetry_mode: BvpAotTelemetryMode,

    /// Runtime execution policy for pure Lambdify callbacks. This is kept
    /// separate from AOT execution policy and does not affect preparation.
    pub lambdify_execution_policy: BvpLambdifyExecutionPolicy,
}

/// Owned sparse AOT-ready BVP problem produced by the legacy BVP symbolic
/// builder.
///
/// This bridge object keeps the selected symbolic payload in owned form so it
/// can later expose a borrow-based [`PreparedSparseProblem`] without leaking
/// temporary vectors. ExprLegacy stores residual/Jacobian expressions in the
/// compatibility fields; AtomView stores the native codegen payload and leaves
/// those fields empty.
#[derive(Debug, Clone)]
pub struct BvpPreparedSparseAotProblem {
    pub residual_fn_name: String,
    pub jacobian_fn_name: String,
    pub variable_names: Vec<String>,
    pub param_names: Vec<String>,
    pub residuals: Vec<Expr>,
    pub sparse_entries: Vec<(usize, usize, Expr)>,
    pub shape: (usize, usize),
    pub residual_strategy: ResidualChunkingStrategy,
    pub jacobian_strategy: SparseChunkingStrategy,
    pub bandwidth: Option<(usize, usize)>,
    pub symbolic_assembly_backend: BvpSymbolicAssemblyBackend,
    pub atom_discretized_system: Option<DiscretizedBvpAtomSystem>,

    /// Prepared AtomView codegen payload for the native AOT route.
    ///
    /// This is deliberately separate from the public Expr compatibility
    /// fields above. AtomView AOT callers keep this payload and leave the
    /// compatibility vectors empty, so no Expr round-trip is required before
    /// code generation or manifest construction.
    pub atom_codegen: Option<PreparedSparseAtomBvpCodegen>,

    /// Preparation timings produced together with the native AtomView
    /// codegen payload. Keeping this beside the owned payload prevents the
    /// artifact adapter from replacing real preparation metrics with a
    /// default zero breakdown.
    pub atom_codegen_prep_breakdown: Option<AtomBvpCodegenPrepBreakdown>,

    /// Validated AtomView AOT plan owned by this prepared bridge.
    ///
    /// Keeping this beside the emission payload prevents diagnostics and
    /// lifecycle code from rebuilding and revalidating the symbolic plan on
    /// every query. `None` is intentional for the retained ExprLegacy route.
    pub atom_aot_plan: Option<AtomAotPreparedPlan>,
}

/// Identifies which symbolic payload owns the AOT preparation route.
///
/// The marker is intentionally kept on the BVP bridge rather than inferred
/// from the selected compiler. Rust/C/Zig are compiler choices; this enum
/// answers the separate question of whether AOT preparation came from the
/// native AtomView graph or from the retained ExprLegacy adapter.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum BvpAotPreparationRoute {
    ExprLegacy,
    AtomViewNative,
}

impl BvpPreparedSparseAotProblem {
    /// Returns the symbolic preparation route used by this owned bridge.
    pub fn aot_preparation_route(&self) -> BvpAotPreparationRoute {
        if self.atom_codegen.is_some() {
            BvpAotPreparationRoute::AtomViewNative
        } else {
            BvpAotPreparationRoute::ExprLegacy
        }
    }

    /// Returns the typed telemetry mode carried by an AtomView-native plan.
    /// ExprLegacy compatibility payloads intentionally report `Off` here.
    pub fn aot_telemetry_mode(&self) -> BvpAotTelemetryMode {
        self.atom_codegen.as_ref().map_or(
            BvpAotTelemetryMode::Off,
            PreparedSparseAtomBvpCodegen::aot_telemetry_mode,
        )
    }

    /// Returns the current typed AOT preparation snapshot, if AtomView owns
    /// the payload. The snapshot is deliberately absent for ExprLegacy.
    pub fn aot_telemetry_snapshot(
        &self,
    ) -> Option<crate::symbolic::bvp::aot_telemetry::BvpAotTelemetrySnapshot> {
        self.atom_aot_plan
            .as_ref()
            .map(AtomAotPreparedPlan::telemetry_snapshot)
    }

    /// Returns a shared typed telemetry handle for the native AOT runtime.
    ///
    /// Callback closures and solver diagnostics intentionally share this
    /// handle so the post-solve snapshot includes warm residual/Jacobian
    /// calls, not only cold preparation stages.
    pub fn aot_telemetry_handle(
        &self,
    ) -> Option<crate::symbolic::bvp::aot_telemetry::BvpAotTelemetry> {
        self.atom_aot_plan
            .as_ref()
            .map(|plan| plan.telemetry().clone())
    }

    /// Records a measured external AOT lifecycle stage on the native
    /// AtomView plan. ExprLegacy deliberately remains a compatibility route
    /// and therefore has no native typed lifecycle stream to update.
    pub(crate) fn record_aot_cold_stage_duration(&self, stage: BvpAotColdStage, elapsed: Duration) {
        if let Some(plan) = self.atom_aot_plan.as_ref() {
            plan.record_cold_stage_duration(stage, elapsed);
        }
    }

    /// Returns the validated AtomView AOT plan without rebuilding it.
    pub fn atom_aot_plan(&self) -> Option<&AtomAotPreparedPlan> {
        self.atom_aot_plan.as_ref()
    }

    /// Exposes the prepared payload through an explicit typed route adapter.
    ///
    /// This is the migration boundary for new lifecycle code. It prevents
    /// callers from treating an optional Expr/Atom field combination as a
    /// valid route and gives incomplete AtomView bridges a typed error.
    pub(crate) fn try_aot_adapter(&self) -> Result<BvpAotAdapter<'_>, BvpAotAdapterError> {
        match (&self.atom_codegen, &self.atom_aot_plan) {
            (Some(codegen), Some(plan)) => Ok(BvpAotAdapter::AtomView(AtomViewAotAdapter {
                codegen,
                plan,
            })),
            (Some(_), None) => Err(BvpAotAdapterError::AtomViewPlanMissing),
            (None, Some(_)) => Err(BvpAotAdapterError::AtomViewCodegenMissing),
            (None, None) => Ok(BvpAotAdapter::ExprLegacy(ExprLegacyAotAdapter {
                residuals: &self.residuals,
                sparse_entries: &self.sparse_entries,
                shape: self.shape,
            })),
        }
    }
}

/// Fine-grained timing breakdown for BVP generated AOT crate assembly.
///
/// This is primarily used to diagnose where time is spent between the symbolic
/// sparse problem and the final emitted Rust source:
/// - building/reusing the sparse Jacobian representation,
/// - lowering it into a `CodegenModule`,
/// - and rendering the final source string.
#[derive(Clone, Debug, Default)]
pub struct BvpGeneratedAotCrateBreakdown {
    pub atom_input_abi_prepare_ms: f64,
    pub jacobian_prepare_ms: f64,
    pub atom_sparse_lookup_prepare_ms: f64,
    pub atom_sparse_jacobian_build_ms: f64,
    pub atom_finalize_codegen_plan_ms: f64,
    pub atom_sparse_nnz: usize,
    pub atom_residual_view_collect_ms: f64,
    pub atom_residual_lower_many_ms: f64,
    pub atom_residual_peephole_ms: f64,
    pub atom_residual_reuse_temps_ms: f64,
    pub atom_residual_push_ms: f64,
    pub atom_sparse_view_collect_ms: f64,
    pub atom_sparse_lower_many_ms: f64,
    pub atom_sparse_peephole_ms: f64,
    pub atom_sparse_reuse_temps_ms: f64,
    pub atom_sparse_push_ms: f64,
    pub prepared_module_init_ms: f64,
    pub prepared_residual_blocks_ms: f64,
    pub prepared_jacobian_blocks_ms: f64,
    pub module_build_ms: f64,
    pub source_probe_emit_ms: f64,
    pub language_source_emit_ms: f64,
    pub c_header_emit_ms: f64,
    pub artifact_packaging_ms: f64,
    pub source_emit_ms: f64,
    pub source_kb: f64,
    pub module_blocks: usize,
    pub total_block_instructions: usize,
    pub total_block_temps: usize,
    pub max_block_instructions: usize,
    pub total_block_outputs: usize,
}

/// Owned BVP-side backend selection result for the sparse main path.
///
/// This keeps the selected branch together with the owned sparse prepared
/// problem bridge, allowing the BVP module to consult lifecycle metadata
/// without leaking borrow-based `PreparedProblem` values across layers.
#[derive(Debug, Clone)]
pub struct BvpSelectedSparseBackend {
    pub prepared_problem: BvpPreparedSparseAotProblem,
    pub aot_preparation_route: BvpAotPreparationRoute,
    pub requested_backend: BackendKind,
    pub effective_backend: SelectedBackendKind,
    pub matrix_backend: MatrixBackend,
    pub aot_resolution: Option<ResolvedAotArtifact>,
}

impl BvpSelectedSparseBackend {
    /// Returns `true` when the selected path points to a compiled AOT artifact.
    pub fn is_compiled_aot(&self) -> bool {
        self.effective_backend == SelectedBackendKind::AotCompiled
    }

    /// Returns the manifest-derived problem key for the currently selected
    /// matrix backend.
    pub fn problem_key(&self) -> String {
        self.prepared_problem
            .problem_key_for_matrix_backend(self.matrix_backend)
    }

    /// Returns the symbolic preparation route used by the selected payload.
    pub fn preparation_route(&self) -> BvpAotPreparationRoute {
        self.aot_preparation_route
    }
}

/// BVP-side sparse backend execution plan.
///
/// This is the first solver-adjacent use layer on top of backend selection:
/// - lambdify variants mean `jac_function` / `residiual_function` are ready,
/// - AOT variants mean lifecycle metadata is resolved and the caller can
///   continue into the compiled-backend branch,
/// - numeric remains a reserved integration point.
#[derive(Debug, Clone)]
pub enum BvpSparseExecutionPlan {
    NumericRequested(BvpSelectedSparseBackend),
    LambdifyReady(BvpSelectedSparseBackend),
    AotCompiled(BvpSelectedSparseBackend),
    AotRegisteredButNotBuilt(BvpSelectedSparseBackend),
    AotMissing(BvpSelectedSparseBackend),
}

impl BvpSparseExecutionPlan {
    pub fn selected(&self) -> &BvpSelectedSparseBackend {
        match self {
            Self::NumericRequested(selected)
            | Self::LambdifyReady(selected)
            | Self::AotCompiled(selected)
            | Self::AotRegisteredButNotBuilt(selected)
            | Self::AotMissing(selected) => selected,
        }
    }

    pub fn selected_mut(&mut self) -> &mut BvpSelectedSparseBackend {
        match self {
            Self::NumericRequested(selected)
            | Self::LambdifyReady(selected)
            | Self::AotCompiled(selected)
            | Self::AotRegisteredButNotBuilt(selected)
            | Self::AotMissing(selected) => selected,
        }
    }
}

/// Transitional BVP-side sparse provider bridge for the main faer-oriented path.
///
/// This object is intentionally thin:
/// - for `LambdifyReady`, it can evaluate residuals and sparse Jacobian values
///   through the already prepared legacy callbacks;
/// - for `Aot*`, it exposes the resolved metadata and sparse structure needed
///   by the next lifecycle stage that will attach compiled callbacks.
///
/// It lets the solver-facing integration use one object shape before the final
/// in-process compiled AOT runtime hook is introduced.
pub struct BvpSparseSolverProvider<'a> {
    jacobian: &'a mut Jacobian,
    execution: BvpSparseExecutionPlan,
    structure: SparseJacobianStructure,
    parallel_config: Option<ParallelExecutorConfig>,
}

impl<'a> BvpSparseSolverProvider<'a> {
    /// Returns the already-selected sparse execution plan.
    ///
    /// Outer integration layers use this to inspect whether the current path is
    /// lambdify-ready, compiled AOT, or still missing a built artifact.
    pub fn execution_plan(&self) -> &BvpSparseExecutionPlan {
        &self.execution
    }

    /// Effective backend after policy + resolver selection.
    pub fn effective_backend(&self) -> SelectedBackendKind {
        self.execution.selected().effective_backend
    }

    /// Backend family originally requested by the upper configuration layer.
    pub fn requested_backend(&self) -> BackendKind {
        self.execution.selected().requested_backend
    }

    /// Returns `true` when residual/Jacobian callbacks can be executed right now.
    ///
    /// That is true for:
    /// - `LambdifyReady`
    /// - `AotCompiled` with a registered linked runtime backend
    pub fn is_runtime_callable(&self) -> bool {
        matches!(self.execution, BvpSparseExecutionPlan::LambdifyReady(_))
            || matches!(self.execution, BvpSparseExecutionPlan::AotCompiled(_))
                && try_resolve_linked_sparse_backend(&self.execution.selected().problem_key())
                    .ok()
                    .flatten()
                    .is_some()
    }

    /// Returns resolved lifecycle metadata for compiled AOT artifacts, when present.
    pub fn resolved_aot_artifact(&self) -> Option<&ResolvedAotArtifact> {
        self.execution.selected().aot_resolution.as_ref()
    }

    /// Number of residual outputs in the current sparse BVP system.
    pub fn residual_len(&self) -> usize {
        self.execution.selected().prepared_problem.shape.0
    }

    /// Sparse Jacobian shape `(rows, cols)` for the current prepared problem.
    pub fn jacobian_shape(&self) -> (usize, usize) {
        self.execution.selected().prepared_problem.shape
    }

    /// Sparse structural pattern shared by lambdify and linked AOT branches.
    pub fn jacobian_structure(&self) -> &SparseJacobianStructure {
        &self.structure
    }

    /// Casts the raw dense solver vector into the matrix/vector backend
    /// expected by the legacy callback layer.
    fn typed_variables_from_args(&self, args: &[f64]) -> Box<dyn VectorType> {
        let dense = DVector::from_column_slice(args);
        let backend = BvpMatrixBackend::from_provider_matrix_backend(
            self.execution.selected().matrix_backend,
        );
        Vectors_type_casting(&dense, backend.legacy_method().to_string())
    }

    fn validate_callback_input(
        &self,
        stage: &str,
        args: &[f64],
    ) -> Result<(), BvpBackendIntegrationError> {
        let expected = self.jacobian_shape().1;
        if args.len() != expected {
            return Err(BvpBackendIntegrationError::CallbackInputLengthMismatch {
                stage: stage.to_string(),
                expected,
                actual: args.len(),
            });
        }
        if let Some(index) = args.iter().position(|value| !value.is_finite()) {
            return Err(BvpBackendIntegrationError::NonFiniteCallbackValue {
                stage: format!("{stage} input"),
                index,
            });
        }
        let expected_parameters = self.execution.selected().prepared_problem.param_names.len();
        if let Some(parameters) = self.jacobian.parameter_values.as_deref() {
            if parameters.len() != expected_parameters {
                return Err(BvpBackendIntegrationError::ParameterCountMismatch {
                    expected: expected_parameters,
                    actual: parameters.len(),
                });
            }
            if let Some(index) = parameters.iter().position(|value| !value.is_finite()) {
                return Err(BvpBackendIntegrationError::NonFiniteCallbackValue {
                    stage: format!("{stage} parameters"),
                    index,
                });
            }
        } else if expected_parameters != 0 {
            return Err(BvpBackendIntegrationError::ParameterCountMismatch {
                expected: expected_parameters,
                actual: 0,
            });
        }
        Ok(())
    }

    #[inline]
    fn record_aot_runtime_error(&self) {
        if let Some(plan) = self.execution.selected().prepared_problem.atom_aot_plan() {
            plan.telemetry().record_error();
        }
    }

    /// Fallible residual callback boundary for the prepared sparse runtime.
    ///
    /// This is the production entrypoint. The historical residual_into
    /// method below remains a panic-on-error compatibility wrapper.
    pub fn try_residual_into(
        &self,
        args: &[f64],
        out: &mut [f64],
    ) -> Result<(), BvpBackendIntegrationError> {
        let input_result = self.validate_callback_input("residual", args);
        if input_result.is_err() {
            self.record_aot_runtime_error();
        }
        input_result?;
        if out.len() != self.residual_len() {
            self.record_aot_runtime_error();
            return Err(BvpBackendIntegrationError::CallbackShapeMismatch {
                stage: "residual".to_string(),
                expected_rows: self.residual_len(),
                expected_columns: 1,
                actual_rows: out.len(),
                actual_columns: 1,
            });
        }

        let callback_result = catch_unwind(AssertUnwindSafe(|| match &self.execution {
            BvpSparseExecutionPlan::LambdifyReady(_) => {
                let typed = self.typed_variables_from_args(args);
                let residual = self.jacobian.residiual_function.call(1.0, &*typed);
                let dense = residual.to_DVectorType();
                if dense.len() != out.len() {
                    return Err(BvpBackendIntegrationError::CallbackShapeMismatch {
                        stage: "residual".to_string(),
                        expected_rows: out.len(),
                        expected_columns: 1,
                        actual_rows: dense.len(),
                        actual_columns: 1,
                    });
                }
                out.copy_from_slice(dense.as_slice());
                Ok(())
            }
            BvpSparseExecutionPlan::AotCompiled(selected) => {
                let linked = try_resolve_linked_sparse_backend(&selected.problem_key())
                    .map_err(|error| {
                        BvpBackendIntegrationError::PipelinePanicked(error.to_string())
                    })?
                    .ok_or_else(
                        || BvpBackendIntegrationError::CompiledAotRuntimeUnavailable {
                            problem_key: selected.problem_key(),
                        },
                    )?;
                let flat_args =
                    flatten_runtime_args(self.jacobian.parameter_values.as_deref(), args);
                let parallel_config = self.parallel_config;
                let chunks = linked_execution_chunk_count(&linked, parallel_config, false);
                let callback = |flat_args: &[f64], callback_out: &mut [f64]| {
                    let callback_result = match parallel_config {
                        Some(config) => try_eval_linked_residual_outputs(
                            &linked,
                            Some(config),
                            flat_args,
                            callback_out,
                        ),
                        None => linked.try_residual_eval(flat_args, callback_out),
                    };
                    callback_result
                        .map_err(|error| map_linked_callback_error("residual", error))
                };
                if let Some(plan) = selected.prepared_problem.atom_aot_plan() {
                    plan.record_worker_batch(
                        if chunks > 1 {
                            rayon::current_num_threads()
                        } else {
                            1
                        },
                        chunks,
                    );
                    plan.try_execute_residual_callback_with_chunks(
                        &flat_args,
                        out,
                        chunks,
                        callback,
                    )
                } else if selected.prepared_problem.symbolic_assembly_backend
                    == BvpSymbolicAssemblyBackend::ExprLegacy
                {
                    callback(&flat_args, out)
                } else {
                    Err(BvpBackendIntegrationError::CompiledAotRuntimeUnavailable {
                        problem_key: selected.problem_key(),
                    })
                }
            }
            _ => Err(BvpBackendIntegrationError::CompiledAotRuntimeUnavailable {
                problem_key: self.execution.selected().problem_key(),
            }),
        }))
        .map_err(
            |panic| BvpBackendIntegrationError::CallbackExecutionFailed {
                stage: "residual".to_string(),
                message: panic_message(panic),
            },
        )
        .and_then(|result| result);
        callback_result?;

        if let Some(index) = out.iter().position(|value| !value.is_finite()) {
            self.record_aot_runtime_error();
            return Err(BvpBackendIntegrationError::NonFiniteCallbackValue {
                stage: "residual".to_string(),
                index,
            });
        }
        Ok(())
    }

    /// Evaluates residual outputs into the caller-provided buffer.
    ///
    /// The provider keeps one solver-facing contract while internally routing
    /// either to legacy lambdified callbacks or to linked compiled AOT
    /// callbacks.
    pub fn residual_into(&self, args: &[f64], out: &mut [f64]) {
        if let Err(error) = self.try_residual_into(args, out) {
            panic!("BvpSparseSolverProvider residual callback failed: {error:?}");
        }
    }

    /// Fallible sparse Jacobian-values callback boundary for the prepared
    /// runtime. Values follow the fixed structural order in
    /// jacobian_structure().
    pub fn try_jacobian_values_into(
        &mut self,
        args: &[f64],
        values_out: &mut [f64],
    ) -> Result<(), BvpBackendIntegrationError> {
        let input_result = self.validate_callback_input("Jacobian", args);
        if input_result.is_err() {
            self.record_aot_runtime_error();
        }
        input_result?;
        if values_out.len() != self.structure.nnz() {
            self.record_aot_runtime_error();
            return Err(BvpBackendIntegrationError::CallbackShapeMismatch {
                stage: "Jacobian".to_string(),
                expected_rows: self.structure.nnz(),
                expected_columns: 1,
                actual_rows: values_out.len(),
                actual_columns: 1,
            });
        }

        let callback_result = catch_unwind(AssertUnwindSafe(|| match &mut self.execution {
            BvpSparseExecutionPlan::LambdifyReady(_) => {
                let typed = self.typed_variables_from_args(args);
                let Some(jacobian) = self.jacobian.jac_function.as_mut() else {
                    return Err(BvpBackendIntegrationError::CallbackExecutionFailed {
                        stage: "Jacobian".to_string(),
                        message: "LambdifyReady plan has no Jacobian callback".to_string(),
                    });
                };
                let dense = jacobian.call(1.0, &*typed).to_DMatrixType();
                let (rows, cols) = self.jacobian_shape();
                if dense.nrows() != rows || dense.ncols() != cols {
                    return Err(BvpBackendIntegrationError::CallbackShapeMismatch {
                        stage: "Jacobian".to_string(),
                        expected_rows: rows,
                        expected_columns: cols,
                        actual_rows: dense.nrows(),
                        actual_columns: dense.ncols(),
                    });
                }
                for (value, (&row, &col)) in values_out.iter_mut().zip(
                    self.structure
                        .row_indices
                        .iter()
                        .zip(self.structure.col_indices.iter()),
                ) {
                    *value = dense[(row, col)];
                }
                Ok(())
            }
            BvpSparseExecutionPlan::AotCompiled(selected) => {
                let linked = try_resolve_linked_sparse_backend(&selected.problem_key())
                    .map_err(|error| {
                        BvpBackendIntegrationError::PipelinePanicked(error.to_string())
                    })?
                    .ok_or_else(
                        || BvpBackendIntegrationError::CompiledAotRuntimeUnavailable {
                            problem_key: selected.problem_key(),
                        },
                    )?;
                let flat_args =
                    flatten_runtime_args(self.jacobian.parameter_values.as_deref(), args);
                let parallel_config = self.parallel_config;
                let chunks = linked_execution_chunk_count(&linked, parallel_config, true);
                let callback = |flat_args: &[f64], callback_out: &mut [f64]| {
                    let callback_result = match parallel_config {
                        Some(config) => try_eval_linked_sparse_values(
                            &linked,
                            Some(config),
                            flat_args,
                            callback_out,
                        ),
                        None => linked.try_jacobian_values_eval(flat_args, callback_out),
                    };
                    callback_result
                        .map_err(|error| map_linked_callback_error("Jacobian", error))
                };
                if let Some(plan) = selected.prepared_problem.atom_aot_plan() {
                    plan.record_worker_batch(
                        if chunks > 1 {
                            rayon::current_num_threads()
                        } else {
                            1
                        },
                        chunks,
                    );
                    plan.try_execute_jacobian_callback_with_chunks(
                        &flat_args,
                        values_out,
                        chunks,
                        callback,
                    )
                } else if selected.prepared_problem.symbolic_assembly_backend
                    == BvpSymbolicAssemblyBackend::ExprLegacy
                {
                    callback(&flat_args, values_out)
                } else {
                    Err(BvpBackendIntegrationError::CompiledAotRuntimeUnavailable {
                        problem_key: selected.problem_key(),
                    })
                }
            }
            _ => Err(BvpBackendIntegrationError::CompiledAotRuntimeUnavailable {
                problem_key: self.execution.selected().problem_key(),
            }),
        }))
        .map_err(
            |panic| BvpBackendIntegrationError::CallbackExecutionFailed {
                stage: "Jacobian".to_string(),
                message: panic_message(panic),
            },
        )
        .and_then(|result| result);
        callback_result?;

        if let Some(index) = values_out.iter().position(|value| !value.is_finite()) {
            self.record_aot_runtime_error();
            return Err(BvpBackendIntegrationError::NonFiniteCallbackValue {
                stage: "Jacobian".to_string(),
                index,
            });
        }
        Ok(())
    }

    /// Evaluates sparse Jacobian explicit values into the caller-provided buffer.
    ///
    /// The output order always follows `jacobian_structure()`, regardless of
    /// whether the active branch is lambdify or linked compiled AOT.
    pub fn jacobian_values_into(&mut self, args: &[f64], values_out: &mut [f64]) {
        if let Err(error) = self.try_jacobian_values_into(args, values_out) {
            panic!("BvpSparseSolverProvider Jacobian callback failed: {error:?}");
        }
    }
}

/// Owned solver-facing sparse BVP bundle.
///
/// This is the assembled object the outer solver infrastructure wants to see:
/// equations/settings go in, and on success it receives one bundle containing:
/// - the selected backend branch,
/// - optional residual/Jacobian callbacks ready to use,
/// - and all solver metadata that used to be read piecemeal from `Jacobian`.
///
/// The solver does not need to know whether the callbacks came from lambdify,
/// AOT, or a future numeric path. It only inspects whether runtime callbacks
/// are available and uses the exposed metadata consistently.
pub struct BvpSparseSolverBundle {
    pub execution: BvpSparseExecutionPlan,
    pub residual_function: Option<Box<dyn Fun>>,
    pub jacobian_function: Option<Box<dyn Jac>>,
    pub runtime_diagnostics: HashMap<String, String>,
    /// Typed wall-clock stages for the symbolic/backend preparation handoff.
    pub generation_telemetry: Option<BvpGenerationTelemetrySnapshot>,
    pub banded_linear_solver_config: LinearSolverConfig,
    pub variable_string: Vec<String>,
    pub parameter_values: Option<Vec<f64>>,
    pub bounds_vec: Option<Vec<(f64, f64)>>,
    pub rel_tolerance_vec: Option<Vec<f64>>,
    pub bandwidth: Option<(usize, usize)>,
    pub bc_position_and_value: Vec<(usize, usize, f64)>,
    pub sparse_structure: SparseJacobianStructure,
    pub atom_discretization_telemetry:
        Option<crate::symbolic::bvp::telemetry::BvpAtomDiscretizationTelemetrySnapshot>,
    /// Runtime callback telemetry for the ExprLegacy Lambdify route.
    pub legacy_lambdify_telemetry: Option<crate::symbolic::bvp::telemetry::BvpLambdifyTelemetry>,
    /// Runtime callback telemetry for the AtomView Lambdify route.
    pub atom_lambdify_telemetry: Option<crate::symbolic::bvp::telemetry::BvpLambdifyTelemetry>,
    /// Immutable snapshot of the direct no-Mutex Banded callback telemetry.
    pub direct_banded_jacobian_telemetry:
        Option<crate::symbolic::bvp::telemetry::BvpDirectJacobianTelemetry>,
    /// Shared typed AOT telemetry handle for live post-solve diagnostics.
    pub aot_telemetry: Option<crate::symbolic::bvp::aot_telemetry::BvpAotTelemetry>,
    /// Shared numeric binding for prepared Lambdify callbacks. It is exposed
    /// only for the Lambdify runtime handoff; linked AOT callbacks keep their
    /// own runtime binding contract and must not be marked rebindable here.
    pub(crate) parameter_binding: Option<BvpParameterBindingHandle>,
}

/// Transitional owned bundle for legacy lambdify/numeric BVP callbacks.
///
/// This keeps the remaining compatibility path assembled in one object instead
/// of forcing solver handoff layers to read multiple `Jacobian` fields
/// directly. It is intentionally smaller than [`BvpSparseSolverBundle`] and
/// exists only while non-sparse solver paths are still converging on the same
/// provider model.
pub struct BvpLegacySolverBundle {
    pub residual_function: Box<dyn Fun>,
    pub jacobian_function: Option<Box<dyn Jac>>,
    pub variable_string: Vec<String>,
    pub bounds_vec: Option<Vec<(f64, f64)>>,
    pub rel_tolerance_vec: Option<Vec<f64>>,
    pub bandwidth: Option<(usize, usize)>,
    pub bc_position_and_value: Vec<(usize, usize, f64)>,
    /// Runtime callback telemetry for the ExprLegacy Lambdify route.
    pub legacy_lambdify_telemetry: Option<crate::symbolic::bvp::telemetry::BvpLambdifyTelemetry>,
    /// Runtime callback telemetry for the AtomView Lambdify route.
    pub atom_lambdify_telemetry: Option<crate::symbolic::bvp::telemetry::BvpLambdifyTelemetry>,
    /// Immutable snapshot of the direct no-Mutex Banded callback telemetry.
    pub direct_banded_jacobian_telemetry:
        Option<crate::symbolic::bvp::telemetry::BvpDirectJacobianTelemetry>,
    /// Shared numeric binding for already compiled Lambdify callbacks.
    pub(crate) parameter_binding: Option<BvpParameterBindingHandle>,
}

/// Fallible wrapper error for the staged BVP backend-selection / AOT path.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum BvpBackendIntegrationError {
    /// User supplied a malformed BVP problem before backend preparation.
    InvalidProblem { field: String, message: String },
    /// User supplied an unsupported or internally inconsistent solver option.
    InvalidSolverConfiguration {
        field: String,
        value: String,
        message: String,
    },
    /// The BVP pipeline panicked while building or selecting a backend.
    PipelinePanicked(String),
    /// The solver was configured with an unsupported logging level string.
    InvalidLogLevel { level: String },
    /// The solver could not create the requested log file.
    LogFileCreationFailed { path: String, message: String },
    /// A compiled AOT backend was required by the outer policy, but the
    /// selection layer could not provide a compiled artifact.
    CompiledAotRequiredButUnavailable {
        problem_key: String,
        effective_backend: SelectedBackendKind,
    },
    /// The lifecycle policy asked the solver layer to build a missing AOT
    /// artifact automatically, but that orchestration is not wired here yet.
    AutomaticAotBuildRequested { problem_key: String },
    /// The lifecycle policy asked the solver layer to rebuild an AOT artifact,
    /// but rebuild orchestration is not wired here yet.
    AutomaticAotRebuildRequested { problem_key: String },
    /// The lifecycle layer attempted to build a generated AOT crate but the
    /// nested Cargo build failed.
    AutomaticAotBuildFailed {
        problem_key: String,
        message: String,
    },
    /// A compiled AOT artifact was selected, but callable runtime callbacks
    /// were not available in the current process.
    CompiledAotRuntimeUnavailable { problem_key: String },
    /// An owned linear factor rejected a solve request or could not apply its
    /// factor to the RHS. The dimensions are retained for diagnostics at the
    /// solver boundary instead of being hidden in a panic message.
    LinearSolveFailed {
        backend: String,
        matrix_rows: usize,
        matrix_columns: usize,
        rhs_len: usize,
        message: String,
    },
    /// A residual or Jacobian callback returned a value with an invalid shape
    /// for the current prepared solver state.
    CallbackShapeMismatch {
        stage: String,
        expected_rows: usize,
        expected_columns: usize,
        actual_rows: usize,
        actual_columns: usize,
    },
    /// A runtime callback received a variable vector with the wrong length.
    CallbackInputLengthMismatch {
        stage: String,
        expected: usize,
        actual: usize,
    },
    /// A prepared callback has a different numeric parameter schema.
    ParameterCountMismatch { expected: usize, actual: usize },
    /// A residual or Jacobian callback panicked at the typed solver boundary.
    CallbackExecutionFailed { stage: String, message: String },
    /// A residual callback returned a non-finite value before damping or a
    /// linear solve could consume it.
    NonFiniteCallbackValue { stage: String, index: usize },
    /// A prepared solve was requested after a problem input changed. The caller
    /// must regenerate the residual/Jacobian callbacks before reusing the
    /// prepared-solver entry point.
    PreparedRuntimeInvalidated { reason: String },
    /// AtomView was selected as the symbolic AOT route, but its owned plan
    /// could not be validated. This must not be converted into the retained
    /// ExprLegacy compatibility payload: doing so would hide a native-route
    /// defect and silently reintroduce an Atom -> Expr round-trip.
    AtomAotPlanPreparationFailed { message: String },
    /// A compiler/materializer/linker lifecycle failure with preserved
    /// partial artifact inspection and retry/quarantine diagnostics.
    AotLifecycleFailure {
        diagnostics: crate::symbolic::codegen::codegen_aot_lifecycle::AotFailureDiagnostics,
    },
}

impl fmt::Display for BvpBackendIntegrationError {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::InvalidProblem { field, message } => {
                write!(f, "invalid BVP problem field `{field}`: {message}")
            }
            Self::InvalidSolverConfiguration {
                field,
                value,
                message,
            } => write!(
                f,
                "invalid solver configuration `{field}={value}`: {message}"
            ),
            Self::PipelinePanicked(message) => {
                write!(f, "BVP backend pipeline panicked: {message}")
            }
            Self::InvalidLogLevel { level } => write!(f, "invalid BVP log level `{level}`"),
            Self::LogFileCreationFailed { path, message } => {
                write!(f, "cannot create BVP log file `{path}`: {message}")
            }
            Self::CompiledAotRequiredButUnavailable {
                problem_key,
                effective_backend,
            } => write!(
                f,
                "compiled AOT backend required but unavailable for `{problem_key}` (effective backend: {effective_backend:?})"
            ),
            Self::AutomaticAotBuildRequested { problem_key } => {
                write!(f, "automatic AOT build requested for `{problem_key}`")
            }
            Self::AutomaticAotRebuildRequested { problem_key } => {
                write!(f, "automatic AOT rebuild requested for `{problem_key}`")
            }
            Self::AutomaticAotBuildFailed {
                problem_key,
                message,
            } => write!(
                f,
                "automatic AOT build failed for `{problem_key}`: {message}"
            ),
            Self::CompiledAotRuntimeUnavailable { problem_key } => {
                write!(f, "compiled AOT runtime unavailable for `{problem_key}`")
            }
            Self::LinearSolveFailed {
                backend,
                matrix_rows,
                matrix_columns,
                rhs_len,
                message,
            } => write!(
                f,
                "linear solve failed for {backend} ({matrix_rows}x{matrix_columns}, rhs={rhs_len}): {message}"
            ),
            Self::CallbackShapeMismatch {
                stage,
                expected_rows,
                expected_columns,
                actual_rows,
                actual_columns,
            } => write!(
                f,
                "{stage} callback shape mismatch: expected {expected_rows}x{expected_columns}, got {actual_rows}x{actual_columns}"
            ),
            Self::CallbackInputLengthMismatch {
                stage,
                expected,
                actual,
            } => write!(
                f,
                "{stage} callback input length mismatch: expected {expected}, got {actual}"
            ),
            Self::ParameterCountMismatch { expected, actual } => write!(
                f,
                "callback parameter count mismatch: expected {expected}, got {actual}"
            ),
            Self::CallbackExecutionFailed { stage, message } => {
                write!(f, "{stage} callback execution failed: {message}")
            }
            Self::NonFiniteCallbackValue { stage, index } => {
                write!(
                    f,
                    "{stage} callback returned a non-finite value at index {index}"
                )
            }
            Self::PreparedRuntimeInvalidated { reason } => {
                write!(f, "prepared BVP runtime invalidated: {reason}")
            }
            Self::AtomAotPlanPreparationFailed { message } => {
                write!(f, "AtomView AOT plan preparation failed: {message}")
            }
            Self::AotLifecycleFailure { diagnostics } => write!(
                f,
                "AOT lifecycle {:?}/{:?} failed for `{}`: {}",
                diagnostics.stage, diagnostics.kind, diagnostics.artifact_key, diagnostics.detail
            ),
        }
    }
}

impl std::error::Error for BvpBackendIntegrationError {}

impl From<AotLifecycleError> for BvpBackendIntegrationError {
    fn from(error: AotLifecycleError) -> Self {
        Self::AotLifecycleFailure {
            diagnostics: error.diagnostics,
        }
    }
}

impl From<AtomAotRuntimeError> for BvpBackendIntegrationError {
    fn from(error: AtomAotRuntimeError) -> Self {
        match error {
            AtomAotRuntimeError::InputLength { expected, actual } => {
                Self::CallbackInputLengthMismatch {
                    stage: "AtomView AOT callback".to_string(),
                    expected,
                    actual,
                }
            }
            AtomAotRuntimeError::OutputLength {
                stage,
                expected,
                actual,
            } => Self::CallbackShapeMismatch {
                stage: stage.to_string(),
                expected_rows: expected,
                expected_columns: 1,
                actual_rows: actual,
                actual_columns: 1,
            },
            AtomAotRuntimeError::NonFiniteInput { index }
            | AtomAotRuntimeError::NonFiniteOutput { index, .. } => Self::NonFiniteCallbackValue {
                stage: "AtomView AOT callback".to_string(),
                index,
            },
            AtomAotRuntimeError::CallbackPanicked { stage, message } => {
                Self::CallbackExecutionFailed {
                    stage: stage.to_string(),
                    message,
                }
            }
            AtomAotRuntimeError::RuntimeNotLinked => Self::CallbackExecutionFailed {
                stage: "AtomView AOT callback".to_string(),
                message: "prepared plan has no linked runtime".to_string(),
            },
            AtomAotRuntimeError::LinkedCallbackFailure { stage, message } => {
                Self::CallbackExecutionFailed {
                    stage: stage.to_string(),
                    message,
                }
            }
        }
    }
}

impl BvpSparseSolverBundle {
    pub fn effective_backend(&self) -> SelectedBackendKind {
        self.execution.selected().effective_backend
    }

    pub fn requested_backend(&self) -> BackendKind {
        self.execution.selected().requested_backend
    }

    pub fn is_runtime_callable(&self) -> bool {
        self.residual_function.is_some() && self.jacobian_function.is_some()
    }

    pub fn resolved_aot_artifact(&self) -> Option<&ResolvedAotArtifact> {
        self.execution.selected().aot_resolution.as_ref()
    }

    pub fn residual_len(&self) -> usize {
        self.execution.selected().prepared_problem.shape.0
    }

    pub fn jacobian_shape(&self) -> (usize, usize) {
        self.execution.selected().prepared_problem.shape
    }

    pub fn residual_call(&self, p: f64, y: &dyn VectorType) -> Option<Box<dyn VectorType>> {
        self.residual_function.as_ref().map(|fun| fun.call(p, y))
    }

    pub fn jacobian_call(&mut self, p: f64, y: &dyn VectorType) -> Option<Box<dyn MatrixType>> {
        self.jacobian_function.as_mut().map(|jac| jac.call(p, y))
    }

    pub fn into_runtime_callbacks(self) -> Option<(Box<dyn Fun>, Box<dyn Jac>)> {
        match (self.residual_function, self.jacobian_function) {
            (Some(fun), Some(jac)) => Some((fun, jac)),
            _ => None,
        }
    }

    /// Returns solver-facing runtime diagnostics for generated callback execution.
    pub fn runtime_diagnostics(&self) -> &HashMap<String, String> {
        &self.runtime_diagnostics
    }

    /// Refreshes diagnostics for the linked sparse AOT runtime binding.
    ///
    /// The values mirror the same branch logic used by the linked callback
    /// evaluator below, so story/perf tests can tell whether a requested
    /// parallel route actually ran in parallel or deliberately fell back.
    pub fn refresh_linked_runtime_diagnostics(
        &mut self,
        execution_policy: &str,
        parallel_config: Option<ParallelExecutorConfig>,
    ) {
        self.runtime_diagnostics =
            linked_runtime_diagnostics_for_bundle(self, execution_policy, parallel_config);
    }

    /// Rebuilds linked compiled-AOT runtime callbacks with an optional
    /// parallel execution policy.
    ///
    /// This keeps the outer solver-facing bundle shape stable while allowing
    /// solver-level execution settings to choose between whole-callback and
    /// chunk-aware linked AOT execution.
    pub fn rebind_linked_runtime_callbacks(
        &mut self,
        parameter_values: Option<&[f64]>,
        parallel_config: Option<ParallelExecutorConfig>,
    ) -> bool {
        if let Some(plan) = self
            .execution
            .selected_mut()
            .prepared_problem
            .atom_aot_plan
            .as_mut()
        {
            // Drop the previous artifact/runtime before resolving the new
            // binding. A failed lookup must never leave stale callbacks live.
            plan.invalidate_linked_runtime();
        }
        self.residual_function = None;
        self.jacobian_function = None;

        if !matches!(self.execution, BvpSparseExecutionPlan::AotCompiled(_)) {
            return false;
        }

        let (problem_key, artifact, matrix_backend, prepared_problem) = {
            let selected = self.execution.selected();
            (
                selected.problem_key(),
                selected.aot_resolution.clone(),
                selected.matrix_backend,
                selected.prepared_problem.clone(),
            )
        };
        let linked_runtime = try_resolve_linked_sparse_backend(&problem_key)
            .ok()
            .flatten();
        let Some(linked_runtime) = linked_runtime else {
            return false;
        };
        let Some((residual, jacobian)) = linked_runtime_callbacks_for_matrix_backend(
            &prepared_problem,
            matrix_backend,
            parameter_values.or(self.parameter_values.as_deref()),
            parallel_config,
            self.banded_linear_solver_config,
        ) else {
            return false;
        };

        if let Some(plan) = self
            .execution
            .selected_mut()
            .prepared_problem
            .atom_aot_plan
            .as_mut()
        {
            plan.bind_linked_runtime(artifact, linked_runtime);
        }
        self.residual_function = Some(residual);
        self.jacobian_function = Some(jacobian);
        true
    }
}

fn insert_linked_stage_runtime_diagnostics(
    diagnostics: &mut HashMap<String, String>,
    prefix: &str,
    workload: usize,
    chunk_count: usize,
    requested_jobs: usize,
    should_parallelize: bool,
) {
    let planned_ranges = linked_contiguous_job_ranges(chunk_count, requested_jobs);
    let planned_jobs = planned_ranges.len().max(1);
    let actual_jobs = if should_parallelize && planned_jobs > 1 {
        planned_jobs
    } else {
        1
    };
    let fallback = actual_jobs == 1;
    let fallback_reason = if chunk_count <= 1 {
        "single_chunk"
    } else if requested_jobs <= 1 {
        "single_requested_job"
    } else if !should_parallelize {
        "fallback_policy_or_small_workload"
    } else {
        "none"
    };

    diagnostics.insert(format!("{prefix}.chunk_count"), chunk_count.to_string());
    diagnostics.insert(
        format!("{prefix}.requested_jobs"),
        requested_jobs.to_string(),
    );
    diagnostics.insert(format!("{prefix}.planned_jobs"), planned_jobs.to_string());
    diagnostics.insert(format!("{prefix}.actual_jobs"), actual_jobs.to_string());
    diagnostics.insert(format!("{prefix}.fallback"), fallback.to_string());
    diagnostics.insert(
        format!("{prefix}.fallback_reason"),
        fallback_reason.to_string(),
    );
    diagnostics.insert(
        format!("{prefix}.work_per_job"),
        workload.div_ceil(actual_jobs).to_string(),
    );
    diagnostics.insert(
        format!("{prefix}.work_per_chunk"),
        if chunk_count == 0 {
            "0".to_string()
        } else {
            workload.div_ceil(chunk_count).to_string()
        },
    );
}

fn linked_runtime_diagnostics_for_bundle(
    bundle: &BvpSparseSolverBundle,
    execution_policy: &str,
    parallel_config: Option<ParallelExecutorConfig>,
) -> HashMap<String, String> {
    let selected = bundle.execution.selected();
    let mut diagnostics = HashMap::new();
    diagnostics.insert(
        "aot.runtime.execution_policy".to_string(),
        execution_policy.to_string(),
    );
    diagnostics.insert(
        "aot.runtime.effective_backend".to_string(),
        format!("{:?}", selected.effective_backend),
    );
    diagnostics.insert(
        "aot.runtime.matrix_backend".to_string(),
        format!("{:?}", selected.matrix_backend),
    );
    diagnostics.insert(
        "aot.runtime.preparation_route".to_string(),
        format!("{:?}", selected.preparation_route()),
    );
    diagnostics.insert(
        "aot.runtime.telemetry_mode".to_string(),
        format!("{:?}", selected.prepared_problem.aot_telemetry_mode()),
    );
    diagnostics.insert(
        "aot.runtime.parallel_requested".to_string(),
        parallel_config.is_some().to_string(),
    );
    if !matches!(selected.effective_backend, SelectedBackendKind::AotCompiled) {
        diagnostics.insert(
            "aot.runtime.binding".to_string(),
            "non_aot_or_lambdify".to_string(),
        );
        return diagnostics;
    }

    // Auto calibration and plan construction can initialize Rayon and perform
    // a machine-specific baseline measurement. Never pay that one-time cost
    // while merely reporting a Lambdify/non-AOT runtime, because no parallel
    // compiled callback will consume the plan on that route.
    let auto_plan = selected.prepared_problem.auto_parallel_plan();
    diagnostics.insert(
        "aot.auto.execution_mode".to_string(),
        format!("{:?}", auto_plan.execution_mode),
    );
    diagnostics.insert(
        "aot.auto.workers".to_string(),
        auto_plan.workers.to_string(),
    );
    diagnostics.insert(
        "aot.auto.min_work_per_job".to_string(),
        auto_plan.min_work_per_job.to_string(),
    );
    diagnostics.insert(
        "aot.auto.residual.reason".to_string(),
        auto_plan.residual_stage.reason.as_str().to_string(),
    );
    diagnostics.insert(
        "aot.auto.residual.work_per_job".to_string(),
        auto_plan.residual_stage.work_per_job.to_string(),
    );
    diagnostics.insert(
        "aot.auto.residual.work_per_chunk".to_string(),
        auto_plan.residual_stage.work_per_chunk.to_string(),
    );
    diagnostics.insert(
        "aot.auto.sparse_jacobian.reason".to_string(),
        auto_plan.sparse_stage.reason.as_str().to_string(),
    );
    diagnostics.insert(
        "aot.auto.sparse_jacobian.work_per_job".to_string(),
        auto_plan.sparse_stage.work_per_job.to_string(),
    );
    diagnostics.insert(
        "aot.auto.sparse_jacobian.work_per_chunk".to_string(),
        auto_plan.sparse_stage.work_per_chunk.to_string(),
    );

    let problem_key = selected
        .prepared_problem
        .problem_key_for_matrix_backend(selected.matrix_backend);
    let Some(linked) = try_resolve_linked_sparse_backend(&problem_key)
        .ok()
        .flatten()
    else {
        diagnostics.insert(
            "aot.runtime.binding".to_string(),
            "compiled_artifact_not_linked".to_string(),
        );
        return diagnostics;
    };
    diagnostics.insert("aot.runtime.binding".to_string(), "linked".to_string());
    diagnostics.insert("aot.runtime.problem_key".to_string(), problem_key);
    diagnostics.insert(
        "aot.runtime.residual_len".to_string(),
        linked.residual_len.to_string(),
    );
    let jacobian_output_len = linked.jacobian_output_len().unwrap_or(linked.nnz);
    diagnostics.insert(
        "aot.runtime.jacobian_output_len".to_string(),
        jacobian_output_len.to_string(),
    );
    diagnostics.insert(
        "aot.runtime.jacobian_layout".to_string(),
        format!("{:?}", linked.jacobian_layout),
    );

    match parallel_config {
        Some(config) => {
            let residual_requested_jobs = linked_max_jobs(config, false);
            let residual_should_parallelize = linked_should_parallelize(
                config,
                linked.residual_len,
                linked.residual_chunks.len(),
                false,
            );
            insert_linked_stage_runtime_diagnostics(
                &mut diagnostics,
                "aot.runtime.residual",
                linked.residual_len,
                linked.residual_chunks.len(),
                residual_requested_jobs,
                residual_should_parallelize,
            );

            let sparse_requested_jobs = linked_max_jobs(config, true);
            let sparse_should_parallelize = linked_should_parallelize(
                config,
                jacobian_output_len,
                linked.jacobian_value_chunks.len(),
                true,
            );
            insert_linked_stage_runtime_diagnostics(
                &mut diagnostics,
                "aot.runtime.sparse_jacobian",
                jacobian_output_len,
                linked.jacobian_value_chunks.len(),
                sparse_requested_jobs,
                sparse_should_parallelize,
            );
        }
        None => {
            insert_linked_stage_runtime_diagnostics(
                &mut diagnostics,
                "aot.runtime.residual",
                linked.residual_len,
                linked.residual_chunks.len(),
                1,
                false,
            );
            insert_linked_stage_runtime_diagnostics(
                &mut diagnostics,
                "aot.runtime.sparse_jacobian",
                jacobian_output_len,
                linked.jacobian_value_chunks.len(),
                1,
                false,
            );
        }
    }

    diagnostics
}

impl Jacobian {
    /// Returns the numeric binding shared by prepared callback closures.
    /// Parameter names are part of the compiled ABI; only values may be
    /// replaced without rebuilding the symbolic evaluator.
    fn parameter_binding_for_callbacks(&mut self) -> BvpParameterBindingHandle {
        if let Some(binding) = &self.parameter_binding {
            return binding.clone();
        }
        let binding = BvpParameterBindingHandle::new(self.parameter_values.clone());
        self.parameter_binding = Some(binding.clone());
        binding
    }

    /// Consumes the legacy lambdify/numeric callback fields into one assembled
    /// compatibility bundle.
    ///
    /// This is a transitional helper for non-sparse solver paths while they
    /// are still being migrated away from direct `Jacobian` field access.
    pub fn into_legacy_solver_bundle(self) -> BvpLegacySolverBundle {
        let legacy_lambdify_telemetry = self.legacy_lambdify_telemetry;
        let atom_lambdify_telemetry = self.atom_lambdify_telemetry;
        let direct_banded_jacobian_telemetry = self.direct_banded_jacobian_telemetry;
        BvpLegacySolverBundle {
            residual_function: self.residiual_function,
            jacobian_function: self.jac_function,
            variable_string: self.variable_string,
            bounds_vec: self.bounds,
            rel_tolerance_vec: self.rel_tolerance_vec,
            bandwidth: self.bandwidth,
            bc_position_and_value: self.BC_pos_n_values,
            legacy_lambdify_telemetry,
            atom_lambdify_telemetry,
            direct_banded_jacobian_telemetry,
            parameter_binding: self.parameter_binding,
        }
    }

    /// Builds a legacy lambdify/numeric BVP pipeline and returns the assembled
    /// compatibility bundle in one step.
    ///
    /// This keeps the remaining non-sparse solver handoff on the symbolic
    /// layer, so outer solver glue does not need to know which internal
    /// `Jacobian` fields carry callbacks and metadata.
    #[allow(clippy::too_many_arguments)]
    pub fn generate_legacy_solver_bundle_with_params(
        mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        border_conditions: HashMap<String, Vec<(usize, f64)>>,
        bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
    ) -> BvpLegacySolverBundle {
        self.generate_BVP_with_params(
            eq_system,
            values,
            arg,
            params,
            t0,
            param,
            n_steps,
            h,
            mesh,
            border_conditions,
            bounds,
            rel_tolerance,
            scheme,
            method,
            bandwidth,
        );
        self.into_legacy_solver_bundle()
    }
}

impl BvpPreparedSparseAotProblem {
    /// Infers the node-major block layout expected by the native banded BVP
    /// solver path.
    ///
    /// Variable names in the discretized BVP follow a node-major convention
    /// like `y_0, z_0, y_1, z_1, ...`. Counting distinct base names across the
    /// flattened variable list gives the number of variables per node.
    fn infer_node_major_layout(&self) -> Result<NodeMajorLayout, String> {
        if self.variable_names.is_empty() || self.shape.0 == 0 || self.shape.0 != self.shape.1 {
            return Err("prepared BVP problem must describe a non-empty square system".into());
        }

        fn base_variable_name(name: &str) -> &str {
            name.rsplit_once('_')
                .and_then(|(base, suffix)| suffix.parse::<usize>().ok().map(|_| base))
                .unwrap_or(name)
        }

        let mut seen = HashSet::new();
        let mut vars_per_node = 0usize;
        for name in &self.variable_names {
            if seen.insert(base_variable_name(name).to_string()) {
                vars_per_node += 1;
            }
        }

        if vars_per_node == 0 || self.shape.1 % vars_per_node != 0 {
            return Err(format!(
                "node-major layout inference failed: unknowns={} vars_per_node={vars_per_node}",
                self.shape.1
            ));
        }

        let n_nodes = self.shape.1 / vars_per_node;
        NodeMajorLayout::new(n_nodes, vars_per_node).map_err(|err| err.to_string())
    }

    /// Returns the flattened symbolic input ordering used across the sparse
    /// AOT path: `[params..., variables...]`.
    fn flattened_input_names(&self) -> Vec<&str> {
        let mut names = Vec::with_capacity(
            self.param_names
                .len()
                .saturating_add(self.variable_names.len()),
        );
        names.extend(self.param_names.iter().map(|name| name.as_str()));
        names.extend(self.variable_names.iter().map(|name| name.as_str()));
        names
    }

    /// Rebuilds borrowed sparse Jacobian entries from the owned cache.
    ///
    /// This is the bridge between the owned BVP representation and the shared
    /// borrow-based codegen/runtime plan layer.
    fn borrowed_sparse_entries(&self) -> Vec<SparseExprEntry<'_>> {
        self.sparse_entries
            .iter()
            .map(|(row, col, expr)| SparseExprEntry {
                row: *row,
                col: *col,
                expr,
            })
            .collect()
    }

    /// Rebuilds borrowed banded Jacobian entries from the owned sparse cache.
    ///
    /// The banded runtime/codegen path stores values in one explicit stream
    /// with matching per-value diagonal metadata.  The native assembly code
    /// does not require diagonal-major order; it only requires that values,
    /// diagonal offsets, and diagonal positions share the same order.
    ///
    /// We therefore keep row/column locality here.  This gives the codegen
    /// lowerer better common-subexpression locality than diagonal-major order,
    /// especially when the stream is later split into AOT chunks.
    fn inferred_bandwidth_from_sparse_entries(&self) -> (usize, usize) {
        self.sparse_entries
            .iter()
            .fold((0usize, 0usize), |(kl, ku), (row, col, _)| {
                if row >= col {
                    (kl.max(row - col), ku)
                } else {
                    (kl, ku.max(col - row))
                }
            })
    }

    /// Returns the explicit band width when it was cached during symbolic
    /// assembly, or infers it lazily from the sparse entry structure when the
    /// legacy path did not populate `self.bandwidth`.
    fn effective_bandwidth(&self) -> (usize, usize) {
        self.bandwidth
            .unwrap_or_else(|| self.inferred_bandwidth_from_sparse_entries())
    }

    fn borrowed_banded_entries(&self) -> Vec<BandedExprEntry<'_>> {
        let (kl, ku) = self.effective_bandwidth();
        let mut entries: Vec<(usize, usize, isize, &Expr)> = self
            .sparse_entries
            .iter()
            .map(|(row, col, expr)| {
                let diag_offset = *col as isize - *row as isize;
                (*row, *col, diag_offset, expr)
            })
            .collect();

        entries.sort_by(|lhs, rhs| lhs.0.cmp(&rhs.0).then_with(|| lhs.1.cmp(&rhs.1)));

        let mut banded_entries = Vec::with_capacity(entries.len());
        for (row, col, diag_offset, expr) in entries {
            assert!(
                diag_offset >= -(kl as isize) && diag_offset <= ku as isize,
                "sparse Jacobian entry ({row}, {col}) lies outside declared banded width"
            );
            let diag_position = if diag_offset >= 0 { row } else { col };
            banded_entries.push(BandedExprEntry {
                row,
                col,
                diag_offset,
                diag_position,
                expr,
            });
        }
        banded_entries
    }

    /// Builds the residual runtime plan used by code generation, manifests,
    /// and linked runtime rebinding.
    fn residual_runtime_plan(&self) -> ResidualRuntimePlan<'_> {
        let chunk_size = match self.residual_strategy {
            ResidualChunkingStrategy::Whole => self.residuals.len().max(1),
            ResidualChunkingStrategy::ByTargetChunkCount { target_chunks } => {
                assert!(target_chunks > 0, "target_chunks must be positive");
                self.residuals.len().max(1).div_ceil(target_chunks).max(1)
            }
            ResidualChunkingStrategy::ByOutputCount {
                max_outputs_per_chunk,
            } => {
                assert!(
                    max_outputs_per_chunk > 0,
                    "max_outputs_per_chunk must be positive"
                );
                max_outputs_per_chunk
            }
        };

        let input_names = self.flattened_input_names();
        let chunks = self
            .residuals
            .chunks(chunk_size)
            .enumerate()
            .map(|(chunk_index, residuals)| {
                let output_offset = chunk_index * chunk_size;
                let fn_name = if chunk_index == 0 && residuals.len() == self.residuals.len() {
                    self.residual_fn_name.clone()
                } else {
                    format!("{}_chunk_{}", self.residual_fn_name, chunk_index)
                };

                let outputs = residuals
                    .iter()
                    .map(|expr| PlannedOutput {
                        expr,
                        coordinate: None,
                    })
                    .collect();

                ResidualChunkPlan {
                    fn_name: fn_name.clone(),
                    output_offset,
                    residuals,
                    plan: CodegenTaskPlan {
                        fn_name: std::borrow::Cow::Owned(fn_name),
                        kind: CodegenTaskKind::Residual,
                        input_names: input_names.clone(),
                        outputs,
                        layout: CodegenOutputLayout::Vector {
                            len: residuals.len(),
                        },
                    },
                }
            })
            .collect();

        ResidualRuntimePlan {
            fn_name: self.residual_fn_name.as_str(),
            output_len: self.residuals.len(),
            input_names,
            chunks,
        }
    }

    /// Builds the sparse Jacobian runtime plan used by code generation,
    /// manifests, and linked runtime rebinding.
    fn sparse_runtime_plan(&self) -> SparseJacobianRuntimePlan<'_> {
        let all_entries = self.borrowed_sparse_entries();
        let input_names = self.flattened_input_names();

        let chunked_entries: Vec<(usize, Vec<SparseExprEntry<'_>>)> = match self.jacobian_strategy {
            SparseChunkingStrategy::Whole => vec![(0usize, all_entries.clone())],
            SparseChunkingStrategy::ByTargetChunkCount { target_chunks } => {
                assert!(target_chunks > 0, "target_chunks must be positive");
                let max_entries_per_chunk = all_entries.len().max(1).div_ceil(target_chunks).max(1);
                chunk_sparse_entries_by_nnz(&all_entries, max_entries_per_chunk)
            }
            SparseChunkingStrategy::ByNonZeroCount {
                max_entries_per_chunk,
            } => {
                assert!(
                    max_entries_per_chunk > 0,
                    "max_entries_per_chunk must be positive"
                );
                chunk_sparse_entries_by_nnz(&all_entries, max_entries_per_chunk)
            }
            SparseChunkingStrategy::ByRowCount { rows_per_chunk } => {
                assert!(rows_per_chunk > 0, "rows_per_chunk must be positive");
                chunk_sparse_entries_by_rows(&all_entries, rows_per_chunk, self.shape.0)
            }
        };

        let chunks = chunked_entries
            .into_iter()
            .enumerate()
            .map(|(chunk_index, (entry_offset, entries))| {
                let fn_name = if chunk_index == 0 && entries.len() == all_entries.len() {
                    self.jacobian_fn_name.clone()
                } else {
                    format!("{}_chunk_{}", self.jacobian_fn_name, chunk_index)
                };
                let outputs = entries
                    .iter()
                    .map(|entry| PlannedOutput {
                        expr: entry.expr,
                        coordinate: Some((entry.row, entry.col)),
                    })
                    .collect();

                SparseJacobianValuesChunkPlan {
                    fn_name: fn_name.clone(),
                    value_offset: entry_offset,
                    entries: entries.clone(),
                    plan: CodegenTaskPlan {
                        fn_name: std::borrow::Cow::Owned(fn_name),
                        kind: CodegenTaskKind::SparseJacobianValues,
                        input_names: input_names.clone(),
                        outputs,
                        layout: CodegenOutputLayout::SparseValues {
                            rows: self.shape.0,
                            cols: self.shape.1,
                            nnz: entries.len(),
                        },
                    },
                }
            })
            .collect();

        SparseJacobianRuntimePlan {
            fn_name: self.jacobian_fn_name.as_str(),
            input_names,
            structure: SparseJacobianStructure {
                rows: self.shape.0,
                cols: self.shape.1,
                row_indices: all_entries.iter().map(|entry| entry.row).collect(),
                col_indices: all_entries.iter().map(|entry| entry.col).collect(),
            },
            chunks,
        }
    }

    /// Derives a diagonal-oriented chunking strategy for the native banded
    /// runtime plan.
    ///
    /// Historically this bridge stored only sparse chunking hints. For the
    /// banded export we map those hints into a diagonal-based strategy so the
    /// generated callbacks remain bandwidth-aware and avoid fragmenting one
    /// diagonal across many tiny chunks.
    fn banded_chunking_strategy(&self) -> BandedChunkingStrategy {
        match self.jacobian_strategy {
            SparseChunkingStrategy::Whole => BandedChunkingStrategy::Whole,
            SparseChunkingStrategy::ByTargetChunkCount { target_chunks } => {
                BandedChunkingStrategy::ByTargetChunkCount { target_chunks }
            }
            SparseChunkingStrategy::ByNonZeroCount {
                max_entries_per_chunk,
            } => {
                let chunks_per_worker = max_entries_per_chunk.max(1);
                let (kl, ku) = self.effective_bandwidth();
                recommended_banded_chunking_for_parallelism(
                    kl.saturating_add(ku).saturating_add(1),
                    chunks_per_worker,
                )
            }
            SparseChunkingStrategy::ByRowCount { rows_per_chunk } => {
                let target_chunks = self.shape.0.max(1).div_ceil(rows_per_chunk.max(1));
                BandedChunkingStrategy::ByTargetChunkCount {
                    target_chunks: target_chunks.max(1),
                }
            }
        }
    }

    /// Builds the banded Jacobian runtime plan used by manifests and backend
    /// selection.
    fn banded_runtime_plan(&self) -> BandedJacobianRuntimePlan<'_> {
        let (kl, ku) = self.effective_bandwidth();
        let all_entries = self.borrowed_banded_entries();
        let input_names = self.flattened_input_names();
        let (chunk_entries, chunk_offsets): (Vec<Vec<BandedExprEntry<'_>>>, Vec<usize>) = match self
            .banded_chunking_strategy()
        {
            BandedChunkingStrategy::Whole => {
                if all_entries.is_empty() {
                    (Vec::new(), Vec::new())
                } else {
                    (vec![all_entries.clone()], vec![0])
                }
            }
            BandedChunkingStrategy::ByTargetChunkCount { target_chunks } => {
                assert!(target_chunks > 0, "target_chunks must be positive");
                // Target-count chunking is the user-facing/AOT story path.
                // Keep the banded value stream contiguous here: diagonal
                // bucket chunking was correct, but release diagnostics
                // showed it duplicated 1.65x..2.05x straight-line IR on
                // combustion grids because cross-diagonal subexpressions
                // could no longer be reused inside one block.
                let entries_per_chunk = all_entries.len().max(1).div_ceil(target_chunks).max(1);
                let mut chunks = Vec::new();
                let mut offsets = Vec::new();
                for (chunk_index, entries) in all_entries.chunks(entries_per_chunk).enumerate() {
                    offsets.push(chunk_index * entries_per_chunk);
                    chunks.push(entries.to_vec());
                }
                (chunks, offsets)
            }
            BandedChunkingStrategy::ByDiagonalCount {
                diagonals_per_chunk,
            } => {
                assert!(
                    diagonals_per_chunk > 0,
                    "diagonals_per_chunk must be positive"
                );
                let mut chunk_entries: Vec<Vec<BandedExprEntry<'_>>> = Vec::new();
                let mut chunk_offsets: Vec<usize> = Vec::new();

                for (entry_index, entry) in all_entries.iter().copied().enumerate() {
                    let bucket = ((entry.diag_offset + kl as isize) as usize) / diagonals_per_chunk;
                    if bucket >= chunk_entries.len() {
                        chunk_entries.resize_with(bucket + 1, Vec::new);
                        chunk_offsets.resize(bucket + 1, 0);
                    }
                    if chunk_entries[bucket].is_empty() {
                        chunk_offsets[bucket] = entry_index;
                    }
                    chunk_entries[bucket].push(entry);
                }
                (chunk_entries, chunk_offsets)
            }
        };

        let chunk_count = chunk_entries
            .iter()
            .filter(|entries| !entries.is_empty())
            .count();
        let chunks = chunk_entries
            .into_iter()
            .enumerate()
            .filter(|(_, entries)| !entries.is_empty())
            .map(|(chunk_index, entries)| {
                let fn_name = if chunk_index == 0 && chunk_count == 1 {
                    self.jacobian_fn_name.clone()
                } else {
                    format!("{}_chunk_{}", self.jacobian_fn_name, chunk_index)
                };
                let outputs = entries
                    .iter()
                    .map(|entry| PlannedOutput {
                        expr: entry.expr,
                        coordinate: Some((entry.row, entry.col)),
                    })
                    .collect();

                crate::symbolic::codegen::codegen_runtime_api::BandedJacobianValuesChunkPlan {
                    fn_name: fn_name.clone(),
                    value_offset: chunk_offsets[chunk_index],
                    entries: entries.clone(),
                    plan: CodegenTaskPlan {
                        fn_name: std::borrow::Cow::Owned(fn_name),
                        kind: CodegenTaskKind::BandedJacobianValues,
                        input_names: input_names.clone(),
                        outputs,
                        layout: CodegenOutputLayout::SparseValues {
                            rows: self.shape.0,
                            cols: self.shape.1,
                            nnz: entries.len(),
                        },
                    },
                }
            })
            .collect();

        BandedJacobianRuntimePlan {
            fn_name: self.jacobian_fn_name.as_str(),
            input_names,
            structure: crate::symbolic::codegen::codegen_runtime_api::BandedJacobianStructure {
                rows: self.shape.0,
                cols: self.shape.1,
                kl,
                ku,
                diagonal_offsets: all_entries.iter().map(|entry| entry.diag_offset).collect(),
                diagonal_positions: all_entries
                    .iter()
                    .map(|entry| entry.diag_position)
                    .collect(),
            },
            compact: false,
            chunks,
        }
    }

    /// Converts the owned BVP bridge object into the shared sparse prepared
    /// problem shape used by provider/build/registry layers.
    pub fn as_prepared_problem(&self) -> PreparedSparseProblem<'_> {
        PreparedSparseProblem::new(
            BackendKind::Aot,
            MatrixBackend::SparseCol,
            self.residual_runtime_plan(),
            self.sparse_runtime_plan(),
        )
    }

    /// Returns sparse structure metadata without constructing an Expr runtime
    /// plan when this bridge owns an AtomView-native AOT payload.
    pub fn sparse_structure(&self) -> SparseJacobianStructure {
        if let Some(atom_codegen) = &self.atom_codegen {
            return SparseJacobianStructure {
                rows: atom_codegen.shape.0,
                cols: atom_codegen.shape.1,
                row_indices: atom_codegen
                    .sparse_entries
                    .iter()
                    .map(|entry| entry.row)
                    .collect(),
                col_indices: atom_codegen
                    .sparse_entries
                    .iter()
                    .map(|entry| entry.col)
                    .collect(),
            };
        }
        self.as_prepared_problem().jacobian_structure().clone()
    }

    /// Converts the owned BVP bridge object into a shared banded prepared
    /// problem when a native banded Jacobian backend is requested.
    pub fn as_prepared_banded_problem(&self) -> PreparedBandedProblem<'_> {
        PreparedBandedProblem::new(
            BackendKind::Aot,
            MatrixBackend::Banded,
            self.residual_runtime_plan(),
            self.banded_runtime_plan(),
        )
    }

    /// Returns the Banded coordinate structure owned by the active route.
    ///
    /// AtomView-native bridges deliberately keep their compatibility `Expr`
    /// vectors empty.  Reading `as_prepared_banded_problem().jacobian_structure`
    /// in that case would therefore produce an empty structure even though the
    /// generated artifact has a complete explicit-entry stream.  Keep this
    /// small metadata projection native and use it only at the solver callback
    /// boundary; legacy callers still receive the historical runtime plan.
    pub fn banded_structure(&self) -> BandedJacobianStructure {
        if let Some(atom_codegen) = &self.atom_codegen {
            let (kl, ku) = self
                .bandwidth
                .unwrap_or_else(|| infer_atom_bandwidth(atom_codegen));
            return BandedJacobianStructure {
                rows: atom_codegen.shape.0,
                cols: atom_codegen.shape.1,
                kl,
                ku,
                diagonal_offsets: atom_codegen
                    .sparse_entries
                    .iter()
                    .map(|entry| entry.col as isize - entry.row as isize)
                    .collect(),
                diagonal_positions: atom_codegen
                    .sparse_entries
                    .iter()
                    .map(|entry| {
                        let offset = entry.col as isize - entry.row as isize;
                        if offset >= 0 { entry.row } else { entry.col }
                    })
                    .collect(),
            };
        }
        self.as_prepared_banded_problem()
            .jacobian_structure()
            .clone()
    }

    /// Converts the owned bridge object into a backend-aware prepared problem.
    ///
    /// This is the main handoff for correctness/performance tests: the same
    /// symbolic cache can now be emitted as either sparse or banded prepared
    /// runtime metadata without rebuilding the BVP from scratch.
    pub fn as_prepared_problem_for_matrix_backend(
        &self,
        matrix_backend: MatrixBackend,
    ) -> PreparedProblem<'_> {
        match matrix_backend {
            MatrixBackend::Banded => PreparedProblem::banded(self.as_prepared_banded_problem()),
            matrix_backend => PreparedProblem::sparse(PreparedSparseProblem::new(
                BackendKind::Aot,
                matrix_backend,
                self.residual_runtime_plan(),
                self.sparse_runtime_plan(),
            )),
        }
    }

    /// Fallible manifest boundary for the selected route.
    ///
    /// AtomView must expose its complete native adapter here. In particular,
    /// an incomplete plan is not allowed to fall through to the historical
    /// Expr manifest because that would change the route identity silently.
    pub fn try_manifest_for_matrix_backend(
        &self,
        matrix_backend: MatrixBackend,
    ) -> Result<PreparedProblemManifest, BvpBackendIntegrationError> {
        if self.symbolic_assembly_backend == BvpSymbolicAssemblyBackend::AtomView {
            self.validate_atom_native_codegen_route(matrix_backend)?;
            let codegen = match self.try_aot_adapter() {
                Ok(BvpAotAdapter::AtomView(adapter)) => {
                    return adapter.prepared_manifest(matrix_backend).map_err(|error| {
                        BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                            message: error.to_string(),
                        }
                    });
                }
                Err(BvpAotAdapterError::AtomViewPlanMissing) => {
                    self.atom_codegen.clone().ok_or_else(|| {
                        BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                            message: "AtomView route has no retained codegen payload".to_string(),
                        }
                    })?
                }
                Err(error) => {
                    return Err(BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                        message: error.to_string(),
                    });
                }
                Ok(BvpAotAdapter::ExprLegacy(_)) => {
                    let Some(discretized) = self.atom_discretized_system.as_ref() else {
                        return Err(BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                            message: "AtomView route has neither a prepared codegen payload nor a discretized system".to_string(),
                        });
                    };
                    prepare_sparse_bvp_codegen_from_discretized_system_with_breakdown(
                        discretized,
                        &self.residual_fn_name,
                        &self.jacobian_fn_name,
                        self.param_names.clone(),
                        self.bandwidth,
                        self.residual_strategy,
                        self.jacobian_strategy,
                    )
                    .0
                }
            };
            let codegen = if matrix_backend == MatrixBackend::Banded
                && !matches!(
                    codegen.matrix_layout(),
                    AtomAotMatrixLayout::BandedCompact { .. }
                )
            {
                let (kl, ku) = self
                    .bandwidth
                    .unwrap_or_else(|| infer_atom_bandwidth(&codegen));
                codegen.with_banded_layout(kl, ku).map_err(|error| {
                    BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                        message: error.to_string(),
                    }
                })?
            } else {
                codegen
            };
            return codegen.prepared_aot_manifest(matrix_backend).map_err(|error| {
                BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                    message: error.to_string(),
                }
            });
        }
        let prepared = self.as_prepared_problem_for_matrix_backend(matrix_backend);
        Ok(PreparedProblemManifest::from(&prepared))
    }

    /// Builds the manifest for the selected route.
    ///
    /// This is retained for compatibility with the historical infallible API.
    /// New callers should use [`Self::try_manifest_for_matrix_backend`] so a
    /// malformed AtomView route is returned as a typed error.
    pub fn manifest_for_matrix_backend(
        &self,
        matrix_backend: MatrixBackend,
    ) -> PreparedProblemManifest {
        self.try_manifest_for_matrix_backend(matrix_backend)
            .unwrap_or_else(|error| {
                panic!("BVP manifest preparation failed at the compatibility boundary: {error:?}")
            })
    }

    /// Returns the manifest-derived problem key used by lifecycle and linked-backend registries.
    pub fn problem_key(&self) -> String {
        self.problem_key_for_matrix_backend(MatrixBackend::SparseCol)
    }

    /// Returns the manifest-derived problem key for a specific matrix backend.
    pub fn problem_key_for_matrix_backend(&self, matrix_backend: MatrixBackend) -> String {
        self.manifest_for_matrix_backend(matrix_backend)
            .problem_key()
    }

    /// Validates that an AtomView AOT request has a complete native payload
    /// before entering the compatibility tuple-returning codegen APIs.
    ///
    /// The older APIs cannot return a typed error and are retained for source
    /// compatibility. New AOT callers should use the `try_*` methods below so
    /// an incomplete plan or invalid Banded width cannot be hidden by an
    /// ExprLegacy fallback.
    fn validate_atom_native_codegen_route(
        &self,
        matrix_backend: MatrixBackend,
    ) -> Result<(), BvpBackendIntegrationError> {
        if self.symbolic_assembly_backend != BvpSymbolicAssemblyBackend::AtomView {
            return Ok(());
        }
        let codegen = match self.try_aot_adapter() {
            Ok(BvpAotAdapter::AtomView(adapter)) => Some(adapter.codegen),
            Err(BvpAotAdapterError::AtomViewPlanMissing) => self.atom_codegen.as_ref(),
            _ => None,
        };
        let Some(codegen) = codegen else {
            // Older prepared Jacobians may retain the native discretization
            // without materializing an AOT owner yet. The codegen builder can
            // prepare that route locally; reject only when no native payload
            // remains at all.
            if self.atom_discretized_system.is_some() {
                return Ok(());
            }
            return Err(BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                message: "AtomView route has neither a prepared codegen payload nor a discretized system".to_string(),
            });
        };
        if matrix_backend == MatrixBackend::Banded {
            let (kl, ku) = self
                .bandwidth
                .unwrap_or_else(|| infer_atom_bandwidth(codegen));
            codegen
                .clone()
                .with_banded_layout(kl, ku)
                .map_err(
                    |error| BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                        message: error.to_string(),
                    },
                )?;
        }
        Ok(())
    }

    /// Fallible AtomView-aware codegen entrypoint for new AOT callers.
    pub fn try_codegen_module_with_breakdown_for_matrix_backend(
        &self,
        module_name: &str,
        matrix_backend: MatrixBackend,
    ) -> Result<(CodegenModule, BvpGeneratedAotCrateBreakdown), BvpBackendIntegrationError> {
        self.validate_atom_native_codegen_route(matrix_backend)?;
        Ok(self.codegen_module_with_breakdown_for_matrix_backend(module_name, matrix_backend))
    }

    /// Builds the emitted `CodegenModule` for this sparse BVP bridge and
    /// returns a timing/size breakdown before crate materialization.
    pub fn codegen_module_with_breakdown(
        &self,
        module_name: &str,
    ) -> (CodegenModule, BvpGeneratedAotCrateBreakdown) {
        self.codegen_module_with_breakdown_for_matrix_backend(module_name, MatrixBackend::SparseCol)
    }

    /// Builds the emitted `CodegenModule` for the selected matrix backend and
    /// returns a timing/size breakdown before crate materialization.
    pub fn codegen_module_with_breakdown_for_matrix_backend(
        &self,
        module_name: &str,
        matrix_backend: MatrixBackend,
    ) -> (CodegenModule, BvpGeneratedAotCrateBreakdown) {
        self.codegen_module_with_breakdown_for_matrix_backend_and_atom_profile(
            module_name,
            matrix_backend,
            AtomOptimizationProfile::Full,
        )
    }

    /// Builds the emitted `CodegenModule` with an explicit AtomView lowering
    /// optimization profile.
    ///
    /// The default public path uses `Full`, preserving the historical AOT
    /// behavior. Diagnostic/performance tests can pass `FastBootstrap` to
    /// measure whether IR cleanup passes pay back their cold-start cost.
    pub fn codegen_module_with_breakdown_for_matrix_backend_and_atom_profile(
        &self,
        module_name: &str,
        matrix_backend: MatrixBackend,
        atom_profile: AtomOptimizationProfile,
    ) -> (CodegenModule, BvpGeneratedAotCrateBreakdown) {
        // Compatibility wrapper: the fallible `try_*` entrypoint is the
        // production boundary, while this old tuple API must fail loudly
        // instead of silently selecting ExprLegacy after a native-route error.
        if let Err(error) = self.validate_atom_native_codegen_route(matrix_backend) {
            panic!("AtomView AOT codegen compatibility boundary rejected the route: {error:?}");
        }
        self.codegen_module_with_breakdown_for_matrix_backend_and_atom_profile_internal(
            module_name,
            matrix_backend,
            atom_profile,
            true,
        )
    }

    fn codegen_module_with_breakdown_for_matrix_backend_and_atom_profile_internal(
        &self,
        module_name: &str,
        matrix_backend: MatrixBackend,
        atom_profile: AtomOptimizationProfile,
        probe_source: bool,
    ) -> (CodegenModule, BvpGeneratedAotCrateBreakdown) {
        if let Some((atom_codegen, atom_breakdown, jacobian_prepare_ms)) =
            self.prepare_atom_codegen_for_matrix_backend(matrix_backend)
        {
            let module_begin = Instant::now();
            let (module, atom_module_breakdown) = atom_codegen
                .codegen_module_with_breakdown_and_optimization_profile(module_name, atom_profile);
            let module_build_ms = module_begin.elapsed().as_secs_f64() * 1_000.0;
            if let Some(plan) = self.atom_aot_plan.as_ref() {
                record_aot_stage_ms(
                    plan,
                    BvpAotColdStage::Lowering,
                    atom_module_breakdown.input_abi_prepare_ms
                        + atom_module_breakdown.residual_view_collect_ms
                        + atom_module_breakdown.residual_lower_many_ms
                        + atom_module_breakdown.residual_peephole_ms
                        + atom_module_breakdown.residual_reuse_temps_ms
                        + atom_module_breakdown.residual_push_ms
                        + atom_module_breakdown.sparse_view_collect_ms
                        + atom_module_breakdown.sparse_lower_many_ms
                        + atom_module_breakdown.sparse_peephole_ms
                        + atom_module_breakdown.sparse_reuse_temps_ms
                        + atom_module_breakdown.sparse_push_ms,
                );
            }

            let (source_probe_emit_ms, source_emit_ms, source_kb) = if probe_source {
                let emit_begin = Instant::now();
                let source_kb = module.emit_source().len() as f64 / 1024.0;
                let source_emit_ms = emit_begin.elapsed().as_secs_f64() * 1_000.0;
                if let Some(plan) = self.atom_aot_plan.as_ref() {
                    record_aot_stage_ms(plan, BvpAotColdStage::SourceEmission, source_emit_ms);
                }
                (source_emit_ms, source_emit_ms, source_kb)
            } else {
                // The artifact path emits the language source below.  Do
                // not pay for a diagnostic size probe there as well.
                (0.0, 0.0, 0.0)
            };

            let breakdown = BvpGeneratedAotCrateBreakdown {
                jacobian_prepare_ms,
                atom_input_abi_prepare_ms: atom_module_breakdown.input_abi_prepare_ms,
                atom_sparse_lookup_prepare_ms: atom_breakdown.sparse_lookup_prepare_ms,
                atom_sparse_jacobian_build_ms: atom_breakdown.sparse_jacobian_build_ms,
                atom_finalize_codegen_plan_ms: atom_breakdown.finalize_codegen_plan_ms,
                atom_sparse_nnz: atom_breakdown.sparse_nnz,
                atom_residual_view_collect_ms: atom_module_breakdown.residual_view_collect_ms,
                atom_residual_lower_many_ms: atom_module_breakdown.residual_lower_many_ms,
                atom_residual_peephole_ms: atom_module_breakdown.residual_peephole_ms,
                atom_residual_reuse_temps_ms: atom_module_breakdown.residual_reuse_temps_ms,
                atom_residual_push_ms: atom_module_breakdown.residual_push_ms,
                atom_sparse_view_collect_ms: atom_module_breakdown.sparse_view_collect_ms,
                atom_sparse_lower_many_ms: atom_module_breakdown.sparse_lower_many_ms,
                atom_sparse_peephole_ms: atom_module_breakdown.sparse_peephole_ms,
                atom_sparse_reuse_temps_ms: atom_module_breakdown.sparse_reuse_temps_ms,
                atom_sparse_push_ms: atom_module_breakdown.sparse_push_ms,
                prepared_module_init_ms: 0.0,
                prepared_residual_blocks_ms: 0.0,
                prepared_jacobian_blocks_ms: 0.0,
                module_build_ms,
                source_probe_emit_ms,
                language_source_emit_ms: 0.0,
                c_header_emit_ms: 0.0,
                artifact_packaging_ms: 0.0,
                source_emit_ms,
                source_kb,
                module_blocks: module.block_count(),
                total_block_instructions: module.total_block_instruction_count(),
                total_block_temps: module.total_block_temp_count(),
                max_block_instructions: module.max_block_instruction_count(),
                total_block_outputs: module.total_block_output_count(),
            };
            return (module, breakdown);
        }

        let prepared = self.as_prepared_problem_for_matrix_backend(matrix_backend);
        let (module, module_breakdown) =
            crate::symbolic::codegen::codegen_aot_driver::codegen_module_from_prepared_problem_with_breakdown(
                module_name,
                &prepared,
            );
        let module_build_ms = module_breakdown.total_ms;

        let (source_probe_emit_ms, source_emit_ms, source_kb) = if probe_source {
            let emit_begin = Instant::now();
            let source_kb = module.emit_source().len() as f64 / 1024.0;
            let source_emit_ms = emit_begin.elapsed().as_secs_f64() * 1_000.0;
            (source_emit_ms, source_emit_ms, source_kb)
        } else {
            // The artifact path emits the language source below.  Keep module-only
            // diagnostics available, but avoid double source generation in the
            // real AOT bootstrap route.
            (0.0, 0.0, 0.0)
        };

        let breakdown = BvpGeneratedAotCrateBreakdown {
            jacobian_prepare_ms: 0.0,
            atom_input_abi_prepare_ms: 0.0,
            atom_sparse_lookup_prepare_ms: 0.0,
            atom_sparse_jacobian_build_ms: 0.0,
            atom_finalize_codegen_plan_ms: 0.0,
            atom_sparse_nnz: 0,
            atom_residual_view_collect_ms: 0.0,
            atom_residual_lower_many_ms: 0.0,
            atom_residual_peephole_ms: 0.0,
            atom_residual_reuse_temps_ms: 0.0,
            atom_residual_push_ms: 0.0,
            atom_sparse_view_collect_ms: 0.0,
            atom_sparse_lower_many_ms: 0.0,
            atom_sparse_peephole_ms: 0.0,
            atom_sparse_reuse_temps_ms: 0.0,
            atom_sparse_push_ms: 0.0,
            prepared_module_init_ms: module_breakdown.module_init_ms,
            prepared_residual_blocks_ms: module_breakdown.residual_blocks_ms,
            prepared_jacobian_blocks_ms: module_breakdown.jacobian_blocks_ms,
            module_build_ms,
            source_probe_emit_ms,
            language_source_emit_ms: 0.0,
            c_header_emit_ms: 0.0,
            artifact_packaging_ms: 0.0,
            source_emit_ms,
            source_kb,
            module_blocks: module.block_count(),
            total_block_instructions: module.total_block_instruction_count(),
            total_block_temps: module.total_block_temp_count(),
            max_block_instructions: module.max_block_instruction_count(),
            total_block_outputs: module.total_block_output_count(),
        };
        (module, breakdown)
    }

    /// Prepares the AtomView payload before any generic Expr-based problem is
    /// constructed. `None` means that the explicit ExprLegacy adapter owns
    /// this route, or that Banded layout validation failed.
    fn prepare_atom_codegen_for_matrix_backend(
        &self,
        matrix_backend: MatrixBackend,
    ) -> Option<(
        PreparedSparseAtomBvpCodegen,
        AtomBvpCodegenPrepBreakdown,
        f64,
    )> {
        self.try_prepare_atom_codegen_for_matrix_backend_with_layout(matrix_backend, false)
            .ok()
            .flatten()
    }

    /// Prepares AtomView codegen with an optional native compact Banded ABI.
    ///
    /// The flag is deliberately private: public callers select the compact
    /// route through the dedicated artifact builder below, while all existing
    /// codegen and compatibility paths retain the historical explicit-entry
    /// layout.
    fn try_prepare_atom_codegen_for_matrix_backend_with_layout(
        &self,
        matrix_backend: MatrixBackend,
        native_banded_compact: bool,
    ) -> Result<
        Option<(
            PreparedSparseAtomBvpCodegen,
            AtomBvpCodegenPrepBreakdown,
            f64,
        )>,
        BvpBackendIntegrationError,
    > {
        if self.symbolic_assembly_backend != BvpSymbolicAssemblyBackend::AtomView {
            return Ok(None);
        }
        let retained_atom_codegen = match self.try_aot_adapter() {
            Ok(BvpAotAdapter::AtomView(adapter)) => Some(adapter.codegen.clone()),
            Err(BvpAotAdapterError::AtomViewPlanMissing) => self.atom_codegen.clone(),
            _ => None,
        };
        if let Some(atom_codegen) = retained_atom_codegen {
            let atom_breakdown = self.atom_codegen_prep_breakdown.clone().unwrap_or_default();
            let jacobian_begin = Instant::now();
            let atom_codegen = if matrix_backend == MatrixBackend::Banded {
                let (kl, ku) = self
                    .bandwidth
                    .unwrap_or_else(|| infer_atom_bandwidth(&atom_codegen));
                let result = if native_banded_compact {
                    atom_codegen.with_native_banded_layout(kl, ku)
                } else {
                    atom_codegen.with_banded_layout(kl, ku)
                };
                match result {
                    Ok(codegen) => codegen,
                    Err(error) => {
                        debug!(
                            "AtomView Banded AOT layout validation failed; using the typed Expr adapter: {error}"
                        );
                        if native_banded_compact {
                            return Err(BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                                message: error.to_string(),
                            });
                        }
                        return Ok(None);
                    }
                }
            } else {
                atom_codegen
            };
            return Ok(Some((
                atom_codegen,
                atom_breakdown,
                jacobian_begin.elapsed().as_secs_f64() * 1_000.0,
            )));
        }
        let Some(discretized) = self.atom_discretized_system.as_ref() else {
            return Ok(None);
        };
        let jacobian_begin = Instant::now();
        let requested_bandwidth = self.bandwidth;
        let (atom_codegen, atom_breakdown) =
            prepare_sparse_bvp_codegen_from_discretized_system_with_breakdown(
                discretized,
                &self.residual_fn_name,
                &self.jacobian_fn_name,
                self.param_names.clone(),
                requested_bandwidth,
                self.residual_strategy,
                self.jacobian_strategy,
            );
        let atom_codegen = if matrix_backend == MatrixBackend::Banded {
            let (kl, ku) = self
                .bandwidth
                .unwrap_or_else(|| infer_atom_bandwidth(&atom_codegen));
            let result = if native_banded_compact {
                atom_codegen.with_native_banded_layout(kl, ku)
            } else {
                atom_codegen.with_banded_layout(kl, ku)
            };
            match result {
                Ok(codegen) => codegen,
                Err(error) => {
                    debug!(
                        "AtomView Banded AOT layout validation failed; using the typed Expr adapter: {error}"
                    );
                    if native_banded_compact {
                        return Err(BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                            message: error.to_string(),
                        });
                    }
                    return Ok(None);
                }
            }
        } else {
            atom_codegen
        };
        Ok(Some((
            atom_codegen,
            atom_breakdown,
            jacobian_begin.elapsed().as_secs_f64() * 1_000.0,
        )))
    }

    /// Builds an Atom manifest through the retained route adapter whenever
    /// possible. The direct codegen fallback is retained for old bridges that
    /// predate the owned-plan field.
    fn prepared_atom_manifest_for_matrix_backend(
        &self,
        atom_codegen: &PreparedSparseAtomBvpCodegen,
        matrix_backend: MatrixBackend,
    ) -> Option<PreparedProblemManifest> {
        // Build the manifest from the actual codegen object. This matters for
        // opt-in compact Banded artifacts: the local layout may intentionally
        // differ from the adapter's historical explicit-entry metadata.
        atom_codegen.prepared_aot_manifest(matrix_backend).ok()
    }

    /// Builds one backend-selected AOT artifact and returns a timing breakdown
    /// for the major stages of artifact assembly.
    pub fn generated_aot_artifact_with_breakdown(
        &self,
        artifact_name: impl Into<String>,
        module_name: &str,
        backend: AotCodegenBackend,
    ) -> (GeneratedAotArtifact, BvpGeneratedAotCrateBreakdown) {
        self.generated_aot_artifact_with_breakdown_for_matrix_backend(
            artifact_name,
            module_name,
            backend,
            MatrixBackend::SparseCol,
        )
    }

    /// Builds one backend-selected AOT artifact for the requested matrix
    /// backend and returns a timing breakdown for the major stages of artifact
    /// assembly.
    pub fn generated_aot_artifact_with_breakdown_for_matrix_backend(
        &self,
        artifact_name: impl Into<String>,
        module_name: &str,
        backend: AotCodegenBackend,
        matrix_backend: MatrixBackend,
    ) -> (GeneratedAotArtifact, BvpGeneratedAotCrateBreakdown) {
        self.generated_aot_artifact_with_breakdown_for_matrix_backend_and_atom_profile(
            artifact_name,
            module_name,
            backend,
            matrix_backend,
            AtomOptimizationProfile::Full,
        )
    }

    /// Fallible AtomView-aware artifact entrypoint for new AOT callers.
    ///
    /// The returned artifact is still unbuilt; compiler and linker failures
    /// are reported by the subsequent materialization/build API. This method
    /// only guarantees that the selected symbolic route and matrix layout are
    /// complete before source generation begins.
    pub fn try_generated_aot_artifact_with_breakdown_for_matrix_backend(
        &self,
        artifact_name: impl Into<String>,
        module_name: &str,
        backend: AotCodegenBackend,
        matrix_backend: MatrixBackend,
    ) -> Result<(GeneratedAotArtifact, BvpGeneratedAotCrateBreakdown), BvpBackendIntegrationError>
    {
        self.validate_atom_native_codegen_route(matrix_backend)?;
        self.generated_aot_artifact_with_breakdown_for_matrix_backend_and_atom_profile_internal(
            artifact_name,
            module_name,
            backend,
            matrix_backend,
            AtomOptimizationProfile::Full,
            false,
        )
    }

    /// Builds one AOT artifact with an explicit AtomView lowering optimization
    /// profile. This is intentionally opt-in; production callers keep the
    /// existing `Full` profile unless they choose a faster bootstrap mode.
    pub fn generated_aot_artifact_with_breakdown_for_matrix_backend_and_atom_profile(
        &self,
        artifact_name: impl Into<String>,
        module_name: &str,
        backend: AotCodegenBackend,
        matrix_backend: MatrixBackend,
        atom_profile: AtomOptimizationProfile,
    ) -> (GeneratedAotArtifact, BvpGeneratedAotCrateBreakdown) {
        self.generated_aot_artifact_with_breakdown_for_matrix_backend_and_atom_profile_internal(
            artifact_name,
            module_name,
            backend,
            matrix_backend,
            atom_profile,
            false,
        )
        .unwrap_or_else(|error| {
            panic!("BVP AOT artifact compatibility boundary rejected the route: {error:?}")
        })
    }

    /// Builds an AtomView-native full-slot compact Banded artifact.
    ///
    /// This is an explicit opt-in because compact callbacks currently require
    /// one whole Jacobian block. The historical public builder above remains
    /// unchanged and continues to emit explicit-entry Banded artifacts.
    pub fn generated_native_banded_aot_artifact_with_breakdown(
        &self,
        artifact_name: impl Into<String>,
        module_name: &str,
        backend: AotCodegenBackend,
        atom_profile: AtomOptimizationProfile,
    ) -> (GeneratedAotArtifact, BvpGeneratedAotCrateBreakdown) {
        self.try_generated_native_banded_aot_artifact_with_breakdown(
            artifact_name,
            module_name,
            backend,
            atom_profile,
        )
        .unwrap_or_else(|error| {
            panic!("native compact Banded AOT compatibility boundary rejected the route: {error:?}")
        })
    }

    /// Fallible AtomView-native compact Banded artifact builder.
    ///
    /// Unlike the historical tuple API, this entrypoint never falls back to
    /// ExprLegacy when compact preparation or manifest validation fails.
    pub fn try_generated_native_banded_aot_artifact_with_breakdown(
        &self,
        artifact_name: impl Into<String>,
        module_name: &str,
        backend: AotCodegenBackend,
        atom_profile: AtomOptimizationProfile,
    ) -> Result<(GeneratedAotArtifact, BvpGeneratedAotCrateBreakdown), BvpBackendIntegrationError>
    {
        self.validate_atom_native_codegen_route(MatrixBackend::Banded)?;
        self.generated_aot_artifact_with_breakdown_for_matrix_backend_and_atom_profile_internal(
            artifact_name,
            module_name,
            backend,
            MatrixBackend::Banded,
            atom_profile,
            true,
        )
    }

    fn generated_aot_artifact_with_breakdown_for_matrix_backend_and_atom_profile_internal(
        &self,
        artifact_name: impl Into<String>,
        module_name: &str,
        backend: AotCodegenBackend,
        matrix_backend: MatrixBackend,
        atom_profile: AtomOptimizationProfile,
        native_banded_compact: bool,
    ) -> Result<(GeneratedAotArtifact, BvpGeneratedAotCrateBreakdown), BvpBackendIntegrationError>
    {
        // Compatibility wrapper: new callers should use the fallible method
        // above so route/layout failures remain typed at the API boundary.
        if let Err(error) = self.validate_atom_native_codegen_route(matrix_backend) {
            return Err(error);
        }
        let artifact_name = artifact_name.into();
        let lifecycle_key = artifact_name.clone();
        if let Some((atom_codegen, atom_breakdown, jacobian_prepare_ms)) = self
            .try_prepare_atom_codegen_for_matrix_backend_with_layout(
                matrix_backend,
                native_banded_compact,
            )?
        {
            if let Some(plan) = self.atom_aot_plan.as_ref() {
                plan.record_lifecycle_with_log(BvpAotLifecycleEvent::Planned, &lifecycle_key);
            }
            let module_begin = Instant::now();
            let (mut module, atom_module_breakdown) = atom_codegen
                .codegen_module_with_breakdown_and_optimization_profile(module_name, atom_profile);
            let module_build_ms = module_begin.elapsed().as_secs_f64() * 1_000.0;
            if let Some(plan) = self.atom_aot_plan.as_ref() {
                record_aot_stage_ms(
                    plan,
                    BvpAotColdStage::Lowering,
                    atom_module_breakdown.residual_view_collect_ms
                        + atom_module_breakdown.residual_lower_many_ms
                        + atom_module_breakdown.residual_peephole_ms
                        + atom_module_breakdown.residual_reuse_temps_ms
                        + atom_module_breakdown.residual_push_ms
                        + atom_module_breakdown.sparse_view_collect_ms
                        + atom_module_breakdown.sparse_lower_many_ms
                        + atom_module_breakdown.sparse_peephole_ms
                        + atom_module_breakdown.sparse_reuse_temps_ms
                        + atom_module_breakdown.sparse_push_ms,
                );
            }
            module.set_language(match backend {
                AotCodegenBackend::Rust => {
                    crate::symbolic::codegen::CodegenIR::CodegenLanguage::Rust
                }
                AotCodegenBackend::C => crate::symbolic::codegen::CodegenIR::CodegenLanguage::C,
                AotCodegenBackend::Zig => crate::symbolic::codegen::CodegenIR::CodegenLanguage::Zig,
            });
            let emit_begin = Instant::now();
            let module_source = module.emit_source();
            let language_source_emit_ms = emit_begin.elapsed().as_secs_f64() * 1_000.0;
            if let Some(plan) = self.atom_aot_plan.as_ref() {
                plan.record_lifecycle_with_log(BvpAotLifecycleEvent::SourceEmitted, &lifecycle_key);
                record_aot_stage_ms(
                    plan,
                    BvpAotColdStage::SourceEmission,
                    language_source_emit_ms,
                );
            }
            let source_kb = module_source.len() as f64 / 1024.0;
            let manifest_begin = Instant::now();
            if let Some(manifest) =
                self.prepared_atom_manifest_for_matrix_backend(&atom_codegen, matrix_backend)
            {
                let manifest_ms = manifest_begin.elapsed().as_secs_f64() * 1_000.0;
                let package_begin = Instant::now();
                let (artifact, c_header_emit_ms) = match backend {
                    AotCodegenBackend::Rust => (
                        GeneratedAotArtifact::Rust(GeneratedAotCrate::new(
                            artifact_name,
                            module_name,
                            module_source,
                            manifest,
                        )),
                        0.0,
                    ),
                    AotCodegenBackend::C => {
                        let c_header_begin = Instant::now();
                        let c_header = module.emit_c_header();
                        let artifact = GeneratedAotArtifact::C(
                            crate::symbolic::codegen::c_backend::codegen_c_aot_library::GeneratedCAotLibrary::new(
                                artifact_name,
                                module_source,
                                c_header,
                                manifest,
                            ),
                        );
                        (
                            artifact,
                            c_header_begin.elapsed().as_secs_f64() * 1_000.0,
                        )
                    }
                    AotCodegenBackend::Zig => (
                        GeneratedAotArtifact::Zig(
                            crate::symbolic::codegen::zig_backend::codegen_zig_aot_library::GeneratedZigAotLibrary::new(
                                artifact_name,
                                module_source,
                                manifest,
                            ),
                        ),
                        0.0,
                    ),
                };
                let artifact_packaging_ms =
                    manifest_ms + package_begin.elapsed().as_secs_f64() * 1_000.0;
                if let Some(plan) = self.atom_aot_plan.as_ref() {
                    record_aot_stage_ms(
                        plan,
                        BvpAotColdStage::Materialization,
                        artifact_packaging_ms,
                    );
                }
                let breakdown = BvpGeneratedAotCrateBreakdown {
                    jacobian_prepare_ms,
                    atom_input_abi_prepare_ms: atom_module_breakdown.input_abi_prepare_ms,
                    atom_sparse_lookup_prepare_ms: atom_breakdown.sparse_lookup_prepare_ms,
                    atom_sparse_jacobian_build_ms: atom_breakdown.sparse_jacobian_build_ms,
                    atom_finalize_codegen_plan_ms: atom_breakdown.finalize_codegen_plan_ms,
                    atom_sparse_nnz: atom_breakdown.sparse_nnz,
                    atom_residual_view_collect_ms: atom_module_breakdown.residual_view_collect_ms,
                    atom_residual_lower_many_ms: atom_module_breakdown.residual_lower_many_ms,
                    atom_residual_peephole_ms: atom_module_breakdown.residual_peephole_ms,
                    atom_residual_reuse_temps_ms: atom_module_breakdown.residual_reuse_temps_ms,
                    atom_residual_push_ms: atom_module_breakdown.residual_push_ms,
                    atom_sparse_view_collect_ms: atom_module_breakdown.sparse_view_collect_ms,
                    atom_sparse_lower_many_ms: atom_module_breakdown.sparse_lower_many_ms,
                    atom_sparse_peephole_ms: atom_module_breakdown.sparse_peephole_ms,
                    atom_sparse_reuse_temps_ms: atom_module_breakdown.sparse_reuse_temps_ms,
                    atom_sparse_push_ms: atom_module_breakdown.sparse_push_ms,
                    module_build_ms,
                    language_source_emit_ms,
                    c_header_emit_ms,
                    artifact_packaging_ms,
                    source_emit_ms: language_source_emit_ms,
                    source_kb,
                    module_blocks: module.block_count(),
                    total_block_instructions: module.total_block_instruction_count(),
                    total_block_temps: module.total_block_temp_count(),
                    max_block_instructions: module.max_block_instruction_count(),
                    total_block_outputs: module.total_block_output_count(),
                    ..Default::default()
                };
                if let Some(plan) = self.atom_aot_plan.as_ref() {
                    plan.record_lifecycle_with_log(
                        BvpAotLifecycleEvent::Materialized,
                        &lifecycle_key,
                    );
                }
                return Ok((artifact, breakdown));
            }
            if native_banded_compact {
                return Err(BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                    message:
                        "native compact Banded preparation did not produce an AtomView codegen plan"
                            .to_string(),
                });
            }
            debug!("AtomView AOT manifest validation failed; using ExprLegacy manifest adapter");
        }
        if native_banded_compact {
            return Err(BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                message: "native compact Banded route cannot fall back to ExprLegacy".to_string(),
            });
        }
        let prepared = self.as_prepared_problem_for_matrix_backend(matrix_backend);
        let (mut module, mut breakdown) = self
            .codegen_module_with_breakdown_for_matrix_backend_and_atom_profile_internal(
                module_name,
                matrix_backend,
                atom_profile,
                false,
            );
        module.set_language(match backend {
            AotCodegenBackend::Rust => crate::symbolic::codegen::CodegenIR::CodegenLanguage::Rust,
            AotCodegenBackend::C => crate::symbolic::codegen::CodegenIR::CodegenLanguage::C,
            AotCodegenBackend::Zig => crate::symbolic::codegen::CodegenIR::CodegenLanguage::Zig,
        });
        let emit_begin = Instant::now();
        let module_source = module.emit_source();
        let language_source_emit_ms = emit_begin.elapsed().as_secs_f64() * 1_000.0;
        breakdown.language_source_emit_ms = language_source_emit_ms;
        breakdown.source_emit_ms = language_source_emit_ms;
        breakdown.source_kb = module_source.len() as f64 / 1024.0;

        let manifest_begin = Instant::now();
        let manifest = PreparedProblemManifest::from(&prepared);
        let manifest_ms = manifest_begin.elapsed().as_secs_f64() * 1_000.0;

        let package_begin = Instant::now();
        let artifact = match backend {
            AotCodegenBackend::Rust => GeneratedAotArtifact::Rust(GeneratedAotCrate::new(
                artifact_name,
                module_name,
                module_source,
                manifest,
            )),
            AotCodegenBackend::C => {
                let header_begin = Instant::now();
                let c_header = module.emit_c_header();
                breakdown.c_header_emit_ms = header_begin.elapsed().as_secs_f64() * 1_000.0;
                GeneratedAotArtifact::C(
                    crate::symbolic::codegen::c_backend::codegen_c_aot_library::GeneratedCAotLibrary::new(
                    artifact_name,
                    module_source,
                    c_header,
                    manifest,
                    ),
                )
            }
            AotCodegenBackend::Zig => GeneratedAotArtifact::Zig(
                crate::symbolic::codegen::zig_backend::codegen_zig_aot_library::GeneratedZigAotLibrary::new(
                    artifact_name,
                    module_source,
                    manifest,
                ),
            ),
        };
        breakdown.artifact_packaging_ms =
            manifest_ms + package_begin.elapsed().as_secs_f64() * 1_000.0;

        Ok((artifact, breakdown))
    }

    /// Builds a generated Rust AOT crate from this BVP bridge object.
    ///
    /// This wrapper keeps existing BVP lifecycle callers stable while the
    /// underlying bridge moves to a backend-agnostic artifact API.
    pub fn generated_aot_crate(
        &self,
        crate_name: impl Into<String>,
        module_name: &str,
    ) -> GeneratedAotCrate {
        self.generated_aot_crate_with_breakdown(crate_name, module_name)
            .0
    }

    /// Builds a generated Rust AOT crate and returns a timing breakdown for the
    /// major stages of crate assembly.
    pub fn generated_aot_crate_with_breakdown(
        &self,
        crate_name: impl Into<String>,
        module_name: &str,
    ) -> (GeneratedAotCrate, BvpGeneratedAotCrateBreakdown) {
        let (artifact, breakdown) = self.generated_aot_artifact_with_breakdown(
            crate_name,
            module_name,
            AotCodegenBackend::Rust,
        );
        (
            artifact
                .into_rust_crate()
                .expect("Rust backend must return GeneratedAotCrate"),
            breakdown,
        )
    }

    /// Derives a complete machine-aware `Auto` recommendation for the current
    /// sparse BVP problem.
    ///
    /// The recommendation combines:
    /// - measured rayon overhead on the current machine,
    /// - coarse workload size of the prepared residual/Jacobian,
    /// - recommended residual/sparse chunking for future code generation,
    /// - and grouped runtime jobs for already compiled linked backends.
    ///
    /// This is intentionally an opt-in recommendation: existing explicit
    /// chunking defaults stay stable so previously generated AOT artifacts keep
    /// their current `problem_key`.
    pub fn auto_parallel_plan(&self) -> SparseAutoParallelPlan {
        recommended_sparse_auto_parallel_plan(
            self.atom_codegen
                .as_ref()
                .map_or(self.residuals.len(), |codegen| codegen.residuals.len()),
            self.shape.0,
            self.atom_codegen
                .as_ref()
                .map_or(self.sparse_entries.len(), |codegen| {
                    codegen.sparse_entries.len()
                }),
        )
    }

    /// Derives a machine-aware runtime parallel config for the current sparse
    /// BVP problem.
    ///
    /// Returns `None` when the measured machine overhead and current workload
    /// suggest that sequential execution is still the cheaper runtime path.
    pub fn auto_parallel_executor_config(&self) -> Option<ParallelExecutorConfig> {
        self.auto_parallel_plan().executor_config
    }
}

fn infer_atom_bandwidth(codegen: &PreparedSparseAtomBvpCodegen) -> (usize, usize) {
    codegen
        .sparse_entries
        .iter()
        .fold((0usize, 0usize), |(kl, ku), entry| {
            if entry.row >= entry.col {
                (kl.max(entry.row - entry.col), ku)
            } else {
                (kl, ku.max(entry.col - entry.row))
            }
        })
}

/// Resolves an in-process linked sparse AOT backend and wraps it into the
/// legacy solver callback traits used by the BVP runtime.
///
/// Dataflow:
/// 1. resolve a previously registered compiled backend by `problem_key`;
/// 2. prepend runtime parameter values to the Newton unknown vector;
/// 3. evaluate either whole-module callbacks or chunked callbacks;
/// 4. reassemble sparse values back into `SparseColMat` for the solver.
///
/// When no linked backend is registered yet, `None` is returned and the caller
/// remains on the lambdify/fallback branch.
fn linked_runtime_callbacks_for_matrix_backend(
    prepared_problem: &BvpPreparedSparseAotProblem,
    matrix_backend: MatrixBackend,
    parameter_values: Option<&[f64]>,
    parallel_config: Option<ParallelExecutorConfig>,
    banded_linear_solver_config: LinearSolverConfig,
) -> Option<(Box<dyn Fun>, Box<dyn Jac>)> {
    let linked = try_resolve_linked_sparse_backend(
        &prepared_problem.problem_key_for_matrix_backend(matrix_backend),
    )
    .ok()
    .flatten()?;
    let parameter_values = parameter_values.unwrap_or(&[]).to_vec();
    let shape = prepared_problem.shape;
    let residual_len = prepared_problem.shape.0;

    let residual_link = linked.clone();
    let residual_parameter_values = parameter_values.clone();
    let residual_parallel_config = parallel_config;
    let residual_aot_telemetry = prepared_problem
        .atom_aot_plan()
        .map(|plan| plan.telemetry().clone());
    let residual_fun = convert_to_fun(Box::new(move |_x: f64, vec: &dyn VectorType| {
        let telemetry_started = residual_aot_telemetry
            .as_ref()
            .and_then(|telemetry| telemetry.start_timing());
        let input_start = std::time::Instant::now();
        let dense = vec.to_DVectorType();
        let flat_args =
            flatten_runtime_args(Some(residual_parameter_values.as_slice()), dense.as_slice());
        record_callback_stage_time("Callback Residual Input Prep", input_start.elapsed());

        let mut out = vec![0.0; residual_len];
        let values_start = std::time::Instant::now();
        eval_linked_residual_outputs(
            &residual_link,
            residual_parallel_config,
            &flat_args,
            &mut out,
        );
        record_callback_stage_time("Callback Residual Values", values_start.elapsed());
        if let Some(telemetry) = &residual_aot_telemetry {
            telemetry.record_residual(
                telemetry_started,
                linked_execution_chunk_count(&residual_link, residual_parallel_config, false),
            );
        }

        let boxing_start = std::time::Instant::now();
        let result = Box::new(ColRef::from_slice(out.as_slice()).to_owned()) as Box<dyn VectorType>;
        record_callback_stage_time("Callback Residual Boxing", boxing_start.elapsed());
        result
    }));

    let jacobian_link = linked;
    let jacobian_parameter_values = parameter_values.clone();
    let jacobian_parallel_config = parallel_config;
    let jacobian_aot_telemetry = prepared_problem
        .atom_aot_plan()
        .map(|plan| plan.telemetry().clone());
    let jacobian: Box<dyn Jac> = match matrix_backend {
        MatrixBackend::Banded => {
            let structure = prepared_problem.banded_structure();
            let layout = prepared_problem
                .infer_node_major_layout()
                .expect("linked banded backend requires node-major compatible layout");
            let compact_layout = jacobian_link.jacobian_layout;
            convert_to_jac(Box::new(move |_x: f64, vec: &dyn VectorType| {
                let telemetry_started = jacobian_aot_telemetry
                    .as_ref()
                    .and_then(|telemetry| telemetry.start_timing());
                let input_start = std::time::Instant::now();
                let dense = vec.to_DVectorType();
                let flat_args = flatten_runtime_args(
                    Some(jacobian_parameter_values.as_slice()),
                    dense.as_slice(),
                );
                record_callback_stage_time("Callback Jacobian Input Prep", input_start.elapsed());

                let output_len = match compact_layout {
                    LinkedJacobianLayout::ExplicitValues => structure.nnz(),
                    LinkedJacobianLayout::BandedCompact { rows, cols, kl, ku } => {
                        BandedCompactJacobianStructure { rows, cols, kl, ku }
                            .storage_len()
                            .expect("linked compact Banded layout must be valid")
                    }
                };
                let mut values = vec![0.0; output_len];
                let values_start = std::time::Instant::now();
                eval_linked_sparse_values(
                    &jacobian_link,
                    jacobian_parallel_config,
                    &flat_args,
                    &mut values,
                );
                record_callback_stage_time("Callback Jacobian Values", values_start.elapsed());
                if let Some(telemetry) = &jacobian_aot_telemetry {
                    telemetry.record_jacobian(
                        telemetry_started,
                        linked_execution_chunk_count(
                            &jacobian_link,
                            jacobian_parallel_config,
                            true,
                        ),
                    );
                }

                let assembly_start = std::time::Instant::now();
                let assembly = match compact_layout {
                    LinkedJacobianLayout::ExplicitValues => {
                        structure.assemble_banded_assembly(values.as_slice())
                    }
                    LinkedJacobianLayout::BandedCompact { rows, cols, kl, ku } => {
                        let compact = BandedCompactJacobianStructure { rows, cols, kl, ku };
                        assert_eq!(
                            (rows, cols),
                            (structure.rows, structure.cols),
                            "linked compact Banded layout shape must match prepared structure"
                        );
                        compact
                            .try_to_assembly(values.as_slice())
                            .expect("linked compact Banded values must assemble")
                    }
                };
                let matrix =
                    BandedMatrixType::new(assembly, layout.clone(), banded_linear_solver_config);
                let boxed = Box::new(matrix) as Box<dyn MatrixType>;
                record_callback_stage_time(
                    "Callback Jacobian Matrix Assembly",
                    assembly_start.elapsed(),
                );
                boxed
            }))
        }
        _ => {
            // AtomView-native AOT keeps the compatibility Expr payload empty.
            // Read the fixed sparse coordinates from the prepared Atom codegen
            // payload instead of reconstructing them through the Expr bridge.
            let structure = prepared_problem.sparse_structure();
            convert_to_jac(Box::new(move |_x: f64, vec: &dyn VectorType| {
                let input_start = std::time::Instant::now();
                let dense = vec.to_DVectorType();
                let flat_args = flatten_runtime_args(
                    Some(jacobian_parameter_values.as_slice()),
                    dense.as_slice(),
                );
                record_callback_stage_time("Callback Jacobian Input Prep", input_start.elapsed());

                let mut values = vec![0.0; structure.nnz()];
                let values_start = std::time::Instant::now();
                eval_linked_sparse_values(
                    &jacobian_link,
                    jacobian_parallel_config,
                    &flat_args,
                    &mut values,
                );
                record_callback_stage_time("Callback Jacobian Values", values_start.elapsed());

                let assembly_start = std::time::Instant::now();
                let triplets = structure
                    .row_indices
                    .iter()
                    .zip(structure.col_indices.iter())
                    .zip(values.iter())
                    .map(|((&row, &col), &value)| Triplet::new(row, col, value))
                    .collect::<Vec<_>>();
                let matrix =
                    SparseColMat::try_new_from_triplets(shape.0, shape.1, triplets.as_slice())
                        .expect("linked sparse backend must return valid sparse values");
                let boxed = Box::new(matrix) as Box<dyn MatrixType>;
                record_callback_stage_time(
                    "Callback Jacobian Matrix Assembly",
                    assembly_start.elapsed(),
                );
                boxed
            }))
        }
    };

    Some((residual_fun, jacobian))
}

/// Splits a contiguous list of chunk descriptors into contiguous job ranges.
///
/// This keeps offsets monotonic inside each job, which later allows each worker
/// to fill one compact, disjoint slice of the final output buffer directly.
fn linked_contiguous_job_ranges(len: usize, max_jobs: usize) -> Vec<std::ops::Range<usize>> {
    if len == 0 {
        return Vec::new();
    }
    let target_jobs = len.min(max_jobs.max(1));
    let items_per_job = len.div_ceil(target_jobs);
    let mut ranges = Vec::with_capacity(target_jobs);
    let mut start = 0usize;
    while start < len {
        let end = (start + items_per_job).min(len);
        ranges.push(start..end);
        start = end;
    }
    ranges
}

/// Decides whether the linked compiled runtime should fan out work in parallel.
///
/// The decision depends on both the configured fallback policy and the actual
/// workload size so that tiny problems still stay on the cheaper sequential
/// callback path.
fn linked_should_parallelize(
    config: ParallelExecutorConfig,
    workload: usize,
    chunk_count: usize,
    sparse: bool,
) -> bool {
    if chunk_count <= 1 {
        return false;
    }
    match config.fallback_policy {
        ParallelFallbackPolicy::Auto => {
            let recommendation = if sparse {
                auto_parallel_recommendation(0, 0, workload, chunk_count)
            } else {
                auto_parallel_recommendation(workload, chunk_count, 0, 0)
            };
            recommendation.should_parallelize
        }
        ParallelFallbackPolicy::Never => true,
        ParallelFallbackPolicy::Thresholds {
            min_residual_outputs,
            min_sparse_values,
        } => {
            if sparse {
                workload >= min_sparse_values
            } else {
                workload >= min_residual_outputs
            }
        }
    }
}

/// Returns the effective number of parallel jobs allowed for linked compiled
/// residual or sparse-value evaluation.
fn linked_max_jobs(config: ParallelExecutorConfig, sparse: bool) -> usize {
    let workers = rayon::current_num_threads().max(1);
    let worker_jobs = workers.saturating_mul(config.jobs_per_worker.max(1));
    if sparse {
        config.max_sparse_jobs.unwrap_or(worker_jobs).max(1)
    } else {
        config.max_residual_jobs.unwrap_or(worker_jobs).max(1)
    }
}

/// Returns the number of generated callbacks used by one linked evaluation.
/// This mirrors the dispatch gate below so warm AOT telemetry reports actual
/// chunk work instead of the legacy placeholder value `1`.
fn linked_execution_chunk_count(
    linked: &crate::symbolic::codegen::codegen_aot_runtime_link::LinkedSparseAotBackend,
    parallel_config: Option<ParallelExecutorConfig>,
    sparse: bool,
) -> usize {
    let Some(config) = parallel_config else {
        return 1;
    };
    let (workload, chunk_count) = if sparse {
        // Compact Banded callbacks publish full storage length, which can be
        // larger than the symbolic explicit-entry count kept in `nnz`.
        (
            linked.jacobian_output_len().unwrap_or(linked.nnz),
            linked.jacobian_value_chunks.len(),
        )
    } else {
        (linked.residual_len, linked.residual_chunks.len())
    };
    if chunk_count == 0 || !linked_should_parallelize(config, workload, chunk_count, sparse) {
        return 1;
    }
    let jobs = linked_contiguous_job_ranges(chunk_count, linked_max_jobs(config, sparse));
    if jobs.len() > 1 { chunk_count } else { 1 }
}

/// Fallible compiled residual dispatch used by the prepared provider.
///
/// Unlike the historical callback closure below, this path keeps chunk
/// failures typed all the way through Rayon joins. Each worker writes only to
/// its disjoint output slice, so no mutex or temporary aggregation buffer is
/// needed.
fn try_eval_linked_residual_outputs(
    linked: &crate::symbolic::codegen::codegen_aot_runtime_link::LinkedSparseAotBackend,
    parallel_config: Option<ParallelExecutorConfig>,
    flat_args: &[f64],
    out: &mut [f64],
) -> Result<(), LinkedAotCallbackError> {
    if out.len() != linked.residual_len {
        return Err(LinkedAotCallbackError::OutputLength {
            stage: "residual",
            expected: linked.residual_len,
            actual: out.len(),
        });
    }
    let Some(config) = parallel_config else {
        return linked.try_residual_eval(flat_args, out);
    };
    if linked.residual_chunks.is_empty()
        || !linked_should_parallelize(
            config,
            linked.residual_len,
            linked.residual_chunks.len(),
            false,
        )
    {
        return linked.try_residual_eval(flat_args, out);
    }
    let job_ranges =
        linked_contiguous_job_ranges(linked.residual_chunks.len(), linked_max_jobs(config, false));
    if job_ranges.len() <= 1 {
        return linked.try_residual_eval(flat_args, out);
    }
    try_eval_linked_residual_job_range_recursive(
        linked,
        &job_ranges,
        flat_args,
        0,
        job_ranges.len(),
        out,
        0,
    )
}

fn try_eval_linked_residual_job_range_recursive(
    linked: &crate::symbolic::codegen::codegen_aot_runtime_link::LinkedSparseAotBackend,
    job_ranges: &[std::ops::Range<usize>],
    flat_args: &[f64],
    job_start: usize,
    job_end: usize,
    out: &mut [f64],
    base_offset: usize,
) -> Result<(), LinkedAotCallbackError> {
    let global_end = base_offset.saturating_add(out.len());
    match job_end - job_start {
        0 => Ok(()),
        1 => {
            for chunk_index in job_ranges[job_start].clone() {
                let chunk = &linked.residual_chunks[chunk_index];
                let chunk_end = chunk.output_offset.saturating_add(chunk.output_len);
                if chunk.output_offset < base_offset || chunk_end > global_end {
                    return Err(LinkedAotCallbackError::ChunkLayout {
                        stage: "residual",
                        index: chunk_index,
                        offset: chunk.output_offset,
                        len: chunk.output_len,
                        total: global_end,
                    });
                }
                let local_start = chunk.output_offset - base_offset;
                linked.try_residual_chunk_eval(
                    chunk_index,
                    flat_args,
                    &mut out[local_start..local_start + chunk.output_len],
                )?;
            }
            Ok(())
        }
        _ => {
            let mid = job_start + (job_end - job_start) / 2;
            let right_start = linked.residual_chunks[job_ranges[mid].start].output_offset;
            if right_start < base_offset || right_start > global_end {
                return Err(LinkedAotCallbackError::ChunkLayout {
                    stage: "residual",
                    index: job_ranges[mid].start,
                    offset: right_start,
                    len: 0,
                    total: global_end,
                });
            }
            let split_at = right_start - base_offset;
            let (left_out, right_out) = out.split_at_mut(split_at);
            let (left, right) = rayon::join(
                || {
                    try_eval_linked_residual_job_range_recursive(
                        linked,
                        job_ranges,
                        flat_args,
                        job_start,
                        mid,
                        left_out,
                        base_offset,
                    )
                },
                || {
                    try_eval_linked_residual_job_range_recursive(
                        linked,
                        job_ranges,
                        flat_args,
                        mid,
                        job_end,
                        right_out,
                        right_start,
                    )
                },
            );
            left?;
            right
        }
    }
}

/// Fallible compiled sparse-Jacobian dispatch used by the prepared provider.
fn try_eval_linked_sparse_values(
    linked: &crate::symbolic::codegen::codegen_aot_runtime_link::LinkedSparseAotBackend,
    parallel_config: Option<ParallelExecutorConfig>,
    flat_args: &[f64],
    values_out: &mut [f64],
) -> Result<(), LinkedAotCallbackError> {
    let expected_len =
        linked
            .jacobian_output_len()
            .map_err(|error| LinkedAotCallbackError::InvalidLayout {
                stage: "Jacobian",
                message: error.to_string(),
            })?;
    if values_out.len() != expected_len {
        return Err(LinkedAotCallbackError::OutputLength {
            stage: "Jacobian",
            expected: expected_len,
            actual: values_out.len(),
        });
    }
    let Some(config) = parallel_config else {
        return linked.try_jacobian_values_eval(flat_args, values_out);
    };
    if linked.jacobian_value_chunks.is_empty()
        || !linked_should_parallelize(
            config,
            expected_len,
            linked.jacobian_value_chunks.len(),
            true,
        )
    {
        return linked.try_jacobian_values_eval(flat_args, values_out);
    }
    let job_ranges = linked_contiguous_job_ranges(
        linked.jacobian_value_chunks.len(),
        linked_max_jobs(config, true),
    );
    if job_ranges.len() <= 1 {
        return linked.try_jacobian_values_eval(flat_args, values_out);
    }
    try_eval_linked_sparse_job_range_recursive(
        linked,
        &job_ranges,
        flat_args,
        0,
        job_ranges.len(),
        values_out,
        0,
    )
}

fn try_eval_linked_sparse_job_range_recursive(
    linked: &crate::symbolic::codegen::codegen_aot_runtime_link::LinkedSparseAotBackend,
    job_ranges: &[std::ops::Range<usize>],
    flat_args: &[f64],
    job_start: usize,
    job_end: usize,
    values_out: &mut [f64],
    base_offset: usize,
) -> Result<(), LinkedAotCallbackError> {
    let global_end = base_offset.saturating_add(values_out.len());
    match job_end - job_start {
        0 => Ok(()),
        1 => {
            for chunk_index in job_ranges[job_start].clone() {
                let chunk = &linked.jacobian_value_chunks[chunk_index];
                let chunk_end = chunk.value_offset.saturating_add(chunk.value_len);
                if chunk.value_offset < base_offset || chunk_end > global_end {
                    return Err(LinkedAotCallbackError::ChunkLayout {
                        stage: "Jacobian",
                        index: chunk_index,
                        offset: chunk.value_offset,
                        len: chunk.value_len,
                        total: global_end,
                    });
                }
                let local_start = chunk.value_offset - base_offset;
                linked.try_jacobian_chunk_eval(
                    chunk_index,
                    flat_args,
                    &mut values_out[local_start..local_start + chunk.value_len],
                )?;
            }
            Ok(())
        }
        _ => {
            let mid = job_start + (job_end - job_start) / 2;
            let right_start = linked.jacobian_value_chunks[job_ranges[mid].start].value_offset;
            if right_start < base_offset || right_start > global_end {
                return Err(LinkedAotCallbackError::ChunkLayout {
                    stage: "Jacobian",
                    index: job_ranges[mid].start,
                    offset: right_start,
                    len: 0,
                    total: global_end,
                });
            }
            let split_at = right_start - base_offset;
            let (left_out, right_out) = values_out.split_at_mut(split_at);
            let (left, right) = rayon::join(
                || {
                    try_eval_linked_sparse_job_range_recursive(
                        linked,
                        job_ranges,
                        flat_args,
                        job_start,
                        mid,
                        left_out,
                        base_offset,
                    )
                },
                || {
                    try_eval_linked_sparse_job_range_recursive(
                        linked,
                        job_ranges,
                        flat_args,
                        mid,
                        job_end,
                        right_out,
                        right_start,
                    )
                },
            );
            left?;
            right
        }
    }
}

/// Evaluates compiled residual outputs, optionally by dispatching contiguous
/// ranges of residual chunks in parallel.
///
/// Sequential and parallel branches must produce identical values; parallelism
/// only changes scheduling. Parallel jobs write directly into disjoint slices of
/// the caller-provided output buffer, avoiding per-call temporary buffers,
/// mutex aggregation, and copy-back overhead.
fn eval_linked_residual_outputs(
    linked: &crate::symbolic::codegen::codegen_aot_runtime_link::LinkedSparseAotBackend,
    parallel_config: Option<ParallelExecutorConfig>,
    flat_args: &[f64],
    out: &mut [f64],
) {
    let Some(config) = parallel_config else {
        linked
            .try_residual_eval(flat_args, out)
            .unwrap_or_else(|error| panic!("linked AOT residual callback failed: {error}"));
        return;
    };
    if linked.residual_chunks.is_empty()
        || !linked_should_parallelize(
            config,
            linked.residual_len,
            linked.residual_chunks.len(),
            false,
        )
    {
        linked
            .try_residual_eval(flat_args, out)
            .unwrap_or_else(|error| panic!("linked AOT residual callback failed: {error}"));
        return;
    }

    let job_ranges =
        linked_contiguous_job_ranges(linked.residual_chunks.len(), linked_max_jobs(config, false));
    if job_ranges.len() <= 1 {
        linked
            .try_residual_eval(flat_args, out)
            .unwrap_or_else(|error| panic!("linked AOT residual callback failed: {error}"));
        return;
    }

    eval_linked_residual_job_range_recursive(
        linked.residual_chunks.as_slice(),
        job_ranges.as_slice(),
        flat_args,
        0,
        job_ranges.len(),
        out,
        0,
    );
}

#[inline]
fn checked_linked_chunk_slice(
    stage: &'static str,
    index: usize,
    offset: usize,
    len: usize,
    base_offset: usize,
    output_len: usize,
) -> Result<(usize, usize), LinkedAotCallbackError> {
    let local_start =
        offset
            .checked_sub(base_offset)
            .ok_or(LinkedAotCallbackError::ChunkLayout {
                stage,
                index,
                offset,
                len,
                total: output_len,
            })?;
    let local_end = local_start
        .checked_add(len)
        .ok_or(LinkedAotCallbackError::ChunkLayout {
            stage,
            index,
            offset,
            len,
            total: output_len,
        })?;
    if local_end > output_len {
        return Err(LinkedAotCallbackError::ChunkLayout {
            stage,
            index,
            offset,
            len,
            total: output_len,
        });
    }
    Ok((local_start, local_end))
}

fn eval_linked_residual_job_range_recursive(
    chunks: &[crate::symbolic::codegen::codegen_aot_runtime_link::LinkedResidualChunk],
    job_ranges: &[std::ops::Range<usize>],
    flat_args: &[f64],
    job_start: usize,
    job_end: usize,
    out: &mut [f64],
    base_offset: usize,
) {
    match job_end - job_start {
        0 => {}
        1 => {
            for chunk_index in job_ranges[job_start].clone() {
                let chunk = &chunks[chunk_index];
                let (local_start, local_end) = checked_linked_chunk_slice(
                    "residual",
                    chunk_index,
                    chunk.output_offset,
                    chunk.output_len,
                    base_offset,
                    out.len(),
                )
                .unwrap_or_else(|error| panic!("linked AOT residual chunk layout failed: {error}"));
                invoke_linked_callback(
                    "residual chunk",
                    chunk.output_len,
                    &*chunk.eval,
                    flat_args,
                    &mut out[local_start..local_end],
                )
                .unwrap_or_else(|error| {
                    panic!("linked AOT residual chunk callback failed: {error}")
                });
            }
        }
        _ => {
            let mid = job_start + (job_end - job_start) / 2;
            let right_start = chunks[job_ranges[mid].start].output_offset;
            let (split_at, _) = checked_linked_chunk_slice(
                "residual",
                job_ranges[mid].start,
                right_start,
                0,
                base_offset,
                out.len(),
            )
            .unwrap_or_else(|error| panic!("linked AOT residual chunk layout failed: {error}"));
            let (left_out, right_out) = out.split_at_mut(split_at);
            rayon::join(
                || {
                    eval_linked_residual_job_range_recursive(
                        chunks,
                        job_ranges,
                        flat_args,
                        job_start,
                        mid,
                        left_out,
                        base_offset,
                    )
                },
                || {
                    eval_linked_residual_job_range_recursive(
                        chunks,
                        job_ranges,
                        flat_args,
                        mid,
                        job_end,
                        right_out,
                        right_start,
                    )
                },
            );
        }
    }
}

/// Evaluates compiled sparse Jacobian values, optionally by dispatching
/// contiguous ranges of Jacobian chunks in parallel.
///
/// Offsets come from the runtime chunk plan, so the values are written back in
/// exactly the same order as the symbolic sparse structure advertised to the
/// solver and lifecycle layers. Parallel jobs write directly into disjoint
/// slices of the caller-provided value buffer.
fn eval_linked_sparse_values(
    linked: &crate::symbolic::codegen::codegen_aot_runtime_link::LinkedSparseAotBackend,
    parallel_config: Option<ParallelExecutorConfig>,
    flat_args: &[f64],
    values_out: &mut [f64],
) {
    let Some(config) = parallel_config else {
        linked
            .try_jacobian_values_eval(flat_args, values_out)
            .unwrap_or_else(|error| panic!("linked AOT Jacobian callback failed: {error}"));
        return;
    };
    if linked.jacobian_value_chunks.is_empty()
        || !linked_should_parallelize(
            config,
            linked.jacobian_output_len().unwrap_or(linked.nnz),
            linked.jacobian_value_chunks.len(),
            true,
        )
    {
        linked
            .try_jacobian_values_eval(flat_args, values_out)
            .unwrap_or_else(|error| panic!("linked AOT Jacobian callback failed: {error}"));
        return;
    }

    let job_ranges = linked_contiguous_job_ranges(
        linked.jacobian_value_chunks.len(),
        linked_max_jobs(config, true),
    );
    if job_ranges.len() <= 1 {
        linked
            .try_jacobian_values_eval(flat_args, values_out)
            .unwrap_or_else(|error| panic!("linked AOT Jacobian callback failed: {error}"));
        return;
    }

    eval_linked_sparse_job_range_recursive(
        linked.jacobian_value_chunks.as_slice(),
        job_ranges.as_slice(),
        flat_args,
        0,
        job_ranges.len(),
        values_out,
        0,
    );
}

fn eval_linked_sparse_job_range_recursive(
    chunks: &[crate::symbolic::codegen::codegen_aot_runtime_link::LinkedSparseJacobianChunk],
    job_ranges: &[std::ops::Range<usize>],
    flat_args: &[f64],
    job_start: usize,
    job_end: usize,
    values_out: &mut [f64],
    base_offset: usize,
) {
    match job_end - job_start {
        0 => {}
        1 => {
            for chunk_index in job_ranges[job_start].clone() {
                let chunk = &chunks[chunk_index];
                let (local_start, local_end) = checked_linked_chunk_slice(
                    "Jacobian",
                    chunk_index,
                    chunk.value_offset,
                    chunk.value_len,
                    base_offset,
                    values_out.len(),
                )
                .unwrap_or_else(|error| panic!("linked AOT Jacobian chunk layout failed: {error}"));
                invoke_linked_callback(
                    "Jacobian chunk",
                    chunk.value_len,
                    &*chunk.eval,
                    flat_args,
                    &mut values_out[local_start..local_end],
                )
                .unwrap_or_else(|error| {
                    panic!("linked AOT Jacobian chunk callback failed: {error}")
                });
            }
        }
        _ => {
            let mid = job_start + (job_end - job_start) / 2;
            let right_start = chunks[job_ranges[mid].start].value_offset;
            let (split_at, _) = checked_linked_chunk_slice(
                "Jacobian",
                job_ranges[mid].start,
                right_start,
                0,
                base_offset,
                values_out.len(),
            )
            .unwrap_or_else(|error| panic!("linked AOT Jacobian chunk layout failed: {error}"));
            let (left_out, right_out) = values_out.split_at_mut(split_at);
            rayon::join(
                || {
                    eval_linked_sparse_job_range_recursive(
                        chunks,
                        job_ranges,
                        flat_args,
                        job_start,
                        mid,
                        left_out,
                        base_offset,
                    )
                },
                || {
                    eval_linked_sparse_job_range_recursive(
                        chunks,
                        job_ranges,
                        flat_args,
                        mid,
                        job_end,
                        right_out,
                        right_start,
                    )
                },
            );
        }
    }
}

/// Groups sparse symbolic entries into contiguous explicit-value chunks.
///
/// Generated sparse Jacobian callbacks write a compact local values slice
/// (`out[0..chunk_len]`).  The exported whole-AOT wrapper then places that slice
/// at `value_offset` in the global explicit-value buffer.  Because of that ABI,
/// every chunk must correspond to a contiguous segment of the original sparse
/// entry order.
fn chunk_sparse_entries_by_nnz<'a>(
    entries: &[SparseExprEntry<'a>],
    max_entries_per_chunk: usize,
) -> Vec<(usize, Vec<SparseExprEntry<'a>>)> {
    assert!(
        max_entries_per_chunk > 0,
        "max_entries_per_chunk must be positive"
    );

    entries
        .chunks(max_entries_per_chunk)
        .enumerate()
        .map(|(chunk_index, entries)| (chunk_index * max_entries_per_chunk, entries.to_vec()))
        .collect()
}

/// Groups sparse symbolic entries by row buckets when that still preserves
/// contiguous explicit-value slices.
///
/// Sparse entries are often stored in column-major order by the sparse
/// structure builder.  In that case a row bucket can interleave non-contiguous
/// positions from the global values array.  Such a chunk is not valid for the
/// generated local-output ABI, so we fall back to contiguous nnz chunks with the
/// same coarse target count.
fn chunk_sparse_entries_by_rows<'a>(
    entries: &[SparseExprEntry<'a>],
    rows_per_chunk: usize,
    total_rows: usize,
) -> Vec<(usize, Vec<SparseExprEntry<'a>>)> {
    let mut chunk_entries: Vec<Vec<(usize, SparseExprEntry<'a>)>> = Vec::new();
    let mut chunk_offsets: Vec<usize> = Vec::new();

    for (entry_index, entry) in entries.iter().copied().enumerate() {
        let bucket = entry.row / rows_per_chunk;
        if bucket >= chunk_entries.len() {
            chunk_entries.resize_with(bucket + 1, Vec::new);
            chunk_offsets.resize(bucket + 1, 0);
        }
        if chunk_entries[bucket].is_empty() {
            chunk_offsets[bucket] = entry_index;
        }
        chunk_entries[bucket].push((entry_index, entry));
    }

    let non_empty_chunks = chunk_entries
        .iter()
        .filter(|entries| !entries.is_empty())
        .collect::<Vec<_>>();
    let all_chunks_are_contiguous = non_empty_chunks.iter().all(|entries| {
        let start = entries.first().map(|(index, _)| *index).unwrap_or(0);
        entries
            .iter()
            .enumerate()
            .all(|(local_index, (entry_index, _))| *entry_index == start + local_index)
    });

    if !all_chunks_are_contiguous {
        let target_chunks = total_rows.max(1).div_ceil(rows_per_chunk).max(1);
        let max_entries_per_chunk = entries.len().max(1).div_ceil(target_chunks).max(1);
        return chunk_sparse_entries_by_nnz(entries, max_entries_per_chunk);
    }

    chunk_entries
        .into_iter()
        .enumerate()
        .filter(|(_, entries)| !entries.is_empty())
        .map(|(chunk_index, entries)| {
            (
                chunk_offsets[chunk_index],
                entries.into_iter().map(|(_, entry)| entry).collect(),
            )
        })
        .collect()
}

#[cfg(test)]
mod bvp_sparse_chunking_tests {
    use super::*;
    use crate::symbolic::codegen::codegen_aot_runtime_link::{
        LinkedResidualChunk, LinkedSparseAotBackend, LinkedSparseJacobianChunk,
        try_register_linked_sparse_backend, try_unregister_linked_sparse_backend,
    };
    use std::sync::Arc;
    use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
    use std::time::Duration;

    #[test]
    fn bvp_sparse_target_chunking_uses_contiguous_value_segments() {
        let e0 = Expr::parse_expression("y0");
        let e1 = Expr::parse_expression("y1");
        let e2 = Expr::parse_expression("y2");
        let e3 = Expr::parse_expression("y3");
        let entries = vec![
            SparseExprEntry {
                row: 0,
                col: 0,
                expr: &e0,
            },
            SparseExprEntry {
                row: 2,
                col: 0,
                expr: &e1,
            },
            SparseExprEntry {
                row: 1,
                col: 1,
                expr: &e2,
            },
            SparseExprEntry {
                row: 3,
                col: 1,
                expr: &e3,
            },
        ];

        let chunks = chunk_sparse_entries_by_nnz(&entries, 2);

        assert_eq!(chunks.len(), 2);
        assert_eq!(chunks[0].0, 0);
        assert_eq!(chunks[0].1[0].row, 0);
        assert_eq!(chunks[0].1[1].row, 2);
        assert_eq!(chunks[1].0, 2);
        assert_eq!(chunks[1].1[0].row, 1);
        assert_eq!(chunks[1].1[1].row, 3);
    }

    #[test]
    fn bvp_sparse_row_chunking_falls_back_when_value_segments_interleave() {
        let e0 = Expr::parse_expression("y0");
        let e1 = Expr::parse_expression("y1");
        let e2 = Expr::parse_expression("y2");
        let e3 = Expr::parse_expression("y3");
        let entries = vec![
            SparseExprEntry {
                row: 0,
                col: 0,
                expr: &e0,
            },
            SparseExprEntry {
                row: 2,
                col: 0,
                expr: &e1,
            },
            SparseExprEntry {
                row: 1,
                col: 1,
                expr: &e2,
            },
            SparseExprEntry {
                row: 3,
                col: 1,
                expr: &e3,
            },
        ];

        let chunks = chunk_sparse_entries_by_rows(&entries, 2, 4);

        assert_eq!(chunks.len(), 2);
        assert_eq!(chunks[0].0, 0);
        assert_eq!(chunks[0].1[0].row, 0);
        assert_eq!(chunks[0].1[1].row, 2);
        assert_eq!(chunks[1].0, 2);
        assert_eq!(chunks[1].1[0].row, 1);
        assert_eq!(chunks[1].1[1].row, 3);
    }

    #[test]
    fn typed_parallel_aot_dispatch_preserves_sparse_and_residual_layouts() {
        let linked = LinkedSparseAotBackend::new(
            "typed-parallel-dispatch",
            4,
            (4, 4),
            4,
            Arc::new(|_, out| out.fill(-1.0)),
            Arc::new(|_, out| out.fill(-2.0)),
        )
        .with_chunked_evaluators(
            (0..2)
                .map(|chunk_index| {
                    LinkedResidualChunk::new(
                        chunk_index * 2,
                        2,
                        Arc::new(move |args, out| {
                            out[0] = args[0] + (chunk_index * 2) as f64;
                            out[1] = args[0] + (chunk_index * 2 + 1) as f64;
                        }),
                    )
                })
                .collect(),
            (0..2)
                .map(|chunk_index| {
                    LinkedSparseJacobianChunk::new(
                        chunk_index * 2,
                        2,
                        Arc::new(move |args, out| {
                            out[0] = args[0] * 10.0 + (chunk_index * 2) as f64;
                            out[1] = args[0] * 10.0 + (chunk_index * 2 + 1) as f64;
                        }),
                    )
                })
                .collect(),
        );
        let config = ParallelExecutorConfig {
            jobs_per_worker: 1,
            max_residual_jobs: Some(2),
            max_sparse_jobs: Some(2),
            fallback_policy: ParallelFallbackPolicy::Never,
        };

        let mut residual = vec![0.0; 4];
        try_eval_linked_residual_outputs(&linked, Some(config), &[3.0], &mut residual)
            .expect("typed residual dispatch should succeed");
        assert_eq!(residual, vec![3.0, 4.0, 5.0, 6.0]);

        let mut values = vec![0.0; 4];
        try_eval_linked_sparse_values(&linked, Some(config), &[3.0], &mut values)
            .expect("typed sparse dispatch should succeed");
        assert_eq!(values, vec![30.0, 31.0, 32.0, 33.0]);

        let malformed = LinkedSparseAotBackend::new(
            "typed-parallel-malformed",
            4,
            (4, 4),
            4,
            Arc::new(|_, out| out.fill(0.0)),
            Arc::new(|_, out| out.fill(0.0)),
        )
        .with_chunked_evaluators(
            vec![
                LinkedResidualChunk::new(0, 2, Arc::new(|_, out| out.fill(0.0))),
                LinkedResidualChunk::new(3, 2, Arc::new(|_, out| out.fill(0.0))),
            ],
            Vec::new(),
        );
        let malformed_result =
            try_eval_linked_residual_outputs(&malformed, Some(config), &[1.0], &mut residual);
        assert!(
            matches!(
                malformed_result,
                Err(LinkedAotCallbackError::ChunkLayout {
                    stage: "residual",
                    index: 1,
                    ..
                })
            ),
            "unexpected malformed chunk result: {malformed_result:?}"
        );
    }

    #[test]
    fn linked_sparse_parallel_callbacks_write_directly_into_final_buffers() {
        fn slice_inside_final_buffer(slice: &[f64], base: usize, len: usize) -> bool {
            let ptr = slice.as_ptr() as usize;
            let end = base + len * std::mem::size_of::<f64>();
            ptr >= base && ptr < end
        }

        let residual_base = Arc::new(AtomicUsize::new(0));
        let sparse_base = Arc::new(AtomicUsize::new(0));
        let residual_direct = Arc::new(AtomicBool::new(true));
        let sparse_direct = Arc::new(AtomicBool::new(true));

        let residual_chunk =
            |offset: usize, len: usize, base: Arc<AtomicUsize>, direct: Arc<AtomicBool>| {
                LinkedResidualChunk::new(
                    offset,
                    len,
                    Arc::new(move |args: &[f64], out: &mut [f64]| {
                        direct.fetch_and(
                            slice_inside_final_buffer(out, base.load(Ordering::SeqCst), 4),
                            Ordering::SeqCst,
                        );
                        for (i, value) in out.iter_mut().enumerate() {
                            *value = args[0] + (offset + i) as f64;
                        }
                    }),
                )
            };
        let sparse_chunk =
            |offset: usize, len: usize, base: Arc<AtomicUsize>, direct: Arc<AtomicBool>| {
                LinkedSparseJacobianChunk::new(
                    offset,
                    len,
                    Arc::new(move |args: &[f64], out: &mut [f64]| {
                        direct.fetch_and(
                            slice_inside_final_buffer(out, base.load(Ordering::SeqCst), 4),
                            Ordering::SeqCst,
                        );
                        for (i, value) in out.iter_mut().enumerate() {
                            *value = args[0] * 10.0 + (offset + i) as f64;
                        }
                    }),
                )
            };

        let linked = LinkedSparseAotBackend::new(
            "direct-linked-test",
            4,
            (4, 4),
            4,
            Arc::new(|_, out: &mut [f64]| out.fill(-1.0)),
            Arc::new(|_, out: &mut [f64]| out.fill(-2.0)),
        )
        .with_chunked_evaluators(
            vec![
                residual_chunk(
                    0,
                    2,
                    Arc::clone(&residual_base),
                    Arc::clone(&residual_direct),
                ),
                residual_chunk(
                    2,
                    2,
                    Arc::clone(&residual_base),
                    Arc::clone(&residual_direct),
                ),
            ],
            vec![
                sparse_chunk(0, 2, Arc::clone(&sparse_base), Arc::clone(&sparse_direct)),
                sparse_chunk(2, 2, Arc::clone(&sparse_base), Arc::clone(&sparse_direct)),
            ],
        );
        let config = ParallelExecutorConfig {
            jobs_per_worker: 1,
            max_residual_jobs: Some(2),
            max_sparse_jobs: Some(2),
            fallback_policy: ParallelFallbackPolicy::Never,
        };

        let mut residual = vec![0.0; 4];
        residual_base.store(residual.as_mut_ptr() as usize, Ordering::SeqCst);
        eval_linked_residual_outputs(&linked, Some(config), &[3.0], &mut residual);
        assert_eq!(residual, vec![3.0, 4.0, 5.0, 6.0]);
        assert!(
            residual_direct.load(Ordering::SeqCst),
            "linked residual chunk callbacks must receive final output slices, not temporary buffers"
        );

        let mut sparse_values = vec![0.0; 4];
        sparse_base.store(sparse_values.as_mut_ptr() as usize, Ordering::SeqCst);
        eval_linked_sparse_values(&linked, Some(config), &[3.0], &mut sparse_values);
        assert_eq!(sparse_values, vec![30.0, 31.0, 32.0, 33.0]);
        assert!(
            sparse_direct.load(Ordering::SeqCst),
            "linked sparse chunk callbacks must receive final value slices, not temporary buffers"
        );
    }

    #[test]
    fn linked_sparse_parallel_callbacks_actually_overlap_on_rayon_workers() {
        fn bump_max(max_active: &AtomicUsize, current: usize) {
            let mut observed = max_active.load(Ordering::SeqCst);
            while current > observed {
                match max_active.compare_exchange(
                    observed,
                    current,
                    Ordering::SeqCst,
                    Ordering::SeqCst,
                ) {
                    Ok(_) => break,
                    Err(next_observed) => observed = next_observed,
                }
            }
        }

        let active_residual = Arc::new(AtomicUsize::new(0));
        let max_residual = Arc::new(AtomicUsize::new(0));
        let active_sparse = Arc::new(AtomicUsize::new(0));
        let max_sparse = Arc::new(AtomicUsize::new(0));

        let residual_chunk =
            |offset: usize, active: Arc<AtomicUsize>, max_active: Arc<AtomicUsize>| {
                LinkedResidualChunk::new(
                    offset,
                    1,
                    Arc::new(move |args: &[f64], out: &mut [f64]| {
                        let current = active.fetch_add(1, Ordering::SeqCst) + 1;
                        bump_max(&max_active, current);
                        std::thread::sleep(Duration::from_millis(25));
                        out[0] = args[0] + offset as f64;
                        active.fetch_sub(1, Ordering::SeqCst);
                    }),
                )
            };
        let sparse_chunk =
            |offset: usize, active: Arc<AtomicUsize>, max_active: Arc<AtomicUsize>| {
                LinkedSparseJacobianChunk::new(
                    offset,
                    1,
                    Arc::new(move |args: &[f64], out: &mut [f64]| {
                        let current = active.fetch_add(1, Ordering::SeqCst) + 1;
                        bump_max(&max_active, current);
                        std::thread::sleep(Duration::from_millis(25));
                        out[0] = args[0] * 10.0 + offset as f64;
                        active.fetch_sub(1, Ordering::SeqCst);
                    }),
                )
            };

        let linked = LinkedSparseAotBackend::new(
            "parallel-overlap-test",
            4,
            (4, 4),
            4,
            Arc::new(|_, out: &mut [f64]| out.fill(-1.0)),
            Arc::new(|_, out: &mut [f64]| out.fill(-2.0)),
        )
        .with_chunked_evaluators(
            (0..4)
                .map(|offset| {
                    residual_chunk(
                        offset,
                        Arc::clone(&active_residual),
                        Arc::clone(&max_residual),
                    )
                })
                .collect(),
            (0..4)
                .map(|offset| {
                    sparse_chunk(offset, Arc::clone(&active_sparse), Arc::clone(&max_sparse))
                })
                .collect(),
        );
        let config = ParallelExecutorConfig {
            jobs_per_worker: 1,
            max_residual_jobs: Some(4),
            max_sparse_jobs: Some(4),
            fallback_policy: ParallelFallbackPolicy::Never,
        };

        rayon::ThreadPoolBuilder::new()
            .num_threads(4)
            .build()
            .expect("local rayon pool should build")
            .install(|| {
                let mut residual = vec![0.0; 4];
                eval_linked_residual_outputs(&linked, Some(config), &[2.0], &mut residual);
                assert_eq!(residual, vec![2.0, 3.0, 4.0, 5.0]);

                let mut sparse_values = vec![0.0; 4];
                eval_linked_sparse_values(&linked, Some(config), &[2.0], &mut sparse_values);
                assert_eq!(sparse_values, vec![20.0, 21.0, 22.0, 23.0]);
            });

        assert!(
            max_residual.load(Ordering::SeqCst) > 1,
            "linked residual chunk callbacks must overlap; max_active={}",
            max_residual.load(Ordering::SeqCst)
        );
        assert!(
            max_sparse.load(Ordering::SeqCst) > 1,
            "linked sparse chunk callbacks must overlap; max_active={}",
            max_sparse.load(Ordering::SeqCst)
        );
    }

    #[test]
    fn compact_banded_solver_parity_rebind_and_telemetry() {
        let _test_report = crate::Utils::test_reporting::TestReportCapture::new(
            "BVP_Damp_AOT",
            concat!(
                module_path!(),
                "::compact_banded_solver_parity_rebind_and_telemetry"
            ),
        );

        let equations = vec![Expr::parse_expression("z"), Expr::parse_expression("-y")];
        let values = vec!["y".to_string(), "z".to_string()];
        let boundary_conditions = HashMap::from([
            ("y".to_string(), vec![(0usize, 1.0)]),
            ("z".to_string(), vec![(0usize, 0.0)]),
        ]);
        let discretized = discretization_system_bvp_par_atom(
            equations,
            values,
            "x".to_string(),
            0.0,
            Some(4),
            None,
            Some((0..=4).map(|index| index as f64 / 4.0).collect()),
            boundary_conditions,
            None,
            None,
            "forward".to_string(),
        );
        let codegen =
            crate::symbolic::View::bvp_codegen::prepare_sparse_bvp_codegen_from_discretized_system(
                &discretized,
                "eval_compact_residual",
                "eval_compact_jacobian",
                vec!["p".to_string()],
                None,
                ResidualChunkingStrategy::Whole,
                SparseChunkingStrategy::Whole,
            );
        let (kl, ku) = infer_atom_bandwidth(&codegen);
        let codegen = codegen
            .with_aot_telemetry_mode(BvpAotTelemetryMode::Detailed)
            .with_native_banded_layout(kl, ku)
            .expect("whole compact Banded codegen plan should be valid");
        let plan = codegen
            .prepared_aot_plan()
            .expect("compact Banded AOT plan should be valid");
        let (rows, cols) = codegen.shape;
        let storage_len = plan.matrix_layout().value_count();
        assert_eq!(rows, cols);
        assert_eq!(storage_len, (kl + ku + 1) * cols);

        // The solver bridge owns the AtomView plan. Its compatibility Expr
        // vectors intentionally stay empty: compact callbacks use only the
        // prepared native layout and node-major variable names.
        let prepared = BvpPreparedSparseAotProblem {
            residual_fn_name: "eval_compact_residual".to_string(),
            jacobian_fn_name: "eval_compact_jacobian".to_string(),
            variable_names: codegen.variable_names.clone(),
            param_names: codegen.param_names.clone(),
            residuals: Vec::new(),
            sparse_entries: Vec::new(),
            shape: codegen.shape,
            residual_strategy: ResidualChunkingStrategy::Whole,
            jacobian_strategy: SparseChunkingStrategy::Whole,
            bandwidth: Some((kl, ku)),
            symbolic_assembly_backend: BvpSymbolicAssemblyBackend::AtomView,
            atom_discretized_system: Some(discretized),
            atom_codegen: Some(codegen),
            atom_codegen_prep_breakdown: None,
            atom_aot_plan: Some(plan),
        };

        let (artifact, _) = prepared.generated_native_banded_aot_artifact_with_breakdown(
            "compact_banded_builder_gate",
            "generated_compact_banded",
            AotCodegenBackend::Rust,
            AtomOptimizationProfile::FastBootstrap,
        );
        let generated = match artifact {
            GeneratedAotArtifact::Rust(crate_spec) => crate_spec,
            other => panic!("expected Rust artifact, got {:?}", other.backend()),
        };
        assert_eq!(
            generated.manifest.io.jacobian_layout,
            Some(
                crate::symbolic::codegen::codegen_manifest::PreparedJacobianLayout::BandedCompact {
                    kl,
                    ku,
                }
            )
        );
        assert_eq!(generated.manifest.io.jacobian_nnz, Some(storage_len));
        assert!(generated.module_source.contains("eval_compact_jacobian"));

        let problem_key = prepared.problem_key_for_matrix_backend(MatrixBackend::Banded);
        let residual_len = prepared.shape.0;
        let jacobian_backend = LinkedSparseAotBackend::new(
            problem_key.clone(),
            residual_len,
            prepared.shape,
            storage_len,
            Arc::new(move |args, out| {
                let parameter = args.first().copied().unwrap_or_default();
                for (index, value) in out.iter_mut().enumerate() {
                    *value = parameter + index as f64;
                }
            }),
            Arc::new(move |args, out| {
                let parameter = args.first().copied().unwrap_or_default();
                out.fill(0.0);
                for index in 0..cols {
                    out[ku * cols + index] = parameter + 10.0;
                }
            }),
        )
        .with_banded_compact_layout(rows, cols, kl, ku);
        try_register_linked_sparse_backend(jacobian_backend)
            .expect("compact solver callback should register");

        let make_callbacks = |parameter: f64| {
            linked_runtime_callbacks_for_matrix_backend(
                &prepared,
                MatrixBackend::Banded,
                Some(&[parameter]),
                None,
                LinearSolverConfig::faithful_banded(),
            )
            .expect("registered compact solver callback should resolve")
        };

        let (residual_first, mut jacobian_first) = make_callbacks(2.0);
        let variables = DVector::from_element(cols, 0.25);
        let typed_variables = Vectors_type_casting(&variables, "Banded".to_string());
        let residual = residual_first.call(0.0, &*typed_variables).to_DVectorType();
        assert_eq!(residual[0], 2.0);
        assert_eq!(residual[residual_len - 1], 2.0 + (residual_len - 1) as f64);
        let first_matrix = jacobian_first.call(0.0, &*typed_variables);
        let first_banded = first_matrix
            .as_any()
            .downcast_ref::<BandedMatrixType>()
            .expect("compact callback must publish native BandedMatrixType");
        assert_eq!(
            first_banded.assembly.diag(0).unwrap(),
            vec![12.0; cols].as_slice()
        );

        let (residual_second, mut jacobian_second) = make_callbacks(5.0);
        let residual = residual_second
            .call(0.0, &*typed_variables)
            .to_DVectorType();
        assert_eq!(residual[0], 5.0);
        let second_matrix = jacobian_second.call(0.0, &*typed_variables);
        let second_banded = second_matrix
            .as_any()
            .downcast_ref::<BandedMatrixType>()
            .expect("rebound compact callback must publish native BandedMatrixType");
        assert_eq!(
            second_banded.assembly.diag(0).unwrap(),
            vec![15.0; cols].as_slice()
        );
        assert!(
            !std::sync::Arc::ptr_eq(&first_banded.cached_solver, &second_banded.cached_solver,),
            "parameter rebind must not reuse the previous Banded factor owner"
        );

        let snapshot = prepared
            .atom_aot_plan()
            .expect("prepared bridge must retain its AtomView plan")
            .telemetry_snapshot();
        assert_eq!(snapshot.residual_calls, 2);
        assert_eq!(snapshot.jacobian_calls, 2);
        assert_eq!(snapshot.residual_chunks, 2);
        assert_eq!(snapshot.jacobian_chunks, 2);
        assert!(snapshot.residual_elapsed > Duration::ZERO);
        assert!(snapshot.jacobian_elapsed > Duration::ZERO);
        crate::Utils::test_reporting::capture_test_line(format_args!(
            "[BVP compact Banded solver contract] shape={} storage={} residual_calls={} jacobian_calls={} residual_ms={:.3} jacobian_ms={:.3}",
            rows,
            storage_len,
            snapshot.residual_calls,
            snapshot.jacobian_calls,
            snapshot.residual_elapsed.as_secs_f64() * 1_000.0,
            snapshot.jacobian_elapsed.as_secs_f64() * 1_000.0,
        ));

        try_unregister_linked_sparse_backend(&problem_key)
            .expect("compact solver callback should unregister");
    }
}

impl Jacobian {
    /// Runs the shared symbolic BVP preparation stage before backend-specific
    /// lambdify/AOT branching.
    ///
    /// This stage performs discretization, symbolic Jacobian construction,
    /// bandwidth discovery or enforcement, and fills the timer map used by the
    /// compatibility reporting path.
    fn prepare_symbolic_bvp_stage_with_params(
        &mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
    ) -> (
        String,
        Vec<String>,
        BvpMatrixBackend,
        HashMap<String, f64>,
        Instant,
    ) {
        let total_start = Instant::now();
        let mut timer_hash: HashMap<String, f64> = HashMap::new();
        self.set_params(params);
        self.atom_discretized_system = None;
        self.atom_symbolic_jacobian_sparse = None;
        let param = arg.clone();
        let matrix_backend = BvpMatrixBackend::from_legacy_method(method.as_str())
            .unwrap_or_else(|| panic!("unknown method: {}", method));
        self.set_backend_config(BvpBackendConfig::lambdify(matrix_backend));

        let begin = Instant::now();
        match self.symbolic_assembly_backend {
            BvpSymbolicAssemblyBackend::ExprLegacy => self.discretization_system_BVP_par(
                eq_system,
                values,
                arg,
                t0,
                n_steps,
                h,
                mesh,
                BorderConditions,
                Bounds,
                rel_tolerance,
                scheme,
            ),
            BvpSymbolicAssemblyBackend::AtomView => {
                let discretized = discretization_system_bvp_par_atom(
                    eq_system,
                    values,
                    arg,
                    t0,
                    n_steps,
                    h,
                    mesh,
                    BorderConditions,
                    Bounds,
                    rel_tolerance,
                    scheme,
                );
                self.install_atom_discretized_system(discretized);
            }
        }
        timer_hash.insert(
            "discretization time".to_string(),
            begin.elapsed().as_secs_f64(),
        );
        info!("system discretized");

        let indexed_values = self.variable_string.clone();
        // The AtomView Jacobian plan is stored as ordered sparse entries, but
        // those entries are also the correct source for the Dense Lambdify
        // callback. Restricting this branch to Sparse/Banded leaves Dense
        // AtomView with an empty ExprLegacy vector and breaks the modern
        // symbolic-assembly contract.
        let atom_native_jacobian =
            self.symbolic_assembly_backend == BvpSymbolicAssemblyBackend::AtomView;

        if let Some(bandwidth_) = bandwidth {
            let now = Instant::now();
            self.bandwidth = Some(bandwidth_);
            info!("Bandwidth provided: {:?}", self.bandwidth);
            timer_hash.insert(
                "find bandwidth time".to_string(),
                now.elapsed().as_secs_f64(),
            );

            info!("Calculating jacobian");
            let now = Instant::now();
            if atom_native_jacobian {
                self.calc_atomview_sparse_jacobian_with_bandwidth(Some(bandwidth_));
            } else {
                self.calc_jacobian_parallel_smart_optimized_with_given_bandwidth();
            }
            let elapsed = now.elapsed();
            info!("Jacobian calculation time:");
            info!("{:?}", elapsed_time(elapsed));
            timer_hash.insert("symbolic jacobian time".to_string(), elapsed.as_secs_f64());
        } else {
            info!("Calculating jacobian");
            let now = Instant::now();
            if atom_native_jacobian {
                self.calc_atomview_sparse_jacobian_with_bandwidth(None);
            } else {
                self.calc_jacobian_parallel_smart_optimized();
            }
            let elapsed = now.elapsed();
            info!("Jacobian calculation time:");
            info!("{:?}", elapsed_time(elapsed));
            timer_hash.insert("symbolic jacobian time".to_string(), elapsed.as_secs_f64());

            let now = Instant::now();
            if atom_native_jacobian {
                self.find_atom_bandwidths();
            } else {
                self.find_bandwidths();
            }
            info!("Bandwidth calculated:");
            info!("(kl, ku) = {:?}", self.bandwidth);
            timer_hash.insert(
                "find bandwidth time".to_string(),
                now.elapsed().as_secs_f64(),
            );
        }
        if let Some(jacobian_breakdown) = self.last_symbolic_jacobian_timer_snapshot.take() {
            timer_hash.extend(jacobian_breakdown);
        }

        (
            param,
            indexed_values,
            matrix_backend,
            timer_hash,
            total_start,
        )
    }

    /// Finalizes the legacy timing table for `generate_BVP*` style entrypoints.
    ///
    /// The modern lifecycle path logs through structured stages, while this
    /// helper keeps the historical human-readable timing summary stable for the
    /// older BVP APIs and tests that still print it.
    fn finalize_generate_bvp_timer_table(
        &mut self,
        mut timer_hash: HashMap<String, f64>,
        total_end: f64,
    ) {
        let mut raw_snapshot = timer_hash.clone();
        raw_snapshot.insert("total time, sec".to_string(), total_end);
        self.last_generate_timer_snapshot = Some(raw_snapshot);
        for key in [
            "discretization time",
            "symbolic jacobian time",
            "jacobian lambdify time",
            "residual functions lambdify time",
            "find bandwidth time",
            "symbolic jacobian variable sets time",
            "symbolic jacobian row differentiation time",
            "symbolic jacobian dense cache materialize time",
            "symbolic jacobian sparse cache flatten time",
            "sparse AOT preparation time",
            "backend selection time",
            "runtime binding time",
            "lambdify jacobian callback compile time",
            "lambdify residual callback compile time",
        ] {
            if let Some(value) = timer_hash.get_mut(key) {
                normalize_timer_percent(value, total_end);
            }
        }
        timer_hash.insert("total time, sec".to_string(), total_end);

        let mut table = Builder::from(timer_hash.clone()).build();
        table.with(Style::modern_rounded());
        println!("{}", table.to_string());
        info!(
            "\n \n ____________END OF GENERATE BVP ________________________________________________________________"
        );
    }

    /// Creates a new Jacobian instance with default values.
    ///
    /// Initializes all fields to empty/default states and sets up a dummy residual function
    /// that returns the input vector unchanged. This serves as a safe default until the
    /// actual BVP system is discretized and compiled.
    ///
    /// # Returns
    /// A new Jacobian instance ready for BVP discretization and symbolic computation.
    pub fn new() -> Self {
        let fun0: Box<dyn Fn(f64, &DVector<f64>) -> DVector<f64>> =
            Box::new(|_x, y: &DVector<f64>| y.clone());
        let boxed_fun: Box<dyn Fun> = Box::new(FunEnum::Dense(fun0));
        Self {
            vector_of_functions: Vec::new(),
            lambdified_functions: Vec::new(),
            method: String::new(),
            backend_config: BvpBackendConfig::default(),
            banded_linear_solver_config: LinearSolverConfig::default(),
            banded_lambdify_config: BandedLambdifyConfig::default(),
            symbolic_assembly_backend: BvpSymbolicAssemblyBackend::ExprLegacy,
            vector_of_variables: Vec::new(),
            variable_string: Vec::new(),
            parameters_string: Vec::new(),
            parameter_values: None,
            parameter_binding: None,
            symbolic_jacobian: Vec::new(),
            symbolic_jacobian_sparse: Vec::new(),
            atom_symbolic_jacobian_sparse: None,
            lambdified_jac_element: None,
            jac_function: None,
            residiual_function: boxed_fun,
            bounds: None,
            rel_tolerance_vec: None,
            bandwidth: None,
            variables_for_all_disrete: Vec::new(),
            BC_pos_n_values: Vec::new(),
            last_generate_timer_snapshot: None,
            last_symbolic_jacobian_timer_snapshot: None,
            last_symbolic_jacobian_telemetry_snapshot: None,
            last_lambdify_binding_timer_snapshot: None,
            atom_discretized_system: None,
            direct_banded_jacobian_telemetry: None,
            legacy_lambdify_telemetry: None,
            atom_lambdify_telemetry: None,
            lambdify_telemetry_mode: BvpLambdifyTelemetryMode::Off,
            aot_telemetry_mode: BvpAotTelemetryMode::Off,
            lambdify_execution_policy: BvpLambdifyExecutionPolicy::default(),
        }
    }

    /// Sets the native linear solver configuration used by banded runtime callbacks.
    pub fn set_banded_linear_solver_config(&mut self, config: LinearSolverConfig) {
        self.banded_linear_solver_config = config;
    }

    /// Sets banded-lambdify runtime controls (chunking/threshold).
    pub fn set_banded_lambdify_config(&mut self, config: BandedLambdifyConfig) {
        self.banded_lambdify_config = config;
    }

    /// Sets only the Jacobian chunking mode for the banded lambdify path.
    pub fn set_banded_jacobian_chunking(&mut self, chunking: BandedJacobianChunking) {
        self.banded_lambdify_config.jacobian_chunking = chunking;
    }

    /// Selects the runtime telemetry level for ExprLegacy/AtomView Lambdify
    /// callbacks. The setting affects callbacks compiled after this call.
    pub fn set_lambdify_telemetry_mode(&mut self, mode: BvpLambdifyTelemetryMode) {
        self.lambdify_telemetry_mode = mode;
        self.legacy_lambdify_telemetry = None;
        self.atom_lambdify_telemetry = None;
    }

    /// Returns the runtime telemetry mode used by newly installed callbacks.
    pub fn lambdify_telemetry_mode(&self) -> BvpLambdifyTelemetryMode {
        self.lambdify_telemetry_mode
    }

    /// Selects the typed telemetry level for AtomView AOT preparation.
    ///
    /// The setting affects subsequently prepared AOT plans and does not
    /// change the ExprLegacy compatibility adapter.
    pub fn set_aot_telemetry_mode(&mut self, mode: BvpAotTelemetryMode) {
        self.aot_telemetry_mode = mode;
    }

    /// Returns the telemetry level used by subsequently prepared AOT plans.
    pub const fn aot_telemetry_mode(&self) -> BvpAotTelemetryMode {
        self.aot_telemetry_mode
    }

    /// Selects the runtime execution policy for newly compiled Lambdify
    /// residual/Jacobian callbacks.
    pub fn set_lambdify_execution_policy(&mut self, policy: BvpLambdifyExecutionPolicy) {
        self.lambdify_execution_policy = policy;
        self.banded_lambdify_config.execution_policy = policy;
    }

    /// Returns the runtime execution policy used by newly installed callbacks.
    pub fn lambdify_execution_policy(&self) -> BvpLambdifyExecutionPolicy {
        self.lambdify_execution_policy
    }

    /// Creates the owned direct-path snapshot used by the no-Mutex Banded runtime.
    ///
    /// This is the only compatibility boundary from the historical `Jacobian`
    /// host into the direct module. The direct module does not implement methods
    /// on `Jacobian` and therefore cannot accidentally route new work through a
    /// legacy callback or mutable legacy state.
    pub fn direct_banded_problem(&self) -> DirectBandedProblem {
        let problem = DirectBandedProblem::new(
            self.vector_of_functions.clone(),
            self.vector_of_variables.clone(),
            self.variable_string.clone(),
            self.parameters_string.clone(),
            self.parameter_values.clone(),
            self.symbolic_jacobian_sparse.clone(),
            self.bandwidth,
        );
        let problem = self
            .parameter_binding
            .clone()
            .map_or(problem.clone(), |binding| {
                problem.with_parameter_binding(binding)
            });
        match (
            self.atom_discretized_system.as_ref(),
            self.atom_symbolic_jacobian_sparse.as_ref(),
        ) {
            (Some(discretized), Some(jacobian)) => problem
                .with_atom_view_data(discretized.vector_of_functions.clone(), jacobian.clone()),
            _ => problem,
        }
    }

    /// Compatibility adapter for direct Banded structure inference.
    pub fn infer_banded_structure_plan(&self) -> Result<BandedStructurePlan, BandedError> {
        self.direct_banded_problem().infer_banded_structure_plan()
    }

    /// Compatibility adapter for native Banded storage allocation.
    pub fn allocate_banded_assembly_from_plan(
        &self,
        plan: &BandedStructurePlan,
    ) -> Result<BandedAssembly, BandedError> {
        self.direct_banded_problem()
            .allocate_banded_assembly_from_plan(plan)
    }

    /// Compatibility adapter for native Banded storage allocation.
    pub fn allocate_banded_assembly(&self) -> Result<BandedAssembly, BandedError> {
        self.direct_banded_problem().allocate_banded_assembly()
    }

    /// Compatibility adapter for the direct-path structure heuristic.
    pub fn is_block_tridiagonal_candidate(&self) -> bool {
        self.direct_banded_problem()
            .is_block_tridiagonal_candidate()
    }

    /// Compatibility adapter retaining the historical closure-only API.
    pub(crate) fn generate_banded_jacobian_assembly_from_plan_parallel(
        &self,
        plan: &BandedStructurePlan,
        config: &BandedLambdifyConfig,
    ) -> Result<Box<dyn Fn(&[f64]) -> Result<BandedAssembly, BandedError> + Send + Sync>, BandedError>
    {
        self.direct_banded_problem()
            .generate_banded_jacobian_assembly_from_plan_parallel(plan, config)
    }

    /// Compatibility adapter retaining the telemetry-bearing direct runtime.
    pub fn generate_banded_jacobian_runtime_from_plan_parallel(
        &self,
        plan: &BandedStructurePlan,
        config: &BandedLambdifyConfig,
    ) -> Result<DirectBandedJacobianRuntime, BandedError> {
        self.direct_banded_problem()
            .generate_banded_jacobian_runtime_from_plan_parallel(plan, config)
    }

    /// Compatibility adapter for the historical closure-only API.
    pub fn generate_banded_jacobian_assembly_parallel(
        &self,
        config: &BandedLambdifyConfig,
    ) -> Result<Box<dyn Fn(&[f64]) -> Result<BandedAssembly, BandedError> + Send + Sync>, BandedError>
    {
        self.direct_banded_problem()
            .generate_banded_jacobian_assembly_parallel(config)
    }

    /// Compatibility adapter retaining the telemetry-bearing direct runtime.
    pub fn generate_banded_jacobian_runtime_parallel(
        &self,
        config: &BandedLambdifyConfig,
    ) -> Result<DirectBandedJacobianRuntime, BandedError> {
        self.direct_banded_problem()
            .generate_banded_jacobian_runtime_parallel(config)
    }

    /// Compatibility adapter for the direct residual callback.
    pub fn generate_banded_residual_parallel(
        &self,
    ) -> Result<Box<dyn Fn(&[f64]) -> Result<Vec<f64>, BandedError> + Send + Sync>, BandedError>
    {
        self.direct_banded_problem()
            .generate_banded_residual_parallel()
    }

    /// Compatibility adapter for a policy-aware direct Banded residual.
    pub fn generate_banded_residual_with_config(
        &self,
        config: &BandedLambdifyConfig,
    ) -> Result<Box<dyn Fn(&[f64]) -> Result<Vec<f64>, BandedError> + Send + Sync>, BandedError>
    {
        self.direct_banded_problem()
            .generate_banded_residual_with_config(config)
    }

    /// Initializes the Jacobian with a vector of symbolic functions and variable names.
    ///
    /// This method sets up the basic symbolic representation of the system before discretization.
    /// It converts string variable names to symbolic Expr::Var objects for internal processing.
    ///
    /// # Arguments
    /// * `vector_of_functions` - Vector of symbolic expressions representing the system equations
    /// * `variable_string` - Vector of variable names as strings (e.g., ["y", "z"])
    ///
    /// # Example
    /// ```ignore
    /// let mut jac = Jacobian::new();
    /// let funcs = vec![Expr::parse_expression("y + z"), Expr::parse_expression("y - z")];
    /// let vars = vec!["y".to_string(), "z".to_string()];
    /// jac.from_vectors(funcs, vars);
    /// ```
    pub fn from_vectors(&mut self, vector_of_functions: Vec<Expr>, variable_string: Vec<String>) {
        self.vector_of_functions = vector_of_functions;
        self.variable_string = variable_string.clone();
        self.vector_of_variables =
            Expr::parse_vector_expression(variable_string.iter().map(|s| s.as_str()).collect());
        println!(" {:?}", self.vector_of_functions);
        println!(" {:?}", self.vector_of_variables);
    }

    /// Sets symbolic parameter names used during residual/Jacobian evaluation.
    pub fn set_params(&mut self, params: Option<&[&str]>) {
        self.parameters_string = params
            .map(|params| params.iter().map(|name| (*name).to_string()).collect())
            .unwrap_or_default();
        // Changing names changes the callback ABI, so an old compiled
        // evaluator cannot be rebound safely. Numeric-only updates are
        // handled by the shared binding below.
        self.parameter_binding = None;
    }

    /// Sets current numeric parameter values.
    pub fn set_param_values(&mut self, values: Option<Vec<f64>>) {
        if let Some(ref values) = values {
            assert_eq!(
                values.len(),
                self.parameters_string.len(),
                "parameter_values length must match parameters_string length"
            );
        }
        self.parameter_values = values;
        if let Some(binding) = &self.parameter_binding {
            binding.replace(self.parameter_values.clone());
        }
    }

    /// Selects the symbolic assembly backend used during BVP discretization.
    pub fn set_symbolic_assembly_backend(&mut self, backend: BvpSymbolicAssemblyBackend) {
        self.symbolic_assembly_backend = backend;
    }

    /// Returns the currently selected symbolic assembly backend.
    pub fn symbolic_assembly_backend(&self) -> BvpSymbolicAssemblyBackend {
        self.symbolic_assembly_backend
    }

    /// Returns the raw stage timings produced by the most recent symbolic BVP
    /// generation path.
    pub fn last_generate_timer_snapshot(&self) -> Option<&HashMap<String, f64>> {
        self.last_generate_timer_snapshot.as_ref()
    }

    /// Returns typed stage timings for the most recent symbolic Jacobian build.
    pub fn last_symbolic_jacobian_telemetry_snapshot(
        &self,
    ) -> Option<BvpSymbolicPreparationTelemetrySnapshot> {
        self.last_symbolic_jacobian_telemetry_snapshot
    }

    pub(crate) fn set_symbolic_jacobian_timer_snapshot(&mut self, timings: HashMap<String, f64>) {
        self.last_symbolic_jacobian_telemetry_snapshot = Some(
            BvpSymbolicPreparationTelemetrySnapshot::from_seconds_map(&timings),
        );
        self.last_symbolic_jacobian_timer_snapshot = Some(timings);
    }

    /// Returns a typed snapshot for the active direct Banded Jacobian path.
    pub fn direct_banded_jacobian_telemetry_snapshot(
        &self,
    ) -> Option<BvpDirectJacobianTelemetrySnapshot> {
        self.direct_banded_jacobian_telemetry
            .as_ref()
            .map(BvpDirectJacobianTelemetry::snapshot)
    }

    /// Returns a typed snapshot for the active legacy Lambdify callbacks.
    ///
    /// The snapshot keeps legacy measurements comparable with the direct
    /// no-Mutex Banded telemetry without exposing the old timing `HashMap`.
    pub fn legacy_lambdify_telemetry_snapshot(&self) -> Option<LegacyLambdifyTelemetrySnapshot> {
        self.legacy_lambdify_telemetry
            .as_ref()
            .map(LegacyLambdifyTelemetry::snapshot)
    }

    /// Returns the typed runtime snapshot for AtomView Lambdify callbacks.
    pub fn atom_lambdify_telemetry_snapshot(&self) -> Option<LegacyLambdifyTelemetrySnapshot> {
        self.atom_lambdify_telemetry
            .as_ref()
            .map(LegacyLambdifyTelemetry::snapshot)
    }

    fn start_legacy_lambdify_telemetry(&mut self) -> LegacyLambdifyTelemetry {
        if let Some(telemetry) = &self.legacy_lambdify_telemetry {
            return telemetry.clone();
        }
        let telemetry = LegacyLambdifyTelemetry::with_mode(self.lambdify_telemetry_mode);
        self.legacy_lambdify_telemetry = Some(telemetry.clone());
        telemetry
    }

    fn start_atom_lambdify_telemetry(&mut self) -> LegacyLambdifyTelemetry {
        if let Some(telemetry) = &self.atom_lambdify_telemetry {
            return telemetry.clone();
        }
        let telemetry = LegacyLambdifyTelemetry::with_mode(self.lambdify_telemetry_mode);
        self.atom_lambdify_telemetry = Some(telemetry.clone());
        telemetry
    }

    fn install_atom_discretized_system(
        &mut self,
        discretized: crate::symbolic::View::bvp::DiscretizedBvpAtomSystem,
    ) {
        // Keep AtomView data packed through the native Lambdify path. The old
        // compatibility projection expanded every discretized residual and
        // variable back into Expr here, dominating cold preparation for large
        // meshes even though Atom differentiation itself was cheap.
        let timer_table = discretized.render_timer_table();
        self.variable_string = discretized.variable_string.clone();
        self.variables_for_all_disrete = discretized.variables_for_all_discrete.clone();
        self.BC_pos_n_values = discretized.bc_pos_n_values.clone();
        self.bounds = discretized.bounds.clone();
        self.rel_tolerance_vec = discretized.rel_tolerance_vec.clone();
        self.vector_of_functions.clear();
        self.vector_of_variables.clear();
        self.atom_discretized_system = Some(discretized);
        info!("Atom/View BVP discretization timing:\n{}", timer_table);
    }

    /// Builds the sparse Jacobian for an AtomView-discretized system without
    /// round-tripping the residuals through `Expr::diff`.
    ///
    /// The canonical AtomView cache remains packed Atom data. The historical
    /// `Expr` sparse cache is deliberately left empty here and is materialized
    /// only by the explicitly named compatibility accessor. Dense/faer/Banded
    /// Lambdify callbacks use the packed Atom cache directly and do not pay
    /// for that conversion.
    fn calc_atomview_sparse_jacobian_with_bandwidth(&mut self, bandwidth: Option<(usize, usize)>) {
        let discretized = self
            .atom_discretized_system
            .as_ref()
            .expect("AtomView Jacobian preparation requires an installed atom system");
        let mut timings = HashMap::new();

        let variable_sets_begin = Instant::now();
        let prepared = PreparedSparseAtomSystem::from_atoms(
            &discretized.vector_of_functions,
            &discretized.variable_string,
            &discretized.variables_for_all_discrete,
        );
        timings.insert(
            "symbolic jacobian variable sets time".to_string(),
            variable_sets_begin.elapsed().as_secs_f64(),
        );

        let differentiation_begin = Instant::now();
        let sparse_atom_entries = prepared.calc_sparse_jacobian_with_bandwidth(bandwidth);
        self.atom_symbolic_jacobian_sparse = Some(sparse_atom_entries.clone());
        timings.insert(
            "symbolic jacobian row differentiation time".to_string(),
            differentiation_begin.elapsed().as_secs_f64(),
        );

        // Keep the canonical AtomView preparation free of an Atom -> Expr
        // projection. The historical sparse cache is materialized only by
        // the explicit compatibility accessor below, or by the retained
        // ExprLegacy adapter. This is the key boundary for native AOT: a
        // selected AtomView route must not pay for a cache it never reads.
        self.symbolic_jacobian_sparse.clear();
        timings.insert(
            "symbolic jacobian sparse cache flatten time".to_string(),
            0.0,
        );

        let dense_cache_begin = Instant::now();
        self.symbolic_jacobian.clear();
        timings.insert(
            "symbolic jacobian dense cache materialize time".to_string(),
            dense_cache_begin.elapsed().as_secs_f64(),
        );
        self.set_symbolic_jacobian_timer_snapshot(timings);
    }

    /// Returns borrowed sparse symbolic Jacobian entries from the internal
    /// owned sparse cache.
    pub fn symbolic_jacobian_sparse_entries(
        &self,
    ) -> Vec<crate::symbolic::codegen::codegen_tasks::SparseExprEntry<'_>> {
        self.symbolic_jacobian_sparse
            .iter()
            .map(
                |(row, col, expr)| crate::symbolic::codegen::codegen_tasks::SparseExprEntry {
                    row: *row,
                    col: *col,
                    expr,
                },
            )
            .collect()
    }

    /// Returns owned sparse symbolic Jacobian entries.
    ///
    /// This is primarily useful for AOT/performance scaffolding that needs to
    /// hand sparse expressions to downstream planning code without keeping an
    /// immutable borrow of `self` alive across later mutable operations.
    pub fn symbolic_jacobian_sparse_entries_owned(&self) -> Vec<(usize, usize, Expr)> {
        if !self.symbolic_jacobian_sparse.is_empty() {
            return self.symbolic_jacobian_sparse.clone();
        }

        // Explicit compatibility projection. AtomView Lambdify/AOT paths use
        // the packed entries directly and never call this method; callers
        // requesting the old Expr representation opt into the conversion here.
        self.atom_symbolic_jacobian_sparse
            .as_ref()
            .map(|entries| {
                entries
                    .iter()
                    .map(|entry| (entry.row, entry.col, atom_to_expr(&entry.value)))
                    .collect()
            })
            .unwrap_or_default()
    }

    /// Builds the AtomView-native AOT bridge without materializing the
    /// historical `Vec<Expr>` compatibility caches.
    ///
    /// The returned bridge is still named `BvpPreparedSparseAotProblem` for
    /// API compatibility, but its `atom_codegen` field is the authoritative
    /// payload. Expr fields are intentionally empty and must only be read by
    /// explicit ExprLegacy adapters.
    pub(crate) fn prepare_atom_native_aot_problem(
        &self,
        residual_fn_name: &str,
        jacobian_fn_name: &str,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
        matrix_backend: MatrixBackend,
    ) -> Result<Option<BvpPreparedSparseAotProblem>, BvpBackendIntegrationError> {
        if self.symbolic_assembly_backend != BvpSymbolicAssemblyBackend::AtomView {
            return Ok(None);
        }
        let discretized = self.atom_discretized_system.as_ref().ok_or_else(|| {
            BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                message: "AtomView AOT route has no discretized Atom system".to_string(),
            }
        })?;
        let (atom_codegen, atom_breakdown) =
            prepare_sparse_bvp_codegen_from_discretized_system_with_breakdown(
                discretized,
                residual_fn_name,
                jacobian_fn_name,
                self.parameters_string.clone(),
                self.bandwidth,
                residual_strategy,
                jacobian_strategy,
            );
        let atom_codegen = atom_codegen.with_aot_telemetry_mode(self.aot_telemetry_mode);
        let bandwidth = if matrix_backend == MatrixBackend::Banded {
            let (kl, ku) = self
                .bandwidth
                .unwrap_or_else(|| infer_atom_bandwidth(&atom_codegen));
            Some((kl, ku))
        } else {
            self.bandwidth
        };
        // Whole Banded AOT owns a complete native LAPACK-style slot buffer.
        // Build that layout before backend selection so the manifest, artifact
        // key, generated library and linked callback share one ABI contract.
        // Chunked Banded AOT remains on the established explicit-entry ABI
        // until its compact chunk routing has its own production path.
        let native_banded_compact = matrix_backend == MatrixBackend::Banded
            && matches!(jacobian_strategy, SparseChunkingStrategy::Whole);
        let atom_codegen = if matrix_backend == MatrixBackend::Banded {
            let Some((kl, ku)) = bandwidth else {
                return Err(BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                    message: "Banded AOT preparation has no bandwidth".to_string(),
                });
            };
            let layout = if native_banded_compact {
                atom_codegen.with_native_banded_layout(kl, ku)
            } else {
                atom_codegen.with_banded_layout(kl, ku)
            };
            layout.map_err(
                |error| BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                    message: error.to_string(),
                },
            )?
        } else {
            atom_codegen
        };
        let atom_aot_plan = atom_codegen.prepared_aot_plan().map_err(|error| {
            BvpBackendIntegrationError::AtomAotPlanPreparationFailed {
                message: error.to_string(),
            }
        })?;
        record_aot_stage_ms(
            &atom_aot_plan,
            BvpAotColdStage::AtomPreparation,
            atom_breakdown.sparse_lookup_prepare_ms + atom_breakdown.finalize_codegen_plan_ms,
        );
        record_aot_stage_ms(
            &atom_aot_plan,
            BvpAotColdStage::JacobianPreparation,
            atom_breakdown.sparse_jacobian_build_ms,
        );
        let shape = atom_codegen.shape;
        Ok(Some(BvpPreparedSparseAotProblem {
            residual_fn_name: residual_fn_name.to_string(),
            jacobian_fn_name: jacobian_fn_name.to_string(),
            variable_names: atom_codegen.variable_names.clone(),
            param_names: atom_codegen.param_names.clone(),
            residuals: Vec::new(),
            sparse_entries: Vec::new(),
            shape,
            residual_strategy,
            jacobian_strategy,
            bandwidth,
            symbolic_assembly_backend: BvpSymbolicAssemblyBackend::AtomView,
            atom_discretized_system: Some(discretized.clone()),
            atom_codegen: Some(atom_codegen),
            atom_codegen_prep_breakdown: Some(atom_breakdown),
            atom_aot_plan: Some(atom_aot_plan),
        }))
    }

    /// Builds a compatibility AOT bridge from the current BVP state.
    ///
    /// This is the retained ExprLegacy adapter for the older
    /// `PreparedProblem` ABI. If the current state is AtomView, this explicit
    /// method is allowed to materialize `Expr` payloads for that compatibility
    /// ABI. Native AtomView callers must use the typed
    /// [`Self::try_select_sparse_backend`] path instead; it owns an
    /// [`AtomAotPreparedPlan`] and never enters this conversion boundary.
    pub fn prepare_sparse_aot_problem(
        &self,
        residual_fn_name: &str,
        jacobian_fn_name: &str,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
    ) -> BvpPreparedSparseAotProblem {
        let (residual_count, variable_count) = self
            .atom_discretized_system
            .as_ref()
            .map(|system| {
                (
                    system.vector_of_functions.len(),
                    system.vector_of_variables.len(),
                )
            })
            .unwrap_or((
                self.vector_of_functions.len(),
                self.vector_of_variables.len(),
            ));

        // AtomView intentionally does not populate the historical Expr
        // residual cache during ordinary Lambdify preparation.  AOT still
        // consumes the compatibility `Vec<Expr>` contract, so materialize it
        // only at this explicit AOT boundary.  This keeps the cheap
        // Atom-native Lambdify path free of a full Expr round-trip while
        // preserving the legacy code-generation ABI.
        let residuals = if !self.vector_of_functions.is_empty() {
            self.vector_of_functions.clone()
        } else {
            self.atom_discretized_system
                .as_ref()
                .map(|system| {
                    system
                        .vector_of_functions
                        .iter()
                        .map(atom_to_expr)
                        .collect()
                })
                .unwrap_or_default()
        };
        BvpPreparedSparseAotProblem {
            residual_fn_name: residual_fn_name.to_string(),
            jacobian_fn_name: jacobian_fn_name.to_string(),
            variable_names: self.variable_string.clone(),
            param_names: self.parameters_string.clone(),
            residuals,
            sparse_entries: self.symbolic_jacobian_sparse_entries_owned(),
            shape: (residual_count, variable_count),
            residual_strategy,
            jacobian_strategy,
            bandwidth: self.bandwidth,
            symbolic_assembly_backend: self.symbolic_assembly_backend,
            atom_discretized_system: self.atom_discretized_system.clone(),
            atom_codegen: None,
            atom_codegen_prep_breakdown: None,
            atom_aot_plan: None,
        }
    }

    /// Selects the sparse BVP backend branch through the shared lifecycle
    /// selection layer while keeping an owned BVP bridge object.
    ///
    /// This is the main thin integration entrypoint for the BVP module:
    /// - symbolic/discretized BVP data stays owned here,
    /// - the generic lifecycle stack still decides whether AOT is compiled,
    ///   merely registered, or missing,
    /// - the caller can then route into lambdify/numeric/AOT execution paths
    ///   without rebuilding prepared-problem metadata by hand.
    pub fn try_select_sparse_backend(
        &self,
        residual_fn_name: &str,
        jacobian_fn_name: &str,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
    ) -> Result<BvpSelectedSparseBackend, BvpBackendIntegrationError> {
        let requested_matrix_backend = match self.backend_config().matrix_backend {
            BvpMatrixBackend::Dense => MatrixBackend::Dense,
            BvpMatrixBackend::Banded => MatrixBackend::Banded,
            BvpMatrixBackend::FaerSparseCol => MatrixBackend::SparseCol,
            BvpMatrixBackend::SprsCsMat => MatrixBackend::CsMat,
        };
        let mut prepared_problem = self
            .prepare_atom_native_aot_problem(
                residual_fn_name,
                jacobian_fn_name,
                residual_strategy,
                jacobian_strategy,
                requested_matrix_backend,
            )?
            .unwrap_or_else(|| {
                self.prepare_sparse_aot_problem(
                    residual_fn_name,
                    jacobian_fn_name,
                    residual_strategy,
                    jacobian_strategy,
                )
            });
        let selected = if prepared_problem.atom_codegen.is_some() {
            let manifest = prepared_problem.manifest_for_matrix_backend(requested_matrix_backend);
            let selected = select_backend_by_manifest(&manifest, policy, resolver);
            (
                selected.requested_backend,
                selected.effective_backend,
                selected.matrix_backend,
                selected.aot_resolution,
            )
        } else {
            let prepared_binding = prepared_problem.clone();
            let prepared =
                prepared_binding.as_prepared_problem_for_matrix_backend(requested_matrix_backend);
            let selected = select_backend(&prepared, policy, resolver);
            (
                selected.requested_backend,
                selected.effective_backend,
                selected.matrix_backend,
                selected.aot_resolution,
            )
        };

        let aot_preparation_route = prepared_problem.aot_preparation_route();
        if let Some(plan) = prepared_problem.atom_aot_plan.as_mut() {
            plan.bind_artifact(selected.3.clone());
        }
        Ok(BvpSelectedSparseBackend {
            prepared_problem,
            aot_preparation_route,
            requested_backend: selected.0,
            effective_backend: selected.1,
            matrix_backend: selected.2,
            aot_resolution: selected.3,
        })
    }

    /// Compatibility wrapper for the historical infallible backend selector.
    ///
    /// New callers should use [`Self::try_select_sparse_backend`]. In
    /// particular, an AtomView plan failure is never converted into an
    /// ExprLegacy payload before the typed boundary has a chance to report it.
    pub fn select_sparse_backend(
        &self,
        residual_fn_name: &str,
        jacobian_fn_name: &str,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
    ) -> BvpSelectedSparseBackend {
        self.try_select_sparse_backend(
            residual_fn_name,
            jacobian_fn_name,
            residual_strategy,
            jacobian_strategy,
            policy,
            resolver,
        )
        .unwrap_or_else(|error| panic!("BVP AOT backend preparation failed: {error:?}"))
    }

    fn prepare_sparse_backend_execution_timed(
        &mut self,
        arg: &str,
        variable_str: Vec<&str>,
        residual_fn_name: &str,
        jacobian_fn_name: &str,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
        timer_hash: &mut HashMap<String, f64>,
    ) -> Result<BvpSparseExecutionPlan, BvpBackendIntegrationError> {
        let prepare_started = Instant::now();
        let requested_matrix_backend = match self.backend_config().matrix_backend {
            BvpMatrixBackend::Dense => MatrixBackend::Dense,
            BvpMatrixBackend::Banded => MatrixBackend::Banded,
            BvpMatrixBackend::FaerSparseCol => MatrixBackend::SparseCol,
            BvpMatrixBackend::SprsCsMat => MatrixBackend::CsMat,
        };
        let mut prepared_problem = self
            .prepare_atom_native_aot_problem(
                residual_fn_name,
                jacobian_fn_name,
                residual_strategy,
                jacobian_strategy,
                requested_matrix_backend,
            )?
            .unwrap_or_else(|| {
                self.prepare_sparse_aot_problem(
                    residual_fn_name,
                    jacobian_fn_name,
                    residual_strategy,
                    jacobian_strategy,
                )
            });
        timer_hash.insert(
            "sparse AOT preparation time".to_string(),
            prepare_started.elapsed().as_secs_f64(),
        );

        let selection_started = Instant::now();
        let selected = if prepared_problem.atom_codegen.is_some() {
            let manifest = prepared_problem.manifest_for_matrix_backend(requested_matrix_backend);
            let selected = select_backend_by_manifest(&manifest, policy, resolver);
            (
                selected.requested_backend,
                selected.effective_backend,
                selected.matrix_backend,
                selected.aot_resolution,
            )
        } else {
            let prepared_binding = prepared_problem.clone();
            let prepared =
                prepared_binding.as_prepared_problem_for_matrix_backend(requested_matrix_backend);
            let selected = select_backend(&prepared, policy, resolver);
            (
                selected.requested_backend,
                selected.effective_backend,
                selected.matrix_backend,
                selected.aot_resolution,
            )
        };
        timer_hash.insert(
            "backend selection time".to_string(),
            selection_started.elapsed().as_secs_f64(),
        );

        let selected_aot_resolution = selected.3.clone();
        if let Some(plan) = prepared_problem.atom_aot_plan.as_mut() {
            plan.bind_artifact(selected_aot_resolution);
        }
        let selected = BvpSelectedSparseBackend {
            aot_preparation_route: prepared_problem.aot_preparation_route(),
            prepared_problem,
            requested_backend: selected.0,
            effective_backend: selected.1,
            matrix_backend: selected.2,
            aot_resolution: selected.3,
        };

        let binding_started = Instant::now();
        let execution = match selected.effective_backend {
            SelectedBackendKind::Numeric => {
                let _ = (arg, variable_str);
                return Err(BvpBackendIntegrationError::InvalidSolverConfiguration {
                    field: "backend_policy".to_string(),
                    value: "NumericOnly".to_string(),
                    message: "symbolic BVP preparation cannot synthesize a numeric backend"
                        .to_string(),
                });
            }
            SelectedBackendKind::Lambdify => {
                let config = BvpBackendConfig::new(
                    BvpBackendKind::Lambdify,
                    BvpMatrixBackend::from_provider_matrix_backend(selected.matrix_backend),
                );
                self.compile_lambdified_problem_with_config(arg, variable_str, config);
                if let Some(binding_breakdown) = self.last_lambdify_binding_timer_snapshot.take() {
                    timer_hash.extend(binding_breakdown);
                }
                BvpSparseExecutionPlan::LambdifyReady(selected)
            }
            SelectedBackendKind::AotCompiled => {
                self.set_backend_config(BvpBackendConfig::new(
                    BvpBackendKind::Aot,
                    BvpMatrixBackend::from_provider_matrix_backend(selected.matrix_backend),
                ));
                BvpSparseExecutionPlan::AotCompiled(selected)
            }
            SelectedBackendKind::AotRegisteredButNotBuilt => {
                self.set_backend_config(BvpBackendConfig::new(
                    BvpBackendKind::Aot,
                    BvpMatrixBackend::from_provider_matrix_backend(selected.matrix_backend),
                ));
                BvpSparseExecutionPlan::AotRegisteredButNotBuilt(selected)
            }
            SelectedBackendKind::AotMissing => {
                self.set_backend_config(BvpBackendConfig::new(
                    BvpBackendKind::Aot,
                    BvpMatrixBackend::from_provider_matrix_backend(selected.matrix_backend),
                ));
                BvpSparseExecutionPlan::AotMissing(selected)
            }
        };
        timer_hash.insert(
            "runtime binding time".to_string(),
            binding_started.elapsed().as_secs_f64(),
        );
        Ok(execution)
    }

    /// Completes backend preparation after symbolic/discretization work has
    /// already produced its owned timing state.
    ///
    /// Keeping this stage separate lets both the compatibility generator and
    /// the fallible `try_generate_*` boundary share exactly one route. The
    /// latter therefore receives native AtomView plan errors directly instead
    /// of observing a compatibility panic after an Expr fallback.
    fn try_finish_sparse_backend_generation(
        &mut self,
        arg_name: String,
        indexed_values_owned: Vec<String>,
        mut timer_hash: HashMap<String, f64>,
        total_start: Instant,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
    ) -> Result<BvpSparseExecutionPlan, BvpBackendIntegrationError> {
        let execution = self.prepare_sparse_backend_execution_timed(
            arg_name.as_str(),
            indexed_values_owned
                .iter()
                .map(|name| name.as_str())
                .collect(),
            "eval_bvp_residual",
            "eval_bvp_sparse_values",
            residual_strategy,
            jacobian_strategy,
            policy,
            resolver,
            &mut timer_hash,
        );
        let total_end = total_start.elapsed().as_secs_f64();
        self.finalize_generate_bvp_timer_table(timer_hash, total_end);
        execution
    }

    /// Builds a sparse BVP execution plan and, when the selected branch is
    /// lambdify, eagerly prepares the current faer/dense legacy evaluators.
    pub fn prepare_sparse_backend_execution(
        &mut self,
        arg: &str,
        variable_str: Vec<&str>,
        residual_fn_name: &str,
        jacobian_fn_name: &str,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
    ) -> BvpSparseExecutionPlan {
        let selected = self.select_sparse_backend(
            residual_fn_name,
            jacobian_fn_name,
            residual_strategy,
            jacobian_strategy,
            policy,
            resolver,
        );

        match selected.effective_backend {
            SelectedBackendKind::Numeric => {
                let _ = (arg, variable_str);
                panic_symbolic_numeric_backend_requested()
            }
            SelectedBackendKind::Lambdify => {
                let config = BvpBackendConfig::new(
                    BvpBackendKind::Lambdify,
                    BvpMatrixBackend::from_provider_matrix_backend(selected.matrix_backend),
                );
                self.compile_lambdified_problem_with_config(arg, variable_str, config);
                BvpSparseExecutionPlan::LambdifyReady(selected)
            }
            SelectedBackendKind::AotCompiled => {
                self.set_backend_config(BvpBackendConfig::new(
                    BvpBackendKind::Aot,
                    BvpMatrixBackend::from_provider_matrix_backend(selected.matrix_backend),
                ));
                BvpSparseExecutionPlan::AotCompiled(selected)
            }
            SelectedBackendKind::AotRegisteredButNotBuilt => {
                self.set_backend_config(BvpBackendConfig::new(
                    BvpBackendKind::Aot,
                    BvpMatrixBackend::from_provider_matrix_backend(selected.matrix_backend),
                ));
                BvpSparseExecutionPlan::AotRegisteredButNotBuilt(selected)
            }
            SelectedBackendKind::AotMissing => {
                self.set_backend_config(BvpBackendConfig::new(
                    BvpBackendKind::Aot,
                    BvpMatrixBackend::from_provider_matrix_backend(selected.matrix_backend),
                ));
                BvpSparseExecutionPlan::AotMissing(selected)
            }
        }
    }

    /// Wraps a prepared sparse execution plan into a BVP-side transitional
    /// solver provider.
    ///
    /// For now the returned provider is directly callable only for the main
    /// `LambdifyReady` sparse path. AOT branches still expose resolved artifact
    /// metadata and sparse structure through the same object shape so the
    /// higher-level solver integration can use one route while compiled
    /// callback hookup is introduced incrementally.
    pub fn sparse_solver_provider<'a>(
        &'a mut self,
        execution: BvpSparseExecutionPlan,
    ) -> BvpSparseSolverProvider<'a> {
        self.sparse_solver_provider_with_parallel_config(execution, None)
    }

    /// Wraps a prepared plan with an explicit AOT parallel execution policy.
    ///
    /// The policy is intentionally attached to this prepared provider rather
    /// than inferred from a backend string. `None` preserves the historical
    /// sequential callback contract; `Some` enables typed chunk dispatch for
    /// compiled AOT callbacks.
    pub fn sparse_solver_provider_with_parallel_config<'a>(
        &'a mut self,
        execution: BvpSparseExecutionPlan,
        parallel_config: Option<ParallelExecutorConfig>,
    ) -> BvpSparseSolverProvider<'a> {
        let structure = execution.selected().prepared_problem.sparse_structure();

        BvpSparseSolverProvider {
            jacobian: self,
            execution,
            structure,
            parallel_config,
        }
    }

    /// Consumes the current symbolic BVP object into an assembled
    /// solver-facing sparse bundle.
    ///
    /// This is the main "collected" handoff shape for outer BVP/NR solvers:
    /// backend selection, runtime callbacks, sparse structure, and solver
    /// metadata are packaged together instead of being pulled out from several
    /// `Jacobian` fields manually.
    pub fn into_sparse_solver_bundle(
        self,
        execution: BvpSparseExecutionPlan,
    ) -> BvpSparseSolverBundle {
        let legacy_lambdify_telemetry = self.legacy_lambdify_telemetry;
        let atom_lambdify_telemetry = self.atom_lambdify_telemetry;
        let direct_banded_jacobian_telemetry = self.direct_banded_jacobian_telemetry;
        let mut runtime_diagnostics = HashMap::new();
        if let Some(snapshot) = self.last_generate_timer_snapshot.as_ref() {
            for (stage, value_seconds) in snapshot {
                let stage_key = stage
                    .chars()
                    .map(|ch| if ch.is_ascii_alphanumeric() { ch } else { '_' })
                    .collect::<String>();
                runtime_diagnostics.insert(
                    format!("generated.prepare.{stage_key}_ms"),
                    format!("{:.6}", value_seconds * 1_000.0),
                );
            }
        }
        let selected = execution.selected().clone();
        let sparse_structure = selected.prepared_problem.sparse_structure();
        let linked_callbacks = if matches!(execution, BvpSparseExecutionPlan::AotCompiled(_)) {
            linked_runtime_callbacks_for_matrix_backend(
                &selected.prepared_problem,
                selected.matrix_backend,
                self.parameter_values.as_deref(),
                None,
                self.banded_linear_solver_config,
            )
        } else {
            None
        };
        let callbacks_available = matches!(execution, BvpSparseExecutionPlan::LambdifyReady(_))
            || linked_callbacks.is_some();
        let (linked_residual, linked_jacobian) =
            linked_callbacks.map_or((None, None), |(fun, jac)| (Some(fun), Some(jac)));
        let parameter_binding =
            if matches!(selected.effective_backend, SelectedBackendKind::Lambdify) {
                self.parameter_binding
            } else {
                None
            };

        BvpSparseSolverBundle {
            execution,
            residual_function: if callbacks_available {
                if matches!(selected.effective_backend, SelectedBackendKind::Lambdify) {
                    Some(self.residiual_function)
                } else {
                    linked_residual
                }
            } else {
                None
            },
            jacobian_function: if callbacks_available {
                if matches!(selected.effective_backend, SelectedBackendKind::Lambdify) {
                    self.jac_function
                } else {
                    linked_jacobian
                }
            } else {
                None
            },
            runtime_diagnostics,
            generation_telemetry: self
                .last_generate_timer_snapshot
                .as_ref()
                .map(BvpGenerationTelemetrySnapshot::from_seconds_map),
            banded_linear_solver_config: self.banded_linear_solver_config,
            variable_string: self.variable_string,
            parameter_values: self.parameter_values,
            bounds_vec: self.bounds,
            rel_tolerance_vec: self.rel_tolerance_vec,
            bandwidth: self.bandwidth,
            bc_position_and_value: self.BC_pos_n_values,
            sparse_structure,
            atom_discretization_telemetry: selected
                .prepared_problem
                .atom_discretized_system
                .as_ref()
                .map(|system| system.telemetry_snapshot()),
            legacy_lambdify_telemetry,
            atom_lambdify_telemetry,
            direct_banded_jacobian_telemetry,
            aot_telemetry: selected.prepared_problem.aot_telemetry_handle(),
            parameter_binding,
        }
    }

    /// Fallible wrapper around `generate_BVP_with_backend_selection`.
    ///
    /// This is the first end-to-end error boundary for the new BVP/AOT path:
    /// unexpected panics in symbolic preparation, backend selection, or
    /// callback preparation are converted into a typed error so outer solver
    /// orchestration can log or fall back deliberately.
    #[allow(clippy::too_many_arguments)]
    pub fn try_generate_BVP_with_backend_selection(
        &mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
    ) -> Result<BvpSparseExecutionPlan, BvpBackendIntegrationError> {
        info!(
            "BVP backend selection started: backend_policy={policy:?}, method={method}, params={}, steps={:?}",
            params.map_or(0, |p| p.len()),
            n_steps
        );
        let result = catch_unwind(AssertUnwindSafe(|| {
            let (arg_name, indexed_values_owned, _matrix_backend, timer_hash, total_start) = self
                .prepare_symbolic_bvp_stage_with_params(
                    eq_system,
                    values,
                    arg,
                    params,
                    t0,
                    n_steps,
                    h,
                    mesh,
                    BorderConditions,
                    Bounds,
                    rel_tolerance,
                    scheme,
                    method,
                    bandwidth,
                );
            let effective_arg_name = param.unwrap_or(arg_name);
            let equation_count = self
                .atom_discretized_system
                .as_ref()
                .map(|system| system.vector_of_functions.len())
                .unwrap_or(self.vector_of_functions.len());
            let residual_strategy =
                recommended_residual_chunking_for_parallelism(equation_count, 4);
            let jacobian_strategy = recommended_row_chunking_for_parallelism(equation_count, 4);
            self.try_finish_sparse_backend_generation(
                effective_arg_name,
                indexed_values_owned,
                timer_hash,
                total_start,
                residual_strategy,
                jacobian_strategy,
                policy,
                resolver,
            )
        }));

        match result {
            Ok(Ok(execution)) => {
                info!(
                    "BVP backend selection finished: effective_backend={:?}, preparation_route={:?}",
                    execution.selected().effective_backend,
                    execution.selected().preparation_route()
                );
                Ok(execution)
            }
            Ok(Err(error)) => {
                error!("BVP backend selection failed: {error:?}");
                Err(error)
            }
            Err(payload) => {
                let message = if let Some(message) = payload.downcast_ref::<&str>() {
                    (*message).to_string()
                } else if let Some(message) = payload.downcast_ref::<String>() {
                    message.clone()
                } else {
                    "unknown panic payload".to_string()
                };
                error!("BVP backend selection failed: {message}");
                Err(BvpBackendIntegrationError::PipelinePanicked(message))
            }
        }
    }

    /// Fallible wrapper returning the assembled solver-facing sparse bundle.
    #[allow(clippy::too_many_arguments)]
    pub fn try_generate_sparse_solver_bundle_with_backend_selection(
        mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
    ) -> Result<BvpSparseSolverBundle, BvpBackendIntegrationError> {
        let execution = self.try_generate_BVP_with_backend_selection(
            eq_system,
            values,
            arg,
            params,
            t0,
            param,
            n_steps,
            h,
            mesh,
            BorderConditions,
            Bounds,
            rel_tolerance,
            scheme,
            method,
            bandwidth,
            policy,
            resolver,
        )?;
        Ok(self.into_sparse_solver_bundle(execution))
    }

    /// Fallible wrapper around `generate_BVP_with_backend_selection_and_chunking`.
    ///
    /// This variant allows the outer solver layer to override residual and
    /// sparse-Jacobian chunk planning instead of relying on the internal BVP
    /// heuristics.
    #[allow(clippy::too_many_arguments)]
    pub fn try_generate_BVP_with_backend_selection_and_chunking(
        &mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
    ) -> Result<BvpSparseExecutionPlan, BvpBackendIntegrationError> {
        info!(
            "BVP backend selection with explicit chunking started: backend_policy={policy:?}, method={method}, params={}, steps={:?}, residual_strategy={residual_strategy:?}, jacobian_strategy={jacobian_strategy:?}",
            params.map_or(0, |p| p.len()),
            n_steps
        );
        let result = catch_unwind(AssertUnwindSafe(|| {
            let (arg_name, indexed_values_owned, _matrix_backend, timer_hash, total_start) = self
                .prepare_symbolic_bvp_stage_with_params(
                    eq_system,
                    values,
                    arg,
                    params,
                    t0,
                    n_steps,
                    h,
                    mesh,
                    BorderConditions,
                    Bounds,
                    rel_tolerance,
                    scheme,
                    method,
                    bandwidth,
                );
            let effective_arg_name = param.unwrap_or(arg_name);
            self.try_finish_sparse_backend_generation(
                effective_arg_name,
                indexed_values_owned,
                timer_hash,
                total_start,
                residual_strategy,
                jacobian_strategy,
                policy,
                resolver,
            )
        }));

        match result {
            Ok(Ok(execution)) => {
                info!(
                    "BVP backend selection with explicit chunking finished: effective_backend={:?}, preparation_route={:?}",
                    execution.selected().effective_backend,
                    execution.selected().preparation_route()
                );
                Ok(execution)
            }
            Ok(Err(error)) => {
                error!("BVP backend selection with explicit chunking failed: {error:?}");
                Err(error)
            }
            Err(payload) => {
                let message = if let Some(message) = payload.downcast_ref::<&str>() {
                    (*message).to_string()
                } else if let Some(message) = payload.downcast_ref::<String>() {
                    message.clone()
                } else {
                    "unknown panic payload".to_string()
                };
                error!("BVP backend selection with explicit chunking failed: {message}");
                Err(BvpBackendIntegrationError::PipelinePanicked(message))
            }
        }
    }

    /// Fallible wrapper returning the assembled solver-facing sparse bundle
    /// with explicit chunking strategies.
    #[allow(clippy::too_many_arguments)]
    pub fn try_generate_sparse_solver_bundle_with_backend_selection_and_chunking(
        mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
    ) -> Result<BvpSparseSolverBundle, BvpBackendIntegrationError> {
        let execution = self.try_generate_BVP_with_backend_selection_and_chunking(
            eq_system,
            values,
            arg,
            params,
            t0,
            param,
            n_steps,
            h,
            mesh,
            BorderConditions,
            Bounds,
            rel_tolerance,
            scheme,
            method,
            bandwidth,
            policy,
            resolver,
            residual_strategy,
            jacobian_strategy,
        )?;
        Ok(self.into_sparse_solver_bundle(execution))
    }

    fn flattened_argument_names_owned(&self, variable_str: &[&str]) -> Vec<String> {
        let mut names = Vec::with_capacity(
            self.parameters_string
                .len()
                .saturating_add(variable_str.len()),
        );
        names.extend(self.parameters_string.iter().cloned());
        names.extend(variable_str.iter().map(|name| (*name).to_string()));
        names
    }

    /// Sets the modern backend configuration and mirrors the matrix choice into
    /// the legacy `method` string used by the existing BVP solver stack.
    pub fn set_backend_config(&mut self, config: BvpBackendConfig) {
        self.backend_config = config;
        self.method = config.matrix_backend.legacy_method().to_string();
    }

    /// Returns the best-known backend configuration for this instance.
    ///
    /// If the legacy `method` field was changed directly, this method keeps the
    /// matrix backend in sync while preserving the currently selected backend
    /// family (`Numeric`, `Lambdify`, `Aot`).
    pub fn backend_config(&self) -> BvpBackendConfig {
        if let Some(matrix_backend) = BvpMatrixBackend::from_legacy_method(self.method.as_str()) {
            BvpBackendConfig::new(self.backend_config.backend_kind, matrix_backend)
        } else {
            self.backend_config
        }
    }

    fn compile_jacobian_for_matrix_backend(
        &mut self,
        arg: &str,
        variable_str: Vec<&str>,
        matrix_backend: BvpMatrixBackend,
    ) {
        match matrix_backend {
            BvpMatrixBackend::Dense => self.lambdify_jacobian_DMatrix_par(arg, variable_str),
            BvpMatrixBackend::SprsCsMat => self.lambdify_jacobian_CsMat(arg, variable_str),
            BvpMatrixBackend::FaerSparseCol => {
                self.lambdify_jacobian_SparseColMat_parallel2(arg, variable_str)
            }
            BvpMatrixBackend::Banded => self.lambdify_jacobian_Banded(arg, variable_str),
        }
    }

    fn compile_residual_for_matrix_backend(
        &mut self,
        arg: &str,
        variable_str: Vec<&str>,
        matrix_backend: BvpMatrixBackend,
    ) {
        match matrix_backend {
            BvpMatrixBackend::Dense => self.lambdify_residual_DVector(arg, variable_str),
            BvpMatrixBackend::SprsCsMat => self.lambdify_residual_CsVec(arg, variable_str),
            BvpMatrixBackend::FaerSparseCol => {
                self.lambdify_residual_Col_parallel2(arg, variable_str)
            }
            BvpMatrixBackend::Banded => self.lambdify_residual_Banded(arg, variable_str),
        }
    }

    fn compile_lambdified_problem_with_config_unchecked(
        &mut self,
        arg: &str,
        variable_str: Vec<&str>,
        config: BvpBackendConfig,
    ) {
        match config.backend_kind {
            BvpBackendKind::Lambdify => {
                self.set_backend_config(config);
                let jacobian_started = Instant::now();
                self.compile_jacobian_for_matrix_backend(
                    arg,
                    variable_str.clone(),
                    config.matrix_backend,
                );
                let jacobian_elapsed = jacobian_started.elapsed().as_secs_f64();
                let residual_started = Instant::now();
                self.compile_residual_for_matrix_backend(arg, variable_str, config.matrix_backend);
                self.last_lambdify_binding_timer_snapshot = Some(HashMap::from([
                    (
                        "lambdify jacobian callback compile time".to_string(),
                        jacobian_elapsed,
                    ),
                    (
                        "lambdify residual callback compile time".to_string(),
                        residual_started.elapsed().as_secs_f64(),
                    ),
                ]));
            }
            BvpBackendKind::Numeric => {
                let _ = (arg, variable_str, config);
                panic_symbolic_numeric_backend_requested()
            }
            BvpBackendKind::Aot => {
                panic!("AOT backend integration is not wired inside symbolic_functions_BVP yet")
            }
        }
    }

    fn validate_atom_lambdify_config(
        &self,
        config: BvpBackendConfig,
    ) -> Result<(), BvpBackendIntegrationError> {
        if self.symbolic_assembly_backend != BvpSymbolicAssemblyBackend::AtomView {
            return Ok(());
        }

        let parameter_values = self.parameter_values.as_deref();
        let parameter_count = self.parameters_string.len();
        match config.matrix_backend {
            BvpMatrixBackend::Dense | BvpMatrixBackend::FaerSparseCol => {
                let Some(entries) = self.atom_symbolic_jacobian_sparse.as_deref() else {
                    return Ok(());
                };
                let (rows, cols) = self
                    .atom_discretized_system
                    .as_ref()
                    .map(|system| {
                        (
                            system.vector_of_functions.len(),
                            system.vector_of_variables.len(),
                        )
                    })
                    .unwrap_or((
                        self.vector_of_functions.len(),
                        self.vector_of_variables.len(),
                    ));
                crate::symbolic::bvp::atom_lambdify::validate_jacobian_inputs(
                    entries,
                    rows,
                    cols,
                    parameter_values,
                    parameter_count,
                )?;
                if let Some(functions) = self
                    .atom_discretized_system
                    .as_ref()
                    .map(|system| system.vector_of_functions.as_slice())
                {
                    crate::symbolic::bvp::atom_lambdify::validate_residual_inputs(
                        functions,
                        parameter_values,
                        parameter_count,
                    )?;
                }
            }
            BvpMatrixBackend::SprsCsMat | BvpMatrixBackend::Banded => {}
        }
        Ok(())
    }

    /// Fallible Lambdify preparation boundary for typed solver callers.
    ///
    /// AtomView input validation happens before the historical callback trait
    /// objects are installed. This turns missing parameter bindings and bad
    /// sparse coordinates into typed errors instead of callback-time panics.
    pub fn try_compile_lambdified_problem_with_config(
        &mut self,
        arg: &str,
        variable_str: Vec<&str>,
        config: BvpBackendConfig,
    ) -> Result<(), BvpBackendIntegrationError> {
        self.validate_atom_lambdify_config(config)?;
        let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            self.compile_lambdified_problem_with_config_unchecked(arg, variable_str, config)
        }));
        result.map_err(|payload| {
            let message = if let Some(message) = payload.downcast_ref::<&str>() {
                (*message).to_string()
            } else if let Some(message) = payload.downcast_ref::<String>() {
                message.clone()
            } else {
                "unknown Lambdify preparation panic".to_string()
            };
            BvpBackendIntegrationError::PipelinePanicked(message)
        })
    }

    /// Compatibility wrapper over [`Self::try_compile_lambdified_problem_with_config`].
    ///
    /// New code should use the fallible entrypoint; this method preserves the
    /// historical panic-based API for downstream callers.
    pub fn compile_lambdified_problem_with_config(
        &mut self,
        arg: &str,
        variable_str: Vec<&str>,
        config: BvpBackendConfig,
    ) {
        self.try_compile_lambdified_problem_with_config(arg, variable_str, config)
            .unwrap_or_else(|error| panic!("BVP Lambdify preparation failed: {error:?}"));
    }

    /// Computes the symbolic Jacobian matrix with bandwidth optimization and smart sparsity detection.
    ///
    /// This is the most optimized Jacobian computation method, combining bandwidth-aware computation
    /// with smart variable tracking for maximum performance on banded sparse systems.
    ///
    /// # Performance Optimizations
    /// 1. **Bandwidth optimization**: Only computes derivatives within the specified band (kl, ku)
    /// 2. **HashSet variable lookup**: O(1) variable presence checking using pre-computed HashSets
    /// 3. **Smart sparsity**: Skips derivatives for variables not present in each equation
    /// 4. **Parallel row computation**: Uses rayon for concurrent processing of Jacobian rows
    /// 5. **Pre-allocated rows**: Full-width rows initialized with zeros, selective assignment
    ///
    /// # Bandwidth Logic
    /// For equation i, only computes derivatives for variables j where:
    /// - `left_border ≤ j < right_border`
    /// - `left_border = max(0, i - kl - 1)`
    /// - `right_border = min(n_vars, i + ku + 1)`
    ///
    /// # When to Use
    /// - **RECOMMENDED** for large banded sparse systems with known bandwidth
    /// - BVP discretizations with local coupling (most common case)
    /// - Systems where bandwidth << matrix_size for maximum benefit
    /// Compatibility entry point for the ExprLegacy symbolic Jacobian path.
    ///
    /// The implementation lives in `bvp::legacy_symbolic`; this method keeps
    /// the historical public API stable for downstream callers.
    pub fn calc_jacobian_parallel_smart_optimized_with_given_bandwidth(&mut self) {
        crate::symbolic::bvp::legacy_symbolic::calc_jacobian_parallel_smart_optimized(self, true);
    }

    /// Computes the symbolic Jacobian matrix with smart sparsity detection and HashSet optimization.
    ///
    /// This method provides significant performance improvements over naive differentiation by using
    /// variable tracking to skip zero derivatives. It's the optimal choice when bandwidth is unknown.
    ///
    /// # Performance Optimizations
    /// 1. **HashSet variable lookup**: Converts variable lists to HashSets for O(1) contains() checks
    /// 2. **Smart sparsity**: Only computes ∂F_i/∂x_j if variable x_j appears in equation F_i
    /// 3. **Parallel computation**: Uses rayon for concurrent processing across equations
    /// 4. **Pre-allocation**: Vectors allocated with exact capacity to avoid reallocations
    /// 5. **Zero insertion**: Directly inserts Expr::Const(0.0) for missing variables
    ///
    /// # Algorithm Complexity
    /// - **Without optimization**: O(n_eqs × n_vars × diff_cost)
    /// - **With this optimization**: O(n_eqs × avg_vars_per_eq × diff_cost)
    /// - **Typical speedup**: 5-50x for sparse BVP systems
    ///
    /// # When to Use
    /// - **RECOMMENDED** for sparse systems without known bandwidth
    /// - General BVP discretizations where bandwidth detection is expensive
    /// - Systems with highly variable sparsity patterns per equation
    /// Compatibility entry point for smart ExprLegacy Jacobian preparation.
    pub fn calc_jacobian_parallel_smart_optimized(&mut self) {
        crate::symbolic::bvp::legacy_symbolic::calc_jacobian_parallel_smart_optimized(self, false);
    }

    /// Computes the symbolic Jacobian matrix using parallel differentiation with smart sparsity optimization.
    ///
    /// This is the most efficient Jacobian computation method. It uses the `variables_for_all_disrete`
    /// field to only compute partial derivatives for variables that actually appear in each equation,
    /// dramatically reducing computation time for sparse systems.
    ///
    /// # Performance Features
    /// - Parallel computation across equations using rayon
    /// - Smart zero-derivative detection: skips ∂F_i/∂x_j if x_j not in equation i
    /// - Automatic symbolic simplification of computed derivatives
    ///
    /// # Panics
    /// - If `vector_of_functions` is empty
    /// - If `vector_of_variables` is empty  
    /// - If `variables_for_all_disrete` is empty (must call discretization first)
    ///
    /// # Note
    /// Must be called after `discretization_system_BVP_par()` which populates `variables_for_all_disrete`.
    pub fn calc_jacobian_parallel_smart(&mut self) {
        crate::symbolic::bvp::legacy_symbolic::calc_jacobian_parallel_smart(self);
    }
    /// Computes the symbolic Jacobian matrix using standard parallel differentiation.
    ///
    /// This method computes all partial derivatives ∂F_i/∂x_j without sparsity optimization.
    /// Use `calc_jacobian_parallel_smart()` for better performance on sparse systems.
    ///
    /// # Performance Features
    /// - Parallel computation across equations using rayon
    /// - Automatic symbolic simplification of computed derivatives
    ///
    /// # Panics
    /// - If `vector_of_functions` is empty
    /// - If `vector_of_variables` is empty
    ///
    /// # When to Use
    /// - For dense systems where most variables appear in most equations
    /// - When `variables_for_all_disrete` is not available (before discretization)
    pub fn calc_jacobian_parallel(&mut self) {
        crate::symbolic::bvp::legacy_symbolic::calc_jacobian_parallel(self);
    }
    //

    ////////////////////////////////////////////////////////////////////////////////////
    //  GENERIC FUNCTIONS
    ////////////////////////////////////////////////////////////////////////////////////

    /// Compiles residual functions using generic VectorType trait for backend flexibility.
    ///
    /// This method creates a generic residual function that works with any vector type
    /// implementing the VectorType trait. It's part of the experimental generic backend system.
    ///
    /// # Arguments
    /// * `arg` - Independent variable name (typically time "t" or "x")
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Note
    /// This is an experimental generic interface. For production use, prefer the specific
    /// backend methods like `lambdify_residual_DVector()` or `lambdify_residual_Col_parallel2()`.
    pub fn lambdify_residual_VectorType(&mut self, arg: &str, variable_str: Vec<&str>) {
        let vector_of_functions = &self.vector_of_functions;
        fn f(
            vector_of_functions: Vec<Expr>,
            _arg: String,
            variable_str: Vec<String>,
        ) -> Box<dyn Fn(f64, &dyn VectorType) -> Box<dyn VectorType>> {
            Box::new(move |_x: f64, v: &dyn VectorType| -> Box<dyn VectorType> {
                let mut result = v.zeros(vector_of_functions.len());

                // Iterate through functions and assign values
                for (i, func) in vector_of_functions.iter().enumerate() {
                    let func = Expr::lambdify_borrowed_thread_safe(
                        &func,
                        variable_str
                            .iter()
                            .map(|s| s.as_str())
                            .collect::<Vec<_>>()
                            .as_slice(),
                    );
                    let v: Vec<f64> = v.iterate().collect();
                    let value = func(v.as_slice());
                    result = result.assign_value(i, value);
                }
                result
            })
        }

        let fun: Box<dyn Fn(f64, &dyn VectorType) -> Box<dyn VectorType>> = f(
            vector_of_functions.to_owned(),
            arg.to_string(),
            variable_str.iter().map(|s| s.to_string()).collect(),
        );
        let residual_function = convert_to_fun(fun);
        self.residiual_function = residual_function;
    }

    /// Generates a generic parallel Jacobian evaluator using trait objects.
    ///
    /// Creates a closure that evaluates the symbolic Jacobian matrix using generic
    /// VectorType and MatrixType traits. Supports bandwidth optimization for sparse matrices.
    ///
    /// # Arguments
    /// * `jac` - 2D vector of symbolic partial derivatives
    /// * `vector_of_functions_len` - Number of equations in the system
    /// * `vector_of_variables_len` - Number of variables in the system
    /// * `variable_str` - Variable names as owned strings
    /// * `_arg` - Independent variable name (unused in current implementation)
    /// * `bandwidth` - Optional (kl, ku) bandwidth for banded matrix optimization
    ///
    /// # Returns
    /// A boxed closure that takes (time, variables) and returns a matrix of partial derivatives
    ///
    /// # Performance Features
    /// - Parallel evaluation using rayon
    /// - Bandwidth-aware computation for sparse matrices
    /// - Thread-safe triplet collection using Mutex
    /// - Zero-threshold filtering (T = 1e-12) to maintain sparsity
    pub fn jacobian_generate_generic_par(
        jac: Vec<Vec<Expr>>,
        vector_of_functions_len: usize,
        vector_of_variables_len: usize,
        variable_str: Vec<String>,
        _arg: String,
        bandwidth: Option<(usize, usize)>,
    ) -> Box<dyn Fn(f64, &(dyn VectorType + Send + Sync)) -> Box<dyn MatrixType>> {
        //let arg = arg.as_str();
        //let variable_str: Vec<&str> = variable_str.iter().map(|s| s.as_str()).collect();

        Box::new(
            move |_x: f64, v: &(dyn VectorType + Send + Sync)| -> Box<dyn MatrixType> {
                let mut vector_of_derivatives =
                    vec![0.0; vector_of_functions_len * vector_of_variables_len];
                let vector_of_derivatives_mutex = std::sync::Mutex::new(&mut vector_of_derivatives);

                let mut vector_of_triplets = Vec::new();
                let vector_of_triplets_mutex = std::sync::Mutex::new(&mut vector_of_triplets);
                (0..vector_of_functions_len).into_par_iter().for_each(|i| {
                    let (right_border, left_border) = if let Some((kl, ku)) = bandwidth {
                        let right_border = std::cmp::min(i + ku + 1, vector_of_variables_len);
                        let left_border = if i as i32 - (kl as i32) - 1 < 0 {
                            0
                        } else {
                            i - kl - 1
                        };
                        (right_border, left_border)
                    } else {
                        let right_border = vector_of_variables_len;
                        let left_border = 0;
                        (right_border, left_border)
                    };
                    for j in left_border..right_border {
                        // if jac[i][j] != Expr::Const(0.0) { println!("i = {}, j = {}, {}", i, j,  &jac[i][j]);}
                        let partial_func = Expr::lambdify_borrowed_thread_safe(
                            &jac[i][j],
                            variable_str
                                .iter()
                                .map(|s| s.as_str())
                                .collect::<Vec<_>>()
                                .as_slice(),
                        );

                        let v_vec: Vec<f64> = v.iterate().collect();
                        let _vec_type = &v.vec_type();
                        let P = partial_func(v_vec.as_slice());
                        if P.abs() > T {
                            vector_of_derivatives_mutex.lock().unwrap()
                                [i * vector_of_functions_len + j] = P;
                            let triplet = (i, j, P);
                            vector_of_triplets_mutex.lock().unwrap().push(triplet);
                        }
                    }
                }); //par_iter()

                let new_function_jacobian: Box<dyn MatrixType> = v.from_vector(
                    vector_of_functions_len,
                    vector_of_variables_len,
                    &vector_of_derivatives,
                    vector_of_triplets,
                );
                //  panic!("stop here");
                new_function_jacobian
            },
        ) // end of box
    } // end of function

    /// Compiles the symbolic Jacobian using the generic trait-based backend system.
    ///
    /// Prepares the generic Jacobian builder but intentionally does not install
    /// it as the active solver callback.
    ///
    /// # Arguments
    /// * `arg` - Independent variable name
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Note
    /// Production BVP paths should use one of the concrete backend installers.
    pub fn lambdify_jacobian_generic(&mut self, arg: &str, variable_str: Vec<&str>) {
        let symbolic_jacobian = self.symbolic_jacobian.clone();
        let symbolic_jacobian_rc = symbolic_jacobian.clone();

        let vector_of_functions_len = self.vector_of_functions.len();
        let vector_of_variables_len = self.vector_of_variables.len();
        let bandwidth = self.bandwidth;
        let _new_jac = Jacobian::jacobian_generate_generic_par(
            symbolic_jacobian_rc,
            vector_of_functions_len,
            vector_of_variables_len,
            variable_str.iter().map(|s| s.to_string()).collect(),
            arg.to_string(),
            bandwidth,
        );
    }

    ////////////////////////////////////////////////////////////////////////////////////
    ///                             NALGEBRA DENSE
    ////////////////////////////////////////////////////////////////////////////////////

    /// Compiles the symbolic Jacobian to a dense nalgebra DMatrix evaluator with parallel optimization.
    ///
    /// This method creates a high-performance dense matrix Jacobian evaluator using nalgebra's DMatrix.
    /// It pre-compiles all non-zero symbolic derivatives and uses parallel evaluation for optimal performance.
    ///
    /// # Arguments
    /// * `arg` - Independent variable name (typically "t" or "x")
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Performance Features
    /// - **Pre-compilation**: All symbolic derivatives compiled once during setup
    /// - **Parallel evaluation**: Uses rayon for concurrent derivative evaluation
    /// - **Bandwidth optimization**: Respects sparse matrix bandwidth if set
    /// - **Zero filtering**: Only evaluates non-zero symbolic derivatives
    /// - **Thread-safe**: Uses `Send + Sync` closures and Mutex for thread safety
    /// - **Outer loop parallelization**: Uses `flat_map()` pattern for better load balancing
    ///
    /// # Matrix Structure
    /// Creates a dense DMatrix where element (i,j) = ∂F_i/∂x_j evaluated at the given point.
    ///
    /// # When to Use
    /// - Small to medium systems (< 1000 variables)
    /// - Dense or moderately sparse Jacobians
    /// - When nalgebra ecosystem integration is important
    pub fn lambdify_jacobian_DMatrix_par(&mut self, _arg: &str, variable_str: Vec<&str>) {
        let binding = self.parameter_binding_for_callbacks();
        let flattened_argument_names_owned =
            self.flattened_argument_names_owned(variable_str.as_slice());
        if self.symbolic_assembly_backend == BvpSymbolicAssemblyBackend::AtomView {
            if let Some(entries) = self.atom_symbolic_jacobian_sparse.clone() {
                let (rows, cols) = self
                    .atom_discretized_system
                    .as_ref()
                    .map(|system| {
                        (
                            system.vector_of_functions.len(),
                            system.vector_of_variables.len(),
                        )
                    })
                    .unwrap_or((
                        self.vector_of_functions.len(),
                        self.vector_of_variables.len(),
                    ));
                let telemetry = self.start_atom_lambdify_telemetry();
                self.jac_function = Some(atom_lambdify::dense_jacobian_with_binding_and_policy(
                    entries,
                    rows,
                    cols,
                    flattened_argument_names_owned,
                    binding,
                    self.parameters_string.len(),
                    telemetry,
                    self.lambdify_execution_policy,
                ));
                return;
            }
        }
        let telemetry = self.start_legacy_lambdify_telemetry();
        self.jac_function = Some(legacy_lambdify::dense_jacobian_with_binding_and_policy(
            self.symbolic_jacobian.clone(),
            self.vector_of_functions.len(),
            self.vector_of_variables.len(),
            self.bandwidth,
            flattened_argument_names_owned,
            binding,
            self.parameters_string.len(),
            telemetry,
            self.lambdify_execution_policy,
        ));
    }

    /// Compiles the residual functions to a dense nalgebra DVector evaluator with parallel optimization.
    ///
    /// Creates a high-performance residual function evaluator that returns nalgebra DVector results.
    /// All symbolic functions are pre-compiled for maximum efficiency during repeated evaluations.
    ///
    /// # Arguments
    /// * `arg` - Independent variable name (typically "t" or "x")
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Performance Features
    /// - **Pre-compilation**: All residual functions compiled once during setup
    /// - **Parallel evaluation**: Uses rayon for concurrent function evaluation
    /// - **Thread-safe closures**: Uses `Send + Sync` bounds for parallel safety
    /// - **Memory efficient**: Direct vector construction without intermediate collections
    ///
    /// # Output Format
    /// Returns DVector where element i = F_i(t, y) for the i-th residual equation.
    ///
    /// # When to Use
    /// - Dense vector operations with nalgebra
    /// - Small to medium systems where dense storage is acceptable
    /// - Integration with nalgebra-based linear solvers
    pub fn lambdify_residual_DVector(&mut self, _arg: &str, variable_str: Vec<&str>) {
        let binding = self.parameter_binding_for_callbacks();
        let variable_str_owned = self.flattened_argument_names_owned(variable_str.as_slice());
        if self.symbolic_assembly_backend == BvpSymbolicAssemblyBackend::AtomView {
            if let Some(functions) = self
                .atom_discretized_system
                .as_ref()
                .map(|system| system.vector_of_functions.clone())
            {
                let telemetry = self.start_atom_lambdify_telemetry();
                self.residiual_function = atom_lambdify::dense_residual_with_binding_and_policy(
                    functions,
                    variable_str_owned,
                    binding,
                    self.parameters_string.len(),
                    telemetry,
                    self.lambdify_execution_policy,
                );
                return;
            }
        }
        let telemetry = self.start_legacy_lambdify_telemetry();
        self.residiual_function = legacy_lambdify::dense_residual_with_binding_and_policy(
            self.vector_of_functions.clone(),
            variable_str_owned,
            binding,
            self.parameters_string.len(),
            telemetry,
            self.lambdify_execution_policy,
        );
    }

    //////////////////////////////////////////////////////////////////////////////////////////
    ////                             SPRS CRATE SPARSE FUNCTIONS
    ///////////////////////////////////////////////////////////////////////////////////////////

    /// Generates a sparse Jacobian evaluator using the sprs crate's CsMat format.
    ///
    /// Creates a closure that evaluates the symbolic Jacobian matrix and returns it as a
    /// sprs::CsMat (Compressed Sparse Matrix). This is a legacy method with sequential evaluation.
    ///
    /// # Arguments
    /// * `jac` - 2D vector of symbolic partial derivatives
    /// * `vector_of_functions_len` - Number of equations in the system
    /// * `vector_of_variables_len` - Number of variables in the system
    /// * `variable_str` - Variable names as owned strings
    /// * `_arg` - Independent variable name (unused)
    ///
    /// # Returns
    /// A mutable closure that takes (time, sparse_vector) and returns a sparse matrix
    ///
    /// # Performance Notes
    /// - Sequential evaluation (not parallelized)
    /// - Uses sprs::CsMat compressed sparse row format
    /// - Evaluates all matrix elements (no sparsity optimization)
    ///
    /// # When to Use
    /// - Legacy code compatibility with sprs crate
    /// - When mutable closure semantics are required
    /// - Small systems where parallelization overhead isn't justified
    pub fn jacobian_generate_CsMat(
        jac: Vec<Vec<Expr>>,
        vector_of_functions_len: usize,
        vector_of_variables_len: usize,
        variable_str: Vec<String>,
        _arg: String,
    ) -> Box<dyn FnMut(f64, &CsVec<f64>) -> CsMat<f64>> {
        //let arg = arg.as_str();
        //let variable_str: Vec<&str> = variable_str.iter().map(|s| s.as_str()).collect();

        Box::new(move |_x: f64, v: &CsVec<f64>| -> CsMat<f64> {
            let mut new_function_jacobian: CsMat<f64> =
                CsMat::zero((vector_of_functions_len, vector_of_variables_len));
            for i in 0..vector_of_functions_len {
                for j in 0..vector_of_variables_len {
                    // println!("i = {}, j = {}", i, j);
                    let partial_func = Expr::lambdify_borrowed_thread_safe(
                        &jac[i][j],
                        variable_str
                            .iter()
                            .map(|s| s.as_str())
                            .collect::<Vec<_>>()
                            .as_slice(),
                    );
                    //let v_vec: Vec<f64> = v.iter().cloned().collect();
                    let v_vec: Vec<f64> = v.to_dense().to_vec();
                    new_function_jacobian.insert(j, i, partial_func(v_vec.as_slice()));
                }
            }
            new_function_jacobian
        })
    } // end of function
    /// Compiles the symbolic Jacobian to a sprs CsMat evaluator.
    ///
    /// Sets up the Jacobian function using the sprs crate's compressed sparse matrix format.
    /// This method wraps the generated CsMat evaluator in the trait object system.
    ///
    /// # Arguments
    /// * `arg` - Independent variable name
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Backend Details
    /// - Uses sprs::CsMat (Compressed Sparse Row format)
    /// - Sequential evaluation (not parallelized)
    /// - Wrapped in JacEnum::Sparse_1 variant
    ///
    /// # When to Use
    /// - Legacy compatibility with sprs-based code
    /// - When CSR format is specifically required
    /// - Small to medium sparse systems
    pub fn lambdify_jacobian_CsMat(&mut self, arg: &str, variable_str: Vec<&str>) {
        let symbolic_jacobian = self.symbolic_jacobian.clone();
        let symbolic_jacobian_rc = symbolic_jacobian.clone();

        let vector_of_functions_len = self.vector_of_functions.len();
        let vector_of_variables_len = self.vector_of_variables.len();

        let new_jac = Jacobian::jacobian_generate_CsMat(
            symbolic_jacobian_rc,
            vector_of_functions_len,
            vector_of_variables_len,
            variable_str.iter().map(|s| s.to_string()).collect(),
            arg.to_string(),
        );
        //  let mut boxed_jacobian: Box<dyn Fn(f64, &DVector<f64>) -> DMatrix<f64>> = Box::new(|arg, variable_str_| {
        //    DMatrix::from_rows(&new_jac) }) ;

        let boxed_jac: Box<dyn Jac> = Box::new(JacEnum::Sparse_1(new_jac));
        self.jac_function = Some(boxed_jac);
    }

    /// Compiles residual functions to a sprs CsVec evaluator for sparse vector operations.
    ///
    /// Creates a residual function evaluator that works with sprs sparse vectors (CsVec).
    /// This method is designed for systems where the residual vector itself is sparse.
    ///
    /// # Arguments
    /// * `arg` - Independent variable name (used in IVP-style evaluation)
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Backend Details
    /// - Uses sprs::CsVec (Compressed Sparse Vector format)
    /// - Supports IVP-style evaluation with time parameter
    /// - Sequential evaluation of residual functions
    /// - Wrapped in FunEnum::Sparse_1 variant
    ///
    /// # When to Use
    /// - Sparse residual vectors (many zero elements)
    /// - Integration with sprs-based linear algebra
    /// - IVP problems where time-dependent evaluation is needed
    pub fn lambdify_residual_CsVec(&mut self, arg: &str, variable_str: Vec<&str>) {
        let vector_of_functions = &self.vector_of_functions;
        fn f(
            vector_of_functions: Vec<Expr>,
            arg: String,
            variable_str: Vec<String>,
        ) -> Box<dyn Fn(f64, &CsVec<f64>) -> CsVec<f64> + 'static> {
            Box::new(move |x: f64, v: &CsVec<f64>| -> CsVec<f64> {
                //  let mut result: CsVec<f64> = CsVec::new(n, indices, data);
                let mut result: CsVec<f64> = CsVec::empty(vector_of_functions.len());
                for (i, func) in vector_of_functions.iter().enumerate() {
                    let func = Expr::lambdify_IVP_owned(
                        func.to_owned(),
                        arg.as_str(),
                        variable_str
                            .iter()
                            .map(|s| s.as_str())
                            .collect::<Vec<_>>()
                            .clone(),
                    );
                    result.append(i, func(x, v.to_dense().to_vec()));
                }
                result
            }) //enf of box
        } // end of function
        let fun = f(
            vector_of_functions.to_owned(),
            arg.to_string(),
            variable_str.clone().iter().map(|s| s.to_string()).collect(),
        );
        let boxed_fun: Box<dyn Fun> = Box::new(FunEnum::Sparse_1(fun));
        self.residiual_function = boxed_fun;
    }

    //////////////////////////////////////////////////////////////////////////////////////////
    //                            NALGEBRA SPARSE CRATE
    ///////////////////////////////////////////////////////////////////////////////////////////

    /// Generates a sparse Jacobian evaluator using nalgebra's CsMatrix format.
    ///
    /// Creates a closure that evaluates the symbolic Jacobian and returns it as nalgebra's
    /// compressed sparse matrix format. Uses dense intermediate storage for simplicity.
    ///
    /// # Arguments
    /// * `jac` - 2D vector of symbolic partial derivatives
    /// * `vector_of_functions_len` - Number of equations in the system
    /// * `vector_of_variables_len` - Number of variables in the system
    /// * `variable_str` - Variable names as owned strings
    /// * `_arg` - Independent variable name (unused)
    ///
    /// # Returns
    /// A mutable closure that takes (time, dense_vector) and returns a sparse matrix
    ///
    /// # Implementation Details
    /// - Uses DMatrix as intermediate storage, then converts to CsMatrix
    /// - Sequential evaluation (not parallelized)
    /// - Evaluates all matrix elements regardless of sparsity
    ///
    /// # When to Use
    /// - Integration with nalgebra's sparse matrix ecosystem
    /// - When CsMatrix format is specifically required
    /// - Small to medium systems where conversion overhead is acceptable
    pub fn jacobian_generate_CsMatrix(
        jac: Vec<Vec<Expr>>,
        vector_of_functions_len: usize,
        vector_of_variables_len: usize,
        variable_str: Vec<String>,
        _arg: String,
    ) -> Box<dyn FnMut(f64, &DVector<f64>) -> CsMatrix<f64>> {
        //let arg = arg.as_str();
        //let variable_str: Vec<&str> = variable_str.iter().map(|s| s.as_str()).collect();

        Box::new(move |_x: f64, v: &DVector<f64>| -> CsMatrix<f64> {
            let _number_of_possible_non_zero: usize = jac.len();
            //let mut new_function_jacobian: CsMatrix<f64> = CsMatrix::new_uninitialized_generic(Dyn(vector_of_functions_len),
            //Dyn(vector_of_variables_len), number_of_possible_non_zero);
            let mut new_function_jacobian: DMatrix<f64> =
                DMatrix::zeros(vector_of_functions_len, vector_of_variables_len);
            for i in 0..vector_of_functions_len {
                for j in 0..vector_of_variables_len {
                    // println!("i = {}, j = {}", i, j);
                    let partial_func = Expr::lambdify_borrowed_thread_safe(
                        &jac[i][j],
                        variable_str
                            .iter()
                            .map(|s| s.as_str())
                            .collect::<Vec<_>>()
                            .as_slice(),
                    );
                    //let v_vec: Vec<f64> = v.iter().cloned().collect();
                    let v_vec: Vec<f64> = v.iter().cloned().collect();
                    //new_function_jacobian = CsMatrix::from_triplet(vector_of_functions_len,
                    // vector_of_variables_len, &[i], &[j], &[partial_func(x, v_vec.clone())]   );
                    new_function_jacobian[(i, j)] = partial_func(v_vec.as_slice());
                }
            }
            new_function_jacobian.into()
        })
    } // end of function
    /// Compiles the symbolic Jacobian to a nalgebra CsMatrix evaluator.
    ///
    /// Sets up the Jacobian function using nalgebra's compressed sparse matrix format.
    /// This method integrates with nalgebra's sparse matrix ecosystem.
    ///
    /// # Arguments
    /// * `arg` - Independent variable name
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Backend Details
    /// - Uses nalgebra::sparse::CsMatrix format
    /// - Sequential evaluation with dense intermediate storage
    /// - Historical reference only: this route is no longer wired into
    ///   `JacEnum` or solver backend selection.
    ///
    /// # When to Use
    /// - Integration with nalgebra's sparse linear algebra
    /// - When nalgebra CsMatrix format is specifically required
    /// - Medium-sized sparse systems
    pub fn lambdify_jacobian_CsMatrix(&mut self, arg: &str, variable_str: Vec<&str>) {
        let _ = (arg, variable_str);
        panic!(
            "BVP nalgebra CsMatrix (`Sparse_2`) runtime backend was removed because its \
             MatrixType implementation was incomplete. Use `Sparse`/faer, `Banded`, \
             `Sparse_1`/sprs, or `Dense` instead."
        );
    }
    ////////////////////////////////////////////////////////////////////////////////////////
    ///         FAER SPARSE CRATE
    ////////////////////////////////////////////////////////////////////////////////////////

    /// Compiles the symbolic Jacobian to a faer SparseColMat evaluator (parallel version 2 - RECOMMENDED).
    ///
    /// Creates the most efficient parallel sparse Jacobian evaluator using outer loop parallelization.
    /// This is the recommended method for large sparse systems due to superior load balancing.
    ///
    /// # Arguments
    /// * `_arg` - Independent variable name (unused)
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Performance Features
    /// - **Outer loop parallelization**: Uses `flat_map()` for optimal load balancing
    /// - **Pre-compilation**: All non-zero derivatives compiled during setup
    /// - **Thread-safe closures**: Uses `lambdify_borrowed_thread_safe()` for efficiency
    /// - **Bandwidth optimization**: Respects sparse matrix bandwidth if set
    /// - **Parallel evaluation**: Concurrent triplet assembly with Mutex protection
    ///
    /// # Parallelization Strategy
    /// ```ignore
    /// (0..n_rows).into_par_iter().flat_map(|i| {
    ///     (cols_for_row_i).filter_map(|j| compile_and_store(i, j))
    /// })
    /// ```
    ///
    /// # Why This is Best
    /// - Better load balancing than nested loops
    /// - Avoids thread contention in inner loops
    /// - Scales well with number of CPU cores
    /// - Optimal for sparse matrices with varying row densities
    ///
    /// # When to Use
    /// - **RECOMMENDED** for all large sparse systems
    /// - Production code requiring maximum performance
    /// - Systems with > 1000 variables
    pub fn lambdify_jacobian_SparseColMat_parallel2(
        &mut self,
        _arg: &str,
        variable_str: Vec<&str>,
    ) {
        let binding = self.parameter_binding_for_callbacks();
        let flattened_argument_names_owned =
            self.flattened_argument_names_owned(variable_str.as_slice());
        if self.symbolic_assembly_backend == BvpSymbolicAssemblyBackend::AtomView {
            if let Some(entries) = self.atom_symbolic_jacobian_sparse.clone() {
                let (rows, cols) = self
                    .atom_discretized_system
                    .as_ref()
                    .map(|system| {
                        (
                            system.vector_of_functions.len(),
                            system.vector_of_variables.len(),
                        )
                    })
                    .unwrap_or((
                        self.vector_of_functions.len(),
                        self.vector_of_variables.len(),
                    ));
                let telemetry = self.start_atom_lambdify_telemetry();
                self.jac_function = Some(atom_lambdify::sparse_jacobian_with_binding_and_policy(
                    entries,
                    rows,
                    cols,
                    flattened_argument_names_owned,
                    binding,
                    self.parameters_string.len(),
                    telemetry,
                    self.lambdify_execution_policy,
                ));
                return;
            }
        }
        let telemetry = self.start_legacy_lambdify_telemetry();
        let sparse_entries = if !self.symbolic_jacobian_sparse.is_empty() {
            self.symbolic_jacobian_sparse.clone()
        } else {
            // Preserve direct legacy fixtures that construct only the dense symbolic cache.
            self.symbolic_jacobian
                .iter()
                .enumerate()
                .flat_map(|(row, values)| {
                    values
                        .iter()
                        .enumerate()
                        .filter_map(move |(col, expression)| {
                            (!expression.is_zero()).then(|| (row, col, expression.clone()))
                        })
                })
                .collect()
        };
        self.jac_function = Some(
            legacy_lambdify::faer_sparse_jacobian_with_binding_and_policy(
                sparse_entries,
                self.vector_of_functions.len(),
                self.vector_of_variables.len(),
                flattened_argument_names_owned,
                binding,
                self.parameters_string.len(),
                telemetry,
                self.lambdify_execution_policy,
            ),
        );
    }
    ////////////////////////////////RESIDUAL LAMBDIFICATION/////////////////////////////////////////////////////////

    /// Compiles residual functions to a faer Col evaluator (parallel version 2 - RECOMMENDED).
    ///
    /// Creates the most efficient parallel residual function evaluator using borrowed references
    /// for optimal memory usage and performance.
    ///
    /// # Arguments
    /// * `_arg` - Independent variable name (unused)
    /// * `variable_str` - Vector of variable names as string slices
    ///
    /// # Performance Features
    /// - **Parallel compilation**: Functions compiled concurrently using rayon
    /// - **Borrowed references**: Uses `lambdify_borrowed_thread_safe()` for efficiency
    /// - **Parallel evaluation**: Concurrent function evaluation during residual computation
    /// - **Memory efficient**: Avoids unnecessary cloning during compilation
    ///
    /// # Why This is Recommended
    /// - More memory efficient than parallel version 1
    /// - Uses borrowed references instead of cloning expressions
    /// - Better performance for large symbolic expressions
    /// - Optimal balance of compilation and evaluation speed
    ///
    /// # When to Use
    /// - **RECOMMENDED** for all parallel residual evaluation
    /// - Large systems with complex symbolic expressions
    /// - Production code requiring optimal memory usage
    pub fn lambdify_residual_Col_parallel2(&mut self, _arg: &str, variable_str: Vec<&str>) {
        let binding = self.parameter_binding_for_callbacks();
        let variable_str_owned = self.flattened_argument_names_owned(variable_str.as_slice());
        if self.symbolic_assembly_backend == BvpSymbolicAssemblyBackend::AtomView {
            if let Some(functions) = self
                .atom_discretized_system
                .as_ref()
                .map(|system| system.vector_of_functions.clone())
            {
                let telemetry = self.start_atom_lambdify_telemetry();
                self.residiual_function = atom_lambdify::sparse_residual_with_binding_and_policy(
                    functions,
                    variable_str_owned,
                    binding,
                    self.parameters_string.len(),
                    telemetry,
                    self.lambdify_execution_policy,
                );
                return;
            }
        }
        let telemetry = self.start_legacy_lambdify_telemetry();
        self.residiual_function = legacy_lambdify::faer_sparse_residual_with_binding_and_policy(
            self.vector_of_functions.clone(),
            variable_str_owned,
            binding,
            self.parameters_string.len(),
            telemetry,
            self.lambdify_execution_policy,
        );
    }

    /// Compiles the symbolic Jacobian into the native banded storage path.
    ///
    /// This branch reuses the sparse symbolic Jacobian cache from
    /// `symbolic_jacobian_sparse`, compiles each nonzero entry once, and returns
    /// a runtime callback that builds `BandedAssembly` directly. The outer BVP
    /// solver still talks to the legacy trait-object API, but the matrix payload
    /// now carries banded layout metadata all the way to the linear solve.
    pub fn lambdify_jacobian_Banded(&mut self, _arg: &str, _variable_str: Vec<&str>) {
        // Keep parameters in the compiled callback ABI. Numeric values are
        // rebound through the shared prepared handle between solves.
        self.parameter_binding_for_callbacks();
        let plan = self
            .infer_banded_structure_plan()
            .expect("banded backend requires node-major compatible Jacobian structure");
        let mut config = self.banded_lambdify_config.clone();
        config.linear_solver_config = self.banded_linear_solver_config;
        config.telemetry_mode = self.lambdify_telemetry_mode;
        let runtime = self
            .generate_banded_jacobian_runtime_from_plan_parallel(&plan, &config)
            .expect("failed to compile banded jacobian evaluator");
        self.direct_banded_jacobian_telemetry = Some(runtime.telemetry());
        let evaluator = runtime.into_callback();
        let layout = plan.layout;
        let solver_config = config.linear_solver_config;

        let jacobian = convert_to_jac(Box::new(move |_x: f64, vec: &dyn VectorType| {
            // Accept any runtime vector storage that can be losslessly exposed as a dense
            // state vector. This keeps the banded backend compatible with both legacy
            // lambdify callers and generated handoff paths that may still pass faer columns.
            let dense = dense_slice_or_copy(vec);
            let assembly = evaluator(dense.as_ref())
                .expect("banded jacobian evaluation failed for current iterate");
            Box::new(BandedMatrixType::new(assembly, layout, solver_config)) as Box<dyn MatrixType>
        }));

        self.jac_function = Some(jacobian);
    }

    /// Compiles the residual vector for the banded backend.
    ///
    /// Even though the Jacobian storage is banded, the residual itself remains a
    /// plain dense vector. This keeps the outer Newton/damping logic unchanged
    /// while still letting the expensive linear algebra run through the new
    /// banded path.
    pub fn lambdify_residual_Banded(&mut self, _arg: &str, _variable_str: Vec<&str>) {
        // The residual and Jacobian must observe the same prepared binding.
        self.parameter_binding_for_callbacks();
        let mut config = self.banded_lambdify_config.clone();
        config.linear_solver_config = self.banded_linear_solver_config;
        let evaluator = self
            .generate_banded_residual_with_config(&config)
            .expect("failed to compile banded residual evaluator");

        let residual = convert_to_fun(Box::new(move |_x: f64, vec: &dyn VectorType| {
            // Residual evaluation should not care whether the outer runtime stores the
            // iterate as a DVector or a faer column. Normalize once here instead of
            // forcing every caller to guess the exact vector contract.
            let dense = dense_slice_or_copy(vec);
            let values = evaluator(dense.as_ref())
                .expect("banded residual evaluation failed for current iterate");
            Box::new(DVector::from_vec(values)) as Box<dyn VectorType>
        }));

        self.residiual_function = residual;
    }
    ///////////////////////////////////////////////////////////////////////////
    ///              DISCRETIZED FUNCTIONS
    ///////////////////////////////////////////////////////////////////////////

    /// Removes numeric suffix from discretized variable names for efficient processing.
    ///
    /// Extracts the base variable name from discretized variables like "y_0", "y_1", "z_0", etc.
    /// This is used for mapping discretized variables back to their original names for bounds
    /// and tolerance processing.
    ///
    /// # Arguments
    /// * `input` - Variable name with potential numeric suffix (e.g., "y_0", "z_15")
    ///
    /// # Returns
    /// Base variable name without suffix (e.g., "y", "z")
    ///
    /// # Performance
    /// Uses `rfind('_')` for O(n) string processing, much faster than regex alternatives.
    ///
    /// # Examples
    /// ```ignore
    /// assert_eq!(remove_numeric_suffix("y_0"), "y");
    /// assert_eq!(remove_numeric_suffix("temperature_15"), "temperature");
    /// assert_eq!(remove_numeric_suffix("x"), "x"); // No suffix
    /// ```
    pub fn remove_numeric_suffix(input: &str) -> String {
        if let Some(pos) = input.rfind('_') {
            input[..pos].to_string()
        } else {
            input.to_string()
        }
    }
    /// Applies discretization scheme to a single equation at a specific time step.
    ///
    /// Transforms a continuous ODE equation into its discretized form using the specified
    /// numerical scheme. Handles variable renaming and time substitution for the discretization.
    ///
    /// # Arguments
    /// * `matrix_of_names` - 2D matrix of discretized variable names [time_step][variable_index]
    /// * `eq_i` - The symbolic equation to discretize
    /// * `values` - Original variable names (e.g., ["y", "z"])
    /// * `arg` - Independent variable name (typically "t" or "x")
    /// * `j` - Current time step index
    /// * `t` - Time value at step j
    /// * `scheme` - Discretization scheme ("forward" or "trapezoid")
    ///
    /// # Returns
    /// Discretized equation with renamed variables and substituted time value
    ///
    /// # Supported Schemes
    /// - **"forward"**: Forward Euler - f(t_j, y_j)
    /// - **"trapezoid"**: Trapezoidal rule - 0.5 * (f(t_j, y_j) + f(t_{j+1}, y_{j+1}))
    ///
    /// # Variable Renaming
    /// Original variables like "y", "z" become "y_j", "z_j" for the j-th time step.
    fn eq_step(
        matrix_of_names: &[Vec<String>],
        eq_i: &Expr,
        values: &[String],
        arg: &str,
        j: usize,
        t: f64,
        scheme: &str,
    ) -> Expr {
        let vec_of_names_on_step = &matrix_of_names[j];
        let hashmap_for_rename: HashMap<String, String> = values
            .iter()
            .zip(vec_of_names_on_step.iter())
            .map(|(k, v)| (k.to_string(), v.to_string()))
            .collect();
        //set time value of j-th time step
        let eq_step_j = eq_i.rename_variables(&hashmap_for_rename);
        let eq_step_j = eq_step_j.set_variable(arg, t);

        match scheme {
            "forward" => eq_step_j,
            "trapezoid" => {
                let vec_of_names_on_step = &matrix_of_names[j + 1];
                let hashmap_for_rename: HashMap<String, String> = values
                    .iter()
                    .zip(vec_of_names_on_step.iter())
                    .map(|(k, v)| (k.to_string(), v.to_string()))
                    .collect();

                let eq_step_j_plus_1 = eq_i
                    .rename_variables(&hashmap_for_rename)
                    .set_variable(arg, t);
                Expr::Const(0.5) * (eq_step_j + eq_step_j_plus_1)
            }
            _ => panic!("Invalid scheme"),
        }
    }
    //
    /// Creates the time/spatial mesh for BVP discretization.
    ///
    /// Generates the discretization mesh either from explicit mesh points or by creating
    /// a uniform grid. Returns step sizes as symbolic expressions and mesh points.
    ///
    /// # Arguments
    /// * `n_steps` - Number of discretization steps (if uniform mesh)
    /// * `h` - Step size (if uniform mesh)
    /// * `mesh` - Explicit mesh points (overrides n_steps/h if provided)
    /// * `t0` - Starting point of the domain
    ///
    /// # Returns
    /// Tuple of (step_sizes, mesh_points, n_steps) where:
    /// - `step_sizes`: Vector of Expr::Const representing h_i = t_{i+1} - t_i
    /// - `mesh_points`: Vector of f64 mesh coordinates
    /// - `n_steps`: Total number of mesh points
    ///
    /// # Mesh Types
    /// - **Explicit mesh**: Uses provided mesh points with variable step sizes
    /// - **Uniform mesh**: Creates evenly spaced points with constant step size
    ///
    /// # Default Values
    /// - n_steps: 101 (100 intervals + 1 endpoint)
    /// - h: 1.0
    fn create_mesh(
        &self,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        t0: f64,
    ) -> (Vec<Expr>, Vec<f64>, usize) {
        // mesh of t's can be defined directly or by size of step -h, and number of steps
        if let Some(mesh) = mesh {
            let n_steps = mesh.len();
            info!("mesh with n_steps = {} is defined directly", n_steps);
            let H: Vec<Expr> = mesh
                .windows(2)
                .map(|window| Expr::Const(window[1] - window[0]))
                .collect();
            (H, mesh, n_steps)
        } else {
            let n_steps = n_steps.unwrap_or(100) + 1;
            info!(
                "mesh is not defined, creating evenly distributed mesh of length {}",
                n_steps
            );
            let h = h.unwrap_or(1.0);
            let H: Vec<Expr> = vec![Expr::Const(h); n_steps - 1]; // number of intervals = n_steps -1
            let T_list: Vec<f64> = (0..n_steps).map(|i| t0 + (i as f64) * h).collect();
            (H, T_list, n_steps)
        }
    }

    /// Processes variable bounds and tolerances for discretized variables with optimized string operations.
    ///
    /// Efficiently maps bounds and tolerances from original variable names to all discretized
    /// variable instances. Uses fast string processing to extract base names from discretized variables.
    ///
    /// # Arguments
    /// * `Bounds` - Optional bounds map: {"y": (min, max), "z": (min, max)}
    /// * `rel_tolerance` - Optional tolerance map: {"y": tol, "z": tol}
    /// * `flat_list_of_names` - All discretized variable names ["y_0", "y_1", "z_0", "z_1", ...]
    ///
    /// # Performance Optimizations
    /// - **Fast string processing**: Uses `rfind('_')` instead of regex (10-100x faster)
    /// - **Pre-allocation**: Vectors allocated with known capacity
    /// - **Single pass**: Processes all variables in one iteration
    ///
    /// # Output
    /// Sets `self.bounds` and `self.rel_tolerance_vec` with per-variable values:
    /// - bounds[i] = (min, max) for discretized variable i
    /// - rel_tolerance_vec[i] = tolerance for discretized variable i
    ///
    /// # Example
    /// ```ignore
    /// // Input: Bounds = {"y": (0.0, 10.0)}, flat_list = ["y_0", "y_1", "z_0"]
    /// // Output: bounds = [(0.0, 10.0), (0.0, 10.0), default_for_z]
    /// ```
    fn process_bounds_and_tolerances(
        &mut self,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        flat_list_of_names: Vec<String>,
    ) {
        let len = flat_list_of_names.len();

        self.bounds = Bounds.as_ref().map(|bounds_map| {
            let mut vec_of_bounds = Vec::with_capacity(len);
            for name in &flat_list_of_names {
                // Fast string processing: find last underscore and extract base name
                let base_name = if let Some(pos) = name.rfind('_') {
                    &name[..pos]
                } else {
                    name
                };

                if let Some(&bound_pair) = bounds_map.get(base_name) {
                    vec_of_bounds.push(bound_pair);
                }
            }
            vec_of_bounds
        });

        self.rel_tolerance_vec = rel_tolerance.as_ref().map(|tolerance_map| {
            let mut vec_of_tolerance = Vec::with_capacity(len);
            for name in &flat_list_of_names {
                // Fast string processing: find last underscore and extract base name
                let base_name = if let Some(pos) = name.rfind('_') {
                    &name[..pos]
                } else {
                    name
                };

                if let Some(&tolerance) = tolerance_map.get(base_name) {
                    vec_of_tolerance.push(tolerance);
                }
            }
            vec_of_tolerance
        });
    }

    /// High-performance parallel BVP discretization with comprehensive optimization.
    ///
    /// Converts a continuous BVP system into a discretized algebraic system ready for Newton-Raphson
    /// solving. This is the core discretization method with extensive performance optimizations.
    ///
    /// # Arguments
    /// * `eq_system` - Vector of symbolic ODE equations [dy/dt = f1(t,y,z), dz/dt = f2(t,y,z)]
    /// * `values` - Original variable names ["y", "z"]
    /// * `arg` - Independent variable name ("t" for time, "x" for space)
    /// * `t0` - Starting point of the domain
    /// * `n_steps` - Number of discretization steps (creates n_steps+1 points)
    /// * `h` - Step size for uniform mesh
    /// * `mesh` - Explicit mesh points (overrides n_steps/h)
    /// * `BorderConditions` - Boundary conditions: {"y": [(0, value), (1, value)]}
    ///   - 0 = initial condition, 1 = final condition
    /// * `Bounds` - Variable bounds for constrained solving
    /// * `rel_tolerance` - Per-variable relative tolerances
    /// * `scheme` - Discretization scheme ("forward" or "trapezoid")
    ///
    /// # Performance Optimizations
    /// 1. **Parallel discretization**: Uses rayon for concurrent equation processing
    /// 2. **Boundary condition caching**: Pre-computes HashMaps for O(1) BC lookups
    /// 3. **Variable tracking**: Tracks which variables appear in each equation for smart Jacobian
    /// 4. **Memory pre-allocation**: Pre-allocates vectors with known capacity
    /// 5. **Efficient BC application**: Parallel boundary condition substitution
    /// 6. **Fast string processing**: Optimized variable name processing
    ///
    /// # Output
    /// Creates discretized system: y_{i+1} - y_i - h*f(t_i, y_i, z_i) = 0
    /// Sets up all internal data structures for subsequent Jacobian computation.
    ///
    /// # Timing
    /// Provides detailed timing breakdown of each optimization phase.
    pub fn discretization_system_BVP_par(
        &mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        t0: f64,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
    ) {
        let total_start = Instant::now();
        let mut timer_hash: HashMap<String, f64> = HashMap::new();
        let (H, T_list, n_steps) = self.create_mesh(n_steps, h, mesh, t0);
        let (matrix_of_expr, matrix_of_names) = Expr::IndexedVarsMatrix(n_steps, values.clone());
        let bc_handling = Instant::now();
        // Pre-compute boundary conditions lookup for O(1) access
        let bc_lookup: HashMap<String, HashMap<usize, f64>> = BorderConditions
            .into_iter()
            .map(|(k, v)| (k, v.into_iter().collect()))
            .collect();

        // Pre-compute variables to exclude from boundary conditions
        let mut vars_for_boundary_conditions = HashMap::new();
        let mut vars_to_exclude = HashSet::new();
        let mut bc_pos_n_values = Vec::new();

        for (var_name, conditions) in &bc_lookup {
            if let Some(var_idx) = values.iter().position(|v| v == var_name) {
                for (&pos, &value) in conditions {
                    match pos {
                        0 => {
                            // Initial condition
                            let var_name = &matrix_of_names[0][var_idx];
                            vars_for_boundary_conditions.insert(var_name.clone(), value);
                            vars_to_exclude.insert(var_name.clone());
                            let full_pos = 0 * values.len() + var_idx;
                            bc_pos_n_values.push((full_pos, 0 as usize, value));
                        }
                        1 => {
                            // Final condition
                            let var_name = &matrix_of_names[n_steps - 1][var_idx];
                            vars_for_boundary_conditions.insert(var_name.clone(), value);
                            vars_to_exclude.insert(var_name.clone());
                            let full_pos = (n_steps - 1) * values.len() + var_idx;
                            bc_pos_n_values.push((full_pos, 1 as usize, value));
                        }
                        _ => {}
                    }
                }
            }
        }
        timer_hash.insert(
            "bc handling".to_string(),
            bc_handling.elapsed().as_millis() as f64,
        );
        // DISCRETAZING EQUATIONS
        println!("creating discretization equations");
        let discretization_start = Instant::now();

        // Optimized discretization with variable tracking
        let (discreditized_system, variables_for_all_discrete): (Vec<Expr>, Vec<Vec<String>>) = (0
            ..n_steps - 1)
            .into_par_iter()
            .flat_map(|j| {
                let t = T_list[j];
                eq_system
                    .par_iter()
                    .enumerate()
                    .map(|(i, eq_i)| {
                        let eq_step_j =
                            Self::eq_step(&matrix_of_names, eq_i, &values, &arg, j, t, &scheme);
                        let Y_j_plus_1 = &matrix_of_expr[j + 1][i];
                        let Y_j = &matrix_of_expr[j][i];
                        let res_ij = Y_j_plus_1.clone() - Y_j.clone() - H[j].clone() * eq_step_j;

                        // Track variables used in this equation (excluding boundary condition vars)
                        let mut vars_in_equation = Vec::new();
                        let y_j_plus_1_name = &matrix_of_names[j + 1][i];
                        let y_j_name = &matrix_of_names[j][i];

                        if !vars_to_exclude.contains(y_j_plus_1_name) {
                            vars_in_equation.push(y_j_plus_1_name.clone());
                        }
                        if !vars_to_exclude.contains(y_j_name) {
                            vars_in_equation.push(y_j_name.clone());
                        }

                        // Add variables from eq_step_j (from original equation at this time step)
                        for var_idx in 0..values.len() {
                            let var_name = &matrix_of_names[j][var_idx];
                            if !vars_to_exclude.contains(var_name) {
                                vars_in_equation.push(var_name.clone());
                            }
                        }

                        // Defer simplification until after boundary-condition
                        // substitution below.  Simplifying here and then
                        // simplifying the same row again after BC application
                        // performs two full recursive tree walks in the
                        // ExprLegacy path without changing the final form.
                        (res_ij, vars_in_equation)
                    })
                    .collect::<Vec<_>>()
            })
            .unzip();

        self.variables_for_all_disrete = variables_for_all_discrete;
        timer_hash.insert(
            "discretization of equations".to_string(),
            discretization_start.elapsed().as_millis() as f64,
        );
        let start_flat_list = Instant::now();

        // Efficient flat list creation with pre-allocation
        let total_vars = values.len() * n_steps;
        let mut flat_list_of_names = Vec::with_capacity(total_vars);
        let mut flat_list_of_expr = Vec::with_capacity(total_vars);
        for time_idx in 0..n_steps {
            for var_idx in 0..values.len() {
                let name = &matrix_of_names[time_idx][var_idx];
                if !vars_to_exclude.contains(name) {
                    flat_list_of_names.push(name.clone());
                    flat_list_of_expr.push(matrix_of_expr[time_idx][var_idx].clone());
                }
            }
        }

        timer_hash.insert(
            "flat list creation".to_string(),
            start_flat_list.elapsed().as_millis() as f64,
        );
        let BC_application_start = Instant::now();
        // Apply boundary conditions to discretized system in parallel
        let discreditized_system_with_BC: Vec<Expr> = discreditized_system
            .into_par_iter()
            .map(|mut eq_i| {
                for (var_name, &value) in &vars_for_boundary_conditions {
                    eq_i = eq_i.set_variable(var_name, value);
                }
                eq_i.simplify()
            })
            .collect();

        let discreditized_system_flat = discreditized_system_with_BC;
        timer_hash.insert(
            "BC application".to_string(),
            BC_application_start.elapsed().as_millis() as f64,
        );
        let consistency_start = Instant::now();
        // Simplified consistency test using pre-computed variables
        let hashset_of_vars: HashSet<&String> = flat_list_of_names.iter().collect();
        let mut missing_vars = Vec::new();

        for var_list in &self.variables_for_all_disrete {
            for var in var_list {
                if !hashset_of_vars.contains(var) {
                    missing_vars.push(var.clone());
                }
            }
        }

        if !missing_vars.is_empty() {
            missing_vars.sort_unstable();
            missing_vars.dedup();
            panic!("Variables not found in system: {:?}", missing_vars);
        }
        timer_hash.insert(
            "consistency test".to_string(),
            consistency_start.elapsed().as_millis() as f64,
        );

        self.vector_of_functions = discreditized_system_flat;
        self.vector_of_variables = flat_list_of_expr;
        self.variable_string = flat_list_of_names.clone();
        self.BC_pos_n_values = bc_pos_n_values;
        let bounds_and_tolerances_start = Instant::now();
        self.process_bounds_and_tolerances(Bounds, rel_tolerance, flat_list_of_names);
        timer_hash.insert(
            "bounds and tolerances".to_string(),
            bounds_and_tolerances_start.elapsed().as_millis() as f64,
        );

        // timing
        let total_end = total_start.elapsed().as_millis() as f64;
        normalize_timer_percent(timer_hash.get_mut("bc handling").unwrap(), total_end);
        normalize_timer_percent(
            timer_hash.get_mut("discretization of equations").unwrap(),
            total_end,
        );
        normalize_timer_percent(timer_hash.get_mut("BC application").unwrap(), total_end);
        normalize_timer_percent(timer_hash.get_mut("flat list creation").unwrap(), total_end);
        normalize_timer_percent(timer_hash.get_mut("consistency test").unwrap(), total_end);
        normalize_timer_percent(
            timer_hash.get_mut("bounds and tolerances").unwrap(),
            total_end,
        );
        timer_hash.insert("total time, sec".to_string(), total_end);

        let mut table = Builder::from(timer_hash.clone()).build();
        table.with(Style::modern_rounded());
        println!("{}", table.to_string());
    }

    /// Automatically detects the bandwidth of the sparse Jacobian matrix using parallel computation.
    ///
    /// Analyzes the symbolic Jacobian to determine the banded structure, computing the number
    /// of sub- and super-diagonals. This information is crucial for banded matrix optimizations.
    ///
    /// # Algorithm
    /// For each non-zero element at position (i,j):
    /// - If j > i: contributes to super-diagonal width (ku)
    /// - If i > j: contributes to sub-diagonal width (kl)
    ///
    /// # Performance Features
    /// - **Parallel computation**: Uses rayon to process rows concurrently
    /// - **Reduction pattern**: Efficiently combines results from parallel workers
    /// - **Early termination**: Skips zero elements for efficiency
    ///
    /// # Output
    /// Sets `self.bandwidth = Some((kl, ku))` where:
    /// - kl = maximum number of sub-diagonals
    /// - ku = maximum number of super-diagonals
    ///
    /// # When to Use
    /// - Automatically called by `generate_BVP()` if bandwidth not provided
    /// - Essential for banded matrix solvers and storage optimization
    /// - Particularly important for large sparse systems
    fn find_bandwidths(&mut self) {
        crate::symbolic::bvp::legacy_symbolic::find_bandwidths(self);
    }

    /// Computes Banded metadata directly from the AtomView sparse layout.
    ///
    /// The legacy detector reads `symbolic_jacobian_sparse`, which is
    /// intentionally empty for the native AtomView route.  Reusing it would
    /// silently produce the sentinel `(0, 0)` and later reject valid
    /// off-diagonal entries during Banded AOT preparation.
    fn find_atom_bandwidths(&mut self) {
        let entries = self
            .atom_symbolic_jacobian_sparse
            .as_deref()
            .unwrap_or_default();
        let (kl, ku) = entries.iter().fold((0usize, 0usize), |(kl, ku), entry| {
            if entry.row >= entry.col {
                (kl.max(entry.row - entry.col), ku)
            } else {
                (kl, ku.max(entry.col - entry.row))
            }
        });
        self.bandwidth = Some((kl, ku));
    }

    /// **MAIN FUNCTION**: Complete BVP symbolic-to-numerical transformation pipeline.
    ///
    /// This is the primary entry point for BVP solving. It orchestrates the entire process from
    /// symbolic ODE system to high-performance numerical functions ready for Newton-Raphson solving.
    ///
    /// # Complete Workflow
    /// 1. **Discretization**: Converts continuous BVP to algebraic system
    /// 2. **Symbolic Jacobian**: Computes analytical partial derivatives
    /// 3. **Bandwidth Detection**: Analyzes sparsity structure
    /// 4. **Function Compilation**: Creates optimized numerical evaluators
    /// 5. **Performance Timing**: Provides detailed timing breakdown
    ///
    /// # Arguments
    /// * `eq_system` - Vector of symbolic ODE equations
    /// * `values` - Original variable names ["y", "z"]
    /// * `arg` - Independent variable name ("t" or "x")
    /// * `t0` - Domain starting point
    /// * `param` - Parameter name (defaults to `arg` if None)
    /// * `n_steps` - Number of discretization steps
    /// * `h` - Step size for uniform mesh
    /// * `mesh` - Explicit mesh points
    /// * `BorderConditions` - Boundary conditions specification
    /// * `Bounds` - Variable bounds for constrained solving
    /// * `rel_tolerance` - Per-variable relative tolerances
    /// * `scheme` - Discretization scheme ("forward" or "trapezoid")
    /// * `method` - Matrix backend ("Dense", "Sparse", "Sparse_1", or "Banded")
    /// * `bandwidth` - Optional pre-computed bandwidth (auto-detected if None)
    ///
    /// # Supported Matrix Backends
    /// - **"Dense"**: nalgebra DMatrix - best for small, dense systems
    /// - **"Sparse"**: faer SparseColMat - recommended for large sparse systems
    /// - **"Sparse_1"**: sprs CsMat - legacy sparse support
    /// - **"Banded"**: Lapack-style banded storage for narrow-band Jacobians
    ///
    /// # Performance Features
    /// - Comprehensive timing analysis with percentage breakdown
    /// - Automatic bandwidth detection for sparse optimization
    /// - Smart sparsity-aware Jacobian computation
    /// - Multi-backend support for optimal performance
    ///
    /// # Output
    /// Sets up `self.jac_function` and `self.residiual_function` ready for BVP solving.
    /// Prints detailed timing table showing performance breakdown.
    ///
    /// # Example Usage
    /// ```ignore
    /// let mut jacobian = Jacobian::new();
    /// jacobian.generate_BVP(
    ///     equations, variables, "t", 0.0, None, Some(100), None, None,
    ///     boundary_conditions, bounds, tolerances, "trapezoid".to_string(),
    ///     "Sparse".to_string(), None
    /// );
    /// ```
    pub fn generate_BVP(
        &mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
    ) {
        self.generate_BVP_with_params(
            eq_system,
            values,
            arg,
            None,
            t0,
            param,
            n_steps,
            h,
            mesh,
            BorderConditions,
            Bounds,
            rel_tolerance,
            scheme,
            method,
            bandwidth,
        );
    }

    /// Same as `generate_BVP`, but with explicit symbolic parameter names that
    /// affect evaluation without becoming Newton unknowns.
    pub fn generate_BVP_with_params(
        &mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
    ) {
        let (arg_name, indexed_values_owned, matrix_backend, mut timer_hash, total_start) = self
            .prepare_symbolic_bvp_stage_with_params(
                eq_system,
                values,
                arg,
                params,
                t0,
                n_steps,
                h,
                mesh,
                BorderConditions,
                Bounds,
                rel_tolerance,
                scheme,
                method,
                bandwidth,
            );
        let param = param.unwrap_or(arg_name);
        let indexed_values: Vec<&str> = indexed_values_owned.iter().map(|x| x.as_str()).collect();

        let now = Instant::now();
        self.compile_jacobian_for_matrix_backend(
            param.as_str(),
            indexed_values.clone(),
            matrix_backend,
        );
        timer_hash.insert(
            "jacobian lambdify time".to_string(),
            now.elapsed().as_secs_f64(),
        );
        info!("Jacobian lambdified");
        if !matches!(
            matrix_backend,
            BvpMatrixBackend::Banded | BvpMatrixBackend::FaerSparseCol
        ) {
            let n = &self.symbolic_jacobian.len();
            for vec_s in &self.symbolic_jacobian {
                assert_eq!(vec_s.len(), *n, "jacobian not square ");
            }
        } else {
            if self.symbolic_assembly_backend == BvpSymbolicAssemblyBackend::AtomView {
                assert!(
                    self.atom_symbolic_jacobian_sparse.is_some(),
                    "AtomView sparse Jacobian payload must be available for compilation"
                );
            } else {
                assert!(
                    !self.symbolic_jacobian_sparse.is_empty(),
                    "ExprLegacy sparse-first Jacobian cache must be available for compilation"
                );
            }
        }

        let now = Instant::now();
        self.compile_residual_for_matrix_backend(param.as_str(), indexed_values, matrix_backend);
        timer_hash.insert(
            "residual functions lambdify time".to_string(),
            now.elapsed().as_secs_f64(),
        );
        info!("Residuals vector lambdified");

        let total_end = total_start.elapsed().as_secs_f64();
        self.finalize_generate_bvp_timer_table(timer_hash, total_end);
    }

    /// Same symbolic/discretization path as `generate_BVP_with_params`, but the
    /// numerical backend branch is selected through the shared lifecycle layer.
    ///
    /// This is a thin staged integration entrypoint:
    /// - the legacy faer/lambdify path remains the main execution path,
    /// - AOT selection metadata is returned when an artifact is registered or compiled,
    /// - the caller can decide whether to continue with lambdify fallback or a
    ///   compiled AOT branch above this module.
    pub fn generate_BVP_with_backend_selection(
        &mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
    ) -> BvpSparseExecutionPlan {
        let (arg_name, indexed_values_owned, _matrix_backend, timer_hash, total_start) = self
            .prepare_symbolic_bvp_stage_with_params(
                eq_system,
                values,
                arg,
                params,
                t0,
                n_steps,
                h,
                mesh,
                BorderConditions,
                Bounds,
                rel_tolerance,
                scheme,
                method,
                bandwidth,
            );
        let effective_arg_name = param.unwrap_or(arg_name);
        let equation_count = self
            .atom_discretized_system
            .as_ref()
            .map(|system| system.vector_of_functions.len())
            .unwrap_or(self.vector_of_functions.len());
        let residual_strategy = recommended_residual_chunking_for_parallelism(equation_count, 4);
        let jacobian_strategy = recommended_row_chunking_for_parallelism(equation_count, 4);

        self.try_finish_sparse_backend_generation(
            effective_arg_name,
            indexed_values_owned,
            timer_hash,
            total_start,
            residual_strategy,
            jacobian_strategy,
            policy,
            resolver,
        )
        .unwrap_or_else(|error| panic!("BVP AOT backend preparation failed: {error:?}"))
    }

    /// Same as `generate_BVP_with_backend_selection`, but with explicit
    /// residual and sparse-Jacobian chunking strategies supplied by the outer
    /// solver layer.
    #[allow(clippy::too_many_arguments)]
    pub fn generate_BVP_with_backend_selection_and_chunking(
        &mut self,
        eq_system: Vec<Expr>,
        values: Vec<String>,
        arg: String,
        params: Option<&[&str]>,
        t0: f64,
        param: Option<String>,
        n_steps: Option<usize>,
        h: Option<f64>,
        mesh: Option<Vec<f64>>,
        BorderConditions: HashMap<String, Vec<(usize, f64)>>,
        Bounds: Option<HashMap<String, (f64, f64)>>,
        rel_tolerance: Option<HashMap<String, f64>>,
        scheme: String,
        method: String,
        bandwidth: Option<(usize, usize)>,
        policy: BackendSelectionPolicy,
        resolver: Option<&AotResolver>,
        residual_strategy: ResidualChunkingStrategy,
        jacobian_strategy: SparseChunkingStrategy,
    ) -> BvpSparseExecutionPlan {
        let (arg_name, indexed_values_owned, _matrix_backend, timer_hash, total_start) = self
            .prepare_symbolic_bvp_stage_with_params(
                eq_system,
                values,
                arg,
                params,
                t0,
                n_steps,
                h,
                mesh,
                BorderConditions,
                Bounds,
                rel_tolerance,
                scheme,
                method,
                bandwidth,
            );
        let effective_arg_name = param.unwrap_or(arg_name);

        self.try_finish_sparse_backend_generation(
            effective_arg_name,
            indexed_values_owned,
            timer_hash,
            total_start,
            residual_strategy,
            jacobian_strategy,
            policy,
            resolver,
        )
        .unwrap_or_else(|error| panic!("BVP AOT backend preparation failed: {error:?}"))
    }
} // end of impl