rustial_engine/

async_data.rs

//! Async retained data pipeline (worker-equivalent architecture.
//!
//! This module provides the infrastructure for offloading heavy per-frame CPU
//! work (terrain mesh generation, vector tessellation, symbol placement) to
//! background tasks, matching the MapLibre/Mapbox web worker model.
//!
//! ## Architecture
//!
//! The [`DataTaskPool`] trait is an executor-agnostic interface that the host
//! or renderer injects into [`MapState`](crate::MapState).  The engine
//! submits work items via typed spawn methods and polls completed results
//! via [`DataTaskResultReceiver::try_recv`] each frame.
//!
//! When no task pool is set, `MapState` falls back to synchronous inline
//! execution (current behavior), ensuring backward compatibility.
//!
//! ## Task families
//!
//! | Family | Input | Output |
//! |--------|-------|--------|
//! | Terrain | `TerrainTaskInput` (tile, elevation grid, config) | `TerrainTaskOutput` (mesh + hillshade) |
//! | Vector | `VectorTaskInput` (features, style, terrain samples, projection) | `VectorTaskOutput` (tessellated mesh + symbol candidates) |
//!
//! ## Retained model
//!
//! Completed results are stored per cache key and reused across frames until
//! invalidated by source data changes, camera origin movement, or zoom-level
//! style property changes.
//!
//! ## Per-tile vector buckets (MapLibre parity)
//!
//! Vector features can be spatially partitioned into per-tile buckets using
//! [`partition_features_by_tile`].  Each bucket is dispatched, tessellated,
//! and cached independently via [`VectorBucketKey`], matching MapLibre's
//! per-tile bucket architecture.  Only buckets whose tiles are in the
//! visible set are collected for rendering; off-screen buckets are retained
//! in the cache for reuse when the camera pans back.

use crate::camera_projection::CameraProjection;
use crate::layers::{FeatureProvenance, VectorMeshData, VectorStyle};
use crate::symbols::SymbolCandidate;
use crate::terrain::{PreparedHillshadeRaster, TerrainMeshData};
use crate::tile_manager::TileTextureRegion;
use crate::geometry::{Feature, FeatureCollection, Geometry};
use rustial_math::{geo_to_tile, ElevationGrid, GeoCoord, TileId};
use std::collections::HashMap;
use std::sync::Arc;

// ---------------------------------------------------------------------------
// DataTaskPool trait
// ---------------------------------------------------------------------------

/// A result receiver for completed async data tasks.
///
/// Implementations should be non-blocking: [`try_recv`](Self::try_recv)
/// returns `None` when no result is available yet.
pub trait DataTaskResultReceiver<T: Send + 'static>: Send + Sync {
    /// Try to receive a completed result without blocking.
    ///
    /// Returns `Some(result)` if a result is ready, `None` otherwise.
    fn try_recv(&self) -> Option<T>;
}

/// An executor-agnostic task pool for offloading heavy data pipeline work.
///
/// The engine submits work items and polls result channels without depending
/// on a specific async runtime.  Implementations may use:
///
/// - Bevy's `AsyncComputeTaskPool`
/// - A `rayon` thread pool
/// - A `tokio` runtime
/// - A simple `std::thread`-based pool
///
/// # Contract
///
/// - Spawn methods must not block.  They should enqueue the closure
///   for execution on a background thread and return a receiver immediately.
/// - The returned receiver must be `Send + Sync` so it can be stored in
///   `MapState`.
/// - The closure must be `Send + 'static` because it will execute on a
///   different thread.
pub trait DataTaskPool: Send + Sync {
    /// Spawn a terrain mesh generation task.
    fn spawn_terrain(
        &self,
        task: Box<dyn FnOnce() -> TerrainTaskOutput + Send + 'static>,
    ) -> Box<dyn DataTaskResultReceiver<TerrainTaskOutput>>;

    /// Spawn a vector tessellation task.
    fn spawn_vector(
        &self,
        task: Box<dyn FnOnce() -> VectorTaskOutput + Send + 'static>,
    ) -> Box<dyn DataTaskResultReceiver<VectorTaskOutput>>;

    /// Spawn an MVT decode task.
    ///
    /// The closure decodes raw PBF bytes into a [`VectorTileData`] on a
    /// background thread, matching MapLibre's web worker model.
    fn spawn_decode(
        &self,
        task: Box<dyn FnOnce() -> MvtDecodeOutput + Send + 'static>,
    ) -> Box<dyn DataTaskResultReceiver<MvtDecodeOutput>>;
}

// ---------------------------------------------------------------------------
// Built-in std::thread task pool
// ---------------------------------------------------------------------------

/// A simple `std::thread`-based [`DataTaskPool`] implementation.
///
/// Each spawned task runs on a new OS thread.  This is suitable for
/// host applications that do not use Bevy or another async runtime.
///
/// For production use with many concurrent tasks, prefer a thread-pool
/// based implementation (e.g. `rayon` or Bevy's `ComputeTaskPool`).
pub struct ThreadDataTaskPool;

impl ThreadDataTaskPool {
    /// Create a new thread-based task pool.
    pub fn new() -> Self {
        Self
    }
}

impl Default for ThreadDataTaskPool {
    fn default() -> Self {
        Self::new()
    }
}

struct ChannelReceiver<T> {
    rx: std::sync::mpsc::Receiver<T>,
}

impl<T: Send + 'static> DataTaskResultReceiver<T> for ChannelReceiver<T> {
    fn try_recv(&self) -> Option<T> {
        self.rx.try_recv().ok()
    }
}

// Safety: `std::sync::mpsc::Receiver` is `Send` but not `Sync` by default.
// However, we only access it via `try_recv()` which is safe to call from
// any single thread (MapState is behind RwLock with exclusive write access).
// We wrap it to satisfy the Sync bound required by MapState's Send + Sync.
unsafe impl<T: Send> Sync for ChannelReceiver<T> {}

fn spawn_on_thread<T: Send + 'static>(
    task: Box<dyn FnOnce() -> T + Send + 'static>,
) -> Box<dyn DataTaskResultReceiver<T>> {
    let (tx, rx) = std::sync::mpsc::channel();
    std::thread::spawn(move || {
        let result = task();
        let _ = tx.send(result);
    });
    Box::new(ChannelReceiver { rx })
}

impl DataTaskPool for ThreadDataTaskPool {
    fn spawn_terrain(
        &self,
        task: Box<dyn FnOnce() -> TerrainTaskOutput + Send + 'static>,
    ) -> Box<dyn DataTaskResultReceiver<TerrainTaskOutput>> {
        spawn_on_thread(task)
    }

    fn spawn_vector(
        &self,
        task: Box<dyn FnOnce() -> VectorTaskOutput + Send + 'static>,
    ) -> Box<dyn DataTaskResultReceiver<VectorTaskOutput>> {
        spawn_on_thread(task)
    }

    fn spawn_decode(
        &self,
        task: Box<dyn FnOnce() -> MvtDecodeOutput + Send + 'static>,
    ) -> Box<dyn DataTaskResultReceiver<MvtDecodeOutput>> {
        spawn_on_thread(task)
    }
}

// ---------------------------------------------------------------------------
// Task inputs and outputs
// ---------------------------------------------------------------------------

/// Input for an async terrain mesh generation task.
#[derive(Clone)]
pub struct TerrainTaskInput {
    /// Tile to generate terrain for.
    pub tile: TileId,
    /// DEM source tile backing this visible terrain tile.
    pub elevation_source_tile: TileId,
    /// Normalized sub-region inside the DEM source tile.
    pub elevation_region: TileTextureRegion,
    /// Elevation grid data for this tile.
    pub elevation: ElevationGrid,
    /// Mesh grid resolution.
    pub resolution: u16,
    /// Vertical exaggeration factor.
    pub vertical_exaggeration: f64,
    /// Generation counter for cache invalidation.
    pub generation: u64,
}

/// Output from an async terrain mesh generation task.
pub struct TerrainTaskOutput {
    /// The tile this result is for.
    pub tile: TileId,
    /// Generated terrain mesh descriptor.
    pub mesh: TerrainMeshData,
    /// Prepared hillshade raster for this tile.
    pub hillshade: PreparedHillshadeRaster,
    /// Generation counter for cache invalidation.
    pub generation: u64,
}

/// Cache key for a retained terrain task result.
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
pub struct TerrainCacheKey {
    /// Tile id.
    pub tile: TileId,
    /// Elevation data generation.
    pub generation: u64,
    /// Grid resolution.
    pub resolution: u16,
}

/// Input for an async vector tessellation task.
#[derive(Clone)]
pub struct VectorTaskInput {
    /// Unique key for this (layer, data generation) pair.
    pub cache_key: VectorCacheKey,
    /// The feature collection to tessellate.
    pub features: FeatureCollection,
    /// The vector style to apply.
    pub style: VectorStyle,
    /// Originating style/runtime layer id, when known.
    pub query_layer_id: Option<String>,
    /// Originating style source id, when known.
    pub query_source_id: Option<String>,
    /// Originating style source-layer id, when known.
    pub query_source_layer: Option<String>,
    /// Per-feature tile/source-layer provenance.
    pub feature_provenance: Vec<Option<FeatureProvenance>>,
    /// Active camera projection.
    pub projection: CameraProjection,
    /// Per-vertex terrain elevation samples (coord -> elevation).
    /// Empty when terrain is disabled.
    pub terrain_samples: Vec<(GeoCoord, f64)>,
}

/// Output from an async vector tessellation task.
pub struct VectorTaskOutput {
    /// Cache key matching the input.
    pub cache_key: VectorCacheKey,
    /// Tessellated mesh data.
    pub mesh: VectorMeshData,
    /// Symbol candidates extracted during tessellation.
    pub symbol_candidates: Vec<SymbolCandidate>,
}

/// Output from an async MVT decode task.
///
/// Produced by [`DataTaskPool::spawn_decode`] on a background thread and
/// consumed by [`AsyncDataPipeline::poll_decodes`] to promote raw vector
/// payloads in the tile cache to fully decoded [`TileData::Vector`] data.
pub struct MvtDecodeOutput {
    /// The tile that was decoded.
    pub tile: TileId,
    /// Decode result -- either decoded vector data or an error.
    pub result: Result<crate::tile_source::VectorTileData, crate::tile_source::TileError>,
    /// Freshness metadata from the original HTTP response, carried through
    /// so that the promoted cache entry retains correct TTL information.
    pub freshness: crate::tile_source::TileFreshness,
}

/// Cache key for a retained vector task result.
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
pub struct VectorCacheKey {
    /// Layer identifier (name or style layer id).
    pub layer_id: String,
    /// Data generation counter.
    pub data_generation: u64,
    /// Projection used for tessellation.
    pub projection: CameraProjection,
}

// ---------------------------------------------------------------------------
// Per-tile vector bucket key and partitioning
// ---------------------------------------------------------------------------

/// Cache key for a per-tile vector bucket.
///
/// This is the fine-grained cache key that matches MapLibre's per-tile
/// bucket architecture.  Each `(layer, tile, generation, projection)` tuple
/// maps to an independently tessellated and cached mesh.
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
pub struct VectorBucketKey {
    /// Layer identifier.
    pub layer_id: String,
    /// Tile that this bucket covers.
    pub tile: TileId,
    /// Data generation counter for cache invalidation.
    pub data_generation: u64,
    /// Projection used for tessellation.
    pub projection: CameraProjection,
}

/// Return the representative coordinate for a feature's geometry.
///
/// Uses the first coordinate found in the geometry (point position,
/// first line vertex, first polygon vertex, etc.).  This is sufficient
/// for tile assignment: a feature is assigned to exactly one tile based
/// on its representative point, matching MapLibre's bucket-assignment
/// strategy.
fn representative_coord(geometry: &Geometry) -> Option<GeoCoord> {
    match geometry {
        Geometry::Point(p) => Some(p.coord),
        Geometry::LineString(ls) => ls.coords.first().copied(),
        Geometry::Polygon(poly) => poly.exterior.first().copied(),
        Geometry::MultiPoint(mp) => mp.points.first().map(|p| p.coord),
        Geometry::MultiLineString(mls) => {
            mls.lines.first().and_then(|ls| ls.coords.first().copied())
        }
        Geometry::MultiPolygon(mpoly) => {
            mpoly.polygons.first().and_then(|p| p.exterior.first().copied())
        }
        Geometry::GeometryCollection(geoms) => {
            geoms.iter().find_map(representative_coord)
        }
    }
}

/// Partition a feature collection into per-tile sub-collections.
///
/// Each feature is assigned to exactly one tile based on its representative
/// coordinate (first vertex).  Features whose geometry has no coordinates
/// are dropped.
///
/// # Arguments
///
/// * `features` - The source feature collection to partition.
/// * `zoom` - The zoom level at which to compute tile assignments.
///
/// # Returns
///
/// A map from `TileId` to the subset of features that fall within that tile.
pub fn partition_features_by_tile(
    features: &FeatureCollection,
    zoom: u8,
) -> HashMap<TileId, FeatureCollection> {
    let mut buckets: HashMap<TileId, Vec<Feature>> = HashMap::new();
    for feature in &features.features {
        if let Some(coord) = representative_coord(&feature.geometry) {
            let tile_coord = geo_to_tile(&coord, zoom);
            let tile_id = tile_coord.tile_id();
            buckets.entry(tile_id).or_default().push(feature.clone());
        }
    }
    buckets
        .into_iter()
        .map(|(tile, feats)| (tile, FeatureCollection { features: feats }))
        .collect()
}

// ---------------------------------------------------------------------------
// Pending task handles
// ---------------------------------------------------------------------------

struct PendingTerrainTask {
    key: TerrainCacheKey,
    receiver: Box<dyn DataTaskResultReceiver<TerrainTaskOutput>>,
}

struct PendingVectorTask {
    key: VectorCacheKey,
    receiver: Box<dyn DataTaskResultReceiver<VectorTaskOutput>>,
}

struct PendingVectorBucketTask {
    key: VectorBucketKey,
    receiver: Box<dyn DataTaskResultReceiver<VectorBucketOutput>>,
}

struct PendingDecodeTask {
    tile: TileId,
    receiver: Box<dyn DataTaskResultReceiver<MvtDecodeOutput>>,
}

/// Output from an async per-tile vector bucket tessellation task.
struct VectorBucketOutput {
    mesh: VectorMeshData,
    symbol_candidates: Vec<SymbolCandidate>,
}

// ---------------------------------------------------------------------------
// AsyncDataPipeline
// ---------------------------------------------------------------------------

/// Manages async data task dispatch and result polling.
///
/// This is the internal coordinator that [`MapState`](crate::MapState) uses
/// to submit terrain/vector/symbol work to a [`DataTaskPool`] and collect
/// completed results each frame.
pub(crate) struct AsyncDataPipeline {
    pool: Arc<dyn DataTaskPool>,
    pending_terrain: Vec<PendingTerrainTask>,
    pending_vector: Vec<PendingVectorTask>,
    terrain_cache: HashMap<TerrainCacheKey, (TerrainMeshData, PreparedHillshadeRaster)>,
    vector_cache: HashMap<VectorCacheKey, (VectorMeshData, Vec<SymbolCandidate>)>,
    next_vector_generation: u64,
    layer_generations: HashMap<String, u64>,
    // Per-tile vector bucket state
    pending_buckets: Vec<PendingVectorBucketTask>,
    bucket_cache: HashMap<VectorBucketKey, (VectorMeshData, Vec<SymbolCandidate>)>,
    // Background MVT decode state
    pending_decodes: Vec<PendingDecodeTask>,
    /// Completed decode results waiting to be promoted into the tile cache.
    ///
    /// Each entry is a `(TileId, TileResponse)` ready to replace the
    /// `TileData::RawVector` entry in the tile cache with a fully decoded
    /// `TileData::Vector`.
    pub(crate) decoded_tiles: Vec<(TileId, crate::tile_source::TileResponse)>,
}

impl AsyncDataPipeline {
    pub(crate) fn new(pool: Arc<dyn DataTaskPool>) -> Self {
        Self {
            pool,
            pending_terrain: Vec::new(),
            pending_vector: Vec::new(),
            terrain_cache: HashMap::new(),
            vector_cache: HashMap::new(),
            next_vector_generation: 1,
            layer_generations: HashMap::new(),
            pending_buckets: Vec::new(),
            bucket_cache: HashMap::new(),
            pending_decodes: Vec::new(),
            decoded_tiles: Vec::new(),
        }
    }

    // -- Terrain ----------------------------------------------------------

    pub(crate) fn dispatch_terrain(&mut self, input: TerrainTaskInput) {
        let key = TerrainCacheKey {
            tile: input.tile,
            generation: input.generation,
            resolution: input.resolution,
        };
        if self.terrain_cache.contains_key(&key) {
            return;
        }
        if self.pending_terrain.iter().any(|p| p.key == key) {
            return;
        }
        let receiver = self.pool.spawn_terrain(Box::new(move || {
            let mesh = crate::terrain::build_terrain_descriptor_with_source(
                &input.tile,
                input.elevation_source_tile,
                input.elevation_region,
                &input.elevation,
                input.resolution,
                input.vertical_exaggeration,
                input.generation,
            );
            let hillshade = crate::terrain::prepare_hillshade_raster(
                &input.elevation,
                input.vertical_exaggeration,
                input.generation,
            );
            TerrainTaskOutput {
                tile: input.tile,
                mesh,
                hillshade,
                generation: input.generation,
            }
        }));
        self.pending_terrain.push(PendingTerrainTask { key, receiver });
    }

    pub(crate) fn poll_terrain(&mut self) {
        let mut still_pending = Vec::new();
        for task in self.pending_terrain.drain(..) {
            if let Some(output) = task.receiver.try_recv() {
                self.terrain_cache
                    .insert(task.key, (output.mesh, output.hillshade));
            } else {
                still_pending.push(task);
            }
        }
        self.pending_terrain = still_pending;
    }

    #[allow(dead_code)]
    pub(crate) fn collect_terrain(
        &self,
        desired_tiles: &[TileId],
        tile_generations: &HashMap<TileId, u64>,
        resolution: u16,
    ) -> (Vec<TerrainMeshData>, Vec<PreparedHillshadeRaster>) {
        let mut meshes = Vec::with_capacity(desired_tiles.len());
        let mut hillshades = Vec::with_capacity(desired_tiles.len());
        for tile in desired_tiles {
            let generation = tile_generations.get(tile).copied().unwrap_or(0);
            let key = TerrainCacheKey {
                tile: *tile,
                generation,
                resolution,
            };
            if let Some((mesh, hs)) = self.terrain_cache.get(&key) {
                meshes.push(mesh.clone());
                hillshades.push(hs.clone());
            }
        }
        (meshes, hillshades)
    }

    /// Collect all currently cached terrain results regardless of tile key.
    pub(crate) fn collect_all_terrain(&self) -> (Vec<TerrainMeshData>, Vec<PreparedHillshadeRaster>) {
        let mut meshes = Vec::with_capacity(self.terrain_cache.len());
        let mut hillshades = Vec::with_capacity(self.terrain_cache.len());
        for (mesh, hs) in self.terrain_cache.values() {
            meshes.push(mesh.clone());
            hillshades.push(hs.clone());
        }
        (meshes, hillshades)
    }

    pub(crate) fn prune_terrain(&mut self, visible_tiles: &[TileId]) {
        let visible_set: std::collections::HashSet<TileId> =
            visible_tiles.iter().copied().collect();
        self.terrain_cache
            .retain(|key, _| visible_set.contains(&key.tile));
    }

    // -- Per-layer vector (legacy) ----------------------------------------

    pub(crate) fn layer_generation(&mut self, layer_id: &str, features_changed: bool) -> u64 {
        if features_changed {
            let gen = self.next_vector_generation;
            self.next_vector_generation += 1;
            self.layer_generations.insert(layer_id.to_owned(), gen);
            gen
        } else {
            self.layer_generations
                .get(layer_id)
                .copied()
                .unwrap_or_else(|| {
                    let gen = self.next_vector_generation;
                    self.next_vector_generation += 1;
                    self.layer_generations.insert(layer_id.to_owned(), gen);
                    gen
                })
        }
    }

    pub(crate) fn dispatch_vector(&mut self, input: VectorTaskInput) {
        let key = input.cache_key.clone();
        if self.vector_cache.contains_key(&key) {
            return;
        }
        if self.pending_vector.iter().any(|p| p.key == key) {
            return;
        }
        let receiver = self.pool.spawn_vector(Box::new(move || {
            let mut features = input.features;
            if !input.terrain_samples.is_empty() {
                apply_terrain_samples(&mut features, &input.terrain_samples);
            }
            let mut temp_layer = crate::layers::VectorLayer::new(
                &input.cache_key.layer_id,
                FeatureCollection::default(),
                input.style,
            )
            .with_query_metadata(input.query_layer_id, input.query_source_id)
            .with_source_layer(input.query_source_layer);
            temp_layer.set_features_with_provenance(features, input.feature_provenance);
            let mesh = temp_layer.tessellate(input.projection);
            let candidates = temp_layer.symbol_candidates();
            VectorTaskOutput {
                cache_key: input.cache_key,
                mesh,
                symbol_candidates: candidates,
            }
        }));
        self.pending_vector.push(PendingVectorTask { key, receiver });
    }

    pub(crate) fn poll_vector(&mut self) {
        let mut still_pending = Vec::new();
        for task in self.pending_vector.drain(..) {
            if let Some(output) = task.receiver.try_recv() {
                self.vector_cache
                    .insert(task.key, (output.mesh, output.symbol_candidates));
            } else {
                still_pending.push(task);
            }
        }
        self.pending_vector = still_pending;
    }

    pub(crate) fn collect_vectors(
        &self,
        layer_keys: &[VectorCacheKey],
    ) -> (Vec<VectorMeshData>, Vec<SymbolCandidate>) {
        let mut meshes = Vec::new();
        let mut candidates = Vec::new();
        for key in layer_keys {
            if let Some((mesh, cands)) = self.vector_cache.get(key) {
                if !mesh.is_empty() {
                    meshes.push(mesh.clone());
                }
                candidates.extend(cands.iter().cloned());
            }
        }
        (meshes, candidates)
    }

    pub(crate) fn prune_vectors(&mut self, active_layer_ids: &[String]) {
        let active_set: std::collections::HashSet<&str> =
            active_layer_ids.iter().map(String::as_str).collect();
        self.vector_cache
            .retain(|key, _| active_set.contains(key.layer_id.as_str()));
        self.layer_generations
            .retain(|id, _| active_set.contains(id.as_str()));
    }

    #[cfg(test)]
    pub(crate) fn current_layer_generation(&self, layer_id: &str) -> Option<u64> {
        self.layer_generations.get(layer_id).copied()
    }

    // -- Per-tile vector buckets ------------------------------------------

    /// Dispatch per-tile vector tessellation tasks for a partitioned feature set.
    ///
    /// For each `(tile, features)` pair, a background task is spawned that
    /// tessellates only the features belonging to that tile.  Results are
    /// cached by [`VectorBucketKey`] and deduplicated against both the
    /// pending queue and the completed cache.
    pub(crate) fn dispatch_vector_buckets(
        &mut self,
        layer_id: &str,
        data_generation: u64,
        projection: CameraProjection,
        style: &VectorStyle,
        tile_features: &HashMap<TileId, FeatureCollection>,
        query_source_id: Option<&str>,
        query_source_layer: Option<&str>,
        tile_provenance: &HashMap<TileId, Vec<Option<FeatureProvenance>>>,
        terrain_samples: &[(GeoCoord, f64)],
    ) {
        for (tile, features) in tile_features {
            let key = VectorBucketKey {
                layer_id: layer_id.to_owned(),
                tile: *tile,
                data_generation,
                projection,
            };
            if self.bucket_cache.contains_key(&key) {
                continue;
            }
            if self.pending_buckets.iter().any(|p| p.key == key) {
                continue;
            }
            let task_key = key.clone();
            let task_features = features.clone();
            let task_style = style.clone();
            let task_samples = terrain_samples.to_vec();
            let task_layer_id = layer_id.to_owned();
            let task_source_id = query_source_id.map(ToOwned::to_owned);
            let task_source_layer = query_source_layer.map(ToOwned::to_owned);
            let task_provenance = tile_provenance.get(tile).cloned().unwrap_or_default();

            let receiver = self.pool.spawn_vector(Box::new(move || {
                let mut feats = task_features;
                if !task_samples.is_empty() {
                    apply_terrain_samples(&mut feats, &task_samples);
                }
                let mut temp_layer = crate::layers::VectorLayer::new(
                    &task_layer_id,
                    FeatureCollection::default(),
                    task_style,
                )
                .with_query_metadata(Some(task_layer_id.clone()), task_source_id)
                .with_source_layer(task_source_layer);
                temp_layer.set_features_with_provenance(feats, task_provenance);
                let mesh = temp_layer.tessellate(task_key.projection);
                let candidates = temp_layer.symbol_candidates();
                VectorTaskOutput {
                    cache_key: VectorCacheKey {
                        layer_id: task_key.layer_id.clone(),
                        data_generation: task_key.data_generation,
                        projection: task_key.projection,
                    },
                    mesh,
                    symbol_candidates: candidates,
                }
            }));

            // Wrap the VectorTaskOutput receiver into a VectorBucketOutput receiver
            self.pending_buckets.push(PendingVectorBucketTask {
                key,
                receiver: Box::new(BucketReceiverAdapter {
                    inner: receiver,
                }),
            });
        }
    }

    /// Poll all pending per-tile vector bucket tasks.
    pub(crate) fn poll_vector_buckets(&mut self) {
        let mut still_pending = Vec::new();
        for task in self.pending_buckets.drain(..) {
            if let Some(output) = task.receiver.try_recv() {
                self.bucket_cache
                    .insert(task.key, (output.mesh, output.symbol_candidates));
            } else {
                still_pending.push(task);
            }
        }
        self.pending_buckets = still_pending;
    }

    // -- Background MVT decode --------------------------------------------

    /// Submit a raw vector tile payload for background decoding.
    ///
    /// The decode closure runs on the [`DataTaskPool`] and produces an
    /// [`MvtDecodeOutput`] that is collected by [`poll_decodes`](Self::poll_decodes).
    pub(crate) fn dispatch_decode(
        &mut self,
        tile: TileId,
        raw: &crate::tile_source::RawVectorPayload,
        freshness: crate::tile_source::TileFreshness,
    ) {
        // Do not double-dispatch if already pending.
        if self.pending_decodes.iter().any(|t| t.tile == tile) {
            return;
        }

        let bytes = Arc::clone(&raw.bytes);
        let opts = raw.decode_options.clone();
        let task_tile = tile;

        let receiver = self.pool.spawn_decode(Box::new(move || {
            let result = crate::mvt::decode_mvt(&bytes, &task_tile, &opts)
                .map(|layers| crate::tile_source::VectorTileData { layers })
                .map_err(|e| {
                    crate::tile_source::TileError::Decode(format!("MVT decode: {e}"))
                });
            MvtDecodeOutput {
                tile: task_tile,
                result,
                freshness,
            }
        }));

        self.pending_decodes.push(PendingDecodeTask { tile, receiver });
    }

    /// Poll all pending background MVT decode tasks.
    ///
    /// Successfully decoded tiles are collected into [`decoded_tiles`](Self::decoded_tiles)
    /// for the caller to promote into the tile cache.  Failed decodes are
    /// also collected so the tile cache can transition to the Failed state.
    pub(crate) fn poll_decodes(&mut self) {
        let mut still_pending = Vec::new();
        for task in self.pending_decodes.drain(..) {
            if let Some(output) = task.receiver.try_recv() {
                match output.result {
                    Ok(vector_data) => {
                        self.decoded_tiles.push((
                            output.tile,
                            crate::tile_source::TileResponse {
                                data: crate::tile_source::TileData::Vector(vector_data),
                                freshness: output.freshness,
                                not_modified: false,
                            },
                        ));
                    }
                    Err(_err) => {
                        // Failed decodes: the tile manager will handle the
                        // cache transition when it sees the tile is still
                        // in RawVector state on the next frame.
                    }
                }
            } else {
                still_pending.push(task);
            }
        }
        self.pending_decodes = still_pending;
    }

    /// Take all completed decode results, leaving the internal buffer empty.
    pub(crate) fn take_decoded_tiles(
        &mut self,
    ) -> Vec<(TileId, crate::tile_source::TileResponse)> {
        std::mem::take(&mut self.decoded_tiles)
    }

    /// Number of decode tasks currently in flight.
    #[allow(dead_code)]
    pub(crate) fn pending_decode_count(&self) -> usize {
        self.pending_decodes.len()
    }

    /// Collect tessellated vector data for the visible tile set.
    ///
    /// Only buckets whose tile is in `visible_tiles` are included in the
    /// output.  Buckets for off-screen tiles remain in the cache for reuse
    /// when the camera pans back.
    pub(crate) fn collect_vector_buckets(
        &self,
        layer_id: &str,
        data_generation: u64,
        projection: CameraProjection,
        visible_tiles: &[TileId],
    ) -> (Vec<VectorMeshData>, Vec<SymbolCandidate>) {
        let mut meshes = Vec::new();
        let mut candidates = Vec::new();
        for tile in visible_tiles {
            let key = VectorBucketKey {
                layer_id: layer_id.to_owned(),
                tile: *tile,
                data_generation,
                projection,
            };
            if let Some((mesh, cands)) = self.bucket_cache.get(&key) {
                if !mesh.is_empty() {
                    meshes.push(mesh.clone());
                }
                candidates.extend(cands.iter().cloned());
            }
        }
        (meshes, candidates)
    }

    /// Prune per-tile vector buckets for layers that are no longer active.
    ///
    /// Buckets for active layers whose tiles are off-screen are **retained**
    /// (they will be reused when the camera pans back).  Only buckets for
    /// entirely removed layers are evicted.
    #[allow(dead_code)]
    pub(crate) fn prune_vector_buckets(&mut self, active_layer_ids: &[String]) {
        let active_set: std::collections::HashSet<&str> =
            active_layer_ids.iter().map(String::as_str).collect();
        self.bucket_cache
            .retain(|key, _| active_set.contains(key.layer_id.as_str()));
    }

    pub(crate) fn evict_vector_buckets(&mut self, layer_id: &str, tiles: &[TileId]) {
        if tiles.is_empty() {
            return;
        }
        let tile_set: std::collections::HashSet<TileId> = tiles.iter().copied().collect();
        self.bucket_cache
            .retain(|key, _| key.layer_id != layer_id || !tile_set.contains(&key.tile));
        self.pending_buckets
            .retain(|task| task.key.layer_id != layer_id || !tile_set.contains(&task.key.tile));
    }

    /// Prune per-tile vector buckets for a specific layer to only retain
    /// buckets matching the given visible tile set.
    ///
    /// This is a more aggressive pruning strategy that evicts off-screen
    /// tile buckets for a layer, useful when memory pressure is high.
    #[allow(dead_code)]
    pub(crate) fn prune_vector_buckets_for_layer(
        &mut self,
        layer_id: &str,
        visible_tiles: &[TileId],
    ) {
        let visible_set: std::collections::HashSet<TileId> =
            visible_tiles.iter().copied().collect();
        self.bucket_cache.retain(|key, _| {
            key.layer_id != layer_id || visible_set.contains(&key.tile)
        });
    }

    /// Return the number of cached per-tile vector buckets.
    #[allow(dead_code)]
    pub(crate) fn bucket_cache_len(&self) -> usize {
        self.bucket_cache.len()
    }

    // -- General ----------------------------------------------------------

    #[allow(dead_code)]
    pub(crate) fn has_pending_tasks(&self) -> bool {
        !self.pending_terrain.is_empty()
            || !self.pending_vector.is_empty()
            || !self.pending_buckets.is_empty()
    }
}

/// Adapter that converts a `DataTaskResultReceiver<VectorTaskOutput>` into
/// a `DataTaskResultReceiver<VectorBucketOutput>` by attaching the bucket key.
struct BucketReceiverAdapter {
    inner: Box<dyn DataTaskResultReceiver<VectorTaskOutput>>,
}

impl DataTaskResultReceiver<VectorBucketOutput> for BucketReceiverAdapter {
    fn try_recv(&self) -> Option<VectorBucketOutput> {
        self.inner.try_recv().map(|output| VectorBucketOutput {
            mesh: output.mesh,
            symbol_candidates: output.symbol_candidates,
        })
    }
}

fn apply_terrain_samples(features: &mut FeatureCollection, samples: &[(GeoCoord, f64)]) {
    let lookup: HashMap<(i64, i64), f64> = samples
        .iter()
        .map(|(coord, elev)| {
            let key = (
                (coord.lat * 100_000.0).round() as i64,
                (coord.lon * 100_000.0).round() as i64,
            );
            (key, *elev)
        })
        .collect();
    for feature in &mut features.features {
        apply_terrain_to_geometry(&mut feature.geometry, &lookup);
    }
}

fn apply_terrain_to_geometry(
    geometry: &mut crate::geometry::Geometry,
    lookup: &HashMap<(i64, i64), f64>,
) {
    use crate::geometry::Geometry;
    match geometry {
        Geometry::Point(p) => {
            if let Some(&elev) = lookup.get(&coord_key(&p.coord)) {
                p.coord.alt = elev;
            }
        }
        Geometry::LineString(ls) => {
            for coord in &mut ls.coords {
                if let Some(&elev) = lookup.get(&coord_key(coord)) {
                    coord.alt = elev;
                }
            }
        }
        Geometry::Polygon(poly) => {
            for coord in &mut poly.exterior {
                if let Some(&elev) = lookup.get(&coord_key(coord)) {
                    coord.alt = elev;
                }
            }
            for hole in &mut poly.interiors {
                for coord in hole {
                    if let Some(&elev) = lookup.get(&coord_key(coord)) {
                        coord.alt = elev;
                    }
                }
            }
        }
        Geometry::MultiPoint(mp) => {
            for p in &mut mp.points {
                if let Some(&elev) = lookup.get(&coord_key(&p.coord)) {
                    p.coord.alt = elev;
                }
            }
        }
        Geometry::MultiLineString(mls) => {
            for ls in &mut mls.lines {
                for coord in &mut ls.coords {
                    if let Some(&elev) = lookup.get(&coord_key(coord)) {
                        coord.alt = elev;
                    }
                }
            }
        }
        Geometry::MultiPolygon(mpoly) => {
            for poly in &mut mpoly.polygons {
                for coord in &mut poly.exterior {
                    if let Some(&elev) = lookup.get(&coord_key(coord)) {
                        coord.alt = elev;
                    }
                }
                for hole in &mut poly.interiors {
                    for coord in hole {
                        if let Some(&elev) = lookup.get(&coord_key(coord)) {
                            coord.alt = elev;
                        }
                    }
                }
            }
        }
        Geometry::GeometryCollection(geoms) => {
            for g in geoms {
                apply_terrain_to_geometry(g, lookup);
            }
        }
    }
}

fn coord_key(coord: &GeoCoord) -> (i64, i64) {
    (
        (coord.lat * 100_000.0).round() as i64,
        (coord.lon * 100_000.0).round() as i64,
    )
}

// ---------------------------------------------------------------------------
// Async visualization preprocessing pipeline
// ---------------------------------------------------------------------------

/// Output of an asynchronous visualization preprocessing task.
///
/// Contains the preprocessed data ready to be applied to a layer via
/// the corresponding `MapState::set_*` method.
pub enum VisualizationTaskOutput {
    /// Preprocessed scalar field values (for grid scalar / grid extrusion).
    ScalarFieldUpdate {
        /// Layer name to update.
        layer_name: String,
        /// New scalar values.
        values: Vec<f32>,
    },
    /// Preprocessed column instances (for instanced column layers).
    ColumnUpdate {
        /// Layer name to update.
        layer_name: String,
        /// New column instance set.
        columns: crate::visualization::ColumnInstanceSet,
        /// Colour ramp (included in case the ramp changes with the data).
        ramp: crate::visualization::ColorRamp,
    },
    /// Preprocessed point instances (for point cloud layers).
    PointCloudUpdate {
        /// Layer name to update.
        layer_name: String,
        /// New point instance set.
        points: crate::visualization::PointInstanceSet,
        /// Colour ramp.
        ramp: crate::visualization::ColorRamp,
    },
}

/// Asynchronous preprocessing pipeline for large visualization data updates.
///
/// Use this when visualization data preparation (binning, normalisation,
/// aggregation, coordinate transforms) is too heavy for the main thread's
/// per-frame budget.  The pipeline offloads work to the same
/// [`DataTaskPool`] used for terrain and vector tessellation.
///
/// # Usage
///
/// ```rust,no_run
/// # use rustial_engine::{AsyncVisualizationPipeline, ThreadDataTaskPool, VisualizationTaskOutput};
/// # use std::sync::Arc;
/// let pool: Arc<dyn rustial_engine::DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
/// let mut viz_pipeline = AsyncVisualizationPipeline::new(pool);
///
/// // Dispatch a heavy computation on a background thread.
/// viz_pipeline.dispatch(Box::new(|| {
///     // Expensive data processing...
///     let values: Vec<f32> = (0..10_000).map(|i| (i as f32).sin()).collect();
///     VisualizationTaskOutput::ScalarFieldUpdate {
///         layer_name: "heatmap".to_string(),
///         values,
///     }
/// }));
///
/// // Each frame, poll for completed results.
/// for result in viz_pipeline.poll() {
///     match result {
///         VisualizationTaskOutput::ScalarFieldUpdate { layer_name, values } => {
///             // Apply: state.update_scalar_field(&layer_name, values);
///         }
///         _ => {}
///     }
/// }
/// ```
pub struct AsyncVisualizationPipeline {
    _pool: Arc<dyn DataTaskPool>,
    pending: Vec<Box<dyn DataTaskResultReceiver<VisualizationTaskOutput>>>,
}

impl AsyncVisualizationPipeline {
    /// Create a new async visualization pipeline backed by the given task pool.
    pub fn new(pool: Arc<dyn DataTaskPool>) -> Self {
        Self {
            _pool: pool,
            pending: Vec::new(),
        }
    }

    /// Dispatch a visualization preprocessing task to the background thread pool.
    ///
    /// The closure runs on a background thread and should return a
    /// [`VisualizationTaskOutput`] with the preprocessed data.
    pub fn dispatch(
        &mut self,
        task: Box<dyn FnOnce() -> VisualizationTaskOutput + Send + 'static>,
    ) {
        let receiver = spawn_on_thread(task);
        self.pending.push(receiver);
    }

    /// Poll all pending tasks and return completed results.
    ///
    /// Call this once per frame from the main thread.  Incomplete tasks
    /// remain in the pending queue for the next poll.
    pub fn poll(&mut self) -> Vec<VisualizationTaskOutput> {
        let mut completed = Vec::new();
        let mut still_pending = Vec::new();

        for receiver in self.pending.drain(..) {
            match receiver.try_recv() {
                Some(result) => completed.push(result),
                None => still_pending.push(receiver),
            }
        }

        self.pending = still_pending;
        completed
    }

    /// Return the number of tasks currently in-flight.
    pub fn pending_count(&self) -> usize {
        self.pending.len()
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::camera_projection::CameraProjection;
    use crate::geometry::{Feature, Geometry, Point, LineString, Polygon};
    use crate::layers::VectorStyle;
    use rustial_math::{ElevationGrid, GeoCoord, TileId};
    use std::collections::HashMap;

    #[test]
    fn thread_pool_dispatches_and_receives_terrain() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile = TileId::new(0, 0, 0);
        let elev = ElevationGrid::flat(tile, 4, 4);
        pipeline.dispatch_terrain(TerrainTaskInput {
            tile,
            elevation_source_tile: tile,
            elevation_region: TileTextureRegion::FULL,
            elevation: elev,
            resolution: 4,
            vertical_exaggeration: 1.0,
            generation: 1,
        });

        for _ in 0..100 {
            pipeline.poll_terrain();
            if !pipeline.has_pending_tasks() {
                break;
            }
            std::thread::sleep(std::time::Duration::from_millis(10));
        }

        let mut gens = HashMap::new();
        gens.insert(tile, 1u64);
        let (meshes, hillshades) = pipeline.collect_terrain(&[tile], &gens, 4);
        assert_eq!(meshes.len(), 1);
        assert_eq!(hillshades.len(), 1);
        assert_eq!(meshes[0].tile, tile);
    }

    #[test]
    fn thread_pool_dispatches_and_receives_vector() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let features = crate::geometry::FeatureCollection {
            features: vec![Feature {
                geometry: Geometry::Point(Point {
                    coord: GeoCoord::from_lat_lon(0.0, 0.0),
                }),
                properties: HashMap::new(),
            }],
        };

        let key = VectorCacheKey {
            layer_id: "test".into(),
            data_generation: 1,
            projection: CameraProjection::WebMercator,
        };

        pipeline.dispatch_vector(VectorTaskInput {
            cache_key: key.clone(),
            features,
            style: VectorStyle::default(),
            query_layer_id: Some("test".into()),
            query_source_id: None,
            query_source_layer: None,
            feature_provenance: Vec::new(),
            projection: CameraProjection::WebMercator,
            terrain_samples: Vec::new(),
        });

        for _ in 0..100 {
            pipeline.poll_vector();
            if !pipeline.has_pending_tasks() {
                break;
            }
            std::thread::sleep(std::time::Duration::from_millis(10));
        }

        let (meshes, _candidates) = pipeline.collect_vectors(&[key]);
        assert_eq!(meshes.len(), 1);
        assert!(meshes[0].vertex_count() > 0);
    }

    #[test]
    fn duplicate_dispatch_is_deduplicated() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile = TileId::new(0, 0, 0);
        let elev = ElevationGrid::flat(tile, 4, 4);
        let input = TerrainTaskInput {
            tile,
            elevation_source_tile: tile,
            elevation_region: TileTextureRegion::FULL,
            elevation: elev,
            resolution: 4,
            vertical_exaggeration: 1.0,
            generation: 1,
        };

        pipeline.dispatch_terrain(input.clone());
        pipeline.dispatch_terrain(input.clone());
        assert_eq!(pipeline.pending_terrain.len(), 1);
    }

    #[test]
    fn cached_result_prevents_redispatch() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile = TileId::new(0, 0, 0);
        let elev = ElevationGrid::flat(tile, 4, 4);
        let input = TerrainTaskInput {
            tile,
            elevation_source_tile: tile,
            elevation_region: TileTextureRegion::FULL,
            elevation: elev,
            resolution: 4,
            vertical_exaggeration: 1.0,
            generation: 1,
        };

        pipeline.dispatch_terrain(input.clone());

        for _ in 0..100 {
            pipeline.poll_terrain();
            if !pipeline.has_pending_tasks() {
                break;
            }
            std::thread::sleep(std::time::Duration::from_millis(10));
        }

        pipeline.dispatch_terrain(input);
        assert!(pipeline.pending_terrain.is_empty());
    }

    #[test]
    fn prune_terrain_removes_stale_entries() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile_a = TileId::new(1, 0, 0);
        let tile_b = TileId::new(1, 1, 0);
        let elev_a = ElevationGrid::flat(tile_a, 4, 4);
        let elev_b = ElevationGrid::flat(tile_b, 4, 4);
        let mesh_a = crate::terrain::build_terrain_descriptor(&tile_a, &elev_a, 4, 1.0, 1);
        let mesh_b = crate::terrain::build_terrain_descriptor(&tile_b, &elev_b, 4, 1.0, 1);
        let hs_a = crate::terrain::prepare_hillshade_raster(&elev_a, 1.0, 1);
        let hs_b = crate::terrain::prepare_hillshade_raster(&elev_b, 1.0, 1);

        pipeline.terrain_cache.insert(
            TerrainCacheKey { tile: tile_a, generation: 1, resolution: 4 },
            (mesh_a, hs_a),
        );
        pipeline.terrain_cache.insert(
            TerrainCacheKey { tile: tile_b, generation: 1, resolution: 4 },
            (mesh_b, hs_b),
        );

        assert_eq!(pipeline.terrain_cache.len(), 2);
        pipeline.prune_terrain(&[tile_a]);
        assert_eq!(pipeline.terrain_cache.len(), 1);
        assert!(pipeline.terrain_cache.keys().all(|k| k.tile == tile_a));
    }

    #[test]
    fn synchronous_fallback_path_still_works() {
        let mut state = crate::MapState::new();
        state.set_viewport(800, 600);
        state.set_camera_target(GeoCoord::from_lat_lon(0.0, 0.0));
        state.set_camera_distance(1_000.0);
        state.update();
        assert!(state.zoom_level() <= 22);
    }

    #[test]
    fn layer_generation_tracking() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let gen1 = pipeline.layer_generation("layer-a", false);
        let gen2 = pipeline.layer_generation("layer-a", false);
        assert_eq!(gen1, gen2);

        let gen3 = pipeline.layer_generation("layer-a", true);
        assert!(gen3 > gen2);

        let gen4 = pipeline.layer_generation("layer-a", false);
        assert_eq!(gen3, gen4);
    }

    #[test]
    fn collect_all_terrain_returns_cached_entries() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile = TileId::new(0, 0, 0);
        let elev = ElevationGrid::flat(tile, 4, 4);
        pipeline.dispatch_terrain(TerrainTaskInput {
            tile,
            elevation_source_tile: tile,
            elevation_region: TileTextureRegion::FULL,
            elevation: elev,
            resolution: 4,
            vertical_exaggeration: 1.0,
            generation: 1,
        });

        for _ in 0..100 {
            pipeline.poll_terrain();
            if !pipeline.has_pending_tasks() {
                break;
            }
            std::thread::sleep(std::time::Duration::from_millis(10));
        }

        let (meshes, hillshades) = pipeline.collect_all_terrain();
        assert_eq!(meshes.len(), 1);
        assert_eq!(hillshades.len(), 1);
        assert_eq!(meshes[0].tile, tile);
    }

    #[test]
    fn async_pipeline_terrain_via_map_state() {
        use crate::terrain::{FlatElevationSource, TerrainConfig};

        let config = TerrainConfig {
            enabled: true,
            mesh_resolution: 4,
            source: Box::new(FlatElevationSource::new(4, 4)),
            ..TerrainConfig::default()
        };
        let mut state = crate::MapState::with_terrain(config, 100);
        state.set_viewport(800, 600);
        state.set_camera_target(GeoCoord::from_lat_lon(0.0, 0.0));
        state.set_camera_distance(10_000_000.0);

        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        state.set_task_pool(pool);
        assert!(state.has_async_pipeline());

        // First update: dispatches terrain tasks.
        state.update_with_dt(1.0 / 60.0);

        // Allow background tasks to complete.
        for _ in 0..100 {
            std::thread::sleep(std::time::Duration::from_millis(10));
            state.update_with_dt(1.0 / 60.0);
            if !state.terrain_meshes().is_empty() {
                break;
            }
        }

        assert!(
            !state.terrain_meshes().is_empty(),
            "async pipeline should produce terrain meshes"
        );
    }

    #[test]
    fn async_pipeline_vectors_via_map_state() {
        use crate::layers::{VectorLayer, VectorStyle};

        let mut state = crate::MapState::new();
        state.set_viewport(800, 600);
        state.set_camera_target(GeoCoord::from_lat_lon(0.0, 0.0));
        state.set_camera_distance(1_000.0);

        let fc = crate::geometry::FeatureCollection {
            features: vec![Feature {
                geometry: Geometry::Point(Point {
                    coord: GeoCoord::from_lat_lon(0.0, 0.0),
                }),
                properties: HashMap::new(),
            }],
        };
        let vl = VectorLayer::new("test-vec", fc, VectorStyle::default());
        state.push_layer(Box::new(vl));

        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        state.set_task_pool(pool);

        // First update: dispatches vector tasks.
        state.update_with_dt(1.0 / 60.0);

        // Allow background tasks to complete.
        for _ in 0..100 {
            std::thread::sleep(std::time::Duration::from_millis(10));
            state.update_with_dt(1.0 / 60.0);
            if !state.vector_meshes().is_empty() {
                break;
            }
        }

        assert!(
            !state.vector_meshes().is_empty(),
            "async pipeline should produce vector meshes"
        );
    }

    #[test]
    fn async_pipeline_clear_reverts_to_sync() {
        let mut state = crate::MapState::new();
        state.set_viewport(800, 600);
        state.set_camera_target(GeoCoord::from_lat_lon(0.0, 0.0));
        state.set_camera_distance(1_000.0);

        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        state.set_task_pool(pool);
        assert!(state.has_async_pipeline());

        state.clear_task_pool();
        assert!(!state.has_async_pipeline());

        // Synchronous update should still work.
        state.update();
        assert!(state.zoom_level() <= 22);
    }

    #[test]
    fn data_update_interval_ignored_with_async_pipeline() {
        use crate::terrain::{FlatElevationSource, TerrainConfig};

        let config = TerrainConfig {
            enabled: true,
            mesh_resolution: 4,
            source: Box::new(FlatElevationSource::new(4, 4)),
            ..TerrainConfig::default()
        };
        let mut state = crate::MapState::with_terrain(config, 100);
        state.set_viewport(800, 600);
        state.set_camera_target(GeoCoord::from_lat_lon(0.0, 0.0));
        state.set_camera_distance(10_000_000.0);
        state.set_data_update_interval(100.0); // Very high throttle

        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        state.set_task_pool(pool);

        // Even with a very high throttle, async dispatch should still run
        // because the throttle is ignored when an async pipeline is active.
        for _ in 0..100 {
            state.update_with_dt(1.0 / 60.0);
            std::thread::sleep(std::time::Duration::from_millis(10));
            if !state.terrain_meshes().is_empty() {
                break;
            }
        }

        assert!(
            !state.terrain_meshes().is_empty(),
            "async pipeline should produce terrain despite high throttle"
        );
    }

    // =====================================================================
    // Per-tile vector bucket tests
    // =====================================================================

    #[test]
    fn partition_features_by_tile_separates_spatially() {
        let features = FeatureCollection {
            features: vec![
                Feature {
                    geometry: Geometry::Point(Point {
                        coord: GeoCoord::from_lat_lon(10.0, 10.0),
                    }),
                    properties: HashMap::new(),
                },
                Feature {
                    geometry: Geometry::Point(Point {
                        coord: GeoCoord::from_lat_lon(-30.0, -60.0),
                    }),
                    properties: HashMap::new(),
                },
                Feature {
                    geometry: Geometry::Point(Point {
                        coord: GeoCoord::from_lat_lon(10.01, 10.01),
                    }),
                    properties: HashMap::new(),
                },
            ],
        };

        let buckets = partition_features_by_tile(&features, 4);
        let total: usize = buckets.values().map(|fc| fc.len()).sum();
        assert_eq!(total, 3, "all features should be partitioned");
        // The first and third points are very close and should be in the same tile at zoom 4
        let tile_a = geo_to_tile(&GeoCoord::from_lat_lon(10.0, 10.0), 4).tile_id();
        let tile_c = geo_to_tile(&GeoCoord::from_lat_lon(10.01, 10.01), 4).tile_id();
        assert_eq!(tile_a, tile_c);
        assert_eq!(buckets.get(&tile_a).map(|fc| fc.len()), Some(2));
    }

    #[test]
    fn partition_features_handles_linestring_and_polygon() {
        let features = FeatureCollection {
            features: vec![
                Feature {
                    geometry: Geometry::LineString(LineString {
                        coords: vec![
                            GeoCoord::from_lat_lon(45.0, 90.0),
                            GeoCoord::from_lat_lon(46.0, 91.0),
                        ],
                    }),
                    properties: HashMap::new(),
                },
                Feature {
                    geometry: Geometry::Polygon(Polygon {
                        exterior: vec![
                            GeoCoord::from_lat_lon(-10.0, -20.0),
                            GeoCoord::from_lat_lon(-10.0, -19.0),
                            GeoCoord::from_lat_lon(-9.0, -19.0),
                        ],
                        interiors: vec![],
                    }),
                    properties: HashMap::new(),
                },
            ],
        };

        let buckets = partition_features_by_tile(&features, 6);
        let total: usize = buckets.values().map(|fc| fc.len()).sum();
        assert_eq!(total, 2);
    }

    #[test]
    fn partition_features_drops_empty_geometry() {
        let features = FeatureCollection {
            features: vec![Feature {
                geometry: Geometry::GeometryCollection(vec![]),
                properties: HashMap::new(),
            }],
        };

        let buckets = partition_features_by_tile(&features, 4);
        assert!(buckets.is_empty(), "empty geometry should not produce a bucket");
    }

    #[test]
    fn dispatch_and_poll_vector_buckets() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile_a = geo_to_tile(&GeoCoord::from_lat_lon(10.0, 10.0), 4).tile_id();
        let tile_b = geo_to_tile(&GeoCoord::from_lat_lon(-30.0, -60.0), 4).tile_id();

        let features = FeatureCollection {
            features: vec![
                Feature {
                    geometry: Geometry::Point(Point {
                        coord: GeoCoord::from_lat_lon(10.0, 10.0),
                    }),
                    properties: HashMap::new(),
                },
                Feature {
                    geometry: Geometry::Point(Point {
                        coord: GeoCoord::from_lat_lon(-30.0, -60.0),
                    }),
                    properties: HashMap::new(),
                },
            ],
        };

        let tile_features = partition_features_by_tile(&features, 4);
        assert!(tile_features.len() >= 2);

        pipeline.dispatch_vector_buckets(
            "test-layer",
            1,
            CameraProjection::WebMercator,
            &VectorStyle::default(),
            &tile_features,
            None,
            None,
            &HashMap::new(),
            &[],
        );

        for _ in 0..100 {
            pipeline.poll_vector_buckets();
            if !pipeline.has_pending_tasks() {
                break;
            }
            std::thread::sleep(std::time::Duration::from_millis(10));
        }

        // Collect only tile_a
        let (meshes_a, _) = pipeline.collect_vector_buckets(
            "test-layer", 1, CameraProjection::WebMercator, &[tile_a],
        );
        assert_eq!(meshes_a.len(), 1, "should have one mesh for tile_a");

        // Collect only tile_b
        let (meshes_b, _) = pipeline.collect_vector_buckets(
            "test-layer", 1, CameraProjection::WebMercator, &[tile_b],
        );
        assert_eq!(meshes_b.len(), 1, "should have one mesh for tile_b");

        // Collect both
        let (meshes_both, _) = pipeline.collect_vector_buckets(
            "test-layer", 1, CameraProjection::WebMercator, &[tile_a, tile_b],
        );
        assert_eq!(meshes_both.len(), 2, "should have meshes for both tiles");
    }

    #[test]
    fn bucket_dispatch_deduplicates() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let features = FeatureCollection {
            features: vec![Feature {
                geometry: Geometry::Point(Point {
                    coord: GeoCoord::from_lat_lon(10.0, 10.0),
                }),
                properties: HashMap::new(),
            }],
        };
        let tile_features = partition_features_by_tile(&features, 4);

        pipeline.dispatch_vector_buckets(
            "layer", 1, CameraProjection::WebMercator,
            &VectorStyle::default(), &tile_features, None, None, &HashMap::new(), &[],
        );
        let first_count = pipeline.pending_buckets.len();

        pipeline.dispatch_vector_buckets(
            "layer", 1, CameraProjection::WebMercator,
            &VectorStyle::default(), &tile_features, None, None, &HashMap::new(), &[],
        );
        assert_eq!(pipeline.pending_buckets.len(), first_count, "duplicate dispatch should be deduped");
    }

    #[test]
    fn bucket_cache_hit_prevents_redispatch() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let features = FeatureCollection {
            features: vec![Feature {
                geometry: Geometry::Point(Point {
                    coord: GeoCoord::from_lat_lon(10.0, 10.0),
                }),
                properties: HashMap::new(),
            }],
        };
        let tile_features = partition_features_by_tile(&features, 4);

        pipeline.dispatch_vector_buckets(
            "layer", 1, CameraProjection::WebMercator,
            &VectorStyle::default(), &tile_features, None, None, &HashMap::new(), &[],
        );

        for _ in 0..100 {
            pipeline.poll_vector_buckets();
            if !pipeline.has_pending_tasks() {
                break;
            }
            std::thread::sleep(std::time::Duration::from_millis(10));
        }

        assert!(pipeline.bucket_cache_len() > 0);

        pipeline.dispatch_vector_buckets(
            "layer", 1, CameraProjection::WebMercator,
            &VectorStyle::default(), &tile_features, None, None, &HashMap::new(), &[],
        );
        assert!(pipeline.pending_buckets.is_empty(), "cached result should prevent redispatch");
    }

    #[test]
    fn prune_vector_buckets_removes_inactive_layers() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile = TileId::new(4, 0, 0);
        let key_a = VectorBucketKey {
            layer_id: "keep".into(),
            tile,
            data_generation: 1,
            projection: CameraProjection::WebMercator,
        };
        let key_b = VectorBucketKey {
            layer_id: "remove".into(),
            tile,
            data_generation: 1,
            projection: CameraProjection::WebMercator,
        };
        pipeline.bucket_cache.insert(key_a, (VectorMeshData::default(), vec![]));
        pipeline.bucket_cache.insert(key_b, (VectorMeshData::default(), vec![]));
        assert_eq!(pipeline.bucket_cache_len(), 2);

        pipeline.prune_vector_buckets(&["keep".to_string()]);
        assert_eq!(pipeline.bucket_cache_len(), 1);
        assert!(pipeline.bucket_cache.keys().all(|k| k.layer_id == "keep"));
    }

    #[test]
    fn prune_vector_buckets_for_layer_removes_offscreen_tiles() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile_a = TileId::new(4, 0, 0);
        let tile_b = TileId::new(4, 1, 0);
        let tile_c = TileId::new(4, 2, 0);

        for tile in &[tile_a, tile_b, tile_c] {
            pipeline.bucket_cache.insert(
                VectorBucketKey {
                    layer_id: "layer".into(),
                    tile: *tile,
                    data_generation: 1,
                    projection: CameraProjection::WebMercator,
                },
                (VectorMeshData::default(), vec![]),
            );
        }
        assert_eq!(pipeline.bucket_cache_len(), 3);

        pipeline.prune_vector_buckets_for_layer("layer", &[tile_a, tile_c]);
        assert_eq!(pipeline.bucket_cache_len(), 2);
        assert!(pipeline.bucket_cache.keys().all(|k| k.tile == tile_a || k.tile == tile_c));
    }

    #[test]
    fn collect_vector_buckets_only_returns_visible() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile_visible = TileId::new(4, 5, 5);
        let tile_offscreen = TileId::new(4, 10, 10);

        let mesh = VectorMeshData {
            positions: vec![[0.0, 0.0, 0.0], [1.0, 0.0, 0.0], [0.0, 1.0, 0.0]],
            colors: vec![[1.0; 4]; 3],
            indices: vec![0, 1, 2],
            ..Default::default()
        };
        for tile in &[tile_visible, tile_offscreen] {
            pipeline.bucket_cache.insert(
                VectorBucketKey {
                    layer_id: "layer".into(),
                    tile: *tile,
                    data_generation: 1,
                    projection: CameraProjection::WebMercator,
                },
                (mesh.clone(), vec![]),
            );
        }

        let (meshes, _) = pipeline.collect_vector_buckets(
            "layer", 1, CameraProjection::WebMercator, &[tile_visible],
        );
        assert_eq!(meshes.len(), 1, "only visible tile bucket should be collected");
    }

    #[test]
    fn representative_coord_all_geometry_types() {
        let p = Geometry::Point(Point { coord: GeoCoord::from_lat_lon(1.0, 2.0) });
        assert_eq!(representative_coord(&p).map(|c| (c.lat, c.lon)), Some((1.0, 2.0)));

        let ls = Geometry::LineString(LineString {
            coords: vec![GeoCoord::from_lat_lon(3.0, 4.0), GeoCoord::from_lat_lon(5.0, 6.0)],
        });
        assert_eq!(representative_coord(&ls).map(|c| (c.lat, c.lon)), Some((3.0, 4.0)));

        let poly = Geometry::Polygon(Polygon {
            exterior: vec![
                GeoCoord::from_lat_lon(7.0, 8.0),
                GeoCoord::from_lat_lon(7.0, 9.0),
                GeoCoord::from_lat_lon(8.0, 9.0),
            ],
            interiors: vec![],
        });
        assert_eq!(representative_coord(&poly).map(|c| (c.lat, c.lon)), Some((7.0, 8.0)));

        let gc = Geometry::GeometryCollection(vec![]);
        assert!(representative_coord(&gc).is_none());

        let gc_with = Geometry::GeometryCollection(vec![p.clone()]);
        assert!(representative_coord(&gc_with).is_some());
    }

    #[test]
    fn dispatch_and_poll_decode_produces_decoded_tile() {
        use crate::tile_source::{RawVectorPayload, TileFreshness};
        use std::sync::Arc;

        // Build a minimal valid MVT tile.
        fn build_test_mvt() -> Vec<u8> {
            fn encode_varint(mut val: u64) -> Vec<u8> {
                let mut buf = Vec::new();
                loop {
                    let mut byte = (val & 0x7F) as u8;
                    val >>= 7;
                    if val != 0 { byte |= 0x80; }
                    buf.push(byte);
                    if val == 0 { break; }
                }
                buf
            }
            fn encode_tag(field: u32, wt: u8) -> Vec<u8> {
                encode_varint(((field as u64) << 3) | wt as u64)
            }
            fn encode_ld(field: u32, data: &[u8]) -> Vec<u8> {
                let mut b = encode_tag(field, 2);
                b.extend(encode_varint(data.len() as u64));
                b.extend_from_slice(data);
                b
            }
            fn encode_vi(field: u32, val: u64) -> Vec<u8> {
                let mut b = encode_tag(field, 0);
                b.extend(encode_varint(val));
                b
            }
            fn zigzag(n: i32) -> u32 { ((n << 1) ^ (n >> 31)) as u32 }

            let mut geom = Vec::new();
            geom.extend(encode_varint(((1u64) << 3) | 1));
            geom.extend(encode_varint(zigzag(2048) as u64));
            geom.extend(encode_varint(zigzag(2048) as u64));

            let mut feat = Vec::new();
            feat.extend(encode_vi(3, 1));
            feat.extend(encode_ld(4, &geom));

            let mut layer = Vec::new();
            layer.extend(encode_ld(1, b"test"));
            layer.extend(encode_ld(2, &feat));
            layer.extend(encode_vi(5, 4096));
            layer.extend(encode_vi(15, 2));

            encode_ld(3, &layer)
        }

        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile = TileId::new(0, 0, 0);
        let raw = RawVectorPayload {
            tile_id: tile,
            bytes: Arc::new(build_test_mvt()),
            decode_options: crate::mvt::MvtDecodeOptions::default(),
        };

        pipeline.dispatch_decode(tile, &raw, TileFreshness::default());

        // Poll until the decode completes (should be near-instant).
        for _ in 0..50 {
            pipeline.poll_decodes();
            if !pipeline.decoded_tiles.is_empty() {
                break;
            }
            std::thread::sleep(std::time::Duration::from_millis(10));
        }

        let decoded = pipeline.take_decoded_tiles();
        assert_eq!(decoded.len(), 1, "should have one decoded tile");
        let (decoded_id, response) = &decoded[0];
        assert_eq!(*decoded_id, tile);
        assert!(response.data.is_vector(), "decoded data should be Vector");
        let vt = response.data.as_vector().expect("should be Vector");
        assert!(vt.layers.contains_key("test"));
        assert_eq!(vt.layers["test"].len(), 1);
    }

    #[test]
    fn dispatch_decode_deduplicates() {
        use crate::tile_source::{RawVectorPayload, TileFreshness};
        use std::sync::Arc;

        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut pipeline = AsyncDataPipeline::new(pool);

        let tile = TileId::new(0, 0, 0);
        let raw = RawVectorPayload {
            tile_id: tile,
            bytes: Arc::new(vec![]),
            decode_options: crate::mvt::MvtDecodeOptions::default(),
        };

        pipeline.dispatch_decode(tile, &raw, TileFreshness::default());
        pipeline.dispatch_decode(tile, &raw, TileFreshness::default());

        // Only one task should be pending despite two dispatches.
        assert_eq!(pipeline.pending_decodes.len(), 1);
    }

    // -----------------------------------------------------------------------
    // AsyncVisualizationPipeline tests
    // -----------------------------------------------------------------------

    #[test]
    fn async_viz_pipeline_dispatches_and_receives_scalar_update() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut viz = AsyncVisualizationPipeline::new(pool);

        viz.dispatch(Box::new(|| {
            VisualizationTaskOutput::ScalarFieldUpdate {
                layer_name: "density".to_string(),
                values: vec![1.0, 2.0, 3.0, 4.0],
            }
        }));

        assert_eq!(viz.pending_count(), 1);

        // Busy-poll until complete.
        let mut results = Vec::new();
        for _ in 0..200 {
            results = viz.poll();
            if !results.is_empty() {
                break;
            }
            std::thread::sleep(std::time::Duration::from_millis(5));
        }

        assert_eq!(results.len(), 1);
        match &results[0] {
            VisualizationTaskOutput::ScalarFieldUpdate { layer_name, values } => {
                assert_eq!(layer_name, "density");
                assert_eq!(values, &[1.0, 2.0, 3.0, 4.0]);
            }
            _ => panic!("expected ScalarFieldUpdate"),
        }
        assert_eq!(viz.pending_count(), 0);
    }

    #[test]
    fn async_viz_pipeline_dispatches_and_receives_column_update() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut viz = AsyncVisualizationPipeline::new(pool);

        viz.dispatch(Box::new(|| {
            VisualizationTaskOutput::ColumnUpdate {
                layer_name: "bars".to_string(),
                columns: crate::visualization::ColumnInstanceSet::new(vec![
                    crate::visualization::ColumnInstance::new(
                        GeoCoord::from_lat_lon(0.0, 0.0),
                        100.0,
                        10.0,
                    ),
                ]),
                ramp: crate::visualization::ColorRamp::new(vec![
                    crate::visualization::ColorStop { value: 0.0, color: [0.0, 0.0, 1.0, 1.0] },
                    crate::visualization::ColorStop { value: 1.0, color: [1.0, 0.0, 0.0, 1.0] },
                ]),
            }
        }));

        let mut results = Vec::new();
        for _ in 0..200 {
            results = viz.poll();
            if !results.is_empty() { break; }
            std::thread::sleep(std::time::Duration::from_millis(5));
        }

        assert_eq!(results.len(), 1);
        match &results[0] {
            VisualizationTaskOutput::ColumnUpdate { layer_name, columns, .. } => {
                assert_eq!(layer_name, "bars");
                assert_eq!(columns.columns.len(), 1);
            }
            _ => panic!("expected ColumnUpdate"),
        }
    }

    #[test]
    fn async_viz_pipeline_dispatches_and_receives_point_cloud_update() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut viz = AsyncVisualizationPipeline::new(pool);

        viz.dispatch(Box::new(|| {
            VisualizationTaskOutput::PointCloudUpdate {
                layer_name: "scatter".to_string(),
                points: crate::visualization::PointInstanceSet::new(vec![
                    crate::visualization::PointInstance::new(
                        GeoCoord::from_lat_lon(1.0, 2.0),
                        5.0,
                    ),
                    crate::visualization::PointInstance::new(
                        GeoCoord::from_lat_lon(3.0, 4.0),
                        8.0,
                    ),
                ]),
                ramp: crate::visualization::ColorRamp::new(vec![
                    crate::visualization::ColorStop { value: 0.0, color: [0.0, 1.0, 0.0, 1.0] },
                    crate::visualization::ColorStop { value: 1.0, color: [1.0, 1.0, 0.0, 1.0] },
                ]),
            }
        }));

        let mut results = Vec::new();
        for _ in 0..200 {
            results = viz.poll();
            if !results.is_empty() { break; }
            std::thread::sleep(std::time::Duration::from_millis(5));
        }

        assert_eq!(results.len(), 1);
        match &results[0] {
            VisualizationTaskOutput::PointCloudUpdate { layer_name, points, .. } => {
                assert_eq!(layer_name, "scatter");
                assert_eq!(points.points.len(), 2);
            }
            _ => panic!("expected PointCloudUpdate"),
        }
    }

    #[test]
    fn async_viz_pipeline_multiple_tasks_complete_independently() {
        let pool: Arc<dyn DataTaskPool> = Arc::new(ThreadDataTaskPool::new());
        let mut viz = AsyncVisualizationPipeline::new(pool);

        for i in 0..5 {
            let name = format!("layer_{i}");
            viz.dispatch(Box::new(move || {
                VisualizationTaskOutput::ScalarFieldUpdate {
                    layer_name: name,
                    values: vec![i as f32; 4],
                }
            }));
        }

        assert_eq!(viz.pending_count(), 5);

        let mut all_results = Vec::new();
        for _ in 0..200 {
            let batch = viz.poll();
            all_results.extend(batch);
            if all_results.len() >= 5 { break; }
            std::thread::sleep(std::time::Duration::from_millis(10));
        }

        assert_eq!(all_results.len(), 5);
        assert_eq!(viz.pending_count(), 0);
    }
}