car_inference/

lib.rs

//! # car-inference
//!
//! Local model inference for the Common Agent Runtime.
//!
//! Provides on-device inference using Candle with automatic hardware detection:
//! - **macOS**: Metal (Apple Silicon GPU)
//! - **Linux**: CUDA (NVIDIA GPU) or CPU fallback
//!
//! Ships with Qwen3 models downloaded on first use from HuggingFace.
//! Supports remote API models (OpenAI, Anthropic, Google) via the same schema.
//!
//! ## Architecture
//!
//! Models are first-class typed resources described by `ModelSchema` (analogous
//! to `ToolSchema`). The `UnifiedRegistry` holds local and remote models.
//! The `AdaptiveRouter` selects the best model using a three-phase strategy:
//! filter → score → explore. The `OutcomeTracker` learns from results to
//! improve routing over time.
//!
//! ## Dual purpose
//!
//! 1. **Internal** — powers skill learning/repair, semantic memory, policy evaluation
//! 2. **Service** — exposes `infer`, `embed`, `classify` as built-in CAR tools

pub mod adaptive_router;
pub mod backend;
pub mod backend_cache;
pub mod hardware;
pub mod intent;
pub mod key_pool;
pub mod models;
pub mod outcome;
pub mod protocol;
pub mod registry;
pub mod remote;
pub mod router;
pub mod routing_ext;
pub mod schema;
pub mod service;
pub mod stream;
pub mod tasks;
pub mod vllm_mlx;

use std::path::{Path, PathBuf};
use std::sync::Arc;
use std::time::Instant;
use std::time::{SystemTime, UNIX_EPOCH};

use serde::Serialize;
use reqwest::multipart::{Form, Part};
use thiserror::Error;
#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
use tokio::process::Command;
#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
use tokio::sync::Mutex;
use tokio::sync::RwLock;
use tracing::{debug, instrument};

// --- New types ---
pub use adaptive_router::{
    AdaptiveRouter, AdaptiveRoutingDecision, RoutingConfig, RoutingStrategy,
};
pub use intent::{IntentHint, TaskHint};
pub use key_pool::{KeyPool, KeyStats};
pub use outcome::{
    CodeOutcome, InferenceOutcome, InferenceTask, InferredOutcome, ModelProfile, OutcomeTracker,
};
pub use registry::{ModelFilter, ModelInfo, ModelRuntimeRequirement, ModelUpgrade, UnifiedRegistry};
pub use remote::RemoteBackend;
pub use routing_ext::{
    CircuitBreaker, CircuitBreakerRegistry, CircuitState, ImplicitSignal, ImplicitSignalType,
    RoutingMode, SpendControl, SpendLimitExceeded, SpendLimits, SpendStatus,
};
pub use schema::{
    ApiProtocol, BenchmarkScore, CostModel, ModelCapability, ModelSchema, ModelSource,
    PerformanceEnvelope, ProprietaryAuth,
};

// --- Legacy re-exports (kept for backward compatibility) ---
pub use adaptive_router::TaskComplexity;
#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
pub use backend::CandleBackend;
#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
pub use backend::EmbeddingBackend;
#[cfg(all(target_os = "macos", target_arch = "aarch64"))]
pub use backend::MlxBackend;
pub use hardware::HardwareInfo;
pub use models::{ModelRegistry, ModelRole};
pub use router::{ModelRouter, RoutingDecision};
pub use stream::{StreamAccumulator, StreamEvent};
pub use tasks::{
    parse_boxes, BoundingBox, ClassifyRequest, ClassifyResult, ContentBlock, EmbedRequest,
    GenerateImageRequest, GenerateImageResult, GenerateParams, GenerateRequest,
    GenerateVideoRequest, GenerateVideoResult, GroundRequest, GroundResult, Message, RerankRequest,
    RerankResult, RerankedDocument, RoutingWorkload, SynthesizeRequest, SynthesizeResult, ToolCall,
    TranscribeRequest, TranscribeResult, VideoMode,
};
// TokenUsage is defined in this module and already public

#[derive(Error, Debug)]
pub enum InferenceError {
    #[error("model not found: {0}")]
    ModelNotFound(String),

    #[error("model download failed: {0}")]
    DownloadFailed(String),

    #[error("inference failed: {0}")]
    InferenceFailed(String),

    /// A request mode is accepted on the public surface but the
    /// selected backend hasn't wired it yet. Distinct from
    /// `InferenceFailed` so callers can distinguish "backend can't"
    /// from "backend tried and something went wrong".
    #[error("mode {mode} not implemented on backend {backend}: {reason}")]
    UnsupportedMode {
        mode: &'static str,
        backend: &'static str,
        reason: &'static str,
    },

    #[error("tokenization error: {0}")]
    TokenizationError(String),

    #[error("device error: {0}")]
    DeviceError(String),

    #[error("io error: {0}")]
    Io(#[from] std::io::Error),
}

/// Which device to run inference on.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Device {
    Cpu,
    Metal,
    Cuda(usize), // device ordinal
}

impl Device {
    /// Auto-detect the best available device for this platform.
    pub fn auto() -> Self {
        #[cfg(feature = "metal")]
        {
            return Device::Metal;
        }
        #[cfg(feature = "cuda")]
        {
            return Device::Cuda(0);
        }
        #[cfg(not(any(feature = "metal", feature = "cuda")))]
        {
            Device::Cpu
        }
    }
}

/// Configuration for the inference engine.
#[derive(Debug, Clone)]
pub struct InferenceConfig {
    /// Where to store downloaded models. Defaults to ~/.car/models/
    pub models_dir: std::path::PathBuf,
    /// Device override. None = auto-detect.
    pub device: Option<Device>,
    /// Default model for generation tasks.
    pub generation_model: String,
    /// Optional preferred model override for generation tasks.
    pub preferred_generation_model: Option<String>,
    /// Default model for embedding tasks.
    pub embedding_model: String,
    /// Optional preferred model override for embedding tasks.
    pub preferred_embedding_model: Option<String>,
    /// Default model for classification tasks.
    pub classification_model: String,
    /// Optional preferred model override for classification tasks.
    pub preferred_classification_model: Option<String>,
}

impl Default for InferenceConfig {
    fn default() -> Self {
        let models_dir = dirs_next()
            .unwrap_or_else(|| std::path::PathBuf::from("."))
            .join(".car")
            .join("models");

        let hw = HardwareInfo::detect();

        Self {
            models_dir,
            device: None,
            generation_model: hw.recommended_model,
            preferred_generation_model: None,
            embedding_model: "Qwen3-Embedding-0.6B".to_string(),
            preferred_embedding_model: None,
            classification_model: "Qwen3-0.6B".to_string(),
            preferred_classification_model: None,
        }
    }
}

fn dirs_next() -> Option<std::path::PathBuf> {
    std::env::var("HOME").ok().map(std::path::PathBuf::from)
}

/// Token usage statistics from a model response.
#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
pub struct TokenUsage {
    /// Number of tokens in the prompt/input.
    pub prompt_tokens: u64,
    /// Number of tokens in the completion/output.
    pub completion_tokens: u64,
    /// Total tokens (prompt + completion).
    pub total_tokens: u64,
    /// Model's maximum context window size.
    pub context_window: u64,
}

/// Result of an inference call, including trace ID for outcome tracking.
#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
pub struct InferenceResult {
    /// The generated text (empty if tool_calls are present).
    pub text: String,
    /// Tool calls returned by the model (when tools were provided in the request).
    pub tool_calls: Vec<crate::tasks::generate::ToolCall>,
    /// Structured bounding boxes when the model emitted Qwen2.5-VL
    /// grounding spans (`<|box_*|>`, `<|object_ref_*|>`) in its text.
    /// Parsed from the same `text` field — the raw span markers remain
    /// visible in `text` for callers that need to see them verbatim.
    /// Empty vec when the model didn't ground anything (typical for
    /// non-VL models or prompts that only ask for description).
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub bounding_boxes: Vec<crate::tasks::grounding::BoundingBox>,
    /// Trace ID for reporting outcomes back to the tracker.
    pub trace_id: String,
    /// Which model was used.
    pub model_used: String,
    /// Wall-clock latency in ms.
    pub latency_ms: u64,
    /// Time to first token in milliseconds. Populated by the local
    /// generate paths (Candle/MLX) which observe the prefill→first-decode
    /// transition directly. `None` for paths that can't measure it
    /// honestly without streaming — currently the non-streaming remote
    /// paths. Callers needing TTFT on remote models should use
    /// [`InferenceEngine::generate_tracked_stream`] and time the first
    /// `text` event arrival themselves.
    ///
    /// Always serialized (as `null` when `None`) so downstream
    /// validation harnesses can distinguish "wasn't measured" from
    /// "field doesn't exist on this client's protocol version".
    #[serde(default)]
    pub time_to_first_token_ms: Option<u64>,
    /// Token usage statistics from the API response (None for local models).
    pub usage: Option<TokenUsage>,
}

impl InferenceResult {
    /// Returns true if the model chose to call tools instead of generating text.
    pub fn has_tool_calls(&self) -> bool {
        !self.tool_calls.is_empty()
    }
}

#[derive(Debug, Clone, Serialize)]
pub struct SpeechRuntimeHealth {
    pub root: PathBuf,
    pub installed: bool,
    pub python: PathBuf,
    pub stt_command: PathBuf,
    pub tts_command: PathBuf,
    pub configured_python: Option<String>,
    pub detected_python: Option<String>,
}

#[derive(Debug, Clone, Serialize)]
pub struct SpeechModelHealth {
    pub id: String,
    pub name: String,
    pub provider: String,
    pub capability: ModelCapability,
    pub is_local: bool,
    pub available: bool,
    pub cached: bool,
    pub selected_by_default: bool,
    pub source: String,
}

#[derive(Debug, Clone, Serialize)]
pub struct SpeechHealthReport {
    pub runtime: SpeechRuntimeHealth,
    pub local_models: Vec<SpeechModelHealth>,
    pub remote_models: Vec<SpeechModelHealth>,
    pub elevenlabs_configured: bool,
    pub prefer_local: bool,
    pub allow_remote_fallback: bool,
    pub preferred_local_stt: Option<String>,
    pub preferred_local_tts: Option<String>,
    pub preferred_remote_stt: Option<String>,
    pub preferred_remote_tts: Option<String>,
    pub local_stt_default: Option<String>,
    pub local_tts_default: Option<String>,
    pub remote_stt_default: Option<String>,
    pub remote_tts_default: Option<String>,
}

#[derive(Debug, Clone, Serialize)]
pub struct ModelDefaultHealth {
    pub capability: ModelCapability,
    pub configured_model: String,
    pub available: bool,
    pub is_local: bool,
    pub provider: Option<String>,
}

#[derive(Debug, Clone, Serialize)]
pub struct ModelProviderHealth {
    pub provider: String,
    pub configured: bool,
    pub local_models: usize,
    pub remote_models: usize,
    pub available_models: usize,
    pub capabilities: Vec<ModelCapability>,
}

#[derive(Debug, Clone, Serialize)]
pub struct ModelCapabilityHealth {
    pub capability: ModelCapability,
    pub total_models: usize,
    pub available_models: usize,
    pub local_available_models: usize,
    pub remote_available_models: usize,
}

#[derive(Debug, Clone, Serialize)]
pub struct RoutingScenarioHealth {
    pub name: String,
    pub workload: RoutingWorkload,
    pub task_family: String,
    pub has_tools: bool,
    pub has_vision: bool,
    pub prefer_local: bool,
    pub quality_first_cold_start: bool,
    pub bootstrap_min_task_observations: u64,
    pub bootstrap_quality_floor: f64,
    pub model_id: String,
    pub model_name: String,
    pub reason: String,
    pub strategy: RoutingStrategy,
}

#[derive(Debug, Clone, Serialize)]
pub struct ModelBenchmarkPriorHealth {
    pub model_id: String,
    pub model_name: Option<String>,
    pub overall_score: f64,
    pub overall_latency_ms: Option<f64>,
    pub task_scores: std::collections::HashMap<String, f64>,
    pub task_latency_ms: std::collections::HashMap<String, f64>,
    pub source_path: PathBuf,
}

#[derive(Debug, Clone, Serialize)]
pub struct ModelHealthReport {
    pub total_models: usize,
    pub available_models: usize,
    pub local_models: usize,
    pub remote_models: usize,
    pub defaults: Vec<ModelDefaultHealth>,
    pub providers: Vec<ModelProviderHealth>,
    pub capabilities: Vec<ModelCapabilityHealth>,
    pub routing_prefer_local: bool,
    pub routing_quality_first_cold_start: bool,
    pub routing_min_observations: u64,
    pub routing_bootstrap_min_task_observations: u64,
    pub routing_bootstrap_quality_floor: f64,
    pub routing_quality_weight: f64,
    pub routing_latency_weight: f64,
    pub routing_cost_weight: f64,
    pub routing_scenarios: Vec<RoutingScenarioHealth>,
    pub benchmark_priors: Vec<ModelBenchmarkPriorHealth>,
    pub speech: SpeechHealthReport,
}

#[derive(Debug, Clone, Serialize)]
pub struct SpeechInstallReport {
    pub name: String,
    pub hf_repo: String,
    pub snapshot_path: PathBuf,
    pub files_downloaded: usize,
}

#[derive(Debug, Clone, Serialize)]
pub struct SpeechSmokePathReport {
    pub path: String,
    pub tts_model: String,
    pub stt_model: String,
    pub audio_path: PathBuf,
    pub transcript: String,
}

#[derive(Debug, Clone, Serialize, Default)]
pub struct SpeechSmokeReport {
    pub local: Option<SpeechSmokePathReport>,
    pub remote: Option<SpeechSmokePathReport>,
    pub skipped: Vec<String>,
}

#[derive(Debug, Clone, Serialize, Default)]
pub struct SpeechPolicy {
    pub prefer_local: bool,
    pub allow_remote_fallback: bool,
    pub preferred_local_stt: Option<String>,
    pub preferred_local_tts: Option<String>,
    pub preferred_remote_stt: Option<String>,
    pub preferred_remote_tts: Option<String>,
}

/// The main inference engine. Thread-safe, lazily loads models.
///
/// Now includes the unified registry, adaptive router, and outcome tracker
/// for schema-driven model selection with learned performance profiles.
pub struct InferenceEngine {
    pub config: InferenceConfig,
    /// Unified model registry (local + remote).
    pub unified_registry: UnifiedRegistry,
    /// Adaptive router with three-phase selection.
    pub adaptive_router: AdaptiveRouter,
    /// Outcome tracker for learning from results.
    pub outcome_tracker: Arc<RwLock<OutcomeTracker>>,
    /// HTTP client for remote API models.
    remote_backend: RemoteBackend,
    /// Native MLX text-gen / embedding backends keyed by model id.
    /// Same cache shape as `flux_cache` / `ltx_cache` / `kokoro_cache`:
    /// per-entry `Arc<Mutex<MlxBackend>>` so concurrent calls for the
    /// same model serialize, while calls for different models proceed
    /// in parallel. Also bounded by `CAR_INFERENCE_MODEL_CACHE_MB`.
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    mlx_backends: Arc<backend_cache::BackendCache<backend::MlxBackend>>,
    /// LRU-evicting, mutex-serialized cache of loaded Flux image backends.
    /// Avoids reloading the 4–5 GB model on every generate_image call,
    /// and serializes concurrent calls onto the same backend (MLX ops
    /// are not `Sync`).
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    flux_cache: Arc<backend_cache::BackendCache<backend::mlx_flux::FluxBackend>>,
    /// Same for LTX video (~9 GB: transformer + Gemma 3 12B + VAE + vocoder).
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    ltx_cache: Arc<backend_cache::BackendCache<backend::mlx_ltx::LtxBackend>>,
    /// Same for Kokoro TTS (~160 MB). Small but reloading per-utterance
    /// added ~1 s of latency to every `synth` call.
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    kokoro_cache: Arc<backend_cache::BackendCache<backend::mlx_kokoro::KokoroBackend>>,
    // Legacy fields kept for backward compatibility
    pub registry: models::ModelRegistry,
    pub router: ModelRouter,
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    backend: Arc<RwLock<Option<CandleBackend>>>,
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    embedding_backend: Arc<RwLock<Option<EmbeddingBackend>>>,
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    speech_runtime: Arc<Mutex<Option<SpeechRuntime>>>,
    speech_policy: SpeechPolicy,
}

impl InferenceEngine {
    fn preferred_model_for_capability(&self, capability: ModelCapability) -> Option<&str> {
        match capability {
            ModelCapability::Generate => self.config.preferred_generation_model.as_deref(),
            ModelCapability::Embed => self.config.preferred_embedding_model.as_deref(),
            ModelCapability::Classify => self.config.preferred_classification_model.as_deref(),
            _ => None,
        }
    }

    fn request_needs_vision(req: &GenerateRequest) -> bool {
        req.images.as_ref().is_some_and(|images| !images.is_empty())
            || req.messages.as_ref().is_some_and(|messages| {
                messages
                    .iter()
                    .any(|msg| matches!(msg, Message::UserMultimodal { .. }))
            })
    }

    /// True when any content block in the request carries video data.
    /// Backends without a video-tokenization path use this to reject
    /// the request up front with [`InferenceError::UnsupportedMode`]
    /// rather than silently dropping the content.
    fn request_has_video(req: &GenerateRequest) -> bool {
        let images_have_video = req
            .images
            .as_ref()
            .is_some_and(|blocks| blocks.iter().any(ContentBlock::is_video));
        let messages_have_video = req.messages.as_ref().is_some_and(|messages| {
            messages.iter().any(|msg| match msg {
                Message::UserMultimodal { content } => {
                    content.iter().any(ContentBlock::is_video)
                }
                _ => false,
            })
        });
        images_have_video || messages_have_video
    }

    /// True when any content block in the request carries audio data.
    /// Same role as [`request_has_video`] but for the audio path
    /// (Gemma 4 small variants, Gemini).
    fn request_has_audio(req: &GenerateRequest) -> bool {
        let images_have_audio = req
            .images
            .as_ref()
            .is_some_and(|blocks| blocks.iter().any(ContentBlock::is_audio));
        let messages_have_audio = req.messages.as_ref().is_some_and(|messages| {
            messages.iter().any(|msg| match msg {
                Message::UserMultimodal { content } => {
                    content.iter().any(ContentBlock::is_audio)
                }
                _ => false,
            })
        });
        images_have_audio || messages_have_audio
    }

    pub fn new(config: InferenceConfig) -> Self {
        let registry = models::ModelRegistry::new(config.models_dir.clone());
        let hw = HardwareInfo::detect();
        let router = ModelRouter::new(hw.clone());
        let unified_registry = UnifiedRegistry::new(config.models_dir.clone());
        let adaptive_router = AdaptiveRouter::with_default_config(hw);
        let mut tracker = OutcomeTracker::new();
        // Load persisted profiles from previous sessions (#13)
        let profiles_path = config.models_dir.join("outcome_profiles.json");
        if let Ok(n) = tracker.load_from_file(&profiles_path) {
            if n > 0 {
                tracing::info!(loaded = n, "loaded persisted model profiles");
            }
        }
        let mut benchmark_models_loaded = 0usize;
        for path in benchmark_priors_paths(&config.models_dir) {
            match routing_ext::load_benchmark_priors(&path) {
                Ok(priors) if !priors.is_empty() => {
                    benchmark_models_loaded += priors.len();
                    routing_ext::apply_benchmark_priors(&mut tracker, &priors);
                    tracing::info!(
                        path = %path.display(),
                        loaded = priors.len(),
                        "loaded benchmark quality priors"
                    );
                }
                Ok(_) => {}
                Err(error) => {
                    tracing::warn!(path = %path.display(), %error, "failed to load benchmark priors");
                }
            }
        }
        if benchmark_models_loaded > 0 {
            tracing::info!(
                loaded = benchmark_models_loaded,
                "applied benchmark priors to cold-start routing"
            );
        }
        let outcome_tracker = Arc::new(RwLock::new(tracker));

        let remote_backend = RemoteBackend::new();

        Self {
            config,
            unified_registry,
            adaptive_router,
            outcome_tracker,
            remote_backend,
            #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
            mlx_backends: Arc::new(backend_cache::BackendCache::from_env()),
            #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
            flux_cache: Arc::new(backend_cache::BackendCache::from_env()),
            #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
            ltx_cache: Arc::new(backend_cache::BackendCache::from_env()),
            #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
            kokoro_cache: Arc::new(backend_cache::BackendCache::from_env()),
            registry,
            router,
            #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
            backend: Arc::new(RwLock::new(None)),
            #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
            embedding_backend: Arc::new(RwLock::new(None)),
            #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
            speech_runtime: Arc::new(Mutex::new(None)),
            speech_policy: SpeechPolicy {
                prefer_local: cfg!(all(target_os = "macos", target_arch = "aarch64")),
                allow_remote_fallback: true,
                preferred_local_stt: None,
                preferred_local_tts: None,
                preferred_remote_stt: None,
                preferred_remote_tts: None,
            },
        }
    }

    /// Initialize key pool: register keys from all remote models and load persisted stats.
    /// Call this after construction (requires async).
    pub async fn init_key_pool(&self) {
        // Register keys from all remote models in the catalog
        for schema in self.unified_registry.list() {
            if schema.is_remote() {
                self.remote_backend.register_model_keys(schema).await;
            }
        }

        // Load persisted key stats
        let stats_path = self.config.models_dir.join("key_pool_stats.json");
        if let Ok(n) = self.remote_backend.key_pool.load_stats(&stats_path).await {
            if n > 0 {
                tracing::info!(loaded = n, "loaded persisted key pool stats");
            }
        }

        let total = self.remote_backend.key_pool.total_keys().await;
        if total > 0 {
            tracing::info!(keys = total, "key pool initialized");
        }
    }

    /// Get or initialize the generative Candle backend, loading the specified model.
    /// Not used on Apple Silicon where all local inference goes through MLX.
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    async fn ensure_backend(&self, model_name: &str) -> Result<(), InferenceError> {
        let read = self.backend.read().await;
        if read.is_some() {
            return Ok(());
        }
        drop(read);

        let mut write = self.backend.write().await;
        if write.is_some() {
            return Ok(());
        }

        let model_path = self.registry.ensure_model(model_name).await?;
        let device = self.config.device.unwrap_or_else(Device::auto);
        let backend = CandleBackend::load(&model_path, device)?;
        *write = Some(backend);
        Ok(())
    }

    /// Get or initialize the embedding backend.
    /// On Apple Silicon, uses the MLX backend instead of Candle.
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    async fn ensure_embedding_backend(&self) -> Result<(), InferenceError> {
        let read = self.embedding_backend.read().await;
        if read.is_some() {
            return Ok(());
        }
        drop(read);

        let mut write = self.embedding_backend.write().await;
        if write.is_some() {
            return Ok(());
        }

        let embedding_model = self
            .preferred_model_for_capability(ModelCapability::Embed)
            .unwrap_or(&self.config.embedding_model);
        let model_path = self
            .registry
            .ensure_model(embedding_model)
            .await?;
        let device = self.config.device.unwrap_or_else(Device::auto);
        let backend = EmbeddingBackend::load(&model_path, device)?;
        *write = Some(backend);
        Ok(())
    }

    /// On Apple Silicon, ensure the MLX embedding model is loaded.
    /// Returns the schema ID of the embedding model for keying into mlx_backends.
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    async fn ensure_mlx_embedding_backend(&self) -> Result<String, InferenceError> {
        let embedding_model_name = self
            .preferred_model_for_capability(ModelCapability::Embed)
            .unwrap_or(&self.config.embedding_model)
            .to_string();
        let schema = self
            .unified_registry
            .get(&embedding_model_name)
            .or_else(|| self.unified_registry.find_by_name(&embedding_model_name))
            .ok_or_else(|| InferenceError::ModelNotFound(embedding_model_name.clone()))?
            .clone();
        self.ensure_mlx_backend(&schema).await?;
        Ok(schema.id)
    }

    /// Get or initialize the native MLX backend for a specific model.
    /// Returns a shared mutex handle — the caller locks it for the
    /// duration of an inference call so concurrent requests serialize.
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    async fn ensure_mlx_backend(
        &self,
        schema: &ModelSchema,
    ) -> Result<backend_cache::CachedBackend<backend::MlxBackend>, InferenceError> {
        if !Self::supports_native_mlx(schema) {
            return Err(InferenceError::InferenceFailed(format!(
                "native MLX backend does not support {} ({}) yet; use vLLM-MLX or add a family-specific MLX backend",
                schema.name, schema.family
            )));
        }

        let model_dir = self.unified_registry.ensure_local(&schema.id).await?;
        let size = backend_cache::estimate_model_size(&model_dir);
        let cache = Arc::clone(&self.mlx_backends);
        let key = schema.id.clone();
        // Loader runs inside `get_or_load` only on a cache miss. Wrap it
        // in `catch_unwind` because MLX/accelerate occasionally panics
        // at the FFI boundary and we don't want the whole engine to die.
        cache.get_or_load(&key, size, move || {
            std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
                backend::MlxBackend::load(&model_dir)
            }))
            .map_err(|e| {
                InferenceError::InferenceFailed(format!(
                    "MLX backend loading panicked (possible Metal/accelerate exception): {:?}",
                    e
                ))
            })?
        })
    }

    /// Pre-load a set of models into the MLX cache so the first real
    /// inference call doesn't pay the 1–14 s model-load latency. Safe
    /// to call multiple times; already-loaded models are no-ops.
    ///
    /// Handles both text-gen MLX backends (`mlx_backends`) and the
    /// image/video/tts caches. Pass the full `schema.id` values.
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    pub async fn warm_up<S: AsRef<str>>(&self, schema_ids: &[S]) -> Vec<Result<(), InferenceError>> {
        let mut results = Vec::with_capacity(schema_ids.len());
        for id in schema_ids {
            let id = id.as_ref();
            let outcome: Result<(), InferenceError> = async {
                let schema = self
                    .unified_registry
                    .get(id)
                    .cloned()
                    .ok_or_else(|| InferenceError::InferenceFailed(
                        format!("warm_up: unknown schema id {id}"),
                    ))?;
                match schema.capabilities.first().copied() {
                    Some(ModelCapability::ImageGeneration) => {
                        let model_dir =
                            self.unified_registry.ensure_local(&schema.id).await?;
                        let size = backend_cache::estimate_model_size(&model_dir);
                        let _ = self.flux_cache.get_or_load(&schema.id, size, || {
                            backend::mlx_flux::FluxBackend::load(&model_dir)
                        })?;
                    }
                    Some(ModelCapability::VideoGeneration) => {
                        let model_dir =
                            self.unified_registry.ensure_local(&schema.id).await?;
                        let size = backend_cache::estimate_model_size(&model_dir);
                        let _ = self.ltx_cache.get_or_load(&schema.id, size, || {
                            backend::mlx_ltx::LtxBackend::load(&model_dir)
                        })?;
                    }
                    Some(ModelCapability::TextToSpeech) => {
                        let model_dir =
                            self.unified_registry.ensure_local(&schema.id).await?;
                        let size = backend_cache::estimate_model_size(&model_dir);
                        let _ = self.kokoro_cache.get_or_load(&schema.id, size, || {
                            backend::mlx_kokoro::KokoroBackend::load(&model_dir)
                        })?;
                    }
                    _ => {
                        let _ = self.ensure_mlx_backend(&schema).await?;
                    }
                }
                Ok(())
            }
            .await;
            results.push(outcome);
        }
        results
    }

    /// No-op on non-macOS — MLX doesn't run here.
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    pub async fn warm_up<S: AsRef<str>>(&self, _schema_ids: &[S]) -> Vec<Result<(), InferenceError>> {
        Vec::new()
    }

    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    fn supports_native_mlx(schema: &ModelSchema) -> bool {
        matches!(schema.family.as_str(), "qwen3" | "qwen2.5-vl" | "qwen2-vl")
    }

    /// Route a prompt using the adaptive router (new). Returns full decision context.
    pub async fn route_adaptive(&self, prompt: &str) -> AdaptiveRoutingDecision {
        if let Some(model) = self.preferred_model_for_capability(ModelCapability::Generate) {
            let ctx_len = self
                .unified_registry
                .get(model)
                .or_else(|| self.unified_registry.find_by_name(model))
                .map(|s| s.context_length)
                .unwrap_or(0);
            return AdaptiveRoutingDecision {
                model_id: model.to_string(),
                model_name: model.to_string(),
                task: InferenceTask::Generate,
                complexity: TaskComplexity::assess(prompt),
                reason: "preferred generation model override".into(),
                strategy: RoutingStrategy::Explicit,
                predicted_quality: 0.5,
                fallbacks: vec![],
                context_length: ctx_len,
                needs_compaction: false,
            };
        }
        let tracker = self.outcome_tracker.read().await;
        self.adaptive_router
            .route(prompt, &self.unified_registry, &tracker)
    }

    /// Route a prompt to the best model without executing (legacy compat).
    pub fn route(&self, prompt: &str) -> RoutingDecision {
        self.router.route_generate(prompt, &self.registry)
    }

    /// Estimate token count for a request against a specific model's context window.
    /// Returns (estimated_input_tokens, context_window_tokens, fits).
    pub fn estimated_tokens(
        &self,
        req: &GenerateRequest,
        model_id: Option<&str>,
    ) -> (usize, usize, bool) {
        let prompt_tokens = remote::estimate_tokens(&req.prompt);
        let context_tokens = req
            .context
            .as_ref()
            .map(|c| remote::estimate_tokens(c))
            .unwrap_or(0);
        let tools_tokens = req
            .tools
            .as_ref()
            .map(|t| remote::estimate_tokens(&serde_json::to_string(t).unwrap_or_default()))
            .unwrap_or(0);
        let total_input = prompt_tokens + context_tokens + tools_tokens;

        let context_window = model_id
            .and_then(|id| {
                self.unified_registry
                    .get(id)
                    .or_else(|| self.unified_registry.find_by_name(id))
            })
            .map(|s| s.context_length)
            .unwrap_or(0);

        let fits = context_window == 0 || (total_input + req.params.max_tokens) <= context_window;
        (total_input, context_window, fits)
    }

    /// Generate text from a prompt with outcome tracking.
    /// Returns `InferenceResult` with trace_id for reporting outcomes.
    #[instrument(
        name = "inference.generate",
        skip_all,
        fields(
            model = tracing::field::Empty,
            max_tokens = req.params.max_tokens,
            prompt_tokens = tracing::field::Empty,
            completion_tokens = tracing::field::Empty,
            latency_ms = tracing::field::Empty,
        )
    )]
    pub async fn generate_tracked(
        &self,
        req: GenerateRequest,
    ) -> Result<InferenceResult, InferenceError> {
        let start = Instant::now();

        // Route using adaptive router (context-aware)
        let (estimated_input, _, _) = self.estimated_tokens(&req, None);
        let tracker_read = self.outcome_tracker.read().await;
        let has_tools = req.tools.is_some();
        let has_vision = Self::request_needs_vision(&req);
        let preferred_model = self
            .preferred_model_for_capability(ModelCapability::Generate)
            .map(str::to_string);
        let decision = match req.model.clone().or(preferred_model) {
            Some(m) => {
                let ctx_len = self
                    .unified_registry
                    .get(&m)
                    .or_else(|| self.unified_registry.find_by_name(&m))
                    .map(|s| s.context_length)
                    .unwrap_or(0);
                AdaptiveRoutingDecision {
                    model_id: m.clone(),
                    model_name: m.clone(),
                    task: InferenceTask::Generate,
                    complexity: TaskComplexity::assess(&req.prompt),
                    reason: "explicit model".into(),
                    strategy: RoutingStrategy::Explicit,
                    predicted_quality: 0.5,
                    fallbacks: vec![],
                    context_length: ctx_len,
                    needs_compaction: ctx_len > 0 && estimated_input > ctx_len,
                }
            }
            None => match &req.intent {
                Some(hint) => self.adaptive_router.route_context_aware_with_intent(
                    &req.prompt,
                    estimated_input,
                    &self.unified_registry,
                    &tracker_read,
                    has_tools,
                    has_vision,
                    req.params.workload,
                    hint,
                ),
                None => self.adaptive_router.route_context_aware(
                    &req.prompt,
                    estimated_input,
                    &self.unified_registry,
                    &tracker_read,
                    has_tools,
                    has_vision,
                    req.params.workload,
                ),
            },
        };
        drop(tracker_read);

        if decision.needs_compaction {
            tracing::info!(
                model = %decision.model_name,
                prompt_tokens = estimated_input,
                context_window = decision.context_length,
                "prompt exceeds model context window — compaction or truncation needed"
            );
        }

        // Record start
        let trace_id = {
            let mut tracker = self.outcome_tracker.write().await;
            tracker.record_start(&decision.model_id, decision.task, &decision.reason)
        };

        debug!(
            model = %decision.model_name,
            strategy = ?decision.strategy,
            reason = %decision.reason,
            trace = %trace_id,
            "adaptive-routed generate request"
        );

        // Auto-enable extended thinking for complex tasks when the model supports it
        // and the caller hasn't explicitly set budget_tokens.
        let mut req = req;
        if req.params.budget_tokens == 0 && matches!(decision.complexity, TaskComplexity::Complex) {
            let supports_thinking = self
                .unified_registry
                .get(&decision.model_id)
                .map(|s| {
                    s.supported_params
                        .contains(&schema::GenerateParam::ExtendedThinking)
                })
                .unwrap_or(false);
            if supports_thinking {
                req.params.budget_tokens = 8000;
                tracing::info!(model = %decision.model_name, budget = 8000, "auto-enabled extended thinking for complex task");
            }
        }

        // Execute — dispatch to local or remote backend, with fallback on failure
        let mut models_to_try = vec![decision.model_id.clone()];
        models_to_try.extend(decision.fallbacks.iter().cloned());

        let mut last_error = None;

        for candidate_id in &models_to_try {
            // `mut` is needed on the aarch64-macos cfg branch below;
            // other targets don't rebind.
            #[allow(unused_mut)]
            let mut schema = self
                .unified_registry
                .get(candidate_id)
                .or_else(|| self.unified_registry.find_by_name(candidate_id))
                .cloned();

            // On Apple Silicon, redirect GGUF/Candle models to their MLX equivalents.
            #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
            if let Some(ref s) = schema {
                if let Some(mlx_equiv) = self.unified_registry.resolve_mlx_equivalent(s) {
                    tracing::info!(
                        from = %s.id, to = %mlx_equiv.id,
                        "redirecting GGUF model to MLX equivalent on Apple Silicon"
                    );
                    schema = Some(mlx_equiv.clone());
                }
            }

            let candidate_name = schema
                .as_ref()
                .map(|s| s.name.clone())
                .unwrap_or_else(|| candidate_id.clone());

            let is_remote = schema
                .as_ref()
                .map(|s| s.is_remote() || s.is_vllm_mlx())
                .unwrap_or(false);

            let has_tools = req.tools.is_some();

            // Reinforce done tool instructions in context (fixes #10: empty done results)
            let context = if has_tools
                && req.tools.as_ref().map_or(false, |t| {
                    t.iter().any(|tool| {
                        tool.get("function")
                            .and_then(|f| f.get("name"))
                            .and_then(|n| n.as_str())
                            == Some("done")
                    })
                }) {
                let base = req.context.as_deref().unwrap_or("");
                Some(format!(
                    "{base}\n\nIMPORTANT: When calling the `done` tool, the `result` field MUST contain a DETAILED summary of everything you found and did. This is the ONLY output the user sees. Do NOT just say 'completed' — include specific findings, data, and conclusions."
                ))
            } else {
                req.context.clone()
            };

            let result = if is_remote {
                let schema_val = schema.unwrap();
                let _ctx_len = schema_val.context_length;
                // Strip unsupported params based on model schema (#15).
                // Use -1.0 as sentinel: remote backends omit temperature entirely.
                let temperature = if !schema_val.supported_params.is_empty()
                    && !schema_val
                        .supported_params
                        .contains(&crate::schema::GenerateParam::Temperature)
                {
                    -1.0
                } else {
                    req.params.temperature
                };

                // Always use the multi path so token usage is preserved on
                // both tool and non-tool requests. The bare `generate()` helper
                // in remote_backend wraps this same call but drops the usage
                // tuple, which breaks observability for plain text inference
                // (sc-3 in outcome 043).
                self.remote_backend
                    .generate_with_tools_multi(
                        &schema_val,
                        &req.prompt,
                        context.as_deref(),
                        temperature,
                        req.params.max_tokens,
                        req.tools.as_deref(),
                        req.images.as_deref(),
                        req.messages.as_deref(),
                        req.params.tool_choice.as_deref(),
                        req.params.parallel_tool_calls,
                        req.params.budget_tokens,
                        req.cache_control,
                        req.response_format.as_ref(),
                    )
                    .await
                    // Non-streaming remote APIs don't expose a
                    // first-token timestamp. Set TTFT=None and let
                    // streaming-aware callers measure it themselves
                    // via generate_tracked_stream.
                    .map(|(t, c, u)| (t, c, u, None::<u64>))
            } else {
                #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
                {
                    // On Apple Silicon, all local models must go through MLX.
                    // GGUF models were redirected to MLX equivalents above;
                    // if we still have a non-MLX model here, it has no MLX equivalent.
                    let schema_ref = schema
                        .as_ref()
                        .ok_or_else(|| InferenceError::ModelNotFound(candidate_id.clone()))?;

                    // Apple FoundationModels — on-device system model.
                    // Tools, vision/audio/video aren't wired in v1; reject
                    // upstream so the router falls through to a richer
                    // model rather than silently dropping capabilities.
                    if schema_ref.is_foundation_models() {
                        if has_tools {
                            Err(InferenceError::UnsupportedMode {
                                mode: "tool-use",
                                backend: "foundation-models",
                                reason:
                                    "tool calling not yet wired through the FoundationModels \
                                     bridge — route to a remote model with ToolUse capability",
                            })
                        } else if Self::request_has_video(&req)
                            || Self::request_has_audio(&req)
                            || req
                                .images
                                .as_ref()
                                .is_some_and(|imgs| !imgs.is_empty())
                        {
                            Err(InferenceError::UnsupportedMode {
                                mode: "multimodal-content",
                                backend: "foundation-models",
                                reason:
                                    "the FoundationModels bridge currently exposes text-only \
                                     generation — route image/audio/video to a remote VL model",
                            })
                        } else {
                            let prompt = req.prompt.clone();
                            let instructions = context.clone();
                            let max_tokens = req.params.max_tokens as u32;
                            let temperature = req.params.temperature;
                            tokio::task::spawn_blocking(move || {
                                crate::backend::foundation_models::generate(
                                    &prompt,
                                    instructions.as_deref(),
                                    max_tokens,
                                    temperature as f32,
                                )
                            })
                            .await
                            .map_err(|e| {
                                InferenceError::InferenceFailed(format!(
                                    "FoundationModels task panicked: {e}"
                                ))
                            })
                            .and_then(|r| r)
                            .map(|text| (text, vec![], None, None))
                        }
                    } else if !schema_ref.is_mlx() {
                        Err(InferenceError::InferenceFailed(format!(
                            "model '{}' has no MLX equivalent; Candle backend disabled on Apple Silicon",
                            schema_ref.id
                        )))
                    } else if schema_ref.tags.iter().any(|t| t == "mlx-vlm-cli") {
                        // Schemas explicitly tagged for the
                        // mlx-vlm CLI shell-out path skip the
                        // wasteful native-MLX text-tower load
                        // entirely — `mlx_vlm.generate` loads its
                        // own weights from the HF cache and
                        // performs vision tokenization that the
                        // native backend does not. Falls through
                        // to the same error message as the
                        // post-load fallback when mlx-vlm is not
                        // installed, so the user-facing failure
                        // is consistent.
                        let has_images = req
                            .images
                            .as_ref()
                            .is_some_and(|imgs| !imgs.is_empty());
                        if !has_images {
                            return Err(InferenceError::UnsupportedMode {
                                mode: "text-only-on-mlx-vlm-id",
                                backend: "mlx-vlm-cli",
                                reason:
                                    "the `mlx-vlm/...` model IDs route exclusively \
                                     through the mlx-vlm CLI for image inference. \
                                     For text-only generation, route to a Qwen3 \
                                     text model (`mlx/qwen3-4b:4bit` etc.) — the \
                                     CLI shell-out has higher latency than the \
                                     in-process MLX text tower.",
                            });
                        }
                        if !crate::backend::mlx_vlm_cli::is_available() {
                            return Err(InferenceError::UnsupportedMode {
                                mode: "image-content-block",
                                backend: "mlx-vlm-cli",
                                reason:
                                    "mlx-vlm CLI not found on PATH. Install with \
                                     `uv tool install mlx-vlm` (or `pip install \
                                     mlx-vlm`) and CAR will route through \
                                     `mlx_vlm.generate` automatically. \
                                     Alternatives: a local vLLM-MLX VLM server, \
                                     or a remote VL model. (#115)",
                            });
                        }
                        let repo = match &schema_ref.source {
                            crate::schema::ModelSource::Mlx { hf_repo, .. } => {
                                hf_repo.clone()
                            }
                            _ => {
                                return Err(InferenceError::InferenceFailed(format!(
                                    "model '{}' is tagged mlx-vlm-cli but its \
                                     source isn't ModelSource::Mlx — registry bug",
                                    schema_ref.id
                                )));
                            }
                        };
                        let imgs = req.images.clone().unwrap_or_default();
                        let temp = req.params.temperature;
                        let max_t = req.params.max_tokens;
                        let prompt = req.prompt.clone();
                        let text = tokio::task::spawn_blocking(move || {
                            crate::backend::mlx_vlm_cli::generate(
                                &repo, &prompt, &imgs, temp, max_t,
                            )
                        })
                        .await
                        .map_err(|e| {
                            InferenceError::InferenceFailed(format!(
                                "mlx_vlm CLI task panicked: {e}"
                            ))
                        })??;
                        let bounding_boxes = parse_boxes(&text);
                        let latency_ms = start.elapsed().as_millis() as u64;
                        {
                            let mut tracker = self.outcome_tracker.write().await;
                            tracker.record_complete(&trace_id, latency_ms, 0, 0);
                        }
                        return Ok(InferenceResult {
                            text,
                            tool_calls: vec![],
                            bounding_boxes,
                            trace_id: trace_id.clone(),
                            model_used: schema_ref.id.clone(),
                            latency_ms,
                            time_to_first_token_ms: None,
                            usage: None,
                        });
                    } else {
                        // Load the backend first so we can ask it what
                        // it's actually able to execute. VL checkpoints
                        // currently load as text-only towers (see the
                        // `language_model.` prefix strip in backend/mlx.rs);
                        // the backend's `supports_capability(Vision)`
                        // returns false until GH #58 wires the vision
                        // tower. A registry-level capability claim is
                        // an aspiration for routing; the backend answer
                        // is the execution contract.
                        let handle = self.ensure_mlx_backend(schema_ref).await?;
                        // Native MLX path doesn't have a video
                        // tokenization pipeline yet. Reject video
                        // content blocks up front with a precise
                        // UnsupportedMode so callers don't silently
                        // get a text-only reply.
                        if Self::request_has_video(&req) {
                            return Err(InferenceError::UnsupportedMode {
                                mode: "video-content-block",
                                backend: "native-mlx-qwen25vl",
                                reason:
                                    "Qwen2.5-VL video understanding is on the request surface \
                                     but the video-tokenization path (frame sampling + merger) \
                                     is not yet wired; route to a remote VL provider for now",
                            });
                        }
                        if Self::request_has_audio(&req) {
                            return Err(InferenceError::UnsupportedMode {
                                mode: "audio-content-block",
                                backend: "native-mlx-qwen25vl",
                                reason:
                                    "audio understanding is on the request surface (Gemma 4 \
                                     E2B/E4B and Gemini accept it) but the native MLX path \
                                     for this model does not — route to Gemini or Gemma-4",
                            });
                        }
                        let has_images = req
                            .images
                            .as_ref()
                            .is_some_and(|imgs| !imgs.is_empty());
                        if has_images {
                            let can_do_vision = {
                                let guard = handle.lock().map_err(|_| {
                                    InferenceError::InferenceFailed(
                                        "MLX backend mutex poisoned".into(),
                                    )
                                })?;
                                guard.supports_capability(
                                    crate::schema::ModelCapability::Vision,
                                )
                            };
                            if !can_do_vision {
                                // The `mlx-vlm-cli`-tagged route handled
                                // above is the primary fix for #115; if
                                // we landed here it means the schema is
                                // a non-tagged `ModelSource::Mlx` (e.g.
                                // a user-registered custom model that
                                // doesn't advertise the CLI route). The
                                // error message points them at the
                                // tagged catalog IDs rather than
                                // claiming nothing local works.
                                return Err(InferenceError::UnsupportedMode {
                                    mode: "image-content-block",
                                    backend: "native-mlx-text",
                                    reason:
                                        "this MLX backend is a plain Qwen3 text tower. \
                                         For local image inference, route to \
                                         `mlx-vlm/qwen3-vl-2b:bf16` or another `mlx-vlm/...` \
                                         catalog ID so CAR shells out to `mlx_vlm.generate`. \
                                         Alternatives: a local vLLM-MLX VLM server, or a \
                                         remote VL model. (#115)",
                                });
                            }
                        }
                        self.generate_mlx(req.clone(), &schema_ref.id)
                            .await
                            .map(|(text, ttft)| (text, vec![], None, ttft))
                    }
                }

                #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
                {
                    match self.ensure_backend(&candidate_name).await {
                        Ok(()) => {
                            let mut write = self.backend.write().await;
                            let backend = write.as_mut().unwrap();
                            tasks::generate::generate(backend, req.clone())
                                .await
                                .map(|(text, ttft)| (text, vec![], None, ttft))
                        }
                        Err(e) => Err(e),
                    }
                }
            };

            match result {
                Ok((text, tool_calls, usage, time_to_first_token_ms)) => {
                    let latency_ms = start.elapsed().as_millis() as u64;
                    let estimated_tokens = usage
                        .as_ref()
                        .map(|u| u.completion_tokens as usize)
                        .unwrap_or_else(|| text.split_whitespace().count());
                    {
                        let mut tracker = self.outcome_tracker.write().await;
                        tracker.record_complete(&trace_id, latency_ms, 0, estimated_tokens);
                    }
                    // Circuit breaker: record success (#25)
                    if let Ok(mut cb) = self.adaptive_router.circuit_breakers.lock() {
                        cb.record_success(candidate_id);
                    }
                    // Auto-persist profiles after each successful call
                    self.auto_save_outcomes().await;

                    // Record deferred span fields now that we have the result
                    let span = tracing::Span::current();
                    span.record("model", candidate_name.as_str());
                    span.record("latency_ms", latency_ms);
                    if let Some(ttft) = time_to_first_token_ms {
                        span.record("ttft_ms", ttft);
                    }
                    if let Some(ref u) = usage {
                        span.record("prompt_tokens", u.prompt_tokens);
                        span.record("completion_tokens", u.completion_tokens);
                    }

                    // Parse Qwen2.5-VL grounding spans out of the
                    // text output. Empty vec on anything else.
                    let bounding_boxes = tasks::grounding::parse_boxes(&text);
                    return Ok(InferenceResult {
                        text,
                        tool_calls,
                        bounding_boxes,
                        trace_id,
                        model_used: candidate_name,
                        latency_ms,
                        time_to_first_token_ms,
                        usage,
                    });
                }
                Err(e) => {
                    tracing::warn!(
                        model = %candidate_name,
                        error = %e,
                        remaining = models_to_try.len().saturating_sub(
                            models_to_try.iter().position(|m| m == candidate_id).unwrap_or(0) + 1
                        ),
                        "model failed, trying next fallback immediately"
                    );
                    // Record failure for this model so the router learns
                    {
                        let mut tracker = self.outcome_tracker.write().await;
                        let fail_trace =
                            tracker.record_start(candidate_id, decision.task, "fallback");
                        tracker.record_failure(&fail_trace, &e.to_string());
                    }
                    // Circuit breaker: record failure (#25).
                    // 4xx errors (client errors) use longer cooldown since they indicate
                    // permanent incompatibility (wrong endpoint, unsupported param).
                    {
                        let err_str = e.to_string();
                        let is_client_error =
                            err_str.contains("API returned 4") && !err_str.contains("429");
                        if let Ok(mut cb) = self.adaptive_router.circuit_breakers.lock() {
                            cb.record_failure(candidate_id);
                            // For persistent 4xx errors, lower the threshold by
                            // recording an extra failure to trip faster
                            if is_client_error {
                                cb.record_failure(candidate_id);
                            }
                        }
                    }
                    // Reset backend so next model can load
                    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
                    {
                        let mut write = self.backend.write().await;
                        *write = None;
                    }
                    last_error = Some(e);
                }
            }
        }

        // All models failed
        let e = last_error.unwrap_or(InferenceError::InferenceFailed(
            "no models available".into(),
        ));
        {
            let mut tracker = self.outcome_tracker.write().await;
            tracker.record_failure(&trace_id, &e.to_string());
        }
        self.auto_save_outcomes().await;
        Err(e)
    }

    /// Stream a response with real-time token output.
    ///
    /// Returns a channel receiver yielding `StreamEvent` variants:
    /// - `TextDelta(String)` — partial text token
    /// - `ToolCallStart { name, index, id }` — tool call begins
    /// - `ToolCallDelta { index, arguments_delta }` — partial tool arguments
    ///
    /// Use `StreamAccumulator` to collect events into a final result.
    ///
    /// ## Local streaming behavior
    ///
    /// Local models (both MLX and Candle backends) emit true incremental `TextDelta`
    /// events as each token is generated. This enables:
    /// - Token-by-token UI updates
    /// - Overlapping speech synthesis with generation (for voice apps)
    /// - Early cancellation when a stop sequence is detected
    ///
    /// The channel capacity is 64 events, providing buffering for burst tokens
    /// without blocking the generation loop.
    ///
    /// ## Example: voice app integration
    ///
    /// ```rust,ignore
    /// let mut rx = engine.generate_tracked_stream(req).await?;
    /// let mut text_buf = String::new();
    /// while let Some(event) = rx.recv().await {
    ///     match event {
    ///         StreamEvent::TextDelta(delta) => {
    ///             text_buf.push_str(&delta);
    ///             // Feed text_buf to TTS when a sentence boundary is reached
    ///         }
    ///         StreamEvent::Done { text, .. } => break,
    ///         _ => {}
    ///     }
    /// }
    /// ```
    pub async fn generate_tracked_stream(
        &self,
        req: GenerateRequest,
    ) -> Result<tokio::sync::mpsc::Receiver<stream::StreamEvent>, InferenceError> {
        let has_tools = req.tools.is_some();
        let has_vision = Self::request_needs_vision(&req);
        let preferred_model = self
            .preferred_model_for_capability(ModelCapability::Generate)
            .map(str::to_string);
        let decision = match req.model.clone().or(preferred_model) {
            Some(m) => {
                let ctx_len = self
                    .unified_registry
                    .get(&m)
                    .or_else(|| self.unified_registry.find_by_name(&m))
                    .map(|s| s.context_length)
                    .unwrap_or(0);
                AdaptiveRoutingDecision {
                    model_id: m.clone(),
                    model_name: m,
                    task: InferenceTask::Generate,
                    complexity: TaskComplexity::assess(&req.prompt),
                    reason: "explicit model".into(),
                    strategy: RoutingStrategy::Explicit,
                    predicted_quality: 0.5,
                    fallbacks: vec![],
                    context_length: ctx_len,
                    needs_compaction: false,
                }
            }
            None => {
                let tracker_read = self.outcome_tracker.read().await;
                if has_vision {
                    self.adaptive_router.route_with_vision(
                        &req.prompt,
                        &self.unified_registry,
                        &tracker_read,
                        has_tools,
                    )
                } else if has_tools {
                    self.adaptive_router.route_with_tools(
                        &req.prompt,
                        &self.unified_registry,
                        &tracker_read,
                    )
                } else {
                    self.adaptive_router
                        .route(&req.prompt, &self.unified_registry, &tracker_read)
                }
            }
        };

        // `mut` is needed on the aarch64-macos cfg branch below;
        // other targets don't rebind.
        #[allow(unused_mut)]
        let mut schema = self
            .unified_registry
            .get(&decision.model_id)
            .or_else(|| self.unified_registry.find_by_name(&decision.model_id))
            .cloned();

        // On Apple Silicon, redirect GGUF/Candle models to their MLX equivalents.
        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        if let Some(ref s) = schema {
            if let Some(mlx_equiv) = self.unified_registry.resolve_mlx_equivalent(s) {
                tracing::info!(
                    from = %s.id, to = %mlx_equiv.id,
                    "redirecting GGUF model to MLX equivalent on Apple Silicon (stream)"
                );
                schema = Some(mlx_equiv.clone());
            }
        }

        let is_remote = schema
            .as_ref()
            .map(|s| s.is_remote() || s.is_vllm_mlx())
            .unwrap_or(false);

        if is_remote {
            let schema = schema.unwrap();
            // Register keys for this model
            self.remote_backend.register_model_keys(&schema).await;

            self.remote_backend
                .generate_stream(
                    &schema,
                    &req.prompt,
                    req.context.as_deref(),
                    req.params.temperature,
                    req.params.max_tokens,
                    req.tools.as_deref(),
                    req.images.as_deref(),
                    req.params.tool_choice.as_deref(),
                    req.params.parallel_tool_calls,
                    req.response_format.as_ref(),
                )
                .await
        } else {
            let schema = schema.ok_or_else(|| {
                InferenceError::ModelNotFound(decision.model_id.clone())
            })?;
            let (tx, rx) = tokio::sync::mpsc::channel(64);

            #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
            {
                // FoundationModels streaming — text-only, no tools.
                if schema.is_foundation_models() {
                    let prompt = req.prompt.clone();
                    let instructions = req.context.clone();
                    let max_tokens = req.params.max_tokens as u32;
                    let temperature = req.params.temperature;
                    let tx_clone = tx.clone();
                    tokio::task::spawn_blocking(move || {
                        // Share the accumulator between the streaming
                        // callback and the post-stream Done event so
                        // the FoundationModels path matches Candle/MLX
                        // shape — `Done.text` is the full assembled
                        // generation, not an empty sentinel that
                        // forces consumers to reassemble.
                        let accum = std::sync::Arc::new(std::sync::Mutex::new(String::new()));
                        let accum_cb = accum.clone();
                        let cb = crate::backend::foundation_models::StreamCallback::new(
                            move |delta: &str| {
                                if let Ok(mut g) = accum_cb.lock() {
                                    g.push_str(delta);
                                }
                                tx_clone
                                    .blocking_send(stream::StreamEvent::TextDelta(
                                        delta.to_string(),
                                    ))
                                    .is_ok()
                            },
                        );
                        let result = crate::backend::foundation_models::stream(
                            &prompt,
                            instructions.as_deref(),
                            max_tokens,
                            temperature as f32,
                            cb,
                        );
                        let final_text = accum
                            .lock()
                            .map(|g| g.clone())
                            .unwrap_or_default();
                        let _ = tx.blocking_send(stream::StreamEvent::Done {
                            text: final_text,
                            tool_calls: vec![],
                        });
                        result
                    });
                    return Ok(rx);
                }
                // On Apple Silicon, all local models must go through MLX.
                if !schema.is_mlx() {
                    return Err(InferenceError::InferenceFailed(format!(
                        "model '{}' has no MLX equivalent; Candle backend disabled on Apple Silicon",
                        schema.id
                    )));
                }
                let backend = self.ensure_mlx_backend(&schema).await?;
                let model_id = schema.id.clone();
                let cache = Arc::clone(&self.mlx_backends);
                // MLX ops are blocking (GPU-bound) and `MutexGuard<MlxBackend>`
                // isn't `Send`, so run the whole generation on a blocking
                // worker. `tx.blocking_send` bridges tokens back to the
                // async stream consumer without holding the guard across
                // an `.await`.
                tokio::task::spawn_blocking(move || {
                    let _ = Self::stream_local_mlx(backend, cache, model_id, req, tx);
                });
                return Ok(rx);
            }

            #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
            {
                self.ensure_backend(&schema.name).await?;
                let backend = self.backend.clone();
                tokio::spawn(async move {
                    let _ = Self::stream_local_candle(backend, req, tx).await;
                });
                Ok(rx)
            }
        }
    }

    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    async fn stream_local_candle(
        backend_lock: Arc<RwLock<Option<CandleBackend>>>,
        req: GenerateRequest,
        tx: tokio::sync::mpsc::Sender<stream::StreamEvent>,
    ) -> Result<(), InferenceError> {
        let mut write = backend_lock.write().await;
        let backend = write
            .as_mut()
            .ok_or_else(|| InferenceError::InferenceFailed("backend not initialized".into()))?;
        backend.clear_kv_cache();

        let formatted = tasks::generate::apply_chat_template(
            &req.prompt,
            req.context.as_deref(),
            req.params.thinking,
        );
        let tokens = backend.encode(&formatted)?;
        let eos = backend.eos_token_id();
        let eos_alt = backend.token_id("<|im_end|>");
        let params = &req.params;

        if tokens.is_empty() {
            let _ = tx
                .send(stream::StreamEvent::Done {
                    text: String::new(),
                    tool_calls: vec![],
                })
                .await;
            return Ok(());
        }

        let max_ctx = backend.context_length().unwrap_or(32768);
        let headroom = params.max_tokens.min(max_ctx / 4);
        let max_prompt = max_ctx.saturating_sub(headroom);
        let tokens = if tokens.len() > max_prompt {
            tokens[tokens.len() - max_prompt..].to_vec()
        } else {
            tokens
        };

        let mut generated = Vec::new();
        let logits = backend.forward(&tokens, 0)?;
        let mut next_token = tasks::generate::sample_token(&logits, params)?;

        for _ in 0..params.max_tokens {
            if eos.map_or(false, |id| next_token == id)
                || eos_alt.map_or(false, |id| next_token == id)
            {
                break;
            }

            generated.push(next_token);
            let delta = backend.decode(&[next_token])?;
            if !delta.is_empty()
                && tx.send(stream::StreamEvent::TextDelta(delta)).await.is_err()
            {
                return Ok(());
            }

            if !params.stop.is_empty() {
                let text_so_far = backend.decode(&generated)?;
                if params.stop.iter().any(|s| text_so_far.contains(s)) {
                    break;
                }
            }

            let pos = tokens.len() + generated.len() - 1;
            let logits = backend.forward(&[next_token], pos)?;
            next_token = tasks::generate::sample_token(&logits, params)?;
        }

        let text = tasks::generate::strip_thinking(&backend.decode(&generated)?, params.thinking);
        let _ = tx
            .send(stream::StreamEvent::Done {
                text,
                tool_calls: vec![],
            })
            .await;
        Ok(())
    }

    /// Blocking streaming generator — runs on a `spawn_blocking` worker
    /// so the sync MLX mutex guard isn't held across an async `.await`.
    /// Uses `tx.blocking_send` to push tokens back to the caller.
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    fn stream_local_mlx(
        handle: backend_cache::CachedBackend<backend::MlxBackend>,
        cache: Arc<backend_cache::BackendCache<backend::MlxBackend>>,
        model_id: String,
        req: GenerateRequest,
        tx: tokio::sync::mpsc::Sender<stream::StreamEvent>,
    ) -> Result<(), InferenceError> {
        let mut guard = handle.lock().map_err(|_| {
            InferenceError::InferenceFailed(format!(
                "MLX backend mutex poisoned for {model_id}"
            ))
        })?;
        let backend: &mut backend::MlxBackend = &mut *guard;
        backend.clear_kv_cache();

        let formatted = tasks::generate::apply_chat_template(
            &req.prompt,
            req.context.as_deref(),
            req.params.thinking,
        );
        let tokens = backend.encode(&formatted)?;
        let eos = backend.eos_token_id();
        let eos_alt = backend.token_id("<|im_end|>");
        let params = &req.params;

        if tokens.is_empty() {
            let _ = tx.blocking_send(stream::StreamEvent::Done {
                text: String::new(),
                tool_calls: vec![],
            });
            return Ok(());
        }

        let max_ctx = backend.context_length();
        let headroom = params.max_tokens.min(max_ctx / 4);
        let max_prompt = max_ctx.saturating_sub(headroom);
        let tokens = if tokens.len() > max_prompt {
            tokens[tokens.len() - max_prompt..].to_vec()
        } else {
            tokens
        };

        let mut generated = Vec::new();
        // Wrap MLX forward calls to catch panics. On panic, drop this
        // backend from the cache — its KV cache may be in an
        // indeterminate state and subsequent callers would inherit it.
        // Outstanding handles continue to work until their Arc drops.
        let logits = match Self::catch_mlx("stream prefill", || backend.forward(&tokens, 0)) {
            Ok(v) => v,
            Err(e) => { cache.invalidate(&model_id); return Err(e); }
        };
        let mut next_token = Self::sample_from_logits(&logits, params)?;

        for _ in 0..params.max_tokens {
            if eos.map_or(false, |id| next_token == id)
                || eos_alt.map_or(false, |id| next_token == id)
            {
                break;
            }

            generated.push(next_token);
            let delta = backend.decode(&[next_token])?;
            if !delta.is_empty()
                && tx.blocking_send(stream::StreamEvent::TextDelta(delta)).is_err()
            {
                return Ok(());
            }

            if !params.stop.is_empty() {
                let text_so_far = backend.decode(&generated)?;
                if params.stop.iter().any(|s| text_so_far.contains(s)) {
                    break;
                }
            }

            let pos = tokens.len() + generated.len() - 1;
            let logits = match Self::catch_mlx("stream forward", || backend.forward(&[next_token], pos)) {
                Ok(v) => v,
                Err(e) => { cache.invalidate(&model_id); return Err(e); }
            };
            next_token = Self::sample_from_logits(&logits, params)?;
        }

        let text = tasks::generate::strip_thinking(&backend.decode(&generated)?, params.thinking);
        let _ = tx.blocking_send(stream::StreamEvent::Done {
            text,
            tool_calls: vec![],
        });
        Ok(())
    }

    /// Route a prompt using the adaptive router without executing inference.
    pub async fn route_context_snapshot(
        &self,
        prompt: &str,
        workload: RoutingWorkload,
        has_tools: bool,
        has_vision: bool,
    ) -> AdaptiveRoutingDecision {
        let tracker = self.outcome_tracker.read().await;
        self.adaptive_router.route_context_aware(
            prompt,
            0,
            &self.unified_registry,
            &tracker,
            has_tools,
            has_vision,
            workload,
        )
    }

    /// Generate text from a prompt (legacy API, no outcome tracking).
    /// When `req.model` is None, uses intelligent routing based on prompt complexity.
    pub async fn generate(&self, req: GenerateRequest) -> Result<String, InferenceError> {
        Ok(self.generate_tracked(req).await?.text)
    }

    /// Encode `text` via the named model's tokenizer. Returns raw token IDs
    /// without any chat-template wrapping or BOS-prepending — pair with
    /// [`Self::detokenize`] for the round-trip property
    /// `detokenize(model, tokenize(model, s)) == s` for any UTF-8 `s`.
    ///
    /// Only local models have a tokenizer the runtime can call directly
    /// (Candle/GGUF on Linux/Windows, MLX on Apple Silicon). For remote
    /// models the call returns
    /// [`InferenceError::UnsupportedMode`] — provider tokenizer endpoints
    /// vary too widely to be portable here, and bundling tiktoken-style
    /// tables would lock the registry to a fixed set of providers.
    pub async fn tokenize(
        &self,
        model: &str,
        text: &str,
    ) -> Result<Vec<u32>, InferenceError> {
        self.assert_local_for_tokenize(model)?;

        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            let schema = self
                .unified_registry
                .get(model)
                .or_else(|| self.unified_registry.find_by_name(model))
                .ok_or_else(|| InferenceError::ModelNotFound(model.to_string()))?
                .clone();
            let handle = self.ensure_mlx_backend(&schema).await?;
            let guard = handle.lock().map_err(|_| {
                InferenceError::InferenceFailed(format!(
                    "MLX backend mutex poisoned for {}", schema.id
                ))
            })?;
            return guard.tokenize_raw(text);
        }

        #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
        {
            self.ensure_backend(model).await?;
            let read = self.backend.read().await;
            let backend = read.as_ref().ok_or_else(|| {
                InferenceError::InferenceFailed(
                    "candle backend missing after ensure_backend".to_string(),
                )
            })?;
            backend.tokenize_raw(text)
        }
    }

    /// Inverse of [`Self::tokenize`]: decode token IDs back to text.
    pub async fn detokenize(
        &self,
        model: &str,
        tokens: &[u32],
    ) -> Result<String, InferenceError> {
        self.assert_local_for_tokenize(model)?;

        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            let schema = self
                .unified_registry
                .get(model)
                .or_else(|| self.unified_registry.find_by_name(model))
                .ok_or_else(|| InferenceError::ModelNotFound(model.to_string()))?
                .clone();
            let handle = self.ensure_mlx_backend(&schema).await?;
            let guard = handle.lock().map_err(|_| {
                InferenceError::InferenceFailed(format!(
                    "MLX backend mutex poisoned for {}", schema.id
                ))
            })?;
            return guard.detokenize_raw(tokens);
        }

        #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
        {
            self.ensure_backend(model).await?;
            let read = self.backend.read().await;
            let backend = read.as_ref().ok_or_else(|| {
                InferenceError::InferenceFailed(
                    "candle backend missing after ensure_backend".to_string(),
                )
            })?;
            backend.detokenize_raw(tokens)
        }
    }

    /// Common pre-flight for [`Self::tokenize`] / [`Self::detokenize`]: bail
    /// early on remote models with the same `UnsupportedMode` taxonomy used
    /// elsewhere on the engine surface.
    fn assert_local_for_tokenize(&self, model: &str) -> Result<(), InferenceError> {
        if let Some(schema) = self
            .unified_registry
            .get(model)
            .or_else(|| self.unified_registry.find_by_name(model))
        {
            if !schema.is_local() {
                return Err(InferenceError::UnsupportedMode {
                    mode: "tokenize/detokenize",
                    backend: "remote",
                    reason:
                        "remote provider tokenizer is not exposed by the runtime; \
                         use a local model (Qwen3 GGUF / MLX) for tokenizer-correctness checks",
                });
            }
        }
        // Unknown model name: let the load step surface ModelNotFound below.
        Ok(())
    }

    /// Wrap an MLX FFI call with catch_unwind to catch Rust panics at the boundary.
    /// NOTE: This catches Rust panics only, NOT C++ exceptions from Metal/MLX.
    /// True C++ exceptions will still abort the process — that requires an upstream
    /// fix in mlx-rs to catch C++ exceptions before they cross the FFI boundary.
    /// On panic, callers MUST remove the backend from the map — it is poisoned.
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    fn catch_mlx<F, T>(context: &str, f: F) -> Result<T, InferenceError>
    where
        F: FnOnce() -> Result<T, InferenceError>,
    {
        std::panic::catch_unwind(std::panic::AssertUnwindSafe(f))
            .map_err(|e| InferenceError::InferenceFailed(
                format!("MLX panicked during {context}: {e:?}")
            ))?
    }

    /// Generate text using the MLX backend.
    /// Mirrors the Candle generate loop but uses MlxBackend::forward which returns Vec<f32>.
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    async fn generate_mlx(
        &self,
        req: GenerateRequest,
        model_id: &str,
    ) -> Result<(String, Option<u64>), InferenceError> {
        let start = std::time::Instant::now();

        let schema = self
            .unified_registry
            .get(model_id)
            .cloned()
            .ok_or_else(|| InferenceError::InferenceFailed(format!(
                "generate_mlx: unknown schema id {model_id}"
            )))?;
        let handle = self.ensure_mlx_backend(&schema).await?;
        let mut guard = handle.lock().map_err(|_| {
            InferenceError::InferenceFailed(format!(
                "MLX backend mutex poisoned for {model_id}"
            ))
        })?;
        let backend: &mut backend::MlxBackend = &mut *guard;
        backend.clear_kv_cache();

        let formatted = tasks::generate::apply_chat_template(
            &req.prompt,
            req.context.as_deref(),
            req.params.thinking,
        );
        let tokens = backend.encode(&formatted)?;
        let eos = backend.eos_token_id();
        let eos_alt = backend.token_id("<|im_end|>");
        let params = &req.params;

        if tokens.is_empty() {
            return Ok((String::new(), None));
        }

        // Truncate to context length
        let max_ctx = backend.context_length();
        let headroom = params.max_tokens.min(max_ctx / 4);
        let max_prompt = max_ctx.saturating_sub(headroom);
        let tokens = if tokens.len() > max_prompt {
            tokens[tokens.len() - max_prompt..].to_vec()
        } else {
            tokens
        };

        let mut generated = Vec::new();

        // Wrap MLX forward calls to catch panics at the FFI boundary.
        // On panic, drop the backend from the cache — it's in an
        // indeterminate state and we don't want subsequent callers to
        // inherit it. Our outstanding `guard` keeps working until we
        // return.
        let logits = match Self::catch_mlx("prefill", || backend.forward(&tokens, 0)) {
            Ok(v) => v,
            Err(e) => {
                drop(guard);
                self.mlx_backends.invalidate(model_id);
                return Err(e);
            }
        };
        let mut next_token = Self::sample_from_logits(&logits, params)?;
        let ttft_ms = Some(start.elapsed().as_millis() as u64);

        for _ in 0..params.max_tokens {
            if eos.map_or(false, |id| next_token == id)
                || eos_alt.map_or(false, |id| next_token == id)
            {
                break;
            }

            generated.push(next_token);

            if !params.stop.is_empty() {
                let text_so_far = backend.decode(&generated)?;
                if params.stop.iter().any(|s| text_so_far.contains(s)) {
                    break;
                }
            }

            let pos = tokens.len() + generated.len() - 1;
            let logits = match Self::catch_mlx("forward", || backend.forward(&[next_token], pos)) {
                Ok(v) => v,
                Err(e) => {
                    drop(guard);
                    self.mlx_backends.invalidate(model_id);
                    return Err(e);
                }
            };
            next_token = Self::sample_from_logits(&logits, params)?;
        }

        let text = backend.decode(&generated)?;
        Ok((tasks::generate::strip_thinking(&text, params.thinking), ttft_ms))
    }

    /// Sample a token from a logits Vec<f32> (for MLX backend).
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    fn sample_from_logits(logits: &[f32], params: &GenerateParams) -> Result<u32, InferenceError> {
        if params.temperature <= 0.0 {
            // Greedy: argmax
            let (idx, _) = logits
                .iter()
                .enumerate()
                .max_by(|a, b| a.1.partial_cmp(b.1).unwrap_or(std::cmp::Ordering::Equal))
                .ok_or_else(|| InferenceError::InferenceFailed("empty logits".into()))?;
            return Ok(idx as u32);
        }

        // Temperature-scaled softmax sampling
        let temp = params.temperature as f32;
        let max_logit = logits.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
        let mut probs: Vec<f32> = logits
            .iter()
            .map(|&l| ((l - max_logit) / temp).exp())
            .collect();
        let sum: f32 = probs.iter().sum();
        for p in &mut probs {
            *p /= sum;
        }

        // Top-p (nucleus) sampling
        if params.top_p < 1.0 {
            let mut indexed: Vec<(usize, f32)> = probs.iter().copied().enumerate().collect();
            indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
            let mut cumsum = 0.0;
            let mut cutoff_idx = indexed.len();
            for (i, &(_, p)) in indexed.iter().enumerate() {
                cumsum += p;
                if cumsum > params.top_p as f32 {
                    cutoff_idx = i + 1;
                    break;
                }
            }
            // Zero out tokens below the cutoff
            let allowed: std::collections::HashSet<usize> =
                indexed[..cutoff_idx].iter().map(|(i, _)| *i).collect();
            for (i, p) in probs.iter_mut().enumerate() {
                if !allowed.contains(&i) {
                    *p = 0.0;
                }
            }
            let sum: f32 = probs.iter().sum();
            if sum > 0.0 {
                for p in &mut probs {
                    *p /= sum;
                }
            }
        }

        // Sample from distribution
        use rand::Rng;
        let mut rng = rand::rng();
        let r: f32 = rng.random();
        let mut cumsum = 0.0;
        for (i, &p) in probs.iter().enumerate() {
            cumsum += p;
            if cumsum >= r {
                return Ok(i as u32);
            }
        }
        Ok((probs.len() - 1) as u32)
    }

    /// Generate embeddings for text using the dedicated embedding model.
    /// On Apple Silicon, uses the native MLX backend; on other platforms, uses Candle.
    pub async fn embed(&self, req: EmbedRequest) -> Result<Vec<Vec<f32>>, InferenceError> {
        let instruction = req
            .instruction
            .as_deref()
            .unwrap_or("Retrieve relevant memory facts");

        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            let model_id = self.ensure_mlx_embedding_backend().await?;
            let schema = self
                .unified_registry
                .get(&model_id)
                .cloned()
                .ok_or_else(|| InferenceError::InferenceFailed(format!(
                    "embed: unknown schema id {model_id}"
                )))?;
            let handle = self.ensure_mlx_backend(&schema).await?;
            let mut guard = handle.lock().map_err(|_| {
                InferenceError::InferenceFailed(format!(
                    "MLX embedding backend mutex poisoned for {model_id}"
                ))
            })?;
            let backend: &mut backend::MlxBackend = &mut *guard;

            let mut results = Vec::with_capacity(req.texts.len());
            for text in &req.texts {
                let embedding = if req.is_query {
                    backend.embed_query(text, instruction)?
                } else {
                    backend.embed_one(text)?
                };
                results.push(embedding);
            }
            return Ok(results);
        }

        #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
        {
            self.ensure_embedding_backend().await?;
            let mut write = self.embedding_backend.write().await;
            let backend = write.as_mut().unwrap();

            let mut results = Vec::with_capacity(req.texts.len());
            for text in &req.texts {
                let embedding = if req.is_query {
                    backend.embed_query(text, instruction)?
                } else {
                    backend.embed_one(text)?
                };
                results.push(embedding);
            }
            Ok(results)
        }
    }

    /// Rerank candidate documents against a query using a cross-encoder
    /// reranker model (Qwen3-Reranker family). Returns documents sorted
    /// by descending relevance.
    ///
    /// ## Scoring
    ///
    /// Qwen3-Reranker is a Qwen3 base LM fine-tuned so that the first
    /// assistant token is `"yes"` or `"no"` given the templated
    /// `<Instruct>/<Query>/<Document>` user turn. We run a short
    /// greedy decode (≤ 3 tokens, so a leading space, BOS artifact, or
    /// the occasional newline don't break us) and score
    /// `yes → 1.0`, `no → 0.0`, anything else → `0.5` with a warning.
    ///
    /// This is a **binary** score — the soft probability
    /// `softmax(logit_yes, logit_no)` would give finer ordering but
    /// requires per-token logit access on [`backend::MlxBackend`],
    /// which isn't exposed publicly yet. Tracked as a follow-up;
    /// binary scores still produce a correct partial ordering, just
    /// with coarser tiebreaks within the {yes} or {no} groups.
    ///
    /// ## Prompt template
    ///
    /// We emit the upstream Qwen3-Reranker chat template verbatim:
    /// a dedicated system prompt fixing the yes/no answer space,
    /// then the user turn with `<Instruct>/<Query>/<Document>`, then
    /// the assistant prefix with a closed empty `<think>` block to
    /// suppress thinking (reranker is not a reasoner — it's a
    /// classifier). Deviating from this template produces sharply
    /// degraded yes/no distributions.
    pub async fn rerank(&self, req: RerankRequest) -> Result<RerankResult, InferenceError> {
        if req.documents.is_empty() {
            return Ok(RerankResult { ranked: Vec::new(), model_used: None });
        }

        let model_name = match req.model.clone() {
            Some(m) => m,
            None => self
                .preferred_model_for_capability(ModelCapability::Rerank)
                .map(str::to_string)
                .ok_or_else(|| {
                    InferenceError::InferenceFailed(
                        "no reranker model available — pull a Qwen3-Reranker model first".into(),
                    )
                })?,
        };

        let schema = self
            .unified_registry
            .find_by_name(&model_name)
            .or_else(|| self.unified_registry.get(&model_name))
            .cloned()
            .ok_or_else(|| {
                InferenceError::InferenceFailed(format!(
                    "rerank: unknown reranker model {model_name}"
                ))
            })?;
        if !schema.has_capability(ModelCapability::Rerank) {
            return Err(InferenceError::InferenceFailed(format!(
                "model {} does not declare the Rerank capability",
                schema.name
            )));
        }

        let instruction = req.instruction.as_deref().unwrap_or(
            "Given a web search query, retrieve relevant passages that answer the query",
        );

        let mut scored: Vec<RerankedDocument> = Vec::with_capacity(req.documents.len());
        for (idx, doc) in req.documents.iter().enumerate() {
            let prompt = rerank_prompt(instruction, &req.query, doc);
            let gen_req = GenerateRequest {
                prompt,
                model: Some(schema.id.clone()),
                params: tasks::generate::GenerateParams {
                    temperature: 0.0,
                    // Three tokens is enough to scan past a leading
                    // space, BOS, or newline that some tokenizers
                    // insert before the real yes/no token.
                    max_tokens: 3,
                    thinking: tasks::generate::ThinkingMode::Off,
                    ..Default::default()
                },
                context: None,
                tools: None,
                images: None,
                messages: None,
                cache_control: false,
                response_format: None,
                intent: None,
            };
            let out = self.generate(gen_req).await?;
            let score = score_from_rerank_output(&out, &schema.name);
            scored.push(RerankedDocument {
                index: idx,
                score,
                document: doc.clone(),
            });
        }

        // Sort descending by score; preserve original index as a
        // deterministic tiebreaker. top_n must truncate after sorting.
        scored.sort_by(|a, b| {
            b.score
                .partial_cmp(&a.score)
                .unwrap_or(std::cmp::Ordering::Equal)
                .then_with(|| a.index.cmp(&b.index))
        });
        if let Some(n) = req.top_n {
            scored.truncate(n);
        }

        Ok(RerankResult {
            ranked: scored,
            model_used: Some(schema.name),
        })
    }

    /// Dedicated endpoint for structured visual grounding.
    ///
    /// Runs a VL generate call under the hood and parses Qwen2.5-VL's
    /// inline `<|object_ref_*|>...<|box_*|>(x1,y1),(x2,y2)` spans into
    /// typed [`BoundingBox`]es. Distinct from the generic
    /// [`InferenceEngine::generate`] + `InferenceResult.bounding_boxes`
    /// path so callers can express "I want boxes" as a first-class
    /// intent — which also lets the router prefer models that declare
    /// the `Grounding` capability.
    pub async fn ground(&self, req: GroundRequest) -> Result<GroundResult, InferenceError> {
        let model_name = match req.model.clone() {
            Some(m) => m,
            None => self
                .preferred_model_for_capability(ModelCapability::Grounding)
                .map(str::to_string)
                .ok_or_else(|| {
                    InferenceError::InferenceFailed(
                        "no grounding-capable model available — pull a Qwen2.5-VL model first"
                            .into(),
                    )
                })?,
        };

        let gen_req = GenerateRequest {
            prompt: req.prompt.clone(),
            model: Some(model_name),
            params: GenerateParams::default(),
            context: None,
            tools: None,
            images: Some(vec![req.image.clone()]),
            messages: None,
            cache_control: false,
            response_format: None,
            intent: None,
        };
        let result = self.generate_tracked(gen_req).await?;
        Ok(GroundResult {
            boxes: result.bounding_boxes,
            raw_text: result.text,
            model_used: Some(result.model_used),
        })
    }

    /// Classify text against candidate labels.
    /// When `req.model` is None, routes to the smallest available model.
    pub async fn classify(
        &self,
        req: ClassifyRequest,
    ) -> Result<Vec<ClassifyResult>, InferenceError> {
        let model = match req
            .model
            .clone()
            .or_else(|| self.preferred_model_for_capability(ModelCapability::Classify).map(str::to_string))
        {
            Some(m) => m,
            None => {
                let m = self.router.route_small(&self.registry);
                debug!(model = %m, "auto-routed classify request");
                m
            }
        };

        // On Apple Silicon, route through the main generate path (which uses MLX)
        // instead of the Candle backend directly.
        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            return self.classify_via_generate(req, &model).await;
        }

        #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
        {
            self.ensure_backend(&model).await?;
            let mut write = self.backend.write().await;
            let backend = write.as_mut().unwrap();
            tasks::classify::classify(backend, req).await
        }
    }

    /// Classify by routing through the main generate path (uses MLX on Apple Silicon).
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    async fn classify_via_generate(
        &self,
        req: ClassifyRequest,
        model: &str,
    ) -> Result<Vec<ClassifyResult>, InferenceError> {
        let labels_str = req
            .labels
            .iter()
            .enumerate()
            .map(|(i, l)| format!("{}. {}", i + 1, l))
            .collect::<Vec<_>>()
            .join("\n");

        let prompt = format!(
            "Classify the following text into one of these categories:\n\
             {labels_str}\n\n\
             Text: {}\n\n\
             Respond with ONLY the category name, nothing else.",
            req.text
        );

        let gen_req = GenerateRequest {
            prompt,
            model: Some(model.to_string()),
            params: tasks::generate::GenerateParams {
                temperature: 0.0,
                max_tokens: 32,
                // Classification is latency-sensitive and single-label;
                // force the fast no-thinking path even on Qwen3.
                thinking: tasks::generate::ThinkingMode::Off,
                ..Default::default()
            },
            context: None,
            tools: None,
            images: None,
            messages: None,
            cache_control: false,
            response_format: None,
            intent: None,
        };

        let response = self.generate(gen_req).await?;
        let response_lower = response.trim().to_lowercase();

        let mut results: Vec<ClassifyResult> = req
            .labels
            .iter()
            .map(|label| {
                let label_lower = label.to_lowercase();
                let score = if response_lower == label_lower {
                    1.0
                } else if response_lower.contains(&label_lower) {
                    0.8
                } else {
                    let label_words: Vec<&str> = label_lower.split_whitespace().collect();
                    let matches = label_words
                        .iter()
                        .filter(|w| response_lower.contains(**w))
                        .count();
                    if label_words.is_empty() {
                        0.0
                    } else {
                        0.5 * (matches as f64 / label_words.len() as f64)
                    }
                };
                ClassifyResult {
                    label: label.clone(),
                    score,
                }
            })
            .collect();

        results.sort_by(|a, b| {
            b.score
                .partial_cmp(&a.score)
                .unwrap_or(std::cmp::Ordering::Equal)
        });

        let total: f64 = results.iter().map(|r| r.score).sum();
        if total > 0.0 {
            for r in &mut results {
                r.score /= total;
            }
        }

        Ok(results)
    }

    /// Transcribe an audio file using the best available STT model.
    pub async fn transcribe(
        &self,
        req: TranscribeRequest,
    ) -> Result<TranscribeResult, InferenceError> {
        let candidates =
            self.speech_candidates(ModelCapability::SpeechToText, req.model.as_deref())?;
        let mut last_error = None;

        for schema in candidates {
            let result = match &schema.source {
                ModelSource::Mlx { .. } => self.transcribe_local_mlx(&schema, &req).await,
                ModelSource::Proprietary { provider, .. } if provider == "elevenlabs" => {
                    self.transcribe_elevenlabs(&schema, &req).await
                }
                _ => Err(InferenceError::InferenceFailed(format!(
                    "speech-to-text not implemented for model source: {}",
                    schema.id
                ))),
            };

            match result {
                Ok(result) => return Ok(result),
                Err(err) => last_error = Some(err),
            }
        }

        Err(last_error.unwrap_or_else(|| {
            InferenceError::InferenceFailed("no speech-to-text models available".into())
        }))
    }

    /// Synthesize speech using the best available TTS model.
    pub async fn synthesize(
        &self,
        req: SynthesizeRequest,
    ) -> Result<SynthesizeResult, InferenceError> {
        let candidates =
            self.speech_candidates(ModelCapability::TextToSpeech, req.model.as_deref())?;
        let mut last_error = None;

        for schema in candidates {
            let result = match &schema.source {
                ModelSource::Mlx { .. } => self.synthesize_local_mlx(&schema, &req).await,
                ModelSource::Proprietary { provider, .. } if provider == "elevenlabs" => {
                    self.synthesize_elevenlabs(&schema, &req).await
                }
                _ => Err(InferenceError::InferenceFailed(format!(
                    "text-to-speech not implemented for model source: {}",
                    schema.id
                ))),
            };

            match result {
                Ok(result) => return Ok(result),
                Err(err) => last_error = Some(err),
            }
        }

        Err(last_error.unwrap_or_else(|| {
            InferenceError::InferenceFailed("no text-to-speech models available".into())
        }))
    }

    /// Generate an image using the best available local MLX image model.
    pub async fn generate_image(
        &self,
        req: GenerateImageRequest,
    ) -> Result<GenerateImageResult, InferenceError> {
        // Dispatch: the native Rust MLX Flux backend now reaches prompt-faithful
        // parity at 512×512×4-step against mflux (verified via the parity harness
        // in tools/parity/ref_flux.py + diff_flux_blocks.py). Default to native;
        // `CAR_IMAGE_BACKEND=external` still routes to the Python mflux fallback
        // for A/B comparison, `CAR_IMAGE_BACKEND=auto` reads env preference then
        // auto-detects.
        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            use crate::backend::external_flux;
            let backend = std::env::var("CAR_IMAGE_BACKEND")
                .unwrap_or_else(|_| "native".to_string());
            let use_external = match backend.as_str() {
                "external" => true,
                "native" => false,
                // `auto` honors external only when the subprocess CLI is on PATH.
                _ => external_flux::is_available() && backend == "auto-external",
            };
            if use_external {
                tracing::info!(
                    "routing image generation to external mflux \
                     (set CAR_IMAGE_BACKEND=native to use the Rust port)"
                );
                let mut req = req;
                req.model = self.resolve_external_hf_repo(
                    req.model.as_deref(),
                    ModelCapability::ImageGeneration,
                );
                return external_flux::generate_image(&req);
            }
            tracing::info!("using native Rust MLX Flux backend");
        }

        let candidates =
            self.media_generation_candidates(ModelCapability::ImageGeneration, req.model.as_deref())?;
        let mut last_error = None;

        for schema in candidates {
            let result = match &schema.source {
                #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
                ModelSource::Mlx { .. } => self.generate_image_native_mlx(&schema, &req).await,
                _ => Err(InferenceError::InferenceFailed(format!(
                    "image generation not implemented for model source: {}",
                    schema.id
                ))),
            };

            match result {
                Ok(result) => return Ok(result),
                Err(err) => last_error = Some(err),
            }
        }

        Err(last_error.unwrap_or_else(|| {
            InferenceError::InferenceFailed("no image generation models available".into())
        }))
    }

    /// Generate one or more variants in a single call.
    ///
    /// Returns `req.variant_count` results (defaulting to 1). The
    /// current MLX Flux backend doesn't support native batching, so
    /// this loops over `generate_image` with the seed advanced per
    /// variant for visual diversity. A future hosted backend
    /// (gpt-image-2, Replicate) can short-circuit this with one
    /// network call producing N coherent images.
    ///
    /// Per-variant errors abort the batch — there's no partial-
    /// success semantics today. Callers needing more lenient
    /// behaviour should call `generate_image` directly in their own
    /// loop.
    ///
    /// Closes #110.
    pub async fn generate_image_batch(
        &self,
        req: GenerateImageRequest,
    ) -> Result<Vec<GenerateImageResult>, InferenceError> {
        let count = req.variant_count.unwrap_or(1).max(1);
        if count == 1 {
            return self.generate_image(req).await.map(|r| vec![r]);
        }
        let base_seed = req.seed.unwrap_or(0);
        let mut results = Vec::with_capacity(count as usize);
        for i in 0..count {
            // Vary the seed per variant so backends that key prompt
            // → output deterministically actually produce different
            // images. Callers wanting reproducible single-seed
            // variants override `seed` per call themselves.
            let mut variant_req = req.clone();
            variant_req.seed = Some(base_seed.wrapping_add(i as u64));
            // Suppress variant_count on the inner call to avoid
            // recursion — generate_image ignores the field today,
            // but this also documents intent.
            variant_req.variant_count = Some(1);
            results.push(self.generate_image(variant_req).await?);
        }
        Ok(results)
    }

    /// Native MLX Flux image generation (no Python shelling).
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    async fn generate_image_native_mlx(
        &self,
        schema: &ModelSchema,
        req: &GenerateImageRequest,
    ) -> Result<GenerateImageResult, InferenceError> {
        let model_dir = self.unified_registry.ensure_local(&schema.id).await?;
        let size = backend_cache::estimate_model_size(&model_dir);
        let cache = Arc::clone(&self.flux_cache);
        let key = schema.id.clone();
        let handle = cache.get_or_load(&key, size, || {
            backend::mlx_flux::FluxBackend::load(&model_dir)
        })?;
        // Run the synchronous, GPU-bound inference on a blocking worker
        // so we don't stall the tokio runtime. The per-model mutex
        // serializes concurrent requests for the same schema.
        let req = req.clone();
        tokio::task::spawn_blocking(move || -> Result<GenerateImageResult, InferenceError> {
            let mut guard = handle.lock().map_err(|_| {
                InferenceError::InferenceFailed("flux backend mutex poisoned".into())
            })?;
            guard.generate(&req)
        })
        .await
        .map_err(|e| InferenceError::InferenceFailed(format!("flux task join: {e}")))?
    }

    /// Native MLX LTX-2.3 video generation (no Python shelling).
    #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
    async fn generate_video_native_mlx(
        &self,
        schema: &ModelSchema,
        req: &GenerateVideoRequest,
    ) -> Result<GenerateVideoResult, InferenceError> {
        let model_dir = self.unified_registry.ensure_local(&schema.id).await?;
        let size = backend_cache::estimate_model_size(&model_dir);
        let cache = Arc::clone(&self.ltx_cache);
        let key = schema.id.clone();
        let handle = cache.get_or_load(&key, size, || {
            backend::mlx_ltx::LtxBackend::load(&model_dir)
        })?;
        let req = req.clone();
        tokio::task::spawn_blocking(move || -> Result<GenerateVideoResult, InferenceError> {
            let mut guard = handle.lock().map_err(|_| {
                InferenceError::InferenceFailed("ltx backend mutex poisoned".into())
            })?;
            guard.generate(&req)
        })
        .await
        .map_err(|e| InferenceError::InferenceFailed(format!("ltx task join: {e}")))?
    }

    /// Generate a video using the best available local MLX video model.
    pub async fn generate_video(
        &self,
        req: GenerateVideoRequest,
    ) -> Result<GenerateVideoResult, InferenceError> {
        // Validate the request shape up front so callers get a clean
        // error rather than a backend failure deep in the stack.
        if let Err(msg) = req.validate() {
            return Err(InferenceError::InferenceFailed(format!(
                "invalid GenerateVideoRequest: {}",
                msg
            )));
        }
        // Backend selection (temporary — until the native Rust MLX port
        // reaches parity with the upstream Python port):
        //   CAR_VIDEO_BACKEND=external  → call the `ltx-2-mlx` Python CLI.
        //   CAR_VIDEO_BACKEND=native    → force the native Rust MLX path.
        //   <unset> / "auto" (default)  → external when the CLI is on PATH,
        //                                  else fall back to native.
        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            use crate::backend::external_ltx;
            // Default flipped to `native` now that the Rust port reaches
            // quality parity with upstream ltx-2-mlx (matching prompt
            // produces a matching subject; see the cascade of fixes in
            // #40 / #45). `external` still honored for A/B comparison.
            let backend = std::env::var("CAR_VIDEO_BACKEND")
                .unwrap_or_else(|_| "native".to_string());
            let use_external = match backend.as_str() {
                "external" => true,
                "native" => false,
                // `auto-external` is an opt-in to the old behavior: use
                // the Python CLI if it happens to be on PATH, else fall
                // back to native.
                "auto-external" => external_ltx::is_available(),
                _ => false,
            };
            if use_external {
                tracing::info!(
                    "routing video generation to external ltx-2-mlx \
                     (set CAR_VIDEO_BACKEND=native to opt into the in-progress Rust port)"
                );
                let mut req = req;
                req.model = self.resolve_external_hf_repo(
                    req.model.as_deref(),
                    ModelCapability::VideoGeneration,
                );
                return external_ltx::generate_video(&req);
            }
            tracing::info!(
                "using native Rust MLX video backend — Gemma left-pad bug was \
                 just fixed; verify output quality before making this the default"
            );
        }

        let candidates =
            self.media_generation_candidates(ModelCapability::VideoGeneration, req.model.as_deref())?;
        let mut last_error = None;

        for schema in candidates {
            let result = match &schema.source {
                #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
                ModelSource::Mlx { .. } => self.generate_video_native_mlx(&schema, &req).await,
                _ => Err(InferenceError::InferenceFailed(format!(
                    "video generation not implemented for model source: {}",
                    schema.id
                ))),
            };

            match result {
                Ok(result) => return Ok(result),
                Err(err) => last_error = Some(err),
            }
        }

        Err(last_error.unwrap_or_else(|| {
            InferenceError::InferenceFailed("no video generation models available".into())
        }))
    }

    /// List all known models and their status (new registry).
    pub fn list_models_unified(&self) -> Vec<ModelInfo> {
        self.unified_registry
            .list()
            .iter()
            .map(|m| ModelInfo::from(*m))
            .collect()
    }

    /// List all known models and their download status (legacy).
    /// List all model schemas from the unified registry (full metadata).
    pub fn list_schemas(&self) -> Vec<ModelSchema> {
        self.unified_registry.list().into_iter().cloned().collect()
    }

    pub fn list_models(&self) -> Vec<models::ModelInfo> {
        self.registry.list_models()
    }

    /// Download a model if not already present.
    pub async fn pull_model(&self, name: &str) -> Result<std::path::PathBuf, InferenceError> {
        let schema = self
            .unified_registry
            .find_by_name(name)
            .or_else(|| self.unified_registry.get(name))
            .ok_or_else(|| InferenceError::ModelNotFound(name.to_string()))?;
        self.unified_registry.ensure_local(&schema.id).await
    }

    /// Remove a downloaded model.
    pub fn remove_model(&self, name: &str) -> Result<(), InferenceError> {
        let schema = self
            .unified_registry
            .get(name)
            .or_else(|| {
                self.unified_registry
                    .list()
                    .into_iter()
                    .find(|schema| schema.name.eq_ignore_ascii_case(name))
            })
            .or_else(|| self.unified_registry.find_by_name(name))
            .ok_or_else(|| InferenceError::ModelNotFound(name.to_string()))?;
        let model_dir = self.unified_registry.models_dir().join(&schema.name);
        if model_dir.exists() {
            std::fs::remove_dir_all(&model_dir)?;
        }
        match &schema.source {
            ModelSource::Mlx { hf_repo, .. } => {
                remove_huggingface_repo_cache(hf_repo)?;
            }
            ModelSource::Local {
                hf_repo,
                tokenizer_repo,
                ..
            } => {
                remove_huggingface_repo_cache(hf_repo)?;
                remove_huggingface_repo_cache(tokenizer_repo)?;
            }
            _ => {}
        }
        Ok(())
    }

    /// Register a model in the unified registry.
    pub fn register_model(&mut self, schema: ModelSchema) {
        self.unified_registry.register(schema);
    }

    /// Discover generic MLX models from a running vLLM-MLX server and register them.
    /// Returns the number of discovered models added or refreshed in the registry.
    pub async fn discover_vllm_mlx_models(&mut self) -> usize {
        let config = vllm_mlx::VllmMlxConfig::default();
        if !config.auto_discover {
            return 0;
        }
        vllm_mlx::discover_and_register(&config, &mut self.unified_registry).await
    }

    /// Get outcome tracker for external use (e.g., memgine integration).
    pub fn outcome_tracker(&self) -> Arc<RwLock<OutcomeTracker>> {
        self.outcome_tracker.clone()
    }

    /// Auto-save outcomes and key pool stats silently (called after every inference call).
    async fn auto_save_outcomes(&self) {
        if let Err(e) = self.save_outcomes().await {
            tracing::debug!("auto-save outcomes failed: {}", e);
        }
        if let Err(e) = self.save_key_pool_stats().await {
            tracing::debug!("auto-save key pool stats failed: {}", e);
        }
    }

    /// Persist outcome profiles to disk for cross-session learning (#13).
    pub async fn save_outcomes(&self) -> Result<(), std::io::Error> {
        let tracker = self.outcome_tracker.read().await;
        let path = self.config.models_dir.join("outcome_profiles.json");
        tracker.save_to_file(&path)
    }

    /// Persist key pool stats to disk.
    pub async fn save_key_pool_stats(&self) -> Result<(), std::io::Error> {
        let path = self.config.models_dir.join("key_pool_stats.json");
        self.remote_backend.key_pool.save_stats(&path).await
    }

    /// Get key pool stats for all endpoints.
    pub async fn key_pool_stats(
        &self,
    ) -> std::collections::HashMap<String, Vec<key_pool::KeyStats>> {
        self.remote_backend.key_pool.all_stats().await
    }

    /// Export model performance profiles for persistence.
    pub async fn export_profiles(&self) -> Vec<ModelProfile> {
        let tracker = self.outcome_tracker.read().await;
        tracker.export_profiles()
    }

    /// Import model performance profiles (from persistence).
    pub async fn import_profiles(&self, profiles: Vec<ModelProfile>) {
        let mut tracker = self.outcome_tracker.write().await;
        tracker.import_profiles(profiles);
    }

    /// Ensure the managed local speech runtime exists and return its root directory.
    /// On Apple Silicon, speech uses native MLX backends and no Python runtime is needed.
    pub async fn prepare_speech_runtime(&self) -> Result<PathBuf, InferenceError> {
        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            // Native MLX speech — no Python runtime needed.
            Ok(self.config.models_dir.clone())
        }
        #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
        {
            Ok(self.ensure_speech_runtime().await?.root)
        }
    }

    /// Override speech routing preferences for the current engine instance.
    pub fn set_speech_policy(&mut self, policy: SpeechPolicy) {
        self.speech_policy = policy;
    }

    pub fn set_routing_config(&mut self, config: RoutingConfig) {
        self.adaptive_router.set_config(config);
    }

    /// Download the curated local speech model set into the shared Hugging Face cache.
    pub async fn install_curated_speech(
        &mut self,
    ) -> Result<Vec<SpeechInstallReport>, InferenceError> {
        let _runtime_root = self.prepare_speech_runtime().await?;
        let schemas = self.list_schemas();
        let mut repos = Vec::new();
        for schema in &schemas {
            if !schema.is_mlx() || !schema.tags.iter().any(|tag| tag == "speech") {
                continue;
            }
            if let ModelSource::Mlx { hf_repo, .. } = &schema.source {
                if !repos.iter().any(|existing: &String| existing == hf_repo) {
                    repos.push(hf_repo.clone());
                }
            }
        }

        let mut installed = Vec::new();
        for repo in repos {
            let (snapshot_path, files_downloaded) = download_hf_repo_snapshot(&repo).await?;
            let name = schemas
                .iter()
                .find(|schema| {
                    matches!(&schema.source, ModelSource::Mlx { hf_repo, .. } if hf_repo == &repo)
                })
                .map(|schema| schema.name.clone())
                .unwrap_or_else(|| repo.clone());
            installed.push(SpeechInstallReport {
                name,
                hf_repo: repo,
                snapshot_path,
                files_downloaded,
            });
        }

        self.unified_registry.refresh_availability();
        Ok(installed)
    }


    /// Report speech runtime, model cache, and remote-provider health.
    pub fn speech_health(&self) -> SpeechHealthReport {
        let local_stt_default = self
            .speech_health_default_name(ModelCapability::SpeechToText, true, false);
        let local_tts_default = self
            .speech_health_default_name(ModelCapability::TextToSpeech, true, false);
        let remote_stt_default = self
            .speech_health_default_name(ModelCapability::SpeechToText, false, true);
        let remote_tts_default = self
            .speech_health_default_name(ModelCapability::TextToSpeech, false, true);

        let mut local_models = Vec::new();
        let mut remote_models = Vec::new();
        for schema in self.list_schemas() {
            let capability = if schema.has_capability(ModelCapability::SpeechToText) {
                Some(ModelCapability::SpeechToText)
            } else if schema.has_capability(ModelCapability::TextToSpeech) {
                Some(ModelCapability::TextToSpeech)
            } else {
                None
            };
            let Some(capability) = capability else {
                continue;
            };

            let selected_by_default = local_stt_default
                .as_ref()
                .is_some_and(|name| name == &schema.name)
                || local_tts_default
                    .as_ref()
                    .is_some_and(|name| name == &schema.name)
                || remote_stt_default
                    .as_ref()
                    .is_some_and(|name| name == &schema.name)
                || remote_tts_default
                    .as_ref()
                    .is_some_and(|name| name == &schema.name);

            let health = SpeechModelHealth {
                id: schema.id.clone(),
                name: schema.name.clone(),
                provider: schema.provider.clone(),
                capability,
                is_local: schema.is_local(),
                available: schema.available,
                cached: speech_model_cached(&schema),
                selected_by_default,
                source: speech_model_source_label(&schema),
            };
            if schema.is_local() {
                local_models.push(health);
            } else {
                remote_models.push(health);
            }
        }

        // On Apple Silicon, native MLX backends are used instead of a Python runtime.
        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        let runtime = SpeechRuntimeHealth {
            root: self.config.models_dir.clone(),
            installed: true,
            python: PathBuf::new(),
            stt_command: PathBuf::new(),
            tts_command: PathBuf::new(),
            configured_python: None,
            detected_python: None,
        };

        #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
        let runtime = {
            let rt = SpeechRuntime::new(speech_runtime_root_from_models_dir(&self.config.models_dir));
            SpeechRuntimeHealth {
                root: rt.root.clone(),
                installed: rt.is_ready(),
                python: rt.python.clone(),
                stt_command: rt.stt_program.clone(),
                tts_command: rt.tts_program.clone(),
                configured_python: std::env::var("CAR_SPEECH_PYTHON")
                    .ok()
                    .filter(|value| !value.trim().is_empty()),
                detected_python: detect_speech_python(),
            }
        };

        SpeechHealthReport {
            runtime,
            local_models,
            remote_models,
            elevenlabs_configured: car_secrets::resolve_env_or_keychain("ELEVENLABS_API_KEY").is_some(),
            prefer_local: self.speech_policy.prefer_local,
            allow_remote_fallback: self.speech_policy.allow_remote_fallback,
            preferred_local_stt: self.speech_policy.preferred_local_stt.clone(),
            preferred_local_tts: self.speech_policy.preferred_local_tts.clone(),
            preferred_remote_stt: self.speech_policy.preferred_remote_stt.clone(),
            preferred_remote_tts: self.speech_policy.preferred_remote_tts.clone(),
            local_stt_default,
            local_tts_default,
            remote_stt_default,
            remote_tts_default,
        }
    }

    /// Report the current model catalog, configured defaults, capability coverage,
    /// and speech runtime/provider health in one place.
    pub async fn model_health(&self) -> ModelHealthReport {
        let schemas = self.list_schemas();
        let total_models = schemas.len();
        let available_models = schemas.iter().filter(|schema| schema.available).count();
        let local_models = schemas.iter().filter(|schema| schema.is_local()).count();
        let remote_models = total_models.saturating_sub(local_models);

        let defaults = vec![
            self.model_default_health(
                ModelCapability::Generate,
                self.preferred_model_for_capability(ModelCapability::Generate)
                    .unwrap_or(&self.config.generation_model),
            ),
            self.model_default_health(
                ModelCapability::Embed,
                self.preferred_model_for_capability(ModelCapability::Embed)
                    .unwrap_or(&self.config.embedding_model),
            ),
            self.model_default_health(
                ModelCapability::Classify,
                self.preferred_model_for_capability(ModelCapability::Classify)
                    .unwrap_or(&self.config.classification_model),
            ),
        ];

        let mut providers = std::collections::BTreeMap::new();
        for schema in &schemas {
            let entry = providers
                .entry(schema.provider.clone())
                .or_insert_with(|| ProviderAccumulator {
                    configured: false,
                    local_models: 0,
                    remote_models: 0,
                    available_models: 0,
                    capabilities: std::collections::HashSet::new(),
                });

            entry.configured |= model_source_configured(schema);
            if schema.is_local() {
                entry.local_models += 1;
            } else {
                entry.remote_models += 1;
            }
            if schema.available {
                entry.available_models += 1;
            }
            for capability in &schema.capabilities {
                entry.capabilities.insert(*capability);
            }
        }

        let providers = providers
            .into_iter()
            .map(|(provider, acc)| ModelProviderHealth {
                provider,
                configured: acc.configured,
                local_models: acc.local_models,
                remote_models: acc.remote_models,
                available_models: acc.available_models,
                capabilities: sort_capabilities(acc.capabilities.into_iter().collect()),
            })
            .collect();

        let capabilities = all_model_capabilities()
            .into_iter()
            .map(|capability| {
                let relevant: Vec<&ModelSchema> = schemas
                    .iter()
                    .filter(|schema| schema.has_capability(capability))
                    .collect();
                let available: Vec<&ModelSchema> = relevant
                    .iter()
                    .copied()
                    .filter(|schema| schema.available)
                    .collect();
                ModelCapabilityHealth {
                    capability,
                    total_models: relevant.len(),
                    available_models: available.len(),
                    local_available_models: available
                        .iter()
                        .filter(|schema| schema.is_local())
                        .count(),
                    remote_available_models: available
                        .iter()
                        .filter(|schema| !schema.is_local())
                        .count(),
                }
            })
            .collect();

        let routing = self.routing_scenarios().await;
        let routing_config = self.adaptive_router.config().clone();
        let benchmark_priors = load_benchmark_prior_health(&self.config.models_dir, &schemas);

        ModelHealthReport {
            total_models,
            available_models,
            local_models,
            remote_models,
            defaults,
            providers,
            capabilities,
            routing_prefer_local: routing_config.prefer_local,
            routing_quality_first_cold_start: routing_config.quality_first_cold_start,
            routing_min_observations: routing_config.min_observations,
            routing_bootstrap_min_task_observations: routing_config.bootstrap_min_task_observations,
            routing_bootstrap_quality_floor: routing_config.bootstrap_quality_floor,
            routing_quality_weight: routing_config.quality_weight,
            routing_latency_weight: routing_config.latency_weight,
            routing_cost_weight: routing_config.cost_weight,
            routing_scenarios: routing,
            benchmark_priors,
            speech: self.speech_health(),
        }
    }

    async fn routing_scenarios(&self) -> Vec<RoutingScenarioHealth> {
        let tracker = self.outcome_tracker.read().await;
        let config = self.adaptive_router.config().clone();
        let scenarios = [
            (
                "interactive_text",
                "Summarize the benefits of local-first AI routing in two sentences.",
                "text",
                RoutingWorkload::Interactive,
                false,
                false,
            ),
            (
                "background_code",
                "Write a Python function named fibonacci(n) that returns the nth Fibonacci number.",
                "code",
                RoutingWorkload::Background,
                false,
                false,
            ),
            (
                "interactive_tool_use",
                "Use the provided weather tool to get the weather for Boston.",
                "tool_use",
                RoutingWorkload::Interactive,
                true,
                false,
            ),
            (
                "interactive_vision",
                "What is in this image? Answer in one word.",
                "vision",
                RoutingWorkload::Interactive,
                false,
                true,
            ),
        ];

        scenarios
            .into_iter()
            .map(|(name, prompt, task_family, workload, has_tools, has_vision)| {
                let decision = self.adaptive_router.route_context_aware(
                    prompt,
                    0,
                    &self.unified_registry,
                    &tracker,
                    has_tools,
                    has_vision,
                    workload,
                );
                let quality_first_cold_start = if has_tools || has_vision {
                    config.quality_first_cold_start
                } else if task_family == "code" && matches!(workload, RoutingWorkload::Background) {
                    false
                } else {
                    config.quality_first_cold_start
                };
                RoutingScenarioHealth {
                    name: name.to_string(),
                    task_family: task_family.to_string(),
                    workload,
                    has_tools,
                    has_vision,
                    prefer_local: if task_family == "speech" {
                        self.speech_policy.prefer_local
                    } else {
                        config.prefer_local
                    },
                    quality_first_cold_start,
                    bootstrap_min_task_observations: config.bootstrap_min_task_observations,
                    bootstrap_quality_floor: config.bootstrap_quality_floor,
                    model_id: decision.model_id,
                    model_name: decision.model_name,
                    reason: decision.reason,
                    strategy: decision.strategy,
                }
            })
            .collect()
    }

    /// Run a real speech smoke test through the configured local and/or remote paths.
    pub async fn smoke_test_speech(
        &self,
        local: bool,
        remote: bool,
    ) -> Result<SpeechSmokeReport, InferenceError> {
        let mut report = SpeechSmokeReport::default();

        if local {
            let tts = self
                .preferred_speech_schema(ModelCapability::TextToSpeech, true, false)
                .ok_or_else(|| {
                    InferenceError::InferenceFailed(
                        "no local text-to-speech model available".into(),
                    )
                })?;
            let stt = self
                .preferred_speech_schema(ModelCapability::SpeechToText, true, false)
                .ok_or_else(|| {
                    InferenceError::InferenceFailed(
                        "no local speech-to-text model available".into(),
                    )
                })?;
            report.local = Some(
                self.run_speech_smoke_path("local", &tts, &stt, "Testing CAR local speech path.")
                    .await?,
            );
        } else {
            report.skipped.push("local".to_string());
        }

        if remote {
            let tts = self
                .preferred_speech_schema(ModelCapability::TextToSpeech, false, true)
                .ok_or_else(|| {
                    InferenceError::InferenceFailed(
                        "no remote text-to-speech model available".into(),
                    )
                })?;
            let stt = self
                .preferred_speech_schema(ModelCapability::SpeechToText, false, true)
                .ok_or_else(|| {
                    InferenceError::InferenceFailed(
                        "no remote speech-to-text model available".into(),
                    )
                })?;
            report.remote = Some(
                self.run_speech_smoke_path("remote", &tts, &stt, "Testing CAR remote speech path.")
                    .await?,
            );
        } else {
            report.skipped.push("remote".to_string());
        }

        Ok(report)
    }

    fn speech_candidates(
        &self,
        capability: ModelCapability,
        explicit: Option<&str>,
    ) -> Result<Vec<ModelSchema>, InferenceError> {
        if let Some(model) = explicit {
            let schema = self
                .unified_registry
                .get(model)
                .or_else(|| self.unified_registry.find_by_name(model))
                .cloned()
                .ok_or_else(|| InferenceError::ModelNotFound(model.to_string()))?;
            if !schema.has_capability(capability) {
                return Err(InferenceError::InferenceFailed(format!(
                    "model {} does not support {:?}",
                    schema.name, capability
                )));
            }
            return Ok(vec![schema]);
        }

        let mut candidates: Vec<ModelSchema> = self
            .unified_registry
            .query(&ModelFilter {
                capabilities: vec![capability],
                ..Default::default()
            })
            .into_iter()
            .cloned()
            .collect();

        if candidates.is_empty() {
            return Err(InferenceError::InferenceFailed(format!(
                "no models registered for capability {:?}",
                capability
            )));
        }

        candidates.sort_by_key(|model| self.speech_sort_key(capability, model));
        if !self.speech_policy.allow_remote_fallback && candidates.iter().any(|model| model.is_local()) {
            candidates.retain(|model| model.is_local());
        }

        Ok(candidates)
    }

    /// Resolve a car-canonical model id (e.g. `mlx/flux-1-lite-8b:q4`) to the
    /// HuggingFace repo (`mlx-community/Flux-1.lite-8B-MLX-Q4`) that the
    /// external Python CLIs expect. Falls back to the input if no schema
    /// matches or the schema is not MLX-sourced.
    fn resolve_external_hf_repo(
        &self,
        explicit: Option<&str>,
        capability: ModelCapability,
    ) -> Option<String> {
        let id = explicit?;
        let schema = self
            .unified_registry
            .get(id)
            .or_else(|| self.unified_registry.find_by_name(id))?;
        if !schema.has_capability(capability) {
            return Some(id.to_string());
        }
        if let ModelSource::Mlx { hf_repo, .. } = &schema.source {
            return Some(hf_repo.clone());
        }
        Some(id.to_string())
    }

    fn media_generation_candidates(
        &self,
        capability: ModelCapability,
        explicit: Option<&str>,
    ) -> Result<Vec<ModelSchema>, InferenceError> {
        if let Some(model) = explicit {
            let schema = self
                .unified_registry
                .get(model)
                .or_else(|| self.unified_registry.find_by_name(model))
                .cloned()
                .ok_or_else(|| InferenceError::ModelNotFound(model.to_string()))?;
            if !schema.has_capability(capability) {
                return Err(InferenceError::InferenceFailed(format!(
                    "model {} does not support {:?}",
                    schema.name, capability
                )));
            }
            return Ok(vec![schema]);
        }

        let mut candidates: Vec<ModelSchema> = self
            .unified_registry
            .query(&ModelFilter {
                capabilities: vec![capability],
                local_only: true,
                ..Default::default()
            })
            .into_iter()
            .cloned()
            .collect();
        candidates.sort_by_key(|schema| (!schema.available, schema.size_mb()));
        if candidates.is_empty() {
            return Err(InferenceError::InferenceFailed(format!(
                "no models registered for capability {:?}",
                capability
            )));
        }
        Ok(candidates)
    }

    fn preferred_speech_schema(
        &self,
        capability: ModelCapability,
        local_only: bool,
        remote_only: bool,
    ) -> Option<ModelSchema> {
        let available_only = remote_only;
        let mut candidates: Vec<ModelSchema> = self
            .unified_registry
            .query(&ModelFilter {
                capabilities: vec![capability],
                available_only,
                ..Default::default()
            })
            .into_iter()
            .filter(|schema| (!local_only || schema.is_local()) && (!remote_only || schema.is_remote()))
            .cloned()
            .collect();
        candidates.sort_by_key(|model| self.speech_sort_key(capability, model));
        candidates.into_iter().next()
    }

    fn speech_health_default_name(
        &self,
        capability: ModelCapability,
        local_only: bool,
        remote_only: bool,
    ) -> Option<String> {
        let preferred = match capability {
            ModelCapability::SpeechToText if local_only => {
                self.speech_policy.preferred_local_stt.as_ref()
            }
            ModelCapability::SpeechToText if remote_only => {
                self.speech_policy.preferred_remote_stt.as_ref()
            }
            ModelCapability::TextToSpeech if local_only => {
                self.speech_policy.preferred_local_tts.as_ref()
            }
            ModelCapability::TextToSpeech if remote_only => {
                self.speech_policy.preferred_remote_tts.as_ref()
            }
            _ => None,
        };

        preferred
            .filter(|name| {
                self.unified_registry.list().iter().any(|schema| {
                    schema.name == **name
                        && schema.has_capability(capability)
                        && (!local_only || schema.is_local())
                        && (!remote_only || schema.is_remote())
                })
            })
            .cloned()
            .or_else(|| {
                self.preferred_speech_schema(capability, local_only, remote_only)
                    .map(|schema| schema.name)
            })
    }

    fn model_default_health(
        &self,
        capability: ModelCapability,
        configured_model: &str,
    ) -> ModelDefaultHealth {
        let schema = self
            .unified_registry
            .find_by_name(configured_model)
            .or_else(|| self.unified_registry.get(configured_model));

        ModelDefaultHealth {
            capability,
            configured_model: configured_model.to_string(),
            available: schema.is_some_and(|model| model.available),
            is_local: schema.is_some_and(ModelSchema::is_local),
            provider: schema.map(|model| model.provider.clone()),
        }
    }

    fn speech_sort_key(
        &self,
        capability: ModelCapability,
        model: &ModelSchema,
    ) -> (u8, u8, u8, u8, u64, u64) {
        let policy_preference = match capability {
            ModelCapability::SpeechToText if model.is_local() => self
                .speech_policy
                .preferred_local_stt
                .as_ref(),
            ModelCapability::SpeechToText => self.speech_policy.preferred_remote_stt.as_ref(),
            ModelCapability::TextToSpeech if model.is_local() => self
                .speech_policy
                .preferred_local_tts
                .as_ref(),
            ModelCapability::TextToSpeech => self.speech_policy.preferred_remote_tts.as_ref(),
            _ => None,
        };
        let local_rank = if self.speech_policy.prefer_local {
            if model.is_local() {
                0
            } else {
                1
            }
        } else if model.is_remote() {
            0
        } else {
            1
        };
        let availability_rank = if model.available {
            0
        } else if model.is_local() {
            1
        } else {
            2
        };
        let policy_rank: u8 = if policy_preference.is_some_and(|preferred| preferred == &model.name) {
            0
        } else {
            1
        };
        let speech_rank = match capability {
            ModelCapability::TextToSpeech => {
                if model.name == "Qwen3-TTS-12Hz-1.7B-Base-5bit" {
                    0
                } else if model.name == "Kokoro-82M-bf16" {
                    1
                } else if model.name == "Kokoro-82M-6bit" {
                    2
                } else {
                    3
                }
            }
            ModelCapability::SpeechToText => {
                if model.name == "Parakeet-TDT-0.6B-v3-MLX" {
                    0
                } else {
                    1
                }
            }
            _ => 0,
        };
        let latency_rank = model.performance.latency_p50_ms.unwrap_or(u64::MAX);
        let size_rank = model.cost.size_mb.unwrap_or(u64::MAX);
        (
            local_rank,
            availability_rank,
            policy_rank,
            speech_rank,
            latency_rank,
            size_rank,
        )
    }

    async fn run_speech_smoke_path(
        &self,
        path: &str,
        tts: &ModelSchema,
        stt: &ModelSchema,
        text: &str,
    ) -> Result<SpeechSmokePathReport, InferenceError> {
        let work_dir = temp_work_dir(&format!("speech-smoke-{path}"))?;
        let audio_path = work_dir.join(format!("{path}.wav"));
        let synth = self
            .synthesize(SynthesizeRequest {
                text: text.to_string(),
                model: Some(tts.name.clone()),
                voice: default_speech_voice(tts),
                language: Some("en".to_string()),
                output_path: Some(audio_path.display().to_string()),
                ..SynthesizeRequest::default()
            })
            .await?;
        let transcript = self
            .transcribe(TranscribeRequest {
                audio_path: synth.audio_path.clone(),
                model: Some(stt.name.clone()),
                language: Some("en".to_string()),
                prompt: None,
                timestamps: false,
            })
            .await?;

        Ok(SpeechSmokePathReport {
            path: path.to_string(),
            tts_model: synth.model_used.unwrap_or_else(|| tts.name.clone()),
            stt_model: transcript.model_used.unwrap_or_else(|| stt.name.clone()),
            audio_path: PathBuf::from(synth.audio_path),
            transcript: transcript.text,
        })
    }

    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    async fn ensure_speech_runtime(&self) -> Result<SpeechRuntime, InferenceError> {
        let mut guard = self.speech_runtime.lock().await;
        if let Some(runtime) = guard.as_ref() {
            if runtime.is_ready() {
                return Ok(runtime.clone());
            }
        }

        let runtime = SpeechRuntime::new(speech_runtime_root_from_models_dir(
            &self.config.models_dir,
        ));
        if !runtime.is_ready() {
            bootstrap_speech_runtime(&runtime).await?;
        }
        if !runtime.is_ready() {
            return Err(InferenceError::InferenceFailed(format!(
                "managed speech runtime is not ready at {}",
                runtime.root.display()
            )));
        }

        *guard = Some(runtime.clone());
        Ok(runtime)
    }

    async fn transcribe_local_mlx(
        &self,
        schema: &ModelSchema,
        req: &TranscribeRequest,
    ) -> Result<TranscribeResult, InferenceError> {
        // Native MLX transcription via Parakeet backend (no Python shelling).
        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            let model_dir = self.unified_registry.ensure_local(&schema.id).await?;
            let parakeet = backend::mlx_parakeet::ParakeetBackend::load(&model_dir)?;
            // Only pay the word-grouping cost when the caller asked.
            let (text, words) = if req.timestamps {
                parakeet
                    .transcribe_detailed(Path::new(&req.audio_path))
                    .map_err(|e| InferenceError::InferenceFailed(format!("native STT: {e}")))?
            } else {
                let t = parakeet
                    .transcribe(Path::new(&req.audio_path))
                    .map_err(|e| InferenceError::InferenceFailed(format!("native STT: {e}")))?;
                (t, Vec::new())
            };
            return Ok(TranscribeResult {
                text,
                model_used: Some(schema.name.clone()),
                language: req.language.clone(),
                words,
            });
        }

        // Non-Apple-Silicon fallback: use Python speech runtime.
        #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
        {
            let runtime = self.ensure_speech_runtime().await?;
            let hf_repo = match &schema.source {
                ModelSource::Mlx { hf_repo, .. } => hf_repo.clone(),
                _ => unreachable!(),
            };
            let output_dir = temp_work_dir("stt")?;
            let output_prefix = output_dir.join("transcript");
            let mut args = vec![
                "--model".to_string(),
                hf_repo,
                "--audio".to_string(),
                req.audio_path.clone(),
                "--output-path".to_string(),
                output_prefix.display().to_string(),
                "--format".to_string(),
                "json".to_string(),
            ];
            if let Some(language) = &req.language {
                args.push("--language".to_string());
                args.push(normalize_lang_code(language));
            }
            if let Some(prompt) = &req.prompt {
                args.push("--context".to_string());
                args.push(prompt.clone());
            }
            if req.timestamps {
                args.push("--verbose".to_string());
            }

            let output = run_mlx_audio_command(&runtime, "stt.generate", &args).await?;
            let text = read_transcription_result(&output_prefix)?
                .or_else(|| extract_text_from_payload(&output.stdout))
                .ok_or_else(|| {
                    InferenceError::InferenceFailed(format!(
                        "mlx-audio transcription returned no text: {}",
                        output.stderr
                    ))
                })?;

            Ok(TranscribeResult {
                text,
                model_used: Some(schema.name.clone()),
                language: req.language.clone(),
                words: Vec::new(),
            })
        }
    }

    async fn synthesize_local_mlx(
        &self,
        schema: &ModelSchema,
        req: &SynthesizeRequest,
    ) -> Result<SynthesizeResult, InferenceError> {
        // Single entry-point check for Qwen3-TTS advanced controls.
        // Hoisted here so that a Kokoro → Kokoro-bf16 fallback chain
        // doesn't double-warn, and so strict callers get one clean
        // error instead of being lied to by partial success.
        let requested = req.requested_advanced_controls();
        let repo_supports_advanced = match &schema.source {
            ModelSource::Mlx { hf_repo, .. } => {
                hf_repo.to_ascii_lowercase().contains("qwen3-tts")
            }
            _ => false,
        };
        if !requested.is_empty() && !repo_supports_advanced {
            if req.strict_capabilities {
                return Err(InferenceError::InferenceFailed(format!(
                    "model {name} does not support Qwen3-TTS advanced controls {requested:?}; \
                     route to a Qwen3-TTS model or set strict_capabilities = false to degrade",
                    name = schema.name,
                )));
            }
            tracing::warn!(
                model = %schema.name,
                fields = ?requested,
                "Qwen3-TTS advanced controls set on non-Qwen3-TTS backend — ignored \
                 (set strict_capabilities=true to error instead)"
            );
        }

        // Native MLX synthesis via Kokoro backend (no Python shelling).
        #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
        {
            // The native Apple-Silicon path is Kokoro-only today; a
            // Qwen3-TTS schema would have routed here but the backend
            // has no cloning support yet. Strict callers are already
            // stopped above; degrade-ok callers get a second, narrower
            // note about the native-vs-Python capability gap.
            if repo_supports_advanced && !requested.is_empty() {
                if req.strict_capabilities {
                    return Err(InferenceError::InferenceFailed(format!(
                        "native MLX TTS backend does not yet implement Qwen3-TTS advanced \
                         controls {requested:?}; run on non-Apple-Silicon to use the Python \
                         mlx-audio fallback, or set strict_capabilities = false"
                    )));
                }
                tracing::warn!(
                    model = %schema.name,
                    fields = ?requested,
                    "Qwen3-TTS advanced controls are not yet implemented in the native MLX TTS \
                     backend; synthesizing without cloning/voice-design"
                );
            }
            let model_dir = self.unified_registry.ensure_local(&schema.id).await?;
            let size = backend_cache::estimate_model_size(&model_dir);
            let cache = Arc::clone(&self.kokoro_cache);
            let key = schema.id.clone();
            let handle = cache.get_or_load(&key, size, || {
                backend::mlx_kokoro::KokoroBackend::load(&model_dir)
            })?;

            let output_path = req.output_path.clone().unwrap_or_else(|| {
                let dir = std::env::temp_dir().join("car_tts");
                let _ = std::fs::create_dir_all(&dir);
                dir.join("output.wav").display().to_string()
            });
            let voice = req.voice.as_deref().unwrap_or("af_heart").to_string();
            let text = req.text.clone();
            let op = tokio::task::spawn_blocking(move || -> Result<PathBuf, InferenceError> {
                let mut guard = handle.lock().map_err(|_| {
                    InferenceError::InferenceFailed("kokoro backend mutex poisoned".into())
                })?;
                guard
                    .synthesize(&text, Some(&voice), Path::new(&output_path))
                    .map_err(|e| InferenceError::InferenceFailed(format!("native TTS: {e}")))
            })
            .await
            .map_err(|e| InferenceError::InferenceFailed(format!("kokoro task join: {e}")))??;

            let final_path =
                materialize_audio_output(&op, req.output_path.as_deref(), &req.format)?;
            return Ok(SynthesizeResult {
                audio_path: final_path.display().to_string(),
                media_type: media_type_for_format(&req.format),
                model_used: Some(schema.name.clone()),
                voice_used: req.voice.clone(),
            });
        }

        // Non-Apple-Silicon fallback: use Python speech runtime.
        #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
        {
            let runtime = self.ensure_speech_runtime().await?;
            let primary_hf_repo = match &schema.source {
                ModelSource::Mlx { hf_repo, .. } => hf_repo.clone(),
                _ => unreachable!(),
            };
            let (produced, model_used) = match self
                .synthesize_local_mlx_repo(&runtime, &primary_hf_repo, schema.name.as_str(), req)
                .await
            {
                Ok(result) => result,
                Err(primary_err)
                    if primary_hf_repo == "mlx-community/Kokoro-82M-6bit"
                        && kokoro_runtime_fallback_enabled() =>
                {
                    let fallback_repo = "mlx-community/Kokoro-82M-bf16";
                    let fallback_name = "Kokoro-82M-bf16";
                    match self
                        .synthesize_local_mlx_repo(&runtime, fallback_repo, fallback_name, req)
                        .await
                    {
                        Ok(result) => result,
                        Err(fallback_err) => {
                            return Err(InferenceError::InferenceFailed(format!(
                                "{primary_err}; fallback {fallback_name} also failed: {fallback_err}"
                            )));
                        }
                    }
                }
                Err(err) => return Err(err),
            };
            let final_path =
                materialize_audio_output(&produced, req.output_path.as_deref(), &req.format)?;

            Ok(SynthesizeResult {
                audio_path: final_path.display().to_string(),
                media_type: media_type_for_format(&req.format),
                model_used: Some(model_used),
                voice_used: req.voice.clone(),
            })
        }
    }

    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    async fn synthesize_local_mlx_repo(
        &self,
        runtime: &SpeechRuntime,
        hf_repo: &str,
        model_name: &str,
        req: &SynthesizeRequest,
    ) -> Result<(PathBuf, String), InferenceError> {
        let output_dir = temp_work_dir("tts")?;
        let mut args = vec![
            "--model".to_string(),
            hf_repo.to_string(),
            "--text".to_string(),
            req.text.clone(),
            "--output_path".to_string(),
            output_dir.display().to_string(),
        ];
        if let Some(voice) = &req.voice {
            args.push("--voice".to_string());
            args.push(voice.clone());
        }
        if let Some(speed) = req.speed {
            args.push("--speed".to_string());
            args.push(speed.to_string());
        }
        let repo_lower = hf_repo.to_ascii_lowercase();
        if repo_lower.contains("kokoro") {
            args.push("--lang_code".to_string());
            args.push(kokoro_lang_code(req.language.as_deref()).to_string());
        } else if let Some(language) = &req.language {
            args.push("--lang_code".to_string());
            args.push(normalize_lang_code(language));
        }

        // Qwen3-TTS advanced controls — reference-audio cloning and
        // voice-design natural-language instruction. The
        // supported/unsupported decision was already made at the
        // `synthesize_local_mlx` entry point; here we only need to
        // forward the flags to the mlx-audio CLI for Qwen3-TTS repos.
        if repo_lower.contains("qwen3-tts") {
            if let Some(ref_audio) = &req.reference_audio_path {
                args.push("--ref_audio".to_string());
                args.push(ref_audio.clone());
            }
            if let Some(ref_text) = &req.reference_text {
                args.push("--ref_text".to_string());
                args.push(ref_text.clone());
            }
            if let Some(instruct) = &req.voice_instruction {
                args.push("--instruct".to_string());
                args.push(instruct.clone());
            }
        }

        let output = if repo_lower.contains("kokoro") {
            let device = std::env::var("CAR_SPEECH_KOKORO_DEVICE")
                .or_else(|_| std::env::var("CAR_SPEECH_MLX_DEVICE"))
                .unwrap_or_else(|_| "cpu".to_string());
            let extra_env = vec![
                // Force MLX device (defaults to CPU to avoid Metal/NSRangeException crashes)
                ("MLX_DEVICE".to_string(), device),
                // Prevent MPS/Metal kernel crashes by enabling CPU fallback
                ("PYTORCH_ENABLE_MPS_FALLBACK".to_string(), "1".to_string()),
            ];
            run_mlx_audio_command_with_env(runtime, "tts.generate", &args, &extra_env).await?
        } else {
            run_mlx_audio_command(runtime, "tts.generate", &args).await?
        };
        let produced = find_audio_file(&output_dir)?.ok_or_else(|| {
            let hint = if repo_lower.contains("kokoro") {
                ". Kokoro models may crash on GPU — try CAR_SPEECH_KOKORO_DEVICE=cpu or use the default Qwen3-TTS model"
            } else {
                ""
            };
            InferenceError::InferenceFailed(format!(
                "mlx-audio synthesis produced no audio file: {}{}",
                output.stderr, hint
            ))
        })?;
        Ok((produced, model_name.to_string()))
    }


    async fn transcribe_elevenlabs(
        &self,
        schema: &ModelSchema,
        req: &TranscribeRequest,
    ) -> Result<TranscribeResult, InferenceError> {
        let (endpoint, api_key) = elevenlabs_auth(schema)?;
        let file_name = Path::new(&req.audio_path)
            .file_name()
            .and_then(|f| f.to_str())
            .unwrap_or("audio.wav")
            .to_string();
        let audio_bytes = tokio::fs::read(&req.audio_path).await?;
        let file_part = Part::bytes(audio_bytes).file_name(file_name);
        let mut form = Form::new()
            .text("model_id", schema.name.clone())
            .part("file", file_part);
        if let Some(language) = &req.language {
            form = form.text("language_code", language.clone());
        }

        let resp = self
            .remote_backend
            .client
            .post(format!(
                "{}/v1/speech-to-text",
                endpoint.trim_end_matches('/')
            ))
            .header("xi-api-key", api_key)
            .multipart(form)
            .send()
            .await
            .map_err(|e| {
                InferenceError::InferenceFailed(format!("ElevenLabs STT request failed: {e}"))
            })?;
        let status = resp.status();
        let body = resp.text().await.map_err(|e| {
            InferenceError::InferenceFailed(format!("read ElevenLabs STT body: {e}"))
        })?;
        if !status.is_success() {
            return Err(InferenceError::InferenceFailed(format!(
                "ElevenLabs STT returned {status}: {body}"
            )));
        }
        let payload: serde_json::Value = serde_json::from_str(&body).map_err(|e| {
            InferenceError::InferenceFailed(format!("parse ElevenLabs STT response: {e}"))
        })?;
        let text = payload
            .get("text")
            .and_then(|v| v.as_str())
            .map(str::to_string)
            .ok_or_else(|| {
                InferenceError::InferenceFailed("ElevenLabs STT response missing text".into())
            })?;

        Ok(TranscribeResult {
            text,
            model_used: Some(schema.name.clone()),
            language: payload
                .get("language_code")
                .and_then(|v| v.as_str())
                .map(str::to_string),
            words: Vec::new(),
        })
    }

    async fn synthesize_elevenlabs(
        &self,
        schema: &ModelSchema,
        req: &SynthesizeRequest,
    ) -> Result<SynthesizeResult, InferenceError> {
        // ElevenLabs doesn't expose a Qwen3-TTS-style cloning or
        // voice-design surface on its `/v1/text-to-speech` endpoint;
        // honor the strict_capabilities contract here too.
        let requested = req.requested_advanced_controls();
        if !requested.is_empty() {
            if req.strict_capabilities {
                return Err(InferenceError::InferenceFailed(format!(
                    "ElevenLabs backend does not support Qwen3-TTS advanced controls \
                     {requested:?}; route to a Qwen3-TTS model or set strict_capabilities = false"
                )));
            }
            tracing::warn!(
                model = %schema.name,
                fields = ?requested,
                "Qwen3-TTS advanced controls ignored by ElevenLabs backend"
            );
        }
        let (endpoint, api_key) = elevenlabs_auth(schema)?;
        let voice_id = req
            .voice
            .clone()
            .unwrap_or_else(|| "JBFqnCBsd6RMkjVDRZzb".to_string());
        let output_format = elevenlabs_output_format(&req.format);
        let url = format!(
            "{}/v1/text-to-speech/{}?output_format={}",
            endpoint.trim_end_matches('/'),
            voice_id,
            output_format
        );

        let mut body = serde_json::json!({
            "text": req.text,
            "model_id": schema.name,
        });
        if let Some(language) = &req.language {
            body["language_code"] = serde_json::Value::String(language.clone());
        }

        let resp = self
            .remote_backend
            .client
            .post(url)
            .header("xi-api-key", api_key)
            .header("Content-Type", "application/json")
            .json(&body)
            .send()
            .await
            .map_err(|e| {
                InferenceError::InferenceFailed(format!("ElevenLabs TTS request failed: {e}"))
            })?;
        let status = resp.status();
        let audio = resp.bytes().await.map_err(|e| {
            InferenceError::InferenceFailed(format!("read ElevenLabs TTS body: {e}"))
        })?;
        if !status.is_success() {
            let err_body = String::from_utf8_lossy(&audio);
            return Err(InferenceError::InferenceFailed(format!(
                "ElevenLabs TTS returned {status}: {err_body}"
            )));
        }

        let final_path = requested_or_temp_output(req.output_path.as_deref(), &req.format)?;
        ensure_parent_dir(&final_path)?;
        tokio::fs::write(&final_path, &audio).await?;

        Ok(SynthesizeResult {
            audio_path: final_path.display().to_string(),
            media_type: media_type_for_format(&req.format),
            model_used: Some(schema.name.clone()),
            voice_used: Some(voice_id),
        })
    }
}

#[derive(Default)]
struct ProviderAccumulator {
    configured: bool,
    local_models: usize,
    remote_models: usize,
    available_models: usize,
    capabilities: std::collections::HashSet<ModelCapability>,
}

// ─── Python Speech Runtime (non-Apple-Silicon only) ──────────────────────────
// On Apple Silicon, speech uses native MLX backends (mlx_parakeet, mlx_kokoro).

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
struct CommandOutput {
    stdout: String,
    stderr: String,
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
#[derive(Debug, Clone)]
struct SpeechRuntime {
    root: PathBuf,
    python: PathBuf,
    stt_program: PathBuf,
    tts_program: PathBuf,
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
impl SpeechRuntime {
    fn new(root: PathBuf) -> Self {
        let bin_dir = root.join("bin");
        Self {
            root,
            python: bin_dir.join("python"),
            stt_program: bin_dir.join("mlx_audio.stt.generate"),
            tts_program: bin_dir.join("mlx_audio.tts.generate"),
        }
    }

    fn is_ready(&self) -> bool {
        self.python.exists() && self.stt_program.exists() && self.tts_program.exists()
    }

    fn command_for(&self, subcommand: &str) -> Result<&Path, InferenceError> {
        match subcommand {
            "stt.generate" => Ok(&self.stt_program),
            "tts.generate" => Ok(&self.tts_program),
            _ => Err(InferenceError::InferenceFailed(format!(
                "unknown speech subcommand: {subcommand}"
            ))),
        }
    }
}


#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
async fn run_mlx_audio_command(
    runtime: &SpeechRuntime,
    subcommand: &str,
    args: &[String],
) -> Result<CommandOutput, InferenceError> {
    run_mlx_audio_command_with_env(runtime, subcommand, args, &[]).await
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
async fn run_mlx_audio_command_with_env(
    runtime: &SpeechRuntime,
    subcommand: &str,
    args: &[String],
    envs: &[(String, String)],
) -> Result<CommandOutput, InferenceError> {
    let program = runtime.command_for(subcommand)?;
    let mut command = Command::new(program);
    command.args(args);
    for (key, value) in envs {
        command.env(key, value);
    }
    let output = command.output().await.map_err(|err| {
        InferenceError::InferenceFailed(format!("{}: {err}", program.display()))
    })?;

    if output.status.success() {
        Ok(CommandOutput {
            stdout: String::from_utf8_lossy(&output.stdout).to_string(),
            stderr: String::from_utf8_lossy(&output.stderr).to_string(),
        })
    } else {
        Err(InferenceError::InferenceFailed(format!(
            "{} exited with {}: {}",
            program.display(),
            output.status,
            String::from_utf8_lossy(&output.stderr)
        )))
    }
}


#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
async fn bootstrap_speech_runtime(runtime: &SpeechRuntime) -> Result<(), InferenceError> {
    std::fs::create_dir_all(&runtime.root)?;
    let python = select_speech_python()?;

    run_command(
        "uv",
        &[
            "venv".to_string(),
            "--python".to_string(),
            python,
            runtime.root.display().to_string(),
        ],
    )
    .await?;

    run_command(
        "uv",
        &[
            "pip".to_string(),
            "install".to_string(),
            "--python".to_string(),
            runtime.python.display().to_string(),
            speech_runtime_mlx_audio_spec(),
            "misaki[en]".to_string(),
            speech_runtime_spacy_model_spec(),
        ],
    )
    .await?;

    Ok(())
}


#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
async fn run_command(program: &str, args: &[String]) -> Result<(), InferenceError> {
    let output = Command::new(program)
        .args(args)
        .output()
        .await
        .map_err(|err| InferenceError::InferenceFailed(format!("{program}: {err}")))?;

    if output.status.success() {
        Ok(())
    } else {
        Err(InferenceError::InferenceFailed(format!(
            "{} exited with {}: {}",
            program,
            output.status,
            String::from_utf8_lossy(&output.stderr)
        )))
    }
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
fn select_speech_python() -> Result<String, InferenceError> {
    if let Ok(path) = std::env::var("CAR_SPEECH_PYTHON") {
        if !path.trim().is_empty() {
            return Ok(path);
        }
    }

    for candidate in ["python3.13", "python3.12", "python3.11"] {
        if command_in_path(candidate) {
            return Ok(candidate.to_string());
        }
    }

    Err(InferenceError::InferenceFailed(
        "no supported Python found for managed speech runtime (tried python3.13, python3.12, python3.11)".into(),
    ))
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
fn detect_speech_python() -> Option<String> {
    if let Ok(path) = std::env::var("CAR_SPEECH_PYTHON") {
        if !path.trim().is_empty() {
            return Some(path);
        }
    }

    ["python3.13", "python3.12", "python3.11"]
        .into_iter()
        .find(|candidate| command_in_path(candidate))
        .map(str::to_string)
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
fn speech_runtime_root_from_models_dir(_models_dir: &Path) -> PathBuf {
    if let Ok(path) = std::env::var("CAR_SPEECH_RUNTIME_DIR") {
        if !path.trim().is_empty() {
            return PathBuf::from(path);
        }
    }

    std::env::var("HOME")
        .map(PathBuf::from)
        .unwrap_or_else(|_| PathBuf::from("."))
        .join(".car")
        .join("speech-runtime")
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
fn command_in_path(name: &str) -> bool {
    std::env::var_os("PATH")
        .map(|paths| {
            std::env::split_paths(&paths).any(|dir| {
                let path = dir.join(name);
                path.exists() && path.is_file()
            })
        })
        .unwrap_or(false)
}

fn speech_model_cached(schema: &ModelSchema) -> bool {
    match &schema.source {
        ModelSource::Mlx { hf_repo, .. } => {
            huggingface_repo_has_snapshot(hf_repo)
        }
        ModelSource::Proprietary { auth, .. } => match auth {
            ProprietaryAuth::ApiKeyEnv { env_var } => std::env::var(env_var).is_ok(),
            ProprietaryAuth::BearerTokenEnv { env_var } => std::env::var(env_var).is_ok(),
            ProprietaryAuth::OAuth2Pkce { .. } => false,
        },
        _ => false,
    }
}

fn remove_huggingface_repo_cache(repo_id: &str) -> Result<(), InferenceError> {
    let repo_dir = std::env::var("HF_HOME")
        .map(PathBuf::from)
        .unwrap_or_else(|_| {
            std::env::var("HOME")
                .map(PathBuf::from)
                .unwrap_or_else(|_| PathBuf::from("."))
                .join(".cache")
                .join("huggingface")
        })
        .join("hub")
        .join(format!("models--{}", repo_id.replace('/', "--")));

    if repo_dir.exists() {
        std::fs::remove_dir_all(repo_dir)?;
    }
    Ok(())
}

fn model_source_configured(schema: &ModelSchema) -> bool {
    match &schema.source {
        ModelSource::RemoteApi {
            api_key_env,
            api_key_envs,
            ..
        } => {
            std::env::var(api_key_env).is_ok()
                || api_key_envs
                    .iter()
                    .any(|env_var| std::env::var(env_var).is_ok())
        }
        ModelSource::Proprietary { auth, .. } => match auth {
            ProprietaryAuth::ApiKeyEnv { env_var } => std::env::var(env_var).is_ok(),
            ProprietaryAuth::BearerTokenEnv { env_var } => std::env::var(env_var).is_ok(),
            ProprietaryAuth::OAuth2Pkce { .. } => false,
        },
        ModelSource::VllmMlx { .. } => std::env::var("VLLM_MLX_ENDPOINT").is_ok() || schema.available,
        ModelSource::Ollama { .. } => schema.available,
        ModelSource::Mlx { .. } | ModelSource::Local { .. } => {
            true
        }
        ModelSource::AppleFoundationModels { .. } => schema.available,
    }
}

fn all_model_capabilities() -> [ModelCapability; 13] {
    [
        ModelCapability::Generate,
        ModelCapability::Embed,
        ModelCapability::Classify,
        ModelCapability::Code,
        ModelCapability::Reasoning,
        ModelCapability::Summarize,
        ModelCapability::ToolUse,
        ModelCapability::MultiToolCall,
        ModelCapability::Vision,
        ModelCapability::SpeechToText,
        ModelCapability::TextToSpeech,
        ModelCapability::ImageGeneration,
        ModelCapability::VideoGeneration,
    ]
}

fn sort_capabilities(mut capabilities: Vec<ModelCapability>) -> Vec<ModelCapability> {
    capabilities.sort_by_key(|capability| {
        all_model_capabilities()
            .iter()
            .position(|candidate| candidate == capability)
            .unwrap_or(usize::MAX)
    });
    capabilities
}

fn speech_model_source_label(schema: &ModelSchema) -> String {
    match &schema.source {
        ModelSource::Mlx { hf_repo, .. } => format!("mlx:{hf_repo}"),
        ModelSource::Proprietary {
            provider, endpoint, ..
        } => format!("proprietary:{provider}:{endpoint}"),
        ModelSource::RemoteApi { endpoint, .. } => format!("remote:{endpoint}"),
        ModelSource::Local { hf_repo, .. } => format!("local:{hf_repo}"),
        ModelSource::VllmMlx {
            endpoint,
            model_name,
        } => format!("vllm-mlx:{endpoint}:{model_name}"),
        ModelSource::Ollama { model_tag, host } => format!("ollama:{host}:{model_tag}"),
        ModelSource::AppleFoundationModels { use_case } => {
            format!("apple-foundation:{}", use_case.as_deref().unwrap_or("default"))
        }
    }
}

/// Build the Qwen3-Reranker chat-template prompt for a single
/// `(query, document)` candidate.
///
/// The format matches upstream `reranker_quick_start.py`: a system
/// message pinning the answer space to yes/no, a user turn with
/// `<Instruct>/<Query>/<Document>`, and an assistant prefix with a
/// closed empty `<think>` block to force non-thinking classification.
fn rerank_prompt(instruction: &str, query: &str, document: &str) -> String {
    const SYSTEM: &str = "Judge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".";
    format!(
        "<|im_start|>system\n{SYSTEM}<|im_end|>\n\
         <|im_start|>user\n<Instruct>: {instruction}\n<Query>: {query}\n<Document>: {document}<|im_end|>\n\
         <|im_start|>assistant\n<think>\n\n</think>\n\n"
    )
}

/// Interpret the first useful token from a Qwen3-Reranker greedy
/// decode as a relevance score. Scans up to the first few tokens so
/// a leading space, BOS artifact, or stray newline doesn't poison the
/// result. Returns 1.0 for "yes", 0.0 for "no", 0.5 on unexpected
/// output (with a warn so the mismatch is visible).
fn score_from_rerank_output(text: &str, model_name: &str) -> f32 {
    // Replace every non-alphanumeric byte with a space, lowercase,
    // and scan the first few whitespace-separated tokens for
    // "yes"/"no". This strips chat-template tags (`<|im_end|>`),
    // punctuation, and underscores cleanly without special-casing.
    let normalized: String = text
        .to_ascii_lowercase()
        .chars()
        .map(|c| if c.is_ascii_alphanumeric() { c } else { ' ' })
        .collect();
    for tok in normalized.split_ascii_whitespace().take(5) {
        match tok {
            "yes" => return 1.0,
            "no" => return 0.0,
            _ => continue,
        }
    }
    tracing::warn!(
        model = %model_name,
        output = %text,
        "rerank: first tokens contain neither `yes` nor `no`; returning neutral 0.5"
    );
    0.5
}

fn default_speech_voice(schema: &ModelSchema) -> Option<String> {
    if schema.provider == "elevenlabs" {
        Some("JBFqnCBsd6RMkjVDRZzb".to_string())
    } else if schema.name == "Kokoro-82M-6bit" || schema.name == "Kokoro-82M-bf16" {
        Some("af_heart".to_string())
    } else if schema.name == "Qwen3-TTS-12Hz-1.7B-Base-5bit" {
        Some("Chelsie".to_string())
    } else {
        None
    }
}

fn huggingface_repo_has_snapshot(repo_id: &str) -> bool {
    find_latest_huggingface_snapshot(repo_id).is_some()
}

fn huggingface_repo_dir(repo_id: &str) -> PathBuf {
    let cache_root = std::env::var("HF_HOME")
        .map(PathBuf::from)
        .unwrap_or_else(|_| {
            std::env::var("HOME")
                .map(PathBuf::from)
                .unwrap_or_else(|_| PathBuf::from("."))
                .join(".cache")
                .join("huggingface")
        })
        .join("hub");
    cache_root.join(format!("models--{}", repo_id.replace('/', "--")))
}

fn find_latest_huggingface_snapshot(repo_id: &str) -> Option<PathBuf> {
    let snapshots = huggingface_repo_dir(repo_id).join("snapshots");
    std::fs::read_dir(snapshots)
        .ok()?
        .filter_map(Result::ok)
        .map(|entry| entry.path())
        .find(|path| path.is_dir() && snapshot_looks_ready(path))
}

fn snapshot_looks_ready(path: &Path) -> bool {
    if path.join("config.json").exists() || path.join("model_index.json").exists() {
        return true;
    }
    snapshot_contains_ext(path, "safetensors")
}

fn snapshot_contains_ext(root: &Path, ext: &str) -> bool {
    let Ok(entries) = std::fs::read_dir(root) else {
        return false;
    };
    entries.filter_map(Result::ok).any(|entry| {
        let path = entry.path();
        if path.is_dir() {
            snapshot_contains_ext(&path, ext)
        } else {
            path.extension()
                .and_then(|value| value.to_str())
                .map(|value| value.eq_ignore_ascii_case(ext))
                .unwrap_or(false)
        }
    })
}

fn count_files_recursive(root: &Path) -> usize {
    let Ok(entries) = std::fs::read_dir(root) else {
        return 0;
    };
    entries
        .filter_map(Result::ok)
        .map(|entry| entry.path())
        .map(|path| {
            if path.is_dir() {
                count_files_recursive(&path)
            } else if path.is_file() {
                1
            } else {
                0
            }
        })
        .sum()
}

async fn download_hf_repo_snapshot(repo_id: &str) -> Result<(PathBuf, usize), InferenceError> {
    let api = hf_hub::api::tokio::ApiBuilder::from_env()
        .with_progress(false)
        .build()
        .map_err(|e| InferenceError::DownloadFailed(format!("init hf api: {e}")))?;
    let repo = api.model(repo_id.to_string());
    let info = repo
        .info()
        .await
        .map_err(|e| InferenceError::DownloadFailed(format!("{repo_id}: {e}")))?;

    let snapshot_path = huggingface_repo_dir(repo_id)
        .join("snapshots")
        .join(&info.sha);
    let mut downloaded = 0usize;
    for sibling in &info.siblings {
        let local_path = snapshot_path.join(&sibling.rfilename);
        if local_path.exists() {
            downloaded += 1;
            continue;
        }
        repo.download(&sibling.rfilename).await.map_err(|e| {
            InferenceError::DownloadFailed(format!("{repo_id}/{}: {e}", sibling.rfilename))
        })?;
        downloaded += 1;
    }

    Ok((snapshot_path, downloaded))
}

fn temp_work_dir(prefix: &str) -> Result<PathBuf, InferenceError> {
    let unique = SystemTime::now()
        .duration_since(UNIX_EPOCH)
        .map_err(|e| InferenceError::InferenceFailed(format!("clock error: {e}")))?
        .as_nanos();
    let dir = std::env::temp_dir().join(format!("car-inference-{prefix}-{unique}"));
    std::fs::create_dir_all(&dir)?;
    Ok(dir)
}

fn ensure_parent_dir(path: &Path) -> Result<(), InferenceError> {
    if let Some(parent) = path.parent() {
        std::fs::create_dir_all(parent)?;
    }
    Ok(())
}

fn requested_or_temp_output(
    output_path: Option<&str>,
    format: &str,
) -> Result<PathBuf, InferenceError> {
    if let Some(path) = output_path {
        return Ok(PathBuf::from(path));
    }
    let dir = temp_work_dir("audio-out")?;
    Ok(dir.join(format!("speech.{format}")))
}

fn requested_or_temp_media_output(
    output_path: Option<&str>,
    format: &str,
    stem: &str,
) -> Result<PathBuf, InferenceError> {
    if let Some(path) = output_path {
        return Ok(PathBuf::from(path));
    }
    let dir = temp_work_dir(&format!("{stem}-out"))?;
    Ok(dir.join(format!("{stem}.{format}")))
}

fn materialize_audio_output(
    produced: &Path,
    requested: Option<&str>,
    format: &str,
) -> Result<PathBuf, InferenceError> {
    if let Some(path) = requested {
        let dest = PathBuf::from(path);
        ensure_parent_dir(&dest)?;
        std::fs::copy(produced, &dest)?;
        Ok(dest)
    } else {
        let dest = requested_or_temp_output(None, format)?;
        ensure_parent_dir(&dest)?;
        std::fs::copy(produced, &dest)?;
        Ok(dest)
    }
}

fn materialize_binary_output(
    produced: &Path,
    requested: Option<&str>,
    format: &str,
    stem: &str,
) -> Result<PathBuf, InferenceError> {
    let dest = requested_or_temp_media_output(requested, format, stem)?;
    ensure_parent_dir(&dest)?;
    std::fs::copy(produced, &dest)?;
    Ok(dest)
}

fn find_generated_file(
    root: &Path,
    extensions: &[&str],
) -> Result<Option<PathBuf>, InferenceError> {
    let entries = std::fs::read_dir(root)?;
    let mut candidates: Vec<PathBuf> = entries
        .filter_map(Result::ok)
        .map(|entry| entry.path())
        .filter(|path| {
            path.is_file()
                && path
                    .extension()
                    .and_then(|ext| ext.to_str())
                    .map(|ext| extensions.iter().any(|candidate| candidate.eq_ignore_ascii_case(ext)))
                    .unwrap_or(false)
        })
        .collect();
    candidates.sort();
    Ok(candidates.pop())
}

fn media_type_for_image_format(format: &str) -> String {
    match format.to_ascii_lowercase().as_str() {
        "jpg" | "jpeg" => "image/jpeg".to_string(),
        "webp" => "image/webp".to_string(),
        _ => "image/png".to_string(),
    }
}

fn media_type_for_video_format(format: &str) -> String {
    match format.to_ascii_lowercase().as_str() {
        "mov" => "video/quicktime".to_string(),
        "gif" => "image/gif".to_string(),
        _ => "video/mp4".to_string(),
    }
}

fn read_transcription_result(output_prefix: &Path) -> Result<Option<String>, InferenceError> {
    let candidates = [
        output_prefix.with_extension("json"),
        output_prefix.to_path_buf(),
    ];

    for path in candidates {
        if path.exists() {
            let contents = std::fs::read_to_string(path)?;
            if let Some(text) = extract_text_from_payload(&contents) {
                return Ok(Some(text));
            }
        }
    }

    Ok(None)
}

fn extract_text_from_payload(payload: &str) -> Option<String> {
    let value: serde_json::Value = serde_json::from_str(payload).ok()?;
    if let Some(text) = value.get("text").and_then(|v| v.as_str()) {
        return Some(text.to_string());
    }
    if let Some(transcripts) = value.get("transcripts").and_then(|v| v.as_array()) {
        let joined = transcripts
            .iter()
            .filter_map(|item| item.get("text").and_then(|v| v.as_str()))
            .collect::<Vec<_>>()
            .join("\n");
        if !joined.is_empty() {
            return Some(joined);
        }
    }
    if let Some(items) = value.as_array() {
        let joined = items
            .iter()
            .filter_map(|item| {
                item.get("text")
                    .or_else(|| item.get("Content"))
                    .and_then(|v| v.as_str())
            })
            .collect::<Vec<_>>()
            .join(" ");
        if !joined.is_empty() {
            return Some(joined);
        }
    }
    None
}

fn find_audio_file(output_dir: &Path) -> Result<Option<PathBuf>, InferenceError> {
    let mut audio_files = Vec::new();
    collect_audio_files(output_dir, &mut audio_files)?;
    audio_files.sort();
    Ok(audio_files.into_iter().next())
}

fn collect_audio_files(dir: &Path, audio_files: &mut Vec<PathBuf>) -> Result<(), InferenceError> {
    for entry in std::fs::read_dir(dir)? {
        let path = entry?.path();
        if path.is_dir() {
            collect_audio_files(&path, audio_files)?;
        } else if matches!(
            path.extension().and_then(|ext| ext.to_str()),
            Some("wav" | "mp3" | "flac" | "pcm" | "m4a")
        ) {
            audio_files.push(path);
        }
    }
    Ok(())
}

fn media_type_for_format(format: &str) -> String {
    match format.to_ascii_lowercase().as_str() {
        "mp3" => "audio/mpeg".to_string(),
        "flac" => "audio/flac".to_string(),
        "pcm" => "audio/L16".to_string(),
        "m4a" => "audio/mp4".to_string(),
        _ => "audio/wav".to_string(),
    }
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
fn kokoro_lang_code(language: Option<&str>) -> &'static str {
    match language.unwrap_or("en").to_ascii_lowercase().as_str() {
        "en-gb" | "british" | "british english" => "b",
        "ja" | "japanese" => "j",
        "zh" | "zh-cn" | "mandarin" | "chinese" => "z",
        "es" | "spanish" => "e",
        "fr" | "french" => "f",
        _ => "a",
    }
}

fn normalize_lang_code(language: &str) -> String {
    match language.to_ascii_lowercase().as_str() {
        "english" | "en-us" | "en_us" => "en".to_string(),
        "spanish" => "es".to_string(),
        "french" => "fr".to_string(),
        "japanese" => "ja".to_string(),
        "chinese" | "mandarin" => "zh".to_string(),
        other => match other {
            "en" | "es" | "fr" | "ja" | "zh" => other.to_string(),
            _ => "en".to_string(),
        },
    }
}

fn elevenlabs_auth(schema: &ModelSchema) -> Result<(String, String), InferenceError> {
    match &schema.source {
        ModelSource::Proprietary {
            endpoint,
            auth: schema::ProprietaryAuth::ApiKeyEnv { env_var },
            ..
        } => {
            let key = car_secrets::resolve_env_or_keychain(env_var).ok_or_else(|| {
                InferenceError::InferenceFailed(format!(
                    "missing API key {env_var}; set the environment variable or \
                     store it with `car secrets put {env_var}`"
                ))
            })?;
            Ok((endpoint.clone(), key))
        }
        _ => Err(InferenceError::InferenceFailed(format!(
            "model {} is not an ElevenLabs proprietary model",
            schema.id
        ))),
    }
}

fn elevenlabs_output_format(format: &str) -> &'static str {
    match format.to_ascii_lowercase().as_str() {
        "mp3" => "mp3_44100_128",
        "pcm" => "pcm_16000",
        _ => "wav_44100",
    }
}

fn benchmark_priors_paths(models_dir: &Path) -> Vec<PathBuf> {
    let mut paths = Vec::new();

    let direct = models_dir.join("benchmark_priors.json");
    if !paths.contains(&direct) {
        paths.push(direct);
    }

    if let Some(parent) = models_dir.parent() {
        let parent_path = parent.join("benchmark_priors.json");
        if !paths.contains(&parent_path) {
            paths.push(parent_path);
        }
    }

    if let Some(path) = std::env::var_os("CAR_BENCHMARK_PRIORS_PATH") {
        let path = PathBuf::from(path);
        if !paths.contains(&path) {
            paths.push(path);
        }
    }

    paths
}

fn load_benchmark_prior_health(
    models_dir: &Path,
    schemas: &[ModelSchema],
) -> Vec<ModelBenchmarkPriorHealth> {
    let mut priors = std::collections::BTreeMap::new();
    for path in benchmark_priors_paths(models_dir) {
        let Ok(loaded) = routing_ext::load_benchmark_priors(&path) else {
            continue;
        };
        for (model_id, prior) in loaded {
            let model_name = schemas
                .iter()
                .find(|schema| schema.id == model_id)
                .map(|schema| schema.name.clone());
            priors.insert(
                model_id.clone(),
                ModelBenchmarkPriorHealth {
                    model_id,
                    model_name,
                    overall_score: prior.overall_score,
                    overall_latency_ms: prior.overall_latency_ms,
                    task_scores: prior.task_scores,
                    task_latency_ms: prior.task_latency_ms,
                    source_path: path.clone(),
                },
            );
        }
    }

    priors.into_values().collect()
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
fn kokoro_runtime_fallback_enabled() -> bool {
    std::env::var("CAR_SPEECH_KOKORO_FALLBACK")
        .ok()
        .map(|value| !matches!(value.trim().to_ascii_lowercase().as_str(), "0" | "false" | "off"))
        .unwrap_or(true)
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
fn speech_runtime_mlx_audio_spec() -> String {
    std::env::var("CAR_SPEECH_RUNTIME_MLX_AUDIO_SPEC")
        .ok()
        .filter(|value| !value.trim().is_empty())
        .unwrap_or_else(|| "mlx-audio==0.4.2".to_string())
}

#[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
fn speech_runtime_spacy_model_spec() -> String {
    std::env::var("CAR_SPEECH_RUNTIME_SPACY_MODEL_SPEC")
        .ok()
        .filter(|value| !value.trim().is_empty())
        .unwrap_or_else(|| {
            "en-core-web-sm @ https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl".to_string()
        })
}

#[cfg(test)]
mod tests {
    use super::*;
    use tempfile::TempDir;

    /// Tests that mutate process-wide env vars must hold this lock to avoid
    /// races with parallel tests (env vars are global mutable state).
    static ENV_MUTEX: std::sync::Mutex<()> = std::sync::Mutex::new(());

    fn test_config(models_dir: PathBuf) -> InferenceConfig {
        InferenceConfig {
            models_dir,
            device: None,
            generation_model: "Qwen3-0.6B".into(),
            preferred_generation_model: None,
            embedding_model: "Qwen3-Embedding-0.6B".into(),
            preferred_embedding_model: None,
            classification_model: "Qwen3-0.6B".into(),
            preferred_classification_model: None,
        }
    }

    #[tokio::test]
    async fn tokenize_rejects_known_remote_model_with_unsupported_mode() {
        // The unified registry's built-in catalog includes remote models like
        // OpenAI / Anthropic ones. Regardless of which exact id ships, we just
        // need any non-local schema to confirm the pre-flight catches it
        // before we try (and fail) to load a non-existent local backend.
        let tmp = TempDir::new().unwrap();
        let engine = InferenceEngine::new(test_config(tmp.path().join("models")));
        let remote_id = engine
            .list_schemas()
            .into_iter()
            .find(|s| !s.is_local())
            .map(|s| s.id)
            .expect(
                "built-in catalog should include at least one remote model schema",
            );

        let err = engine
            .tokenize(&remote_id, "hello")
            .await
            .expect_err("remote tokenize must error");
        match err {
            InferenceError::UnsupportedMode { mode, backend, .. } => {
                assert_eq!(mode, "tokenize/detokenize");
                assert_eq!(backend, "remote");
            }
            other => panic!("expected UnsupportedMode, got {other:?}"),
        }

        let err = engine
            .detokenize(&remote_id, &[1, 2, 3])
            .await
            .expect_err("remote detokenize must error");
        assert!(
            matches!(err, InferenceError::UnsupportedMode { .. }),
            "expected UnsupportedMode, got {err:?}"
        );
    }

    #[test]
    fn engine_loads_benchmark_priors_on_startup() {
        let _env = ENV_MUTEX.lock().unwrap();
        let tmp = TempDir::new().unwrap();
        let priors_path = tmp.path().join("benchmark_priors.json");
        std::fs::write(
            &priors_path,
            serde_json::json!({
                "model_id": "qwen/qwen3-8b:q4_k_m",
                "overall_score": 0.88
            })
            .to_string(),
        )
        .unwrap();

        unsafe {
            std::env::set_var("CAR_BENCHMARK_PRIORS_PATH", &priors_path);
        }

        let engine = InferenceEngine::new(test_config(tmp.path().join("models")));
        let tracker = engine.outcome_tracker.blocking_read();
        let profile = tracker
            .profile("qwen/qwen3-8b:q4_k_m")
            .expect("benchmark prior should create a profile");
        assert!((profile.ema_quality - 0.88).abs() < 0.01);

        unsafe {
            std::env::remove_var("CAR_BENCHMARK_PRIORS_PATH");
        }
    }

    #[test]
    fn benchmark_priors_do_not_override_observed_profiles() {
        let _env = ENV_MUTEX.lock().unwrap();
        let tmp = TempDir::new().unwrap();
        let models_dir = tmp.path().join("models");
        std::fs::create_dir_all(&models_dir).unwrap();

        let observed = vec![ModelProfile {
            model_id: "qwen/qwen3-8b:q4_k_m".into(),
            total_calls: 12,
            success_count: 3,
            fail_count: 9,
            total_latency_ms: 1200,
            total_input_tokens: 0,
            total_output_tokens: 0,
            task_stats: std::collections::HashMap::new(),
            ema_quality: 0.21,
            quality_per_1k_tokens: 0.0,
            updated_at: 1,
        }];
        std::fs::write(
            models_dir.join("outcome_profiles.json"),
            serde_json::to_string(&observed).unwrap(),
        )
        .unwrap();

        let priors_path = tmp.path().join("benchmark_priors.json");
        std::fs::write(
            &priors_path,
            serde_json::json!({
                "model_id": "qwen/qwen3-8b:q4_k_m",
                "overall_score": 0.95
            })
            .to_string(),
        )
        .unwrap();

        unsafe {
            std::env::set_var("CAR_BENCHMARK_PRIORS_PATH", &priors_path);
        }

        let engine = InferenceEngine::new(test_config(models_dir));
        let tracker = engine.outcome_tracker.blocking_read();
        let profile = tracker
            .profile("qwen/qwen3-8b:q4_k_m")
            .expect("observed profile should remain present");
        assert!((profile.ema_quality - 0.21).abs() < 0.01);
        assert_eq!(profile.total_calls, 12);

        unsafe {
            std::env::remove_var("CAR_BENCHMARK_PRIORS_PATH");
        }
    }

    #[test]
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    fn speech_runtime_package_spec_defaults_and_overrides() {
        let _env = ENV_MUTEX.lock().unwrap();
        unsafe {
            std::env::remove_var("CAR_SPEECH_RUNTIME_MLX_AUDIO_SPEC");
        }
        assert_eq!(speech_runtime_mlx_audio_spec(), "mlx-audio==0.4.2");

        unsafe {
            std::env::set_var("CAR_SPEECH_RUNTIME_MLX_AUDIO_SPEC", "mlx-audio==0.4.1");
        }
        assert_eq!(speech_runtime_mlx_audio_spec(), "mlx-audio==0.4.1");

        unsafe {
            std::env::remove_var("CAR_SPEECH_RUNTIME_MLX_AUDIO_SPEC");
        }
    }

    #[test]
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    fn speech_runtime_spacy_model_spec_defaults_and_overrides() {
        let _env = ENV_MUTEX.lock().unwrap();
        unsafe {
            std::env::remove_var("CAR_SPEECH_RUNTIME_SPACY_MODEL_SPEC");
        }
        assert!(
            speech_runtime_spacy_model_spec().starts_with("en-core-web-sm @ https://github.com/")
        );

        unsafe {
            std::env::set_var(
                "CAR_SPEECH_RUNTIME_SPACY_MODEL_SPEC",
                "en-core-web-sm==3.8.0",
            );
        }
        assert_eq!(
            speech_runtime_spacy_model_spec(),
            "en-core-web-sm==3.8.0"
        );

        unsafe {
            std::env::remove_var("CAR_SPEECH_RUNTIME_SPACY_MODEL_SPEC");
        }
    }

    #[test]
    #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
    fn kokoro_runtime_fallback_defaults_on() {
        unsafe {
            std::env::remove_var("CAR_SPEECH_KOKORO_FALLBACK");
        }
        assert!(kokoro_runtime_fallback_enabled());

        unsafe {
            std::env::set_var("CAR_SPEECH_KOKORO_FALLBACK", "false");
        }
        assert!(!kokoro_runtime_fallback_enabled());

        unsafe {
            std::env::remove_var("CAR_SPEECH_KOKORO_FALLBACK");
        }
    }

    #[test]
    fn preferred_local_tts_wins_over_builtin_rank() {
        let tmp = TempDir::new().unwrap();
        let mut engine = InferenceEngine::new(test_config(tmp.path().join("models")));
        engine.set_speech_policy(SpeechPolicy {
            prefer_local: true,
            allow_remote_fallback: false,
            preferred_local_stt: None,
            preferred_local_tts: Some("Kokoro-82M-6bit".into()),
            preferred_remote_stt: None,
            preferred_remote_tts: None,
        });

        let schema = engine
            .preferred_speech_schema(ModelCapability::TextToSpeech, true, false)
            .expect("preferred local TTS should resolve");
        assert_eq!(schema.name, "Kokoro-82M-6bit");
    }

    #[test]
    fn preferred_discovered_vllm_mlx_model_wins_generate_routing() {
        let tmp = TempDir::new().unwrap();
        let mut config = test_config(tmp.path().join("models"));
        config.preferred_generation_model =
            Some("vllm-mlx/mlx-community_gemma-3n-E2B-it-lm-4bit".into());
        let mut engine = InferenceEngine::new(config);
        let schema = crate::vllm_mlx::to_model_schema(
            &crate::vllm_mlx::DiscoveredModel {
                id: "mlx-community/gemma-3n-E2B-it-lm-4bit".into(),
                owned_by: Some("mlx-community".into()),
            },
            "http://127.0.0.1:8001",
        );
        engine.register_model(schema);

        let rt = tokio::runtime::Runtime::new().unwrap();
        let decision = rt.block_on(engine.route_adaptive("say hello in one sentence"));
        assert_eq!(decision.model_id, "vllm-mlx/mlx-community_gemma-3n-E2B-it-lm-4bit");
        assert_eq!(decision.strategy, RoutingStrategy::Explicit);
        assert_eq!(decision.reason, "preferred generation model override");
    }

    /// Issue #43 — InferenceResult must serialize with all fields preserved
    /// (text, tool_calls, trace_id, model_used, latency_ms, usage) using
    /// snake_case field names. The car-server WebSocket handler relies on
    /// `serde_json::to_value(&InferenceResult)` producing this exact shape.
    #[test]
    fn inference_result_serializes_with_full_shape() {
        use crate::tasks::generate::ToolCall;
        use std::collections::HashMap;

        let mut args = HashMap::new();
        args.insert("path".to_string(), serde_json::json!("README.md"));

        let result = InferenceResult {
            text: String::new(),
            bounding_boxes: Vec::new(),
            tool_calls: vec![ToolCall {
                id: None,
                name: "read_file".into(),
                arguments: args,
            }],
            trace_id: "trace-abc".into(),
            model_used: "test-model".into(),
            latency_ms: 1234,
            time_to_first_token_ms: Some(180),
            usage: Some(TokenUsage {
                prompt_tokens: 100,
                completion_tokens: 50,
                total_tokens: 150,
                context_window: 8192,
            }),
        };

        let json = serde_json::to_value(&result).expect("serialize");

        // Required snake_case fields with type-strict assertions
        assert_eq!(json["text"].as_str(), Some(""));
        assert_eq!(json["trace_id"].as_str(), Some("trace-abc"));
        assert_eq!(json["model_used"].as_str(), Some("test-model"));
        assert_eq!(json["latency_ms"].as_u64(), Some(1234));

        // tool_calls is a non-empty array with name + arguments
        let tool_calls = json["tool_calls"].as_array().expect("tool_calls array");
        assert_eq!(tool_calls.len(), 1);
        assert_eq!(tool_calls[0]["name"].as_str(), Some("read_file"));
        assert_eq!(tool_calls[0]["arguments"]["path"].as_str(), Some("README.md"));

        // usage is an object with all four documented fields
        let usage = &json["usage"];
        assert_eq!(usage["prompt_tokens"].as_u64(), Some(100));
        assert_eq!(usage["completion_tokens"].as_u64(), Some(50));
        assert_eq!(usage["total_tokens"].as_u64(), Some(150));
        assert_eq!(usage["context_window"].as_u64(), Some(8192));

        // TTFT propagates through serialization when populated.
        assert_eq!(json["time_to_first_token_ms"].as_u64(), Some(180));
    }

    /// Issue #43 — Lock the top-level WebSocket `infer` response contract.
    /// If a future change adds a field to `InferenceResult`, this test forces
    /// the developer to deliberately update the protocol surface and the
    /// expected key set here, rather than silently leaking new fields onto
    /// the wire.
    #[test]
    fn inference_result_top_level_keys_are_locked() {
        use std::collections::BTreeSet;

        let result = InferenceResult {
            text: "anything".into(),
            bounding_boxes: Vec::new(),
            tool_calls: vec![],
            trace_id: "t".into(),
            model_used: "m".into(),
            latency_ms: 0,
            time_to_first_token_ms: None,
            usage: None,
        };

        let json = serde_json::to_value(&result).expect("serialize");
        let keys: BTreeSet<&str> = json
            .as_object()
            .expect("top-level object")
            .keys()
            .map(String::as_str)
            .collect();

        let expected: BTreeSet<&str> = [
            "text",
            "tool_calls",
            "trace_id",
            "model_used",
            "latency_ms",
            "time_to_first_token_ms",
            "usage",
        ]
        .into_iter()
        .collect();

        assert_eq!(
            keys, expected,
            "infer response top-level keys drifted -- update both the test \
             and the WebSocket protocol documentation if this is intentional"
        );

        // All keys are snake_case (constraint c-2 in outcome 043).
        for key in &keys {
            assert!(
                !key.chars().any(|c| c.is_uppercase()) && !key.contains('-'),
                "key '{}' is not snake_case",
                key
            );
        }
    }

    /// Plain text result (no tools) must still serialize cleanly with text
    /// populated and tool_calls present as an empty array. Backward compat
    /// for clients that only care about `.text`.
    #[test]
    fn inference_result_serializes_plain_text_response() {
        let result = InferenceResult {
            text: "hello world".into(),
            bounding_boxes: Vec::new(),
            tool_calls: vec![],
            trace_id: "trace-xyz".into(),
            model_used: "test-model".into(),
            latency_ms: 42,
            time_to_first_token_ms: None,
            usage: None,
        };

        let json = serde_json::to_value(&result).expect("serialize");
        assert_eq!(json["text"], "hello world");
        assert!(json["tool_calls"].is_array());
        assert_eq!(json["tool_calls"].as_array().unwrap().len(), 0);
        assert_eq!(json["model_used"], "test-model");
        assert!(json["usage"].is_null());
        // Honest "wasn't measured" rather than missing key — the field
        // is always present at the protocol surface.
        assert!(json["time_to_first_token_ms"].is_null());
    }

    #[test]
    fn rerank_prompt_matches_upstream_template_shape() {
        let p = rerank_prompt("retrieve relevant passages", "who runs the treasury?", "doc x");
        assert!(p.contains("<|im_start|>system"));
        assert!(p.contains("Note that the answer can only be \"yes\" or \"no\"."));
        assert!(p.contains("<|im_start|>user\n<Instruct>: retrieve relevant passages"));
        assert!(p.contains("<Query>: who runs the treasury?"));
        assert!(p.contains("<Document>: doc x<|im_end|>"));
        assert!(p.contains("<|im_start|>assistant\n<think>\n\n</think>\n\n"));
    }

    #[test]
    fn rerank_score_yes_and_no_exactly() {
        assert_eq!(score_from_rerank_output("yes", "m"), 1.0);
        assert_eq!(score_from_rerank_output("no", "m"), 0.0);
    }

    #[test]
    fn rerank_score_handles_case_leading_space_and_chat_sentinels() {
        // Real decodes often include leading whitespace, punctuation,
        // or chat-template sentinels around the answer token.
        assert_eq!(score_from_rerank_output(" Yes", "m"), 1.0);
        assert_eq!(score_from_rerank_output("\nno.", "m"), 0.0);
        assert_eq!(score_from_rerank_output("<|im_end|>yes", "m"), 1.0);
    }

    #[test]
    fn rerank_score_scans_up_to_three_tokens() {
        // Tokenizer artifacts can produce a BOS-like leading token
        // before the real answer. Don't miss it.
        assert_eq!(score_from_rerank_output("_bos_ yes", "m"), 1.0);
    }

    #[test]
    fn rerank_score_unexpected_is_neutral() {
        // Plain-base models that aren't reranker-fine-tuned will emit
        // arbitrary completion tokens. Don't partition; go neutral.
        assert_eq!(score_from_rerank_output("maybe", "m"), 0.5);
        assert_eq!(score_from_rerank_output("", "m"), 0.5);
    }
}