rch_common/

types.rs

//! Common types used across RCH components.

use crate::{CompilationKind, toolchain::ToolchainInfo};
use rand::RngExt;
use schemars::JsonSchema;
use serde::{Deserialize, Serialize};
use std::path::PathBuf;

/// Unique identifier for a worker in the fleet.
#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize, JsonSchema)]
pub struct WorkerId(pub String);

impl WorkerId {
    pub fn new(id: impl Into<String>) -> Self {
        Self(id.into())
    }

    pub fn as_str(&self) -> &str {
        &self.0
    }
}

impl std::fmt::Display for WorkerId {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "{}", self.0)
    }
}

/// Status of a worker in the fleet.
///
/// State transitions:
/// ```text
/// HEALTHY ←→ DEGRADED (automatic based on response times)
///    ↓           ↓
///    ↓       UNREACHABLE (automatic on heartbeat failure)
///    ↓
/// DRAINING (via `rch workers drain`) - finishing current jobs
///    ↓
/// DRAINED (automatic when all jobs complete) - idle, ready to disable
///    ↓
/// DISABLED (via `rch workers disable`) - completely offline
///
/// Use `rch workers enable` to return from DRAINING/DRAINED/DISABLED → HEALTHY
/// ```
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
#[serde(rename_all = "snake_case")]
pub enum WorkerStatus {
    /// Worker is healthy and accepting jobs.
    #[default]
    Healthy,
    /// Worker is responding slowly.
    Degraded,
    /// Worker failed to respond to heartbeat.
    Unreachable,
    /// Worker is not accepting new jobs (finishing current).
    Draining,
    /// Worker has finished draining (no active jobs, idle).
    Drained,
    /// Worker is manually disabled.
    Disabled,
}

/// Circuit breaker state for a worker.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
#[serde(rename_all = "snake_case")]
pub enum CircuitState {
    /// Normal operation.
    #[default]
    Closed,
    /// Circuit is open (short-circuit).
    Open,
    /// Circuit is half-open (probing).
    HalfOpen,
}

/// Required runtime for command execution.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default, JsonSchema)]
#[serde(rename_all = "lowercase")]
pub enum RequiredRuntime {
    /// No specific runtime required (default).
    #[default]
    None,
    /// Requires Rust toolchain.
    Rust,
    /// Requires Bun runtime.
    Bun,
    /// Requires Node.js runtime.
    Node,
    /// Requires the Nix package manager (nix binary + populated `/nix/store`).
    Nix,
    /// Requires the Go toolchain (`go` binary).
    Go,
    /// Requires the zig cross-compilation toolchain (`cargo-zigbuild` + `zig`).
    Zig,
}

/// Per-command priority hint for worker selection.
///
/// This is an *input hint* from the caller (typically the hook) and should not
/// change default behavior when omitted.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default, JsonSchema)]
#[serde(rename_all = "lowercase")]
pub enum CommandPriority {
    Low,
    #[default]
    Normal,
    High,
}

impl std::fmt::Display for CommandPriority {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        let value = match self {
            Self::Low => "low",
            Self::Normal => "normal",
            Self::High => "high",
        };
        write!(f, "{}", value)
    }
}

impl std::str::FromStr for CommandPriority {
    type Err = ();

    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s.trim().to_lowercase().as_str() {
            "low" => Ok(Self::Low),
            "normal" => Ok(Self::Normal),
            "high" => Ok(Self::High),
            _ => Err(()),
        }
    }
}

/// Worker selection request sent from hook to daemon.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SelectionRequest {
    /// Project identifier (usually directory name or hash).
    pub project: String,
    /// Full command being executed (optional, for active build tracking).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub command: Option<String>,
    /// Priority hint for this request.
    #[serde(default)]
    pub command_priority: CommandPriority,
    /// Estimated CPU cores needed for this compilation.
    pub estimated_cores: u32,
    /// Additional free GiB required on the worker's reported build filesystem.
    /// A positive value is a hard admission budget, retained by the daemon for
    /// this build's lifetime; zero leaves disk admission to the ordinary policy.
    #[serde(default, skip_serializing_if = "disk_headroom_disabled")]
    pub disk_headroom_gib: u32,
    /// Preferred worker IDs (e.g., from project config).
    #[serde(default)]
    pub preferred_workers: Vec<WorkerId>,
    /// Rust toolchain information for the project.
    #[serde(default)]
    pub toolchain: Option<ToolchainInfo>,
    /// Required runtime for command execution.
    #[serde(default)]
    pub required_runtime: RequiredRuntime,
    /// Classification decision latency in microseconds (for AGENTS.md compliance).
    /// This tracks how long the 5-tier classification took on the hook side.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub classification_duration_us: Option<u64>,
    /// Job-mode admission (bd-g7rpy): set by `rch exec --job` only. When the
    /// one-active-job-per-project-per-worker guard excludes every otherwise-
    /// eligible worker, a job request may QUEUE on that exclusion instead of
    /// failing over to local execution. Compilation requests never set it,
    /// preserving their immediate NoAdmissibleWorkers behavior.
    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
    pub job_mode: bool,
    /// Operator-declared named tools this work requires VERIFIED on the worker
    /// (`rch exec --job --require-tool NAME`, or a project's
    /// `jobs.required_tools`).
    ///
    /// Job mode only — the compilation classifier never sets it, so ordinary
    /// offloaded builds keep their existing admission behavior exactly. Empty
    /// means no gate; a name no worker has verified excludes every worker,
    /// because the caller asked for the gate precisely because the job cannot
    /// run without the tool.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub required_tools: Vec<String>,
    /// Process ID of the hook (for active build tracking).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub hook_pid: Option<u32>,
}

fn disk_headroom_disabled(gib: &u32) -> bool {
    *gib == 0
}

/// Reason for worker selection result.
///
/// Provides context when no worker is available, enabling informative
/// fallback messages in the hook.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
#[serde(rename_all = "snake_case")]
pub enum SelectionReason {
    /// Worker assigned successfully.
    Success,
    /// No workers configured in workers.toml.
    NoWorkersConfigured,
    /// All workers are unreachable (failed health checks).
    AllWorkersUnreachable,
    /// All workers have circuits open (after repeated failures).
    AllCircuitsOpen,
    /// All workers are at capacity (no available slots).
    AllWorkersBusy,
    /// Workers were present but none passed selection health thresholds.
    NoWorkersPassedHealth,
    /// All candidate workers failed hard preflight checks.
    AllWorkersFailedPreflight,
    /// All candidate workers failed repo convergence checks (repos missing/stale/failed).
    AllWorkersFailedConvergence,
    /// Workers exist but admission was blocked by concrete capacity/preflight reasons.
    NoAdmissibleWorkers(String),
    /// No workers match required tags or preferences.
    NoMatchingWorkers,
    /// No workers have the required runtime (e.g., Bun, Node).
    NoWorkersWithRuntime(String),
    /// Internal error during selection.
    SelectionError(String),
    /// Worker assigned via affinity pinning (recent successful build).
    AffinityPinned,
    /// Worker assigned via last-success fallback (all others unavailable).
    AffinityFallback,
}

/// Wire protocol version for daemon `/select-worker` responses.
///
/// This is intentionally separate from the broader CLI API envelope version:
/// selection responses are emitted as a raw [`SelectionResponse`] on the hook
/// hot path, so the client needs a small explicit compatibility marker.
pub const SELECTION_RESPONSE_PROTOCOL_VERSION: u64 = 1;

impl std::fmt::Display for SelectionReason {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            Self::Success => write!(f, "worker assigned successfully"),
            Self::NoWorkersConfigured => write!(f, "no workers configured"),
            Self::AllWorkersUnreachable => write!(f, "all workers unreachable"),
            Self::AllCircuitsOpen => write!(f, "all worker circuits open"),
            Self::AllWorkersBusy => write!(f, "all workers at capacity"),
            Self::NoWorkersPassedHealth => write!(f, "no workers passed health thresholds"),
            Self::AllWorkersFailedPreflight => write!(f, "all workers failed preflight checks"),
            Self::AllWorkersFailedConvergence => {
                write!(f, "all workers failed repo convergence checks")
            }
            Self::NoAdmissibleWorkers(summary) => {
                write!(f, "no admissible workers: {}", summary)
            }
            Self::NoMatchingWorkers => write!(f, "no matching workers found"),
            Self::NoWorkersWithRuntime(rt) => write!(f, "no workers with {} installed", rt),
            Self::SelectionError(e) => write!(f, "selection error: {}", e),
            Self::AffinityPinned => write!(f, "worker assigned via affinity pinning"),
            Self::AffinityFallback => write!(f, "worker assigned via last-success fallback"),
        }
    }
}

// ============================================================================
// Worker Selection Strategy
// ============================================================================

/// Worker selection strategy determining how workers are chosen for jobs.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
#[serde(rename_all = "snake_case")]
pub enum SelectionStrategy {
    /// Legacy/manual control: prioritize worker.priority first, then break ties
    /// using command-aware cache/speed hints.
    Priority,
    /// Select worker with highest SpeedScore.
    /// Best for performance-critical builds with homogeneous workers.
    Fastest,
    /// Balance all factors: SpeedScore, load, health, cache affinity.
    /// Default for general use across diverse worker pools.
    #[default]
    Balanced,
    /// Prefer workers with warm caches for the project.
    /// Best for incremental builds on large codebases.
    CacheAffinity,
    /// Weighted random selection favoring fast workers but ensuring fairness.
    /// Prevents hot-spotting when many workers are available.
    FairFastest,
}

impl std::fmt::Display for SelectionStrategy {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        let name = match self {
            Self::Priority => "priority",
            Self::Fastest => "fastest",
            Self::Balanced => "balanced",
            Self::CacheAffinity => "cache_affinity",
            Self::FairFastest => "fair_fastest",
        };
        write!(f, "{}", name)
    }
}

impl std::str::FromStr for SelectionStrategy {
    type Err = String;

    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s.trim().to_lowercase().as_str() {
            "priority" => Ok(Self::Priority),
            "fastest" => Ok(Self::Fastest),
            "balanced" => Ok(Self::Balanced),
            "cache_affinity" | "cache-affinity" | "cacheaffinity" => Ok(Self::CacheAffinity),
            "fair_fastest" | "fair-fastest" | "fairfastest" => Ok(Self::FairFastest),
            _ => Err(format!(
                "unknown selection strategy '{}', expected one of: priority, fastest, balanced, cache_affinity, fair_fastest",
                s
            )),
        }
    }
}

/// Configuration for the worker selection algorithm.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SelectionConfig {
    /// Selection strategy to use.
    #[serde(default)]
    pub strategy: SelectionStrategy,
    /// Minimum success rate (0.0-1.0) for a worker to be considered.
    #[serde(default = "default_min_success_rate")]
    pub min_success_rate: f64,
    /// Factor weights for the balanced strategy.
    #[serde(default)]
    pub weights: SelectionWeightConfig,
    /// Fairness settings for fair_fastest strategy.
    #[serde(default)]
    pub fairness: FairnessConfig,
    /// Cache affinity settings for project-to-worker pinning.
    #[serde(default)]
    pub affinity: AffinityConfig,

    // Preflight health guards (bd-3eaa)
    /// Maximum load-per-core threshold. Workers exceeding this are skipped.
    /// Set to None to disable load-based filtering.
    #[serde(default = "default_max_load_per_core")]
    pub max_load_per_core: Option<f64>,
    /// Minimum free disk space in GB. Workers below this are skipped.
    /// Set to None to disable the free-space floor; per-slot disk derating still applies.
    #[serde(default = "default_min_free_gb")]
    pub min_free_gb: Option<f64>,
    /// Disk budget in GiB per concurrent slot, after reserving `min_free_gb`.
    /// Must be finite and positive. Unknown disk telemetry preserves configured slots.
    #[serde(
        default = "default_disk_gb_per_slot",
        deserialize_with = "deserialize_disk_gb_per_slot"
    )]
    pub disk_gb_per_slot: f64,
    /// Free-disk ratio at or below which rchd marks a worker Critical and
    /// gives it no slots (bd-xxck3). Read by rchd at startup.
    #[serde(default = "default_disk_critical_free_ratio")]
    pub disk_critical_free_ratio: f64,
    /// When set, the ratio rule above trips only while free space is also at
    /// or below this many GB, so a large disk is judged by absolute headroom
    /// (5% of 900 GB is 45 GB). Unset keeps the ratio rule unconditional.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub disk_critical_ratio_cap_gb: Option<f64>,
}

fn default_disk_critical_free_ratio() -> f64 {
    0.05
}

impl Default for SelectionConfig {
    fn default() -> Self {
        Self {
            strategy: SelectionStrategy::default(),
            min_success_rate: default_min_success_rate(),
            weights: SelectionWeightConfig::default(),
            fairness: FairnessConfig::default(),
            affinity: AffinityConfig::default(),
            max_load_per_core: default_max_load_per_core(),
            min_free_gb: default_min_free_gb(),
            disk_gb_per_slot: default_disk_gb_per_slot(),
            disk_critical_free_ratio: default_disk_critical_free_ratio(),
            disk_critical_ratio_cap_gb: None,
        }
    }
}

fn default_min_success_rate() -> f64 {
    0.8
}

fn default_max_load_per_core() -> Option<f64> {
    Some(2.0) // Skip workers with load/core > 2.0
}

fn default_min_free_gb() -> Option<f64> {
    Some(10.0) // Skip workers with < 10 GB free
}

fn default_disk_gb_per_slot() -> f64 {
    10.0
}

fn deserialize_disk_gb_per_slot<'de, D>(deserializer: D) -> Result<f64, D::Error>
where
    D: serde::Deserializer<'de>,
{
    let budget = f64::deserialize(deserializer)?;
    if budget.is_finite() && budget > 0.0 {
        Ok(budget)
    } else {
        Err(serde::de::Error::custom(
            "disk_gb_per_slot must be finite and greater than zero",
        ))
    }
}

/// Weight configuration for the balanced selection strategy.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SelectionWeightConfig {
    /// Weight for SpeedScore (0.0-1.0).
    #[serde(default = "default_weight_speedscore")]
    pub speedscore: f64,
    /// Weight for available slots (0.0-1.0).
    #[serde(default = "default_weight_slots")]
    pub slots: f64,
    /// Weight for health/success rate (0.0-1.0).
    #[serde(default = "default_weight_health")]
    pub health: f64,
    /// Weight for cache affinity (0.0-1.0).
    #[serde(default = "default_weight_cache")]
    pub cache: f64,
    /// Weight for network latency (0.0-1.0).
    #[serde(default = "default_weight_network")]
    pub network: f64,
    /// Weight for worker priority (0.0-1.0).
    #[serde(default = "default_weight_priority")]
    pub priority: f64,
    /// Weight for disk headroom (0.0-1.0); zero disables disk ranking.
    #[serde(
        default = "default_weight_disk",
        deserialize_with = "deserialize_weight_disk"
    )]
    pub disk: f64,
    /// Penalty multiplier for half-open circuit workers (0.0-1.0).
    #[serde(default = "default_half_open_penalty")]
    pub half_open_penalty: f64,
}

impl Default for SelectionWeightConfig {
    fn default() -> Self {
        Self {
            speedscore: default_weight_speedscore(),
            slots: default_weight_slots(),
            health: default_weight_health(),
            cache: default_weight_cache(),
            network: default_weight_network(),
            priority: default_weight_priority(),
            disk: default_weight_disk(),
            half_open_penalty: default_half_open_penalty(),
        }
    }
}

fn default_weight_speedscore() -> f64 {
    0.5
}
fn default_weight_slots() -> f64 {
    0.4
}
fn default_weight_health() -> f64 {
    0.3
}
fn default_weight_cache() -> f64 {
    0.2
}
fn default_weight_network() -> f64 {
    0.1
}
fn default_weight_priority() -> f64 {
    // Worker priority is a first-class selection factor: operators set high
    // priority on preferred boxes (e.g. fast dedicated workers) and expect them
    // to actually receive work. At the old 0.1 a top-priority worker gained at
    // most +0.1 to its balanced score — too weak to overcome a warm-cache or
    // lower-latency competitor, so the highest-priority workers could sit idle
    // indefinitely. Load/slots still throttle a high-priority worker as it fills.
    0.5
}
fn default_half_open_penalty() -> f64 {
    0.5
}

fn default_weight_disk() -> f64 {
    0.2
}

fn deserialize_weight_disk<'de, D>(deserializer: D) -> Result<f64, D::Error>
where
    D: serde::Deserializer<'de>,
{
    let weight = f64::deserialize(deserializer)?;
    if weight.is_finite() && (0.0..=1.0).contains(&weight) {
        Ok(weight)
    } else {
        Err(serde::de::Error::custom(
            "disk weight must be finite and between zero and one",
        ))
    }
}

/// Fairness settings for the fair_fastest selection strategy.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FairnessConfig {
    /// Lookback window in seconds for tracking recent selections.
    #[serde(default = "default_fairness_lookback_secs")]
    pub lookback_secs: u64,
    /// Maximum consecutive selections for a single worker before penalty.
    #[serde(default = "default_max_consecutive_selections")]
    pub max_consecutive_selections: u32,
}

impl Default for FairnessConfig {
    fn default() -> Self {
        Self {
            lookback_secs: default_fairness_lookback_secs(),
            max_consecutive_selections: default_max_consecutive_selections(),
        }
    }
}

fn default_fairness_lookback_secs() -> u64 {
    300 // 5 minutes
}

fn default_max_consecutive_selections() -> u32 {
    3
}

/// Cache affinity settings for project-to-worker pinning.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AffinityConfig {
    /// Enable affinity pinning (pin projects to last successful worker).
    #[serde(default = "default_affinity_enabled")]
    pub enabled: bool,
    /// Time window in minutes to pin a project to its last successful worker.
    /// After this window expires, normal selection resumes.
    #[serde(default = "default_affinity_pin_minutes")]
    pub pin_minutes: u64,
    /// Enable last-success fallback when no workers match selection criteria.
    /// If enabled, the daemon will attempt the last successful worker for a project
    /// when all other workers are unavailable or fail selection.
    #[serde(default = "default_last_success_fallback")]
    pub enable_last_success_fallback: bool,
    /// Minimum success rate for the last-success worker to be used as fallback.
    /// Must be lower than min_success_rate to be meaningful as a fallback.
    #[serde(default = "default_fallback_min_success_rate")]
    pub fallback_min_success_rate: f64,
}

impl Default for AffinityConfig {
    fn default() -> Self {
        Self {
            enabled: default_affinity_enabled(),
            pin_minutes: default_affinity_pin_minutes(),
            enable_last_success_fallback: default_last_success_fallback(),
            fallback_min_success_rate: default_fallback_min_success_rate(),
        }
    }
}

fn default_affinity_enabled() -> bool {
    true
}

fn default_affinity_pin_minutes() -> u64 {
    60 // 1 hour default
}

fn default_last_success_fallback() -> bool {
    true
}

fn default_fallback_min_success_rate() -> f64 {
    0.5 // More lenient than normal min_success_rate (0.8)
}

/// Details about a selected worker for remote compilation.
#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
pub struct SelectedWorker {
    /// Selected worker ID.
    pub id: WorkerId,
    /// Host address for SSH.
    pub host: String,
    /// SSH user.
    pub user: String,
    /// Path to SSH identity file.
    pub identity_file: String,
    /// Number of slots available on this worker after reservation.
    pub slots_available: u32,
    /// Worker's speed score (0-100).
    pub speed_score: f64,
    /// Operator-declared host OS (e.g. `windows`), if any. Propagated so the
    /// hook can pick the right transport (Windows workers use `C:/rch` +
    /// tar-over-ssh instead of the Unix rsync path). `None` for the historical
    /// unlabelled workers. `#[serde(default)]` keeps old daemons/hooks
    /// wire-compatible.
    #[serde(default)]
    pub declared_os: Option<String>,
}

/// Worker selection response from daemon to hook.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SelectionResponse {
    /// Selected worker details, if available.
    pub worker: Option<SelectedWorker>,
    /// Reason for the selection result.
    pub reason: SelectionReason,
    /// Optional build ID assigned by the daemon (for active build tracking).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub build_id: Option<u64>,
    /// Optional per-worker selector diagnostics for machine-readable triage.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub diagnostics: Option<SelectionDiagnostics>,
}

/// Final selector decision for a worker considered during routing.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
#[serde(rename_all = "snake_case")]
pub enum WorkerSelectionDiagnosticDecision {
    /// Worker passed hard gates and normal health threshold.
    Allow,
    /// Worker failed normal health threshold but passed fallback threshold.
    FallbackCandidate,
    /// Worker was denied by a hard gate or both health thresholds.
    Deny,
}

/// Per-worker selector facts captured at worker-selection time.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
pub struct WorkerSelectionDiagnostic {
    /// Worker ID.
    pub worker_id: WorkerId,
    /// Worker lifecycle/status label.
    pub status: String,
    /// Circuit breaker state label.
    pub circuit_state: String,
    /// Current normalized pressure state.
    pub pressure_state: String,
    /// Stable pressure reason code.
    pub pressure_reason_code: String,
    /// Rolling selector success rate, where available.
    #[serde(skip_serializing_if = "Option::is_none")]
    pub success_rate: Option<f64>,
    /// Normal selection success-rate threshold.
    pub min_success_rate: f64,
    /// Last-success fallback success-rate threshold.
    pub fallback_min_success_rate: f64,
    /// Required runtime for this command.
    pub required_runtime: RequiredRuntime,
    /// Whether this worker has the required runtime.
    pub runtime_available: bool,
    /// Available slots at selection time.
    pub available_slots: u32,
    /// Total configured slots.
    pub total_slots: u32,
    /// Estimated slots requested by this command.
    pub estimated_cores: u32,
    /// Whether this worker was excluded because it already runs this project.
    pub active_project_excluded: bool,
    /// Final per-worker selector decision.
    pub final_decision: WorkerSelectionDiagnosticDecision,
    /// Human-readable final reason.
    pub final_reason: String,
    /// Stable machine-readable reasons collected for this worker.
    pub reason_codes: Vec<String>,
}

/// Machine-readable worker selector diagnostics for a selection request.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
pub struct SelectionDiagnostics {
    /// Required runtime for this command.
    pub required_runtime: RequiredRuntime,
    /// Estimated slots requested by this command.
    pub estimated_cores: u32,
    /// Normal success-rate threshold.
    pub min_success_rate: f64,
    /// Last-success fallback success-rate threshold.
    pub fallback_min_success_rate: f64,
    /// Count of workers excluded because they already run this project.
    pub active_project_exclusion_count: usize,
    /// One diagnostic entry per configured worker.
    pub workers: Vec<WorkerSelectionDiagnostic>,
}

/// Request to release reserved worker slots.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ReleaseRequest {
    /// ID of the worker to release slots on.
    pub worker_id: WorkerId,
    /// Number of slots to release.
    pub slots: u32,
    /// Optional build ID to mark complete.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub build_id: Option<u64>,
    /// Exact client wrapper identity associated with the admitted build.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub local_wrapper_id: Option<String>,
    /// Optional exit code for the build (used to finalize active build tracking).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub exit_code: Option<i32>,
    /// Optional total duration in milliseconds (pipeline or execution).
    ///
    /// If omitted, the daemon will compute duration from the active build's start time.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub duration_ms: Option<u64>,
    /// Optional bytes transferred during the pipeline (upload + artifact download).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub bytes_transferred: Option<u64>,
    /// Optional per-phase timing breakdown for the build pipeline.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub timing: Option<CommandTimingBreakdown>,
    /// The hook classified this non-zero exit as a worker fault (missing
    /// toolchain or system library, CPU-capability signal, full disk), so the
    /// daemon records no cache warmth for it (review of GH #81).
    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
    pub worker_fault: bool,
    /// The failed remote command reported exhausted disk space or quota.
    /// After matching the durable build owner, the daemon temporarily bypasses
    /// this worker until fresh recovery probes establish disk headroom.
    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
    pub worker_disk_full: bool,
    /// Actual remote source and output roots used by the failed attempt.
    /// Recovery must measure these filesystems as well as configured roots.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub worker_disk_roots: Vec<String>,
    /// The worker's source receiver refused the sync because its copy of the
    /// project is a declared live checkout (`RCH_SYNC_LIVE_CHECKOUT_REFUSED`,
    /// bd-fbtws). The daemon keeps that project off the worker for a while.
    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
    pub worker_live_checkout_refused: bool,
}

/// Build execution phase for daemon heartbeat tracking.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
#[serde(rename_all = "snake_case")]
pub enum BuildHeartbeatPhase {
    /// Uploading sources/dependencies to the worker.
    SyncUp,
    /// Running the remote compilation/test command.
    Execute,
    /// Downloading build artifacts from the worker.
    SyncDown,
    /// Build is completing or finalizing.
    Finalize,
}

/// Request to update liveness/progress for an active build.
#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
pub struct BuildHeartbeatRequest {
    /// Active build identifier assigned by the daemon.
    pub build_id: u64,
    /// Worker executing the build.
    pub worker_id: WorkerId,
    /// Hook process ID sending the heartbeat.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub hook_pid: Option<u32>,
    /// Client-minted wrapper identity for local durable-lease correlation.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub local_wrapper_id: Option<String>,
    /// Remote file containing the process-group leader PID for cancellation.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub remote_pgid_file: Option<String>,
    /// Current build execution phase.
    pub phase: BuildHeartbeatPhase,
    /// Optional human-readable progress detail.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub detail: Option<String>,
    /// Optional monotonic progress counter.
    ///
    /// The hook increments this when it observes meaningful progress
    /// (phase transitions, new output, or richer progress signals).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub progress_counter: Option<u64>,
    /// Optional progress estimate in [0,100].
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub progress_percent: Option<f64>,
}

/// Configuration for a remote worker.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct WorkerConfig {
    /// Unique identifier for this worker.
    pub id: WorkerId,
    /// SSH hostname or IP address.
    pub host: String,
    /// SSH username.
    pub user: String,
    /// Path to SSH private key.
    pub identity_file: String,
    /// Total CPU slots available on this worker.
    pub total_slots: u32,
    /// Priority for worker selection (higher = preferred).
    #[serde(default = "default_priority")]
    pub priority: u32,
    /// Optional tags for filtering.
    ///
    /// Mostly descriptive, with one reserved form: `os:<name>` (see
    /// [`crate::declared_os`]) is a hard admission gate, written in
    /// `workers.toml` as `os = "<name>"`.
    #[serde(default)]
    pub tags: Vec<String>,
    /// Operator-declared named tool probes for this worker, written in
    /// `workers.toml` as
    /// `tools = [{ name = "clang", command = ["clang", "--version"] }]`.
    ///
    /// Unlike a tag, which is an unverified naming convention, each entry is
    /// VERIFIED by running its fixed argv on the worker as the configured user;
    /// only a zero exit makes the tool present. That is what lets
    /// `--require-tool` gate selection on evidence instead of on a promise.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub tools: Vec<WorkerToolProbe>,
}

/// One operator-declared tool probe as it appears in `workers.toml`.
///
/// Kept as plain configuration data; [`crate::capability_probe::NamedToolProbe`]
/// is the validated form, and the conversion is where a malformed name is
/// refused.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct WorkerToolProbe {
    /// The name `--require-tool` and `jobs.required_tools` refer to.
    pub name: String,
    /// The exact argv to run. Zero exit = present; anything else = absent.
    pub command: Vec<String>,
}

fn default_priority() -> u32 {
    100
}

/// Tag prefix reserved for the operator-declared host OS.
pub const OS_TAG_PREFIX: &str = "os:";

/// The host OS an operator declared for a worker, from its tags.
///
/// Written in `workers.toml` as `os = "windows"` and normalized into the
/// reserved `os:windows` tag on load, so it rides the existing tag vector
/// through every path that already carries a [`WorkerConfig`].
///
/// Declaring it makes the worker **exclusive**: selection admits it only for
/// commands that require that OS, so a Windows box cannot quietly absorb a
/// `cargo check` dispatched from Linux and return wrong-platform artifacts.
/// Workers without the tag — every historical worker — keep taking anything.
#[must_use]
pub fn declared_os(tags: &[String]) -> Option<String> {
    // `find` over the mapped values, not `find_map` over the tags: a malformed
    // `os:` entry must not mask a valid declaration later in the list, because
    // failing open here is precisely the "Windows worker silently takes Linux
    // jobs" outcome the gate exists to prevent.
    tags.iter()
        .filter_map(|tag| tag.strip_prefix(OS_TAG_PREFIX))
        .map(|os| os.trim().to_ascii_lowercase())
        .find(|os| !os.is_empty())
}

/// Build the reserved tag for a declared OS.
#[must_use]
pub fn os_tag(os: &str) -> String {
    format!("{OS_TAG_PREFIX}{}", os.trim().to_ascii_lowercase())
}

impl Default for WorkerConfig {
    fn default() -> Self {
        Self {
            id: WorkerId::new("default-worker"),
            host: "localhost".to_string(),
            user: "user".to_string(),
            identity_file: "~/.ssh/id_rsa".to_string(),
            total_slots: 4,
            priority: default_priority(),
            tags: Vec::new(),
            tools: Vec::new(),
        }
    }
}

/// Runtime capabilities detected on a worker.
///
/// These are probed during health checks and cached for routing decisions.
/// Commands requiring specific runtimes (e.g., `bun test`) can be routed
/// only to workers with the corresponding capability.
#[derive(Debug, Clone, Default, Serialize, Deserialize, JsonSchema)]
pub struct WorkerCapabilities {
    /// Rust compiler version (from `rustc --version`).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub rustc_version: Option<String>,
    /// Installed rustup toolchains, including their host triple when rustup
    /// reports one (for example `nightly-2026-07-05-x86_64-unknown-linux-gnu`).
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub rustup_toolchains: Vec<String>,
    /// Installed rustup components as `<toolchain>:<component>` facts. Component
    /// names are normalized by removing only the probed host-triple suffix, so
    /// `clippy-preview` can never satisfy a `clippy` requirement.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub rustup_components: Vec<String>,
    /// Non-fatal diagnostics gathered while probing (for example a tool binary
    /// that had to be resolved from a well-known fallback location because the
    /// caller's PATH was minimal). Empty in the common case; clients may
    /// surface these so "no facts" is distinguishable from "facts show
    /// absence" (bd-deft5).
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub probe_warnings: Vec<String>,
    /// Bun runtime version (from `bun --version`).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub bun_version: Option<String>,
    /// Node.js version (from `node --version`).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub node_version: Option<String>,
    /// npm version (from `npm --version`).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub npm_version: Option<String>,
    /// Nix version (from `nix --version`). Presence implies a usable nix binary
    /// with an accompanying `/nix/store`, gating `nix build`/`nix develop -c`
    /// routing to this worker.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub nix_version: Option<String>,
    /// Go toolchain version (from `go version`). Presence gates `go build`/
    /// `go test`/`go vet` routing to this worker.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub go_version: Option<String>,
    /// Zig version (from `zig version`). `cargo zigbuild` shells out to `zig`
    /// as the linker/C toolchain, so the bare `zig` binary is necessary but not
    /// sufficient — see [`WorkerCapabilities::has_zig`].
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub zig_version: Option<String>,
    /// cargo-zigbuild version (from `cargo-zigbuild --version`). Probed as the
    /// hyphenated binary rather than `cargo zigbuild --version`, which the
    /// subcommand's own parser rejects.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub cargo_zigbuild_version: Option<String>,
    /// x86-64 microarchitecture level (1..=4) parsed from `/proc/cpuinfo`
    /// flags on the worker (bd-6qchz / bd-68hon item 4). `Some(2)` marks a
    /// pre-AVX2 box (e.g. ovh-b's Ivy Bridge) whose build scripts SIGILL on
    /// x86-64-v3 codegen; selection deprioritizes such workers proactively
    /// while the reactive SIGILL quarantine remains the primary mechanism.
    /// `None` on non-x86 workers or when the probe cannot read cpuinfo.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub cpu_microarch_level: Option<u8>,
    /// Operator-declared named tools whose probe exited ZERO on this worker.
    ///
    /// The list is evidence, not configuration: a name appears here only
    /// because the worker ran the declared argv and it succeeded. Selection
    /// gates `--require-tool` on this, which is the difference between a
    /// verified capability and a tag somebody wrote by hand.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub tools_present: Vec<String>,
    /// Operator-declared named tools whose probe RAN and did not exit zero.
    ///
    /// Kept separate from "never declared" so an operator can tell a worker
    /// that is missing a tool from a fleet that never probed for it — the two
    /// have different fixes.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub tools_absent: Vec<String>,

    // Health metrics (bd-3eaa)
    /// Number of CPU cores on the worker.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub num_cpus: Option<u32>,
    /// 1-minute load average.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub load_avg_1: Option<f64>,
    /// 5-minute load average.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub load_avg_5: Option<f64>,
    /// 15-minute load average.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub load_avg_15: Option<f64>,
    /// Free space in GB on the tightest mount a build touches: whichever of
    /// the projects root, its alias and `/tmp` has the lowest free ratio.
    /// Drives pressure classification (a full `/tmp` breaks scp/mktemp).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub disk_free_gb: Option<f64>,
    /// Total size in GB of the mount reported in `disk_free_gb`.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub disk_total_gb: Option<f64>,
    /// Free space in GB on the filesystem that holds build trees (the fuller
    /// of the projects root and its alias; `/tmp` excluded). Sizes capacity:
    /// slot derating, disk headroom and the free-space floor (GH #78).
    /// `None` from an `rch-wkr` that predates the field.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub build_disk_free_gb: Option<f64>,
    /// Total size in GB of the mount reported in `build_disk_free_gb`.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub build_disk_total_gb: Option<f64>,
    /// Canonical path-topology preflight status (`/data/projects` + `/dp` alias).
    ///
    /// `Some(false)` indicates a hard preflight failure that should exclude this
    /// worker from remote scheduling until a subsequent successful revalidation.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub projects_root_ok: Option<bool>,
    /// Machine-readable failure reason for path-topology preflight.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub projects_root_issue: Option<String>,
    /// Unix timestamp (ms) when topology preflight last executed.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub projects_root_checked_at_unix_ms: Option<i64>,
}

impl WorkerCapabilities {
    /// Create new empty capabilities.
    pub fn new() -> Self {
        Self::default()
    }

    /// Create mock capabilities with Rust installed.
    ///
    /// Used by mock transport to simulate a worker with Rust toolchain.
    pub fn mock_with_rust() -> Self {
        Self {
            rustc_version: Some("1.85.0-nightly (mock)".to_string()),
            projects_root_ok: Some(true),
            ..Self::default()
        }
    }

    /// Whether this is a known pre-x86-64-v3 (no AVX2) x86 worker — the class
    /// whose build-script/proc-macro binaries SIGILL on v3 codegen (bd-68hon).
    /// `None` (unknown/non-x86) is NOT pre-v3: only affirmative evidence
    /// deprioritizes a worker.
    pub fn is_pre_v3_x86(&self) -> bool {
        matches!(self.cpu_microarch_level, Some(1 | 2))
    }

    /// Check if this worker has Bun installed.
    pub fn has_bun(&self) -> bool {
        self.bun_version.is_some()
    }

    /// Check if this worker has Node.js installed.
    pub fn has_node(&self) -> bool {
        self.node_version.is_some()
    }

    /// Check if this worker has Rust installed.
    pub fn has_rust(&self) -> bool {
        self.rustc_version.is_some()
    }

    /// Check whether `component` is installed for `toolchain`.
    ///
    /// Rustup reports pinned toolchains with a host-triple suffix. The request
    /// carries the project pin without that suffix, so the owner half uses the
    /// same exact-or-`-`-boundary prefix rule as toolchain admission. The
    /// component half is exact.
    pub fn has_rustup_component(&self, toolchain: &str, component: &str) -> bool {
        let owner_prefix = format!("{toolchain}-");
        self.rustup_components.iter().any(|fact| {
            fact.split_once(':').is_some_and(|(owner, name)| {
                name == component && (owner == toolchain || owner.starts_with(&owner_prefix))
            })
        })
    }

    /// Check if this worker has Nix installed (nix binary + `/nix/store`).
    pub fn has_nix(&self) -> bool {
        self.nix_version.is_some()
    }

    /// Check if this worker has the Go toolchain installed.
    pub fn has_go(&self) -> bool {
        self.go_version.is_some()
    }

    /// Check if this worker can run `cargo zigbuild` cross-compiles.
    ///
    /// Requires BOTH halves of the toolchain: the `cargo-zigbuild` cargo
    /// subcommand and the `zig` binary it drives as the linker. Either one
    /// alone fails at build time with an error the remote-result heuristics do
    /// not recognize as infrastructure (`error: no such command: 'zigbuild'` is
    /// not a rustup message), so the exit code would be surfaced verbatim
    /// instead of falling open to local. Gating both here keeps that failure
    /// out of the user's build.
    pub fn has_zig(&self) -> bool {
        self.zig_version.is_some() && self.cargo_zigbuild_version.is_some()
    }

    /// Calculate load per core (1-minute load average / num_cpus).
    /// Returns None if metrics are unavailable.
    pub fn load_per_core(&self) -> Option<f64> {
        match (self.load_avg_1, self.num_cpus) {
            (Some(load), Some(cpus)) if cpus > 0 => Some(load / cpus as f64),
            _ => None,
        }
    }

    /// Check if worker has high load (exceeds threshold).
    /// Returns None if metrics unavailable (fail-open).
    pub fn is_high_load(&self, max_load_per_core: f64) -> Option<bool> {
        self.load_per_core().map(|lpc| lpc > max_load_per_core)
    }

    /// `(free_gb, total_gb)` of the filesystem that holds build trees.
    ///
    /// Capacity questions (how many builds fit, is the free-space floor met)
    /// are about that filesystem, not about a small tmpfs `/tmp` that happens
    /// to be the tightest mount. Falls back to the tightest-mount sample when
    /// the worker does not report a build-disk sample (older `rch-wkr`), which
    /// is the conservative pre-GH-#78 behaviour.
    pub fn build_disk_gb(&self) -> (Option<f64>, Option<f64>) {
        if self.build_disk_free_gb.is_some() {
            (self.build_disk_free_gb, self.build_disk_total_gb)
        } else {
            (self.disk_free_gb, self.disk_total_gb)
        }
    }

    /// Check if the build disk is below the free-space threshold.
    /// Returns None if metrics unavailable (fail-open).
    pub fn is_low_disk(&self, min_free_gb: f64) -> Option<bool> {
        self.build_disk_gb().0.map(|free| free < min_free_gb)
    }

    /// Check if canonical path-topology preflight is healthy.
    ///
    /// Returns `None` when probe data is unavailable (fail-open behavior).
    pub fn is_topology_healthy(&self) -> Option<bool> {
        self.projects_root_ok
    }
}

/// Classify an x86-64 microarchitecture level (1..=4) from a `/proc/cpuinfo`
/// `flags` line (bd-6qchz).
///
/// Levels follow the psABI x86-64-v2/v3/v4 feature baselines, restricted to
/// flags as they appear in Linux cpuinfo. A CPU missing any flag of a level is
/// classified at the level below — e.g. ovh-b's Ivy Bridge E3-1245 V2 has AVX
/// but no `avx2`/`bmi2`/`fma`, so it classifies as 2 and its v3-codegen build
/// scripts SIGILL (the bd-68hon incident class).
#[must_use]
pub fn x86_64_microarch_level_from_flags(flags: &str) -> u8 {
    let set: std::collections::HashSet<&str> = flags.split_whitespace().collect();
    let has_all = |required: &[&str]| required.iter().all(|flag| set.contains(flag));
    // x86-64-v2: CMPXCHG16B, LAHF/SAHF, POPCNT, SSE3/SSSE3/SSE4.x.
    const V2: &[&str] = &["cx16", "lahf_lm", "popcnt", "sse4_1", "sse4_2", "ssse3"];
    // x86-64-v3: AVX/AVX2, BMI1/2, F16C, FMA, MOVBE, XSAVE (LZCNT is `abm`).
    const V3: &[&str] = &[
        "abm", "avx", "avx2", "bmi1", "bmi2", "f16c", "fma", "movbe", "xsave",
    ];
    // x86-64-v4: the AVX-512 F/BW/CD/DQ/VL baseline.
    const V4: &[&str] = &["avx512f", "avx512bw", "avx512cd", "avx512dq", "avx512vl"];
    if !has_all(V2) {
        return 1;
    }
    if !has_all(V3) {
        return 2;
    }
    if !has_all(V4) {
        return 3;
    }
    4
}

/// Path topology configuration.
///
/// Overrides the default canonical and alias project roots used by
/// [`crate::path_topology::PathTopologyPolicy`].  This is useful on platforms
/// such as macOS where SIP prevents using `/data/projects` and the user must
/// choose a different root (e.g. `~/Projects`).
///
/// ```toml
/// [path_topology]
/// canonical_root = "/Users/me/Projects"
/// alias_root = "/Users/me/p"
/// ```
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct PathTopologyConfig {
    /// Canonical project root directory.
    /// Defaults to `/data/projects` when absent.
    #[serde(default)]
    pub canonical_root: Option<String>,

    /// Alias (symlink) root that points at the canonical root.
    /// Defaults to `/dp` when absent.
    #[serde(default)]
    pub alias_root: Option<String>,
}

impl PathTopologyConfig {
    /// Build a [`crate::path_topology::PathTopologyPolicy`] from this config,
    /// falling back to the compiled-in defaults for any field left unset.
    ///
    /// Empty strings are treated as unset (fall back to defaults).
    /// Tilde prefixes (`~/...`) are expanded to the user's home directory.
    pub fn to_policy(&self) -> crate::path_topology::PathTopologyPolicy {
        let canonical = self
            .canonical_root
            .as_deref()
            .filter(|s| !s.is_empty())
            .map(|s| shellexpand::tilde(s).into_owned())
            .unwrap_or_else(|| crate::path_topology::DEFAULT_CANONICAL_PROJECT_ROOT.to_string());
        let alias = self
            .alias_root
            .as_deref()
            .filter(|s| !s.is_empty())
            .map(|s| shellexpand::tilde(s).into_owned())
            .unwrap_or_else(|| crate::path_topology::DEFAULT_ALIAS_PROJECT_ROOT.to_string());
        crate::path_topology::PathTopologyPolicy::new(
            PathBuf::from(canonical),
            PathBuf::from(alias),
        )
    }
}

/// RCH configuration.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct RchConfig {
    #[serde(default)]
    pub general: GeneralConfig,
    #[serde(default)]
    pub compilation: CompilationConfig,
    #[serde(default)]
    pub transfer: TransferConfig,
    #[serde(default)]
    pub environment: EnvironmentConfig,
    #[serde(default)]
    pub circuit: CircuitBreakerConfig,
    #[serde(default)]
    pub output: OutputConfig,
    #[serde(default)]
    pub self_healing: SelfHealingConfig,
    #[serde(default)]
    pub self_test: SelfTestConfig,
    /// Worker selection algorithm configuration.
    #[serde(default)]
    pub selection: SelectionConfig,
    /// Execution allowlist configuration (bd-785w).
    #[serde(default)]
    pub execution: ExecutionConfig,
    /// Worker health alerting configuration (daemon).
    #[serde(default)]
    pub alerts: AlertsConfig,
    /// Fleet operation configuration (bd-rs7w.2).
    #[serde(default)]
    pub fleet: FleetConfig,
    /// Path topology overrides for project root directories.
    #[serde(default)]
    pub path_topology: PathTopologyConfig,
    /// Doctor / reliability subsystem configuration (verdict webhooks).
    #[serde(default)]
    pub doctor: DoctorConfig,
    /// Session-history remediation knobs: the central schema and default policy
    /// for temporary bypass, auto-rejoin, reconciliation, proof, incident
    /// ledger, build-root, pooled targets, telemetry freshness, log retention,
    /// disk pressure, and smoke defaults (bd-...remediation-ocv9i.17.1).
    #[serde(default)]
    pub remediation: crate::remediation_config::RemediationConfig,
    /// Layer 0 configuration pack knobs (bd-bqu38 / M-1 §149). Defaults OFF;
    /// injected as CARGO_PROFILE_* env vars into remote cargo builds.
    #[serde(default)]
    pub layer0: Layer0Config,
    /// Daemon status API over TCP for other machines on the tailnet
    /// (`[api]`, bd-2f5ms). Off unless `bind` is set.
    #[serde(default)]
    pub api: ApiConfig,
    /// Where the fleet dashboard lives (`[dashboard]`), for `rch web`.
    #[serde(default)]
    pub dashboard: DashboardConfig,
    /// Job-mode defaults (`[jobs]`), for `rch exec --job`.
    #[serde(default)]
    pub jobs: JobsConfig,
}

/// Job-mode configuration (`[jobs]`).
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct JobsConfig {
    /// Named tools every job from this project requires a worker to have
    /// VERIFIED, as though each were passed with `--require-tool`.
    ///
    /// This is the project saying "my jobs cannot run on a worker without
    /// these", so it applies on top of any `--require-tool` given on the
    /// command line rather than being replaced by it. It affects job mode
    /// ONLY: ordinary offloaded compilation keeps its existing admission
    /// behavior, because a project-wide requirement that silently narrowed
    /// every build's worker pool would be a surprising way to lose the fleet.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub required_tools: Vec<String>,
}

/// Daemon status API over TCP (`[api]`).
///
/// rchd's rich state (`/status`, `/workers/capabilities`) lives on a `0600`
/// Unix socket, which only the local user can read. This section opens the
/// same JSON on a TCP listener so agents and the fleet dashboard on OTHER
/// machines can ask a dispatcher "why are you local-only right now?" without
/// ssh. It is meant for a tailnet: the bind address must be loopback or a
/// Tailscale address (100.64.0.0/10, fd7a:115c:a1e0::/48) unless
/// `allow_any_addr` is set, and every status route requires the bearer token
/// unless `token` is unset AND `no_token` is explicitly true.
///
/// ```toml
/// [api]
/// bind = "tailscale"       # or "tailscale:9101", "100.68.51.94:9101", "127.0.0.1:9101"
/// token = "…"              # or token_file = "~/.config/rch/api.token"
/// ```
#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq, Eq)]
#[serde(default)]
pub struct ApiConfig {
    /// Where to listen. `"tailscale"` resolves this machine's Tailscale IPv4
    /// at startup (default port 9101); `"tailscale:PORT"`, `"IP:PORT"` or
    /// `"[v6]:PORT"` are taken literally. Unset (empty) = API off.
    pub bind: String,
    /// Bearer token required on `/status` and `/workers/capabilities`.
    pub token: Option<String>,
    /// File holding the bearer token (trimmed). Takes precedence over `token`.
    pub token_file: Option<String>,
    /// Serve the status routes with NO token. Only honoured when no token is
    /// configured; must be set explicitly so an empty token is never an
    /// accidental open door.
    pub no_token: bool,
    /// Permit a bind address outside loopback and the Tailscale ranges.
    pub allow_any_addr: bool,
}

/// Where the fleet dashboard is served (`[dashboard]`).
#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq, Eq)]
#[serde(default)]
pub struct DashboardConfig {
    /// URL `rch web` opens, e.g. `https://rch-fleet.vercel.app`.
    pub url: Option<String>,
}

/// Doctor reliability subsystem configuration.
///
/// Currently carries the verdict-transition webhook surface. Kept as its own
/// section (`[doctor]`) so future doctor knobs land here rather than bloating
/// the top-level config.
#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq, Eq)]
pub struct DoctorConfig {
    /// Verdict-transition webhook destinations.
    #[serde(default)]
    pub webhooks: DoctorWebhooksConfig,
}

/// Webhook notification surface for reliability verdict transitions.
///
/// Fires when `rch doctor --reliability --watch` observes a verdict change
/// (`Healthy → Degraded`, `Degraded → Failing`, etc.). Each endpoint selects
/// which transitions it cares about and how the payload is formatted.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
pub struct DoctorWebhooksConfig {
    /// Named webhook destinations. Empty = no webhook dispatch.
    #[serde(default)]
    pub endpoints: Vec<DoctorWebhookEndpoint>,
    /// Maximum number of in-flight webhook deliveries. Bounds memory and
    /// outbound connection pressure: when the queue is full, a transition's
    /// delivery is dropped (and journaled) rather than growing unboundedly.
    /// Per-process, shared across all endpoints.
    #[serde(default = "default_webhook_max_inflight")]
    pub max_inflight: usize,
}

// Manual `Default` so the in-code default matches the serde default
// (`#[derive(Default)]` would give `max_inflight = 0`, diverging from a
// from-TOML load that fills in 16).
impl Default for DoctorWebhooksConfig {
    fn default() -> Self {
        Self {
            endpoints: Vec::new(),
            max_inflight: default_webhook_max_inflight(),
        }
    }
}

/// A single webhook destination for verdict transitions.
///
/// SECRETS ARE NEVER STORED HERE. Routing keys, bearer tokens, and signing
/// secrets are referenced by the *name* of an environment variable
/// (`*_env` fields) read at dispatch time. There is deliberately no
/// `secret` / `routing_key` / `bearer_token` / `password` field on this
/// struct (enforced by `test_no_inline_secret_fields`).
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
pub struct DoctorWebhookEndpoint {
    /// Operator-facing name (appears in logs and the audit trail).
    pub name: String,
    /// Destination URL.
    pub url: String,
    /// Payload format.
    #[serde(default)]
    pub format: DoctorWebhookFormat,
    /// Transition predicates that arm this endpoint. Empty = never fire.
    /// Supported: `healthy_to_degraded`, `degraded_to_failing`,
    /// `any_to_failing`, `any_to_degraded`, `failing_to_healthy`,
    /// `degraded_to_healthy`, `any_to_healthy`, `any_to_any`.
    #[serde(default)]
    pub on_transitions: Vec<String>,
    /// Maximum retry attempts after the first try fails.
    #[serde(default = "default_webhook_retry_max")]
    pub retry_max: u32,
    /// Base backoff between retries (milliseconds); doubles each attempt.
    #[serde(default = "default_webhook_retry_backoff_ms")]
    pub retry_backoff_ms: u64,
    /// Per-attempt request timeout (milliseconds).
    #[serde(default = "default_webhook_timeout_ms")]
    pub timeout_ms: u64,
    /// Per-endpoint kill switch.
    #[serde(default = "default_true")]
    pub enabled: bool,
    /// Name of the env var holding the PagerDuty Events API routing key.
    /// Required for `format = "pagerduty"`. Read at runtime, never stored.
    #[serde(default)]
    pub routing_key_env: Option<String>,
    /// Name of the env var holding a bearer token sent as
    /// `Authorization: Bearer <token>`. Read at runtime, never stored.
    #[serde(default)]
    pub bearer_token_env: Option<String>,
    /// Name of the env var holding an HMAC-SHA256 signing secret. When set,
    /// the body is signed and the signature sent as `X-RCH-Signature:
    /// sha256=<hex>`. Read at runtime, never stored.
    #[serde(default)]
    pub signing_secret_env: Option<String>,
}

/// Built-in webhook payload formats.
#[derive(Debug, Clone, Copy, Default, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum DoctorWebhookFormat {
    /// Opinionated RCH JSON schema (default).
    #[default]
    GenericJson,
    /// Slack Block Kit payload.
    Slack,
    /// PagerDuty Events API v2 payload.
    Pagerduty,
}

fn default_webhook_max_inflight() -> usize {
    16
}

fn default_webhook_retry_max() -> u32 {
    3
}

fn default_webhook_retry_backoff_ms() -> u64 {
    200
}

fn default_webhook_timeout_ms() -> u64 {
    5000
}

/// Worker health alerting configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AlertsConfig {
    /// Enable or disable alert generation in the daemon.
    #[serde(default = "default_true")]
    pub enabled: bool,
    /// Suppress duplicate alerts for this many seconds.
    #[serde(default = "default_alert_suppress_duplicates_secs")]
    pub suppress_duplicates_secs: u64,
    /// How long to keep a cleared alert visible (with state
    /// `cleared_pending_clean`) before removing it entirely. Prevents
    /// transient, self-healing circuit-breaker warnings from lingering on
    /// `rch status` long after the fleet returned to healthy (bd-3ogaz).
    #[serde(default = "default_alert_cleared_retention_secs")]
    pub cleared_retention_secs: u64,
    /// Webhook configuration for external notifications.
    #[serde(default)]
    pub webhook: Option<WebhookConfig>,
}

/// Webhook configuration for alert dispatch.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct WebhookConfig {
    /// Webhook endpoint URL (e.g., Slack/Discord webhook URL).
    pub url: Option<String>,
    /// Secret for HMAC-SHA256 signing (optional).
    #[serde(default)]
    pub secret: Option<String>,
    /// Timeout in seconds for webhook requests.
    #[serde(default = "default_webhook_timeout_secs")]
    pub timeout_secs: u64,
    /// Number of retries on failure.
    #[serde(default = "default_webhook_retry_count")]
    pub retry_count: u32,
    /// Events to send (empty = all events).
    /// Supported: worker_offline, worker_degraded, circuit_open, all_workers_offline
    #[serde(default)]
    pub events: Vec<String>,
}

fn default_webhook_timeout_secs() -> u64 {
    5
}

fn default_webhook_retry_count() -> u32 {
    3
}

fn default_alert_suppress_duplicates_secs() -> u64 {
    300
}

fn default_alert_cleared_retention_secs() -> u64 {
    300
}

impl Default for AlertsConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            suppress_duplicates_secs: default_alert_suppress_duplicates_secs(),
            cleared_retention_secs: default_alert_cleared_retention_secs(),
            webhook: None,
        }
    }
}

/// Fleet operation configuration.
///
/// Controls timeouts, thresholds, and behavior for fleet management commands
/// like `rch fleet preflight`, `rch fleet status`, and `rch fleet deploy`.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FleetConfig {
    /// SSH connection timeout in seconds.
    #[serde(default = "default_fleet_ssh_connect_timeout_secs")]
    pub ssh_connect_timeout_secs: u64,
    /// SSH command execution timeout in seconds.
    #[serde(default = "default_fleet_ssh_command_timeout_secs")]
    pub ssh_command_timeout_secs: u64,
    /// Minimum required disk space in MB for worker health.
    #[serde(default = "default_fleet_min_disk_space_mb")]
    pub min_disk_space_mb: u64,
    /// Maximum acceptable load average for worker health.
    ///
    /// Not read anywhere yet: setting it changes nothing. Selection already
    /// accounts for load through telemetry pressure and SpeedScore.
    #[serde(default = "default_fleet_max_load_average")]
    pub max_load_average: f64,
    /// Maximum number of concurrent SSH connections.
    #[serde(default = "default_fleet_max_concurrent_workers")]
    pub max_concurrent_workers: usize,
    /// Number of retry attempts for transient failures.
    #[serde(default = "default_fleet_retry_count")]
    pub retry_count: u32,
    /// Delay between retries in milliseconds.
    #[serde(default = "default_fleet_retry_delay_ms")]
    pub retry_delay_ms: u64,
}

fn default_fleet_ssh_connect_timeout_secs() -> u64 {
    10
}

fn default_fleet_ssh_command_timeout_secs() -> u64 {
    30
}

fn default_fleet_min_disk_space_mb() -> u64 {
    500
}

fn default_fleet_max_load_average() -> f64 {
    10.0
}

fn default_fleet_max_concurrent_workers() -> usize {
    10
}

fn default_fleet_retry_count() -> u32 {
    2
}

fn default_fleet_retry_delay_ms() -> u64 {
    1000
}

impl Default for FleetConfig {
    fn default() -> Self {
        Self {
            ssh_connect_timeout_secs: default_fleet_ssh_connect_timeout_secs(),
            ssh_command_timeout_secs: default_fleet_ssh_command_timeout_secs(),
            min_disk_space_mb: default_fleet_min_disk_space_mb(),
            max_load_average: default_fleet_max_load_average(),
            max_concurrent_workers: default_fleet_max_concurrent_workers(),
            retry_count: default_fleet_retry_count(),
            retry_delay_ms: default_fleet_retry_delay_ms(),
        }
    }
}

impl FleetConfig {
    /// Get SSH connect timeout as Duration.
    pub fn ssh_connect_timeout(&self) -> std::time::Duration {
        std::time::Duration::from_secs(self.ssh_connect_timeout_secs)
    }

    /// Get SSH command timeout as Duration.
    pub fn ssh_command_timeout(&self) -> std::time::Duration {
        std::time::Duration::from_secs(self.ssh_command_timeout_secs)
    }

    /// Get retry delay as Duration.
    pub fn retry_delay(&self) -> std::time::Duration {
        std::time::Duration::from_millis(self.retry_delay_ms)
    }
}

/// Box-level role (bd-wywsj): what this MACHINE is for, so "never
/// build locally" is configuration rather than a per-call env var.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum BoxRole {
    /// Orchestrator box: offloadable builds default FAIL-CLOSED +
    /// queue (RCH_REQUIRE_REMOTE semantics) — a saturated fleet waits
    /// for a worker instead of repopulating local rustc. Explicit
    /// per-call env still overrides.
    Dispatcher,
    /// The box IS the compute: fail-open/local; `rch shim install`
    /// refuses here.
    Worker,
    /// Today's behavior (fully backward compatible).
    #[default]
    Hybrid,
}

impl BoxRole {
    /// Stable lowercase token.
    #[must_use]
    pub fn as_str(self) -> &'static str {
        match self {
            BoxRole::Dispatcher => "dispatcher",
            BoxRole::Worker => "worker",
            BoxRole::Hybrid => "hybrid",
        }
    }
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct GeneralConfig {
    /// Whether RCH is enabled.
    #[serde(default = "default_true")]
    pub enabled: bool,
    /// Box-level role (bd-wywsj): dispatcher boxes default fail-closed
    /// + queue for offloadable builds; worker boxes are the compute.
    #[serde(default)]
    pub role: BoxRole,
    /// Force local execution for this project (never offload), even if the command is classified.
    ///
    /// Intended for sensitive projects or cases where local state must be authoritative.
    #[serde(default)]
    pub force_local: bool,
    /// Force remote execution for this project (always attempt offload) when safe.
    ///
    /// This bypasses heuristic gating (e.g. confidence threshold overrides) but still respects
    /// structural safety checks and NEVER_INTERCEPT patterns.
    #[serde(default)]
    pub force_remote: bool,
    /// Log level (trace, debug, info, warn, error).
    #[serde(default = "default_log_level")]
    pub log_level: String,
    /// Path to Unix socket for daemon communication.
    #[serde(default = "default_socket_path")]
    pub socket_path: String,
}

/// Environment variable passthrough configuration.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct EnvironmentConfig {
    /// Allowlist of environment variables to forward to remote workers.
    #[serde(default)]
    pub allowlist: Vec<String>,
}

impl Default for GeneralConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            role: BoxRole::default(),
            force_local: false,
            force_remote: false,
            log_level: "info".to_string(),
            socket_path: default_socket_path(),
        }
    }
}

/// Visibility level for hook output.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default, JsonSchema)]
#[serde(rename_all = "lowercase")]
pub enum OutputVisibility {
    #[default]
    None,
    Summary,
    Verbose,
}

impl std::fmt::Display for OutputVisibility {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        let value = match self {
            OutputVisibility::None => "none",
            OutputVisibility::Summary => "summary",
            OutputVisibility::Verbose => "verbose",
        };
        write!(f, "{}", value)
    }
}

impl std::str::FromStr for OutputVisibility {
    type Err = ();

    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s.trim().to_lowercase().as_str() {
            "none" | "silent" | "quiet" => Ok(Self::None),
            "summary" | "short" => Ok(Self::Summary),
            "verbose" | "debug" => Ok(Self::Verbose),
            _ => Err(()),
        }
    }
}

/// Color mode for remote command output.
///
/// Controls whether ANSI color codes are preserved when streaming
/// output from remote compilation commands.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
#[serde(rename_all = "lowercase")]
pub enum ColorMode {
    /// Force color output regardless of terminal detection.
    /// Sets CARGO_TERM_COLOR=always and similar environment variables.
    #[default]
    Always,
    /// Let the remote command detect terminal capabilities (may lose colors).
    Auto,
    /// Disable color output entirely.
    Never,
}

impl std::fmt::Display for ColorMode {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        let value = match self {
            ColorMode::Always => "always",
            ColorMode::Auto => "auto",
            ColorMode::Never => "never",
        };
        write!(f, "{}", value)
    }
}

impl std::str::FromStr for ColorMode {
    type Err = ();

    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s.trim().to_lowercase().as_str() {
            "always" | "force" | "yes" | "true" => Ok(Self::Always),
            "auto" | "detect" => Ok(Self::Auto),
            "never" | "none" | "no" | "false" => Ok(Self::Never),
            _ => Err(()),
        }
    }
}

fn default_color_mode() -> ColorMode {
    ColorMode::default()
}

/// Hook output configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct OutputConfig {
    /// Output visibility level for hook messages.
    #[serde(default = "default_output_visibility")]
    pub visibility: OutputVisibility,
    /// Whether the first-run success message has been shown.
    #[serde(default)]
    pub first_run_complete: bool,
    /// Color mode for remote command output.
    /// Controls whether ANSI color codes are preserved in remote output.
    #[serde(default = "default_color_mode")]
    pub color_mode: ColorMode,
}

impl Default for OutputConfig {
    fn default() -> Self {
        Self {
            visibility: OutputVisibility::None,
            first_run_complete: false,
            color_mode: ColorMode::default(),
        }
    }
}

// ============================================================================
// Self-Healing Configuration
// ============================================================================

fn default_autostart_cooldown_secs() -> u64 {
    30
}

fn default_autostart_timeout_secs() -> u64 {
    3
}

/// Log level for self-healing diagnostic output.
///
/// Used to filter `tracing` events emitted by the self-healing subsystem
/// (hook auto-start, daemon hook-install). Read from
/// `RCH_SELF_HEALING_LOG_LEVEL=debug|info|warn|error` (case-insensitive).
#[derive(Debug, Clone, Copy, Serialize, Deserialize, Default, PartialEq, Eq)]
#[serde(rename_all = "lowercase")]
pub enum SelfHealingLogLevel {
    Debug,
    #[default]
    Info,
    Warn,
    Error,
}

impl SelfHealingLogLevel {
    /// Parse a human-friendly level name. Returns `None` for unknown input.
    pub fn from_env_str(s: &str) -> Option<Self> {
        match s.trim().to_ascii_lowercase().as_str() {
            "debug" => Some(Self::Debug),
            "info" => Some(Self::Info),
            "warn" | "warning" => Some(Self::Warn),
            "error" => Some(Self::Error),
            _ => None,
        }
    }
}

/// Self-healing behaviors for the hook and daemon.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SelfHealingConfig {
    /// Whether the hook should attempt to auto-start the daemon.
    #[serde(default = "default_true")]
    pub hook_starts_daemon: bool,
    /// Whether the daemon should auto-install Claude Code hooks on startup.
    #[serde(default = "default_true")]
    pub daemon_installs_hooks: bool,
    /// Cooldown between auto-start attempts in seconds.
    #[serde(default = "default_autostart_cooldown_secs")]
    pub auto_start_cooldown_secs: u64,
    /// Time to wait for the daemon socket after spawning in seconds.
    /// Also aliased as `daemon_start_timeout` for backwards compatibility.
    #[serde(
        default = "default_autostart_timeout_secs",
        alias = "daemon_start_timeout"
    )]
    pub auto_start_timeout_secs: u64,
    /// Log level for self-healing diagnostic events. Default: info.
    /// Override via `RCH_SELF_HEALING_LOG_LEVEL=debug|info|warn|error`.
    ///
    /// Parsed but not applied yet: self-healing events log at their own
    /// levels under the normal `RUST_LOG` / `general.log_level` filter.
    #[serde(default)]
    pub self_healing_log_level: SelfHealingLogLevel,
}

impl Default for SelfHealingConfig {
    fn default() -> Self {
        Self {
            hook_starts_daemon: default_true(),
            daemon_installs_hooks: default_true(),
            auto_start_cooldown_secs: default_autostart_cooldown_secs(),
            auto_start_timeout_secs: default_autostart_timeout_secs(),
            self_healing_log_level: SelfHealingLogLevel::default(),
        }
    }
}

impl SelfHealingConfig {
    /// Apply environment variable overrides to this config.
    ///
    /// Supported environment variables:
    /// - `RCH_NO_SELF_HEALING=1` - Master switch to disable all self-healing
    /// - `RCH_HOOK_STARTS_DAEMON=0|1` - Control hook auto-starting daemon
    /// - `RCH_DAEMON_INSTALLS_HOOKS=0|1` - Control daemon auto-installing hooks
    /// - `RCH_AUTO_START_TIMEOUT_SECS=<seconds>` - Max wait for daemon start
    /// - `RCH_AUTO_START_COOLDOWN_SECS=<seconds>` - Min time between auto-starts
    pub fn with_env_overrides(mut self) -> Self {
        // Master disable switch
        if let Ok(val) = std::env::var("RCH_NO_SELF_HEALING")
            && (val == "1" || val.eq_ignore_ascii_case("true"))
        {
            self.hook_starts_daemon = false;
            self.daemon_installs_hooks = false;
            return self;
        }

        // Individual toggles
        if let Ok(val) = std::env::var("RCH_HOOK_STARTS_DAEMON") {
            self.hook_starts_daemon = val != "0" && !val.eq_ignore_ascii_case("false");
        }
        if let Ok(val) = std::env::var("RCH_DAEMON_INSTALLS_HOOKS") {
            self.daemon_installs_hooks = val != "0" && !val.eq_ignore_ascii_case("false");
        }

        // Numeric settings
        if let Ok(val) = std::env::var("RCH_AUTO_START_TIMEOUT_SECS")
            && let Ok(secs) = val.parse()
        {
            self.auto_start_timeout_secs = secs;
        }
        if let Ok(val) = std::env::var("RCH_AUTO_START_COOLDOWN_SECS")
            && let Ok(secs) = val.parse()
        {
            self.auto_start_cooldown_secs = secs;
        }
        // br-4zf3p: log-level override.
        if let Ok(val) = std::env::var("RCH_SELF_HEALING_LOG_LEVEL") {
            if let Some(level) = SelfHealingLogLevel::from_env_str(&val) {
                self.self_healing_log_level = level;
            } else {
                tracing::warn!(
                    target: "rch::self_healing",
                    invalid = %val,
                    "RCH_SELF_HEALING_LOG_LEVEL has invalid value; using default=info"
                );
            }
        }

        self
    }
}

// ============================================================================
// Self-Test Configuration
// ============================================================================

/// Action to take when a scheduled self-test fails.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
#[serde(rename_all = "snake_case")]
pub enum SelfTestFailureAction {
    /// Only emit an alert/log entry.
    #[default]
    Alert,
    /// Disable the failing worker automatically.
    DisableWorker,
    /// Alert and disable the worker.
    AlertAndDisable,
}

/// Which workers to include in self-test runs.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[serde(untagged)]
pub enum SelfTestWorkers {
    /// Special value "all" (or a single worker id string).
    All(String),
    /// Explicit worker list.
    List(Vec<WorkerId>),
}

impl Default for SelfTestWorkers {
    fn default() -> Self {
        Self::All("all".to_string())
    }
}

impl SelfTestWorkers {
    /// Resolve to an explicit list if configured, or None to indicate "all".
    pub fn resolve(&self) -> Option<Vec<WorkerId>> {
        match self {
            SelfTestWorkers::All(value) => {
                if value.eq_ignore_ascii_case("all") {
                    None
                } else {
                    Some(vec![WorkerId::new(value.clone())])
                }
            }
            SelfTestWorkers::List(list) => Some(list.clone()),
        }
    }
}

fn default_self_test_retry_count() -> u32 {
    3
}

fn default_self_test_retry_delay() -> String {
    "5m".to_string()
}

fn default_self_test_enabled() -> bool {
    false
}

/// Self-test scheduling and behavior configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SelfTestConfig {
    /// Enable scheduled self-tests.
    #[serde(default = "default_self_test_enabled")]
    pub enabled: bool,
    /// Cron schedule (e.g., "0 3 * * *").
    #[serde(default)]
    pub schedule: Option<String>,
    /// Interval duration (e.g., "24h").
    #[serde(default)]
    pub interval: Option<String>,
    /// Which workers to test (default: "all").
    #[serde(default)]
    pub workers: SelfTestWorkers,
    /// Action on failure.
    #[serde(default)]
    pub on_failure: SelfTestFailureAction,
    /// Retry count for failed tests.
    #[serde(default = "default_self_test_retry_count")]
    pub retry_count: u32,
    /// Delay between retries.
    #[serde(default = "default_self_test_retry_delay")]
    pub retry_delay: String,
}

impl Default for SelfTestConfig {
    fn default() -> Self {
        Self {
            enabled: default_self_test_enabled(),
            schedule: None,
            interval: None,
            workers: SelfTestWorkers::default(),
            on_failure: SelfTestFailureAction::default(),
            retry_count: default_self_test_retry_count(),
            retry_delay: default_self_test_retry_delay(),
        }
    }
}

/// Circuit breaker configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CircuitBreakerConfig {
    /// Consecutive failures to open circuit.
    #[serde(default = "default_circuit_failure_threshold")]
    pub failure_threshold: u32,
    /// Consecutive successes to close circuit.
    #[serde(default = "default_circuit_success_threshold")]
    pub success_threshold: u32,
    /// Error rate threshold (0.0-1.0) to open circuit.
    #[serde(default = "default_circuit_error_rate_threshold")]
    pub error_rate_threshold: f64,
    /// Rolling window size in seconds.
    #[serde(default = "default_circuit_window_secs")]
    pub window_secs: u64,
    /// Cooldown duration before half-open (seconds).
    #[serde(default = "default_circuit_open_cooldown_secs")]
    pub open_cooldown_secs: u64,
    /// Max concurrent probes in half-open.
    #[serde(default = "default_circuit_half_open_max_probes")]
    pub half_open_max_probes: u32,
    /// Consecutive *command-class* failures to open the circuit.
    ///
    /// Command-class evidence is an operation of the same shape as dispatched
    /// work — a remote binary that must run to completion (the telemetry poll,
    /// or a health probe whose session connected but whose command never
    /// finished). It is tracked separately from the plain liveness probe
    /// because a host under fork exhaustion or swap thrash keeps answering
    /// `echo` over a warm connection while every real command times out;
    /// interleaved liveness successes would otherwise keep resetting
    /// `failure_threshold` and the breaker would never trip.
    #[serde(default = "default_circuit_command_failure_threshold")]
    pub command_failure_threshold: u32,
    /// Upper bound (seconds) for the open cooldown after repeated reopens.
    ///
    /// Each Open -> HalfOpen -> Open cycle without an intervening Closed
    /// doubles the cooldown (`open_cooldown_secs`, x2, x4, ...) up to this
    /// cap, so a flapping worker converges to Open instead of being
    /// re-admitted every cooldown. Set equal to `open_cooldown_secs` to
    /// disable the backoff.
    #[serde(default = "default_circuit_open_cooldown_max_secs")]
    pub open_cooldown_max_secs: u64,
}

impl Default for CircuitBreakerConfig {
    fn default() -> Self {
        Self {
            failure_threshold: default_circuit_failure_threshold(),
            success_threshold: default_circuit_success_threshold(),
            error_rate_threshold: default_circuit_error_rate_threshold(),
            window_secs: default_circuit_window_secs(),
            open_cooldown_secs: default_circuit_open_cooldown_secs(),
            half_open_max_probes: default_circuit_half_open_max_probes(),
            command_failure_threshold: default_circuit_command_failure_threshold(),
            open_cooldown_max_secs: default_circuit_open_cooldown_max_secs(),
        }
    }
}

impl CircuitBreakerConfig {
    /// Effective cooldown (ms) for the given reopen streak, with exponential
    /// backoff capped at `open_cooldown_max_secs`. `reopen_streak` counts
    /// Open transitions since the circuit last closed (1 = first open).
    pub fn cooldown_ms_for_streak(&self, reopen_streak: u32) -> u64 {
        let base_ms = self.open_cooldown_secs.saturating_mul(1000);
        let cap_ms = self
            .open_cooldown_max_secs
            .saturating_mul(1000)
            .max(base_ms);
        // Exponent bounded so the shift cannot overflow; the cap clamps anyway.
        let exponent = reopen_streak.saturating_sub(1).min(16);
        base_ms.saturating_mul(1u64 << exponent).min(cap_ms)
    }
}

/// Circuit breaker statistics for tracking state transitions.
///
/// Maintains a rolling window of successes and failures to determine
/// when the circuit should open, close, or enter half-open state.
#[derive(Debug, Clone, Default)]
pub struct CircuitStats {
    /// Current circuit state.
    state: CircuitState,
    /// Consecutive failure count (reset on success).
    consecutive_failures: u32,
    /// Consecutive success count in half-open state (reset on failure).
    consecutive_successes: u32,
    /// Total successes in the current rolling window.
    window_successes: u32,
    /// Total failures in the current rolling window.
    window_failures: u32,
    /// Timestamp when circuit entered Open state (epoch millis).
    opened_at: Option<u64>,
    /// Timestamp of last state change (epoch millis).
    last_state_change: u64,
    /// Current active probes in half-open state.
    active_probes: u32,
    /// Consecutive command-class failures (telemetry poll exhaustion, probe
    /// sessions that connected but never completed). Reset only by a
    /// command-class success or a circuit close — NOT by a liveness success,
    /// which is the whole point (see `CircuitBreakerConfig::command_failure_threshold`).
    consecutive_command_failures: u32,
    /// Number of Open transitions since the circuit last closed. Drives the
    /// cooldown backoff; 0 while Closed-since-recovery.
    reopen_streak: u32,
    /// Recent health check results (true=success, false=failure).
    /// Used for history visualization in status display.
    recent_results: Vec<bool>,
}

/// Maximum number of recent results to keep for history visualization.
const CIRCUIT_HISTORY_SIZE: usize = 10;

impl CircuitStats {
    /// Create new circuit stats in the closed state.
    pub fn new() -> Self {
        Self {
            state: CircuitState::Closed,
            consecutive_failures: 0,
            consecutive_successes: 0,
            window_successes: 0,
            window_failures: 0,
            opened_at: None,
            last_state_change: Self::now_millis(),
            active_probes: 0,
            consecutive_command_failures: 0,
            reopen_streak: 0,
            recent_results: Vec::with_capacity(CIRCUIT_HISTORY_SIZE),
        }
    }

    /// Consecutive command-class failures observed since the last
    /// command-class success (or circuit close).
    pub fn consecutive_command_failures(&self) -> u32 {
        self.consecutive_command_failures
    }

    /// Open transitions since the circuit last closed (0 when it has not
    /// opened since recovery). Exposed for diagnostics.
    pub fn reopen_streak(&self) -> u32 {
        self.reopen_streak
    }

    /// Cooldown currently in force (ms), including reopen backoff.
    fn effective_cooldown_ms(&self, config: &CircuitBreakerConfig) -> u64 {
        config.cooldown_ms_for_streak(self.reopen_streak.max(1))
    }

    /// Get the current circuit state.
    pub fn state(&self) -> CircuitState {
        self.state
    }

    /// Get the timestamp when the circuit opened (if open).
    pub fn opened_at(&self) -> Option<u64> {
        self.opened_at
    }

    /// Get the timestamp of the last state change.
    pub fn last_state_change(&self) -> u64 {
        self.last_state_change
    }

    /// Get the consecutive failure count.
    pub fn consecutive_failures(&self) -> u32 {
        self.consecutive_failures
    }

    /// Get the error rate in the current window (0.0-1.0).
    ///
    /// Computed over the bounded recent-results window (the last
    /// `CIRCUIT_HISTORY_SIZE` health checks) rather than the unbounded lifetime
    /// `window_successes`/`window_failures` counters. A long-lived daemon polling
    /// distant workers accumulates occasional transient failures over days; with a
    /// lifetime ratio a currently-healthy worker could stay gated (circuit kept
    /// open, depressed health score) until the daemon was restarted. A recency
    /// window lets the rate self-heal as fresh successes age out old failures.
    pub fn error_rate(&self) -> f64 {
        if self.recent_results.is_empty() {
            return 0.0;
        }
        let failures = self.recent_results.iter().filter(|&&ok| !ok).count();
        failures as f64 / self.recent_results.len() as f64
    }

    /// Record a successful operation.
    ///
    /// In closed state: resets consecutive failures, increments window successes.
    /// In half-open state: increments consecutive successes.
    pub fn record_success(&mut self) {
        self.window_successes += 1;
        self.consecutive_failures = 0;
        self.push_result(true);

        if self.state == CircuitState::HalfOpen {
            self.consecutive_successes += 1;
            if self.active_probes > 0 {
                self.active_probes -= 1;
            }
        }
    }

    /// Record a failed operation.
    ///
    /// Increments consecutive failures and window failures.
    /// In half-open state, resets consecutive successes.
    pub fn record_failure(&mut self) {
        self.window_failures += 1;
        self.consecutive_failures += 1;
        self.push_result(false);

        if self.state == CircuitState::HalfOpen {
            self.consecutive_successes = 0;
            if self.active_probes > 0 {
                self.active_probes -= 1;
            }
        }
    }

    /// Check if the circuit should open based on config thresholds.
    ///
    /// Returns true if:
    /// - Consecutive failures exceed threshold, OR
    /// - Error rate exceeds threshold (with minimum sample size)
    pub fn should_open(&self, config: &CircuitBreakerConfig) -> bool {
        if self.state != CircuitState::Closed {
            return false;
        }

        // Open if consecutive failures exceed threshold
        if self.consecutive_failures >= config.failure_threshold {
            return true;
        }

        // Open if error rate exceeds threshold (with a minimum sample size).
        // Counts samples in the bounded recent-results window so the gate
        // reflects recent behavior, matching error_rate() above.
        if self.recent_results.len() >= 5 && self.error_rate() >= config.error_rate_threshold {
            return true;
        }

        false
    }

    /// Check if the circuit should transition to half-open.
    ///
    /// Returns true if the circuit is open and the cooldown period has elapsed.
    pub fn should_half_open(&self, config: &CircuitBreakerConfig) -> bool {
        if self.state != CircuitState::Open {
            return false;
        }

        let now = Self::now_millis();
        let cooldown_ms = self.effective_cooldown_ms(config);

        if let Some(opened_at) = self.opened_at {
            now.saturating_sub(opened_at) >= cooldown_ms
        } else {
            false
        }
    }

    /// Check if the circuit should close.
    ///
    /// Returns true if in half-open state, consecutive successes meet or
    /// exceed the success threshold, AND no command-class failure is
    /// outstanding. A worker whose last real command timed out is not
    /// recovered no matter how many liveness probes it answers; it stays
    /// HalfOpen (degraded, probe-limited) until a command-class success
    /// clears the evidence.
    pub fn should_close(&self, config: &CircuitBreakerConfig) -> bool {
        if self.state != CircuitState::HalfOpen {
            return false;
        }

        self.consecutive_successes >= config.success_threshold
            && self.consecutive_command_failures == 0
    }

    /// Check if a probe request can be made in half-open state.
    ///
    /// Returns true if active probes are below the maximum allowed.
    pub fn can_probe(&self, config: &CircuitBreakerConfig) -> bool {
        if self.state != CircuitState::HalfOpen {
            return false;
        }

        self.active_probes < config.half_open_max_probes
    }

    /// Start a probe request in half-open state.
    ///
    /// Returns true if the probe was started, false if already at max probes.
    pub fn start_probe(&mut self, config: &CircuitBreakerConfig) -> bool {
        if !self.can_probe(config) {
            return false;
        }
        self.active_probes += 1;
        true
    }

    /// Transition the circuit to open state.
    pub fn open(&mut self) {
        if self.state != CircuitState::Open {
            self.state = CircuitState::Open;
            self.opened_at = Some(Self::now_millis());
            self.last_state_change = Self::now_millis();
            self.consecutive_successes = 0;
            self.active_probes = 0;
            // Every open without an intervening close lengthens the next
            // cooldown (see `CircuitBreakerConfig::cooldown_ms_for_streak`).
            self.reopen_streak = self.reopen_streak.saturating_add(1);
        }
    }

    /// Transition the circuit to half-open state.
    pub fn half_open(&mut self) {
        if self.state != CircuitState::HalfOpen {
            self.state = CircuitState::HalfOpen;
            self.last_state_change = Self::now_millis();
            self.consecutive_successes = 0;
            self.active_probes = 0;
        }
    }

    /// Transition the circuit to closed state.
    pub fn close(&mut self) {
        if self.state != CircuitState::Closed {
            self.state = CircuitState::Closed;
            self.last_state_change = Self::now_millis();
            self.opened_at = None;
            self.consecutive_failures = 0;
            self.consecutive_successes = 0;
            self.active_probes = 0;
            self.consecutive_command_failures = 0;
            self.reopen_streak = 0;
            // Reset the window on close, INCLUDING the recent-results history
            // that error_rate() now reads. Otherwise a freshly-recovered circuit
            // would still compute a high error rate from the pre-failure samples
            // and immediately re-open (flapping). Matches the original intent of
            // resetting the error-rate inputs when the circuit recovers.
            self.window_successes = 0;
            self.window_failures = 0;
            self.recent_results.clear();
        }
    }

    /// Reset the rolling window counters.
    ///
    /// Called periodically to ensure the window reflects recent activity.
    /// Also clears the recent-results history that `error_rate()` is computed
    /// from, so a reset truly zeroes the observed error rate.
    pub fn reset_window(&mut self) {
        self.window_successes = 0;
        self.window_failures = 0;
        self.recent_results.clear();
    }

    /// Get recent health check results for history visualization.
    ///
    /// Returns a slice of recent results (true=success, false=failure),
    /// with the most recent result at the end.
    pub fn recent_results(&self) -> &[bool] {
        &self.recent_results
    }

    /// Calculate seconds remaining until circuit auto-transitions to half-open.
    ///
    /// Returns None if circuit is not open or cooldown has already elapsed.
    pub fn recovery_remaining_secs(&self, config: &CircuitBreakerConfig) -> Option<u64> {
        if self.state != CircuitState::Open {
            return None;
        }

        let now = Self::now_millis();
        let cooldown_ms = self.effective_cooldown_ms(config);

        if let Some(opened_at) = self.opened_at {
            let elapsed_ms = now.saturating_sub(opened_at);
            if elapsed_ms < cooldown_ms {
                return Some((cooldown_ms - elapsed_ms) / 1000);
            }
        }
        None
    }

    /// Push a result to the history, maintaining the maximum size.
    fn push_result(&mut self, success: bool) {
        if self.recent_results.len() >= CIRCUIT_HISTORY_SIZE {
            self.recent_results.remove(0);
        }
        self.recent_results.push(success);
    }

    fn now_millis() -> u64 {
        std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .map(|d| d.as_millis() as u64)
            .unwrap_or(0)
    }

    /// Drive the circuit through a single health-check outcome and return the
    /// resulting [`CircuitState`].
    ///
    /// This is the single, self-contained transition engine shared by every
    /// circuit consumer (the daemon health monitor via
    /// `WorkerState::record_health_check`, and the diagnostic `WorkerHealth`
    /// panel via `WorkerHealth::observe`). Keeping the transition logic in ONE
    /// place is what makes `WorkerState.circuit` (which selection reads) and the
    /// diagnostic `WorkerHealth.circuit` agree.
    ///
    /// Behaviour:
    /// - On a **healthy** outcome: if the circuit is Open and its cooldown has
    ///   elapsed it first transitions Open → HalfOpen *before* crediting the
    ///   success, so recovery needs exactly `success_threshold` clean probes
    ///   (crediting a success against an Open circuit would be lost). Then the
    ///   success is recorded and the circuit closes once the threshold is met.
    /// - On an **unhealthy** outcome: the failure is recorded, the circuit opens
    ///   if the closed-state open condition is met, and a HalfOpen probe failure
    ///   reopens the circuit — UNLESS the failure was `transient` (a retryable
    ///   transport blip that survived the in-check retries). A transient
    ///   half-open failure neither reopens nor resets the cooldown, so a single
    ///   network hiccup during recovery cannot bounce a worker back to Open.
    /// - After either branch, an Open circuit past its cooldown is nudged to
    ///   HalfOpen so probing resumes on the next poll even for a still-failing
    ///   worker.
    pub fn apply_health_outcome(
        &mut self,
        healthy: bool,
        transient: bool,
        config: &CircuitBreakerConfig,
    ) -> CircuitState {
        if healthy {
            // Transition to half-open BEFORE crediting the success so the probe
            // counts toward recovery. record_success() only increments the
            // consecutive-success counter while in HalfOpen.
            if self.state == CircuitState::Open && self.should_half_open(config) {
                self.half_open();
            }
            self.record_success();
            if self.should_close(config) {
                self.close();
            }
        } else {
            let was_half_open = self.state == CircuitState::HalfOpen;
            self.record_failure();
            if self.should_open(config) {
                self.open();
            } else if was_half_open && !transient {
                // A real (non-transient) probe failure in half-open reopens the
                // circuit and restarts the cooldown. A transient failure is
                // ignored so a network blip can't undo an in-progress recovery.
                self.open();
            }
        }

        // Even for a still-failing worker, resume probing once the cooldown
        // elapses so a recovered worker is retried promptly.
        if self.state == CircuitState::Open && self.should_half_open(config) {
            self.half_open();
        }

        self.state
    }

    /// Drive the circuit through a single *command-class* outcome and return
    /// the resulting [`CircuitState`].
    ///
    /// Command-class evidence comes from an operation of the same shape as
    /// dispatched work: a remote binary that must run to completion (the
    /// telemetry poll, or a health probe whose SSH session connected but whose
    /// command never finished). It is the signal the plain liveness probe
    /// cannot provide — a host under fork exhaustion or swap thrash keeps
    /// answering `echo` over a warm connection while every real command hangs,
    /// and on that evidence alone the breaker would happily re-admit it.
    ///
    /// Behaviour:
    /// - **Success**: clears the outstanding command-failure count and then
    ///   counts as an ordinary healthy outcome (may advance Open -> HalfOpen ->
    ///   Closed exactly like a liveness success).
    /// - **Failure**: increments the command-failure count and records a plain
    ///   failure. The circuit opens when either the ordinary open condition or
    ///   `command_failure_threshold` is met. In HalfOpen a command failure
    ///   ALWAYS reopens — there is no transient exemption, because "connected
    ///   but the command did not complete" is precisely the condition the
    ///   half-open trial exists to detect. While any command failure is
    ///   outstanding, [`Self::should_close`] refuses to close on liveness
    ///   successes alone.
    pub fn apply_command_outcome(
        &mut self,
        success: bool,
        config: &CircuitBreakerConfig,
    ) -> CircuitState {
        if success {
            self.consecutive_command_failures = 0;
            return self.apply_health_outcome(true, false, config);
        }

        self.consecutive_command_failures = self.consecutive_command_failures.saturating_add(1);
        let was_half_open = self.state == CircuitState::HalfOpen;
        self.record_failure();
        let threshold = config.command_failure_threshold.max(1);
        if was_half_open
            || self.should_open(config)
            || (self.state == CircuitState::Closed
                && self.consecutive_command_failures >= threshold)
        {
            self.open();
        }

        if self.state == CircuitState::Open && self.should_half_open(config) {
            self.half_open();
        }

        self.state
    }
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CompilationConfig {
    /// Minimum confidence score to intercept (0.0-1.0).
    #[serde(default = "default_confidence")]
    pub confidence_threshold: f64,
    /// Run locally when this build's measured local median is below this (ms).
    /// Consulted only by the speed gate, i.e. when `local_slots > 0`.
    #[serde(default = "default_min_local_time")]
    pub min_local_time_ms: u64,
    /// Minimum measured speedup ratio (local_time / remote_time) to offload.
    /// If the measured speedup < threshold, run locally. Default: 1.2 (20% faster).
    /// Consulted only by the speed gate, i.e. when `local_slots > 0`.
    #[serde(default = "default_speedup_threshold")]
    pub remote_speedup_threshold: f64,
    /// Speed gate (bd-2spi4): how many offloadable builds may run concurrently
    /// on THIS machine because its own measured build history says remote is
    /// not worth it. Zero (default) disables the gate, so every offloadable
    /// build is offloaded as before. Set it only on a dispatcher that is faster
    /// than its workers; local runs also need free memory (see the hook).
    #[serde(default)]
    pub local_slots: u32,
    /// Default slot estimate for build commands (cargo build, gcc, etc.).
    #[serde(default = "default_build_slots")]
    pub build_slots: u32,
    /// Default slot estimate for test commands (cargo test, nextest).
    /// Tests typically use more parallelism than builds.
    #[serde(default = "default_test_slots")]
    pub test_slots: u32,
    /// Default slot estimate for check/lint commands (cargo check, clippy).
    /// These are typically faster and use fewer resources.
    #[serde(default = "default_check_slots")]
    pub check_slots: u32,
    /// Per-build additional free-space budget on the worker's reported build
    /// filesystem, in GiB. Include expected output growth and a safety margin.
    /// Zero (default) disables this explicit budget. Positive budgets require
    /// a fresh disk observation and are held in durable daemon ownership until
    /// completion. They are admission accounting, not filesystem quotas.
    #[serde(default)]
    pub disk_headroom_gib: u32,
    /// Timeout in seconds for build commands.
    /// Build commands (cargo build, gcc, etc.) typically complete faster than tests.
    #[serde(default = "default_build_timeout")]
    pub build_timeout_sec: u64,
    /// Timeout in seconds for test commands.
    /// Test commands often need longer timeouts due to test suite execution time.
    #[serde(default = "default_test_timeout")]
    pub test_timeout_sec: u64,
    /// Timeout in seconds for bun commands (bun test, bun typecheck).
    /// Bun has known issues where its internal --timeout flag doesn't work for
    /// CPU-bound hangs, so we use a shorter external timeout by default.
    #[serde(default = "default_bun_timeout")]
    pub bun_timeout_sec: u64,
    /// Whether to wrap remote commands with external timeout protection.
    /// This prevents runaway/stuck processes from consuming worker slots indefinitely.
    /// Default: true. Set to false to disable timeout wrapping entirely.
    #[serde(default = "default_external_timeout_enabled")]
    pub external_timeout_enabled: bool,
    /// Whether a remote build FAILURE may fall back to LOCAL execution on the
    /// orchestrator once every eligible remote worker has been exhausted.
    ///
    /// On a remote failure the hook first retries the build on a *different,
    /// higher-capacity* worker (see `run_exec`'s worker-fallback loop). Local
    /// execution runs the heavy compile on the orchestrator that coordinates the
    /// whole fleet, so it is a genuine last resort and can flood that box. This
    /// flag gates it. Default `true` preserves the historical fail-open behavior
    /// (local only when no worker can serve); set to `false` on an orchestrator
    /// to refuse local fallback and fail the build closed instead. `force_remote`
    /// / proof mode / `RCH_REQUIRE_REMOTE` still override this by refusing local
    /// unconditionally.
    #[serde(default = "default_allow_local_fallback")]
    pub allow_local_fallback: bool,
    /// Per-job rustc parallelism cap applied on the WORKER (issue #49).
    ///
    /// `total_slots` bounds concurrent *jobs* per worker, but cargo defaults
    /// `-j` to the worker's `nproc`, so one offloaded `cargo test --workspace`
    /// forks `nproc` rustc regardless of slots and can swap-thrash a
    /// RAM-light box. This policy injects `CARGO_BUILD_JOBS` into the remote
    /// build environment; see [`RemoteBuildJobs`] for the modes. It never
    /// overrides a `CARGO_BUILD_JOBS` the worker session already carries, an
    /// inline `CARGO_BUILD_JOBS=N` or `-j N` in the command, or a project
    /// `.cargo/config.toml` `[build] jobs` — explicit intent always wins.
    #[serde(default)]
    pub remote_build_jobs: RemoteBuildJobs,
}

/// How rch derives the `CARGO_BUILD_JOBS` it injects into a remote build.
///
/// - `auto` (default): computed on the worker from its live `nproc` and
///   total RAM as `clamp(ram_gib / 8, 2, min(nproc, 8))` — budgeting roughly
///   8 GiB of RAM per concurrent rustc leaves headroom for two to three
///   concurrent jobs at a realistic per-rustc peak of ~3 GiB.
/// - `off`: inject nothing (historical behavior; cargo uses `nproc`).
/// - `<n>`: inject a fixed job count fleet-wide.
///
/// In every mode an explicitly set `CARGO_BUILD_JOBS` on the worker (for
/// example from `/etc/environment`) or in the command is left untouched.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum RemoteBuildJobs {
    /// Derive from the worker's cores and RAM at execution time.
    #[default]
    Auto,
    /// Do not inject `CARGO_BUILD_JOBS`.
    Off,
    /// Inject a fixed job count (must be >= 1).
    Fixed(u32),
}

impl RemoteBuildJobs {
    /// Parse the config/CLI spelling: `auto`, `off`, or a positive integer.
    pub fn parse(value: &str) -> Result<Self, String> {
        let trimmed = value.trim();
        match trimmed.to_ascii_lowercase().as_str() {
            "auto" => Ok(Self::Auto),
            "off" | "none" | "disabled" => Ok(Self::Off),
            _ => match trimmed.parse::<u32>() {
                Ok(0) => Err(
                    "remote_build_jobs must be >= 1 (cargo rejects `-j 0`); use `off` to disable"
                        .to_string(),
                ),
                Ok(n) => Ok(Self::Fixed(n)),
                Err(_) => Err(format!(
                    "'{trimmed}' is not a valid remote_build_jobs value (expected `auto`, `off`, or a positive integer)"
                )),
            },
        }
    }
}

impl std::fmt::Display for RemoteBuildJobs {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            Self::Auto => f.write_str("auto"),
            Self::Off => f.write_str("off"),
            Self::Fixed(n) => write!(f, "{n}"),
        }
    }
}

impl std::str::FromStr for RemoteBuildJobs {
    type Err = String;

    fn from_str(value: &str) -> Result<Self, Self::Err> {
        Self::parse(value)
    }
}

impl Serialize for RemoteBuildJobs {
    fn serialize<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
        match self {
            Self::Auto => serializer.serialize_str("auto"),
            Self::Off => serializer.serialize_str("off"),
            Self::Fixed(n) => serializer.serialize_u32(*n),
        }
    }
}

impl<'de> Deserialize<'de> for RemoteBuildJobs {
    fn deserialize<D: serde::Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
        #[derive(Deserialize)]
        #[serde(untagged)]
        enum Raw {
            Count(u64),
            Spelling(String),
        }
        match Raw::deserialize(deserializer)? {
            Raw::Count(n) => {
                let n = u32::try_from(n).map_err(|_| {
                    serde::de::Error::custom(format!("remote_build_jobs {n} is out of range"))
                })?;
                Self::parse(&n.to_string()).map_err(serde::de::Error::custom)
            }
            Raw::Spelling(s) => Self::parse(&s).map_err(serde::de::Error::custom),
        }
    }
}

impl Default for CompilationConfig {
    fn default() -> Self {
        Self {
            confidence_threshold: 0.85,
            min_local_time_ms: 2000,
            remote_speedup_threshold: default_speedup_threshold(),
            local_slots: 0,
            build_slots: default_build_slots(),
            test_slots: default_test_slots(),
            check_slots: default_check_slots(),
            disk_headroom_gib: 0,
            build_timeout_sec: default_build_timeout(),
            test_timeout_sec: default_test_timeout(),
            bun_timeout_sec: default_bun_timeout(),
            external_timeout_enabled: default_external_timeout_enabled(),
            allow_local_fallback: default_allow_local_fallback(),
            remote_build_jobs: RemoteBuildJobs::default(),
        }
    }
}

fn default_build_slots() -> u32 {
    4
}

fn default_test_slots() -> u32 {
    8
}

fn default_check_slots() -> u32 {
    2
}

/// Default build timeout: 5 minutes (300 seconds).
/// Most builds complete well within this time.
fn default_build_timeout() -> u64 {
    300
}

/// Default test timeout: 30 minutes (1800 seconds).
/// Test suites often take significantly longer than builds.
fn default_test_timeout() -> u64 {
    1800
}

/// Default bun timeout: 10 minutes (600 seconds).
/// Bun's internal --timeout flag doesn't work for CPU-bound hangs,
/// so we use a shorter external timeout to prevent zombies.
fn default_bun_timeout() -> u64 {
    600
}

/// Default: external timeout protection is enabled.
fn default_external_timeout_enabled() -> bool {
    true
}

/// Default: permit last-resort local fallback after remote workers are
/// exhausted. Set `compilation.allow_local_fallback = false` to fail closed and
/// keep heavy compiles off the orchestrator entirely.
fn default_allow_local_fallback() -> bool {
    true
}

impl CompilationConfig {
    /// Returns the appropriate external timeout for the given compilation kind.
    ///
    /// - Bun commands get the shorter bun_timeout (to protect against known hang issues)
    /// - Test commands get the longer test_timeout
    /// - All other commands (builds, checks, clippy) get the build_timeout
    pub fn timeout_for_kind(&self, kind: Option<CompilationKind>) -> std::time::Duration {
        let secs = match kind {
            // Bun has known issues where internal timeout doesn't work for CPU hangs
            Some(CompilationKind::BunTest) | Some(CompilationKind::BunTypecheck) => {
                self.bun_timeout_sec
            }
            // Other test commands use the standard test timeout
            Some(CompilationKind::CargoTest) | Some(CompilationKind::CargoNextest) => {
                self.test_timeout_sec
            }
            // Nix derivation builds can take a long time (fetch + compile large
            // closures); give them the generous test-timeout bucket rather than
            // the short build timeout so legitimate builds are not killed early.
            Some(CompilationKind::NixBuild) => self.test_timeout_sec,
            // Go test suites are long-running like cargo test; give them the
            // test bucket rather than the short build timeout.
            Some(CompilationKind::GoTest) => self.test_timeout_sec,
            // Builds, checks, clippy, and unknown use build timeout
            _ => self.build_timeout_sec,
        };
        std::time::Duration::from_secs(secs)
    }

    /// Returns whether external timeout wrapping is enabled.
    pub fn external_timeout_enabled(&self) -> bool {
        self.external_timeout_enabled
    }
}

/// Version of the Layer 0 configuration pack schema (bd-bqu38 / M-1 §149).
///
/// Bump when a knob is added, removed, or its meaning changes; consumers
/// record the version alongside emitted baselines so regressions are
/// attributable to a specific pack revision.
pub const LAYER0_PACK_VERSION: &str = "2";

/// Layer 0 configuration pack knobs (bd-bqu38).
///
/// Every knob defaults to OFF and toggles independently — the kill/rollback
/// contract from the master plan (§149: removed from defaults if it regresses
/// representative p95, output equivalence, debugger behavior, or
/// compatibility). Knobs are injected into remote cargo builds as
/// `CARGO_PROFILE_*` environment variables, so enabling one never mutates any
/// user-owned `.cargo/config.toml`. Verification lanes (`--clean-overlay`,
/// `--source-content-receipt`) build exactly what the project declares and
/// never receive the pack.
#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq, Eq, JsonSchema)]
#[serde(default)]
pub struct Layer0Config {
    /// Master switch: nothing is injected unless this is on.
    pub enabled: bool,
    /// `profile.release.lto = "thin"` via `CARGO_PROFILE_RELEASE_LTO=thin`.
    pub release_lto_thin: bool,
    /// `profile.release.codegen-units = 1` via
    /// `CARGO_PROFILE_RELEASE_CODEGEN_UNITS=1`.
    pub release_codegen_units_1: bool,
    /// `debug = "line-tables-only"` for the dev and test profiles, via
    /// `CARGO_PROFILE_{DEV,TEST}_DEBUG` (bd-hoxgn). Full DWARF was about 77%
    /// of fleet debug/test binaries (2026-09-25 sweep), so this is the main
    /// lever on worker target size. Backtraces keep file and line; debuggers
    /// lose local variables. A host-level `.cargo/config.toml` cannot carry
    /// it: clean-overlay builds refuse any ancestor Cargo config (RCH-E413).
    pub debug_line_tables_only: bool,
}

impl Layer0Config {
    /// Environment pairs to inject into a remote cargo build.
    ///
    /// Empty unless [`Self::enabled`] — the master switch gates every knob so
    /// an operator can disable the whole pack with one key. Order is stable
    /// (declaration order) for deterministic emission and test assertions.
    #[must_use]
    pub fn active_env(&self) -> Vec<(&'static str, &'static str)> {
        if !self.enabled {
            return Vec::new();
        }
        let mut env = Vec::new();
        if self.release_lto_thin {
            env.push(("CARGO_PROFILE_RELEASE_LTO", "thin"));
        }
        if self.release_codegen_units_1 {
            env.push(("CARGO_PROFILE_RELEASE_CODEGEN_UNITS", "1"));
        }
        if self.debug_line_tables_only {
            env.push(("CARGO_PROFILE_DEV_DEBUG", "line-tables-only"));
            env.push(("CARGO_PROFILE_TEST_DEBUG", "line-tables-only"));
        }
        env
    }
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct TransferConfig {
    /// Durable controller retrieval staging base. Unset uses the job-lease
    /// directory's retrieval subtree. Must be absolute; TMPDIR is not used.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub retrieval_staging_dir: Option<PathBuf>,
    /// Free bytes to retain on the staging filesystem in addition to selected
    /// uncompressed output bytes. The preflight never deletes retained data.
    #[serde(default)]
    pub retrieval_min_free_bytes: u64,
    /// zstd compression level (1-19).
    #[serde(default = "default_compression")]
    pub compression_level: u32,
    /// Patterns to exclude from transfer.
    #[serde(default = "default_excludes")]
    pub exclude_patterns: Vec<String>,
    /// SSH keepalive interval in seconds (`ssh -o ServerAliveInterval=<N>`).
    ///
    /// When unset, build transfers use 15 seconds. Zero defers to OpenSSH configuration.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub ssh_server_alive_interval_secs: Option<u64>,
    /// SSH ControlPersist idle timeout in seconds (`ssh -o ControlPersist=<N>s`).
    ///
    /// - Unset preserves OpenSSH defaults (ControlPersist enabled by default via mux).
    /// - `0` disables persistence (`ControlPersist=no`).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub ssh_control_persist_secs: Option<u64>,
    /// Staging base on the worker for the transfer paths that do NOT mirror
    /// the client's project layout: `--clean-overlay` roots, Windows workers,
    /// and `rch cache warm`. Must be an absolute path; defaults to
    /// [`default_remote_base`] (`/data/tmp/rch`).
    ///
    /// An ordinary `rch exec` does not use this: it mirrors the project under
    /// the worker's canonical root (`[path_topology] canonical_root`), and the
    /// pooled Cargo target store rides along inside that mirror. To move
    /// pooled stores onto another filesystem, set
    /// `[remediation.pooled_target] store_base` (issue #64).
    #[serde(default = "default_remote_base")]
    pub remote_base: String,
    /// Per-attempt timeout for source synchronization, in milliseconds.
    ///
    /// This is deliberately separate from remote compilation and artifact-return
    /// timeouts. When unset, RCH derives a conservative timeout from the source
    /// payload size (30 seconds plus one second per MiB, capped at one hour).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub sync_timeout_ms: Option<u64>,
    /// Silence-based stall timeout for source synchronization, in seconds
    /// (issue #59).
    ///
    /// Aborts a source sync when rsync emits NO output at all — no progress
    /// refresh, no stats, no itemized line — for this long. This is deliberately
    /// far tighter than the wall-clock `sync_timeout_ms` cap (which scales with
    /// payload size up to an hour): a large first-time sync keeps refreshing its
    /// progress line and is never killed, while a dead channel or wedged rsync
    /// is detected within this window instead of holding the reservation until
    /// the wall-clock cap. On a stall the hook releases the worker and re-enters
    /// selection excluding it. `0` disables silence detection.
    #[serde(default = "default_source_sync_silence_timeout_secs")]
    pub source_sync_silence_timeout_secs: u64,
    /// Retry policy for transient network errors during transfer.
    #[serde(default)]
    pub retry: RetryConfig,
    /// Verify artifact integrity using blake3 hashes after transfer (bd-377q).
    ///
    /// Not wired: no transfer path calls `artifact_verify`, so enabling this
    /// has no effect. rsync already checks each transferred file against a
    /// whole-file checksum, and recovery publication verifies delivered
    /// outputs, which is why the extra pass was never connected.
    #[serde(default)]
    pub verify_artifacts: bool,
    /// Maximum file size (bytes) for artifact verification (bd-377q).
    ///
    /// Files larger than this limit are skipped during verification to avoid
    /// excessive I/O. Defaults to 100MB.
    #[serde(default = "default_verify_max_size")]
    pub verify_max_size_bytes: u64,

    // =========================================================================
    // Transfer Optimization (bd-3hho)
    // =========================================================================
    /// Maximum transfer size in MB before skipping remote execution.
    ///
    /// When set, runs `rsync --dry-run --stats` to estimate transfer size.
    /// If the estimated size exceeds this threshold, remote offload is skipped
    /// and the command runs locally. Set to `None` (default) to disable.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub max_transfer_mb: Option<u64>,

    /// Maximum estimated transfer time in milliseconds before skipping remote.
    ///
    /// Uses `estimated_bandwidth_bps` (or measured link speed) to calculate
    /// expected transfer time. If it exceeds this threshold, remote offload
    /// is skipped. Set to `None` (default) to disable.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub max_transfer_time_ms: Option<u64>,

    /// Bandwidth limit for rsync in KB/s.
    ///
    /// Passed as `rsync --bwlimit=<N>` to limit transfer speed and prevent
    /// network saturation. Set to `None` or `0` (default) for unlimited.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub bwlimit_kbps: Option<u64>,

    /// Explicit rsync binary for every local transfer (issue #66).
    ///
    /// Unset (the default) lets rch pick: the PATH `rsync` when it is a
    /// modern (3.1+) rsync, otherwise a modern rsync from the well-known
    /// Homebrew/MacPorts locations, otherwise the PATH binary driven with the
    /// openrsync/2.6.9-compatible argv. Set this to force one binary
    /// (`~` is expanded; a bare name is looked up on PATH). The
    /// `RCH_RSYNC_BIN` environment variable overrides this setting.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub rsync_bin: Option<String>,

    /// Estimated link bandwidth in bytes per second.
    ///
    /// Used for transfer time estimation when `max_transfer_time_ms` is set.
    /// If not provided, defaults to 10 MB/s (10485760 bytes/sec).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub estimated_bandwidth_bps: Option<u64>,

    // =========================================================================
    // Adaptive Compression (bd-243w)
    // =========================================================================
    /// Enable adaptive zstd compression based on transfer size (bd-243w).
    ///
    /// When enabled, automatically selects compression level based on estimated
    /// payload size:
    /// - < 10MB: level 1 (fast, minimal compression)
    /// - 10-200MB: level 3 (balanced, default)
    /// - > 200MB: level 7 (slower, better compression)
    ///
    /// When disabled (default), uses the fixed `compression_level` setting.
    #[serde(default)]
    pub adaptive_compression: bool,

    /// Minimum compression level for adaptive mode (bd-243w).
    ///
    /// Even in adaptive mode, compression level won't go below this.
    /// Defaults to 1.
    #[serde(default = "default_min_compression")]
    pub min_compression_level: u32,

    /// Maximum compression level for adaptive mode (bd-243w).
    ///
    /// Even in adaptive mode, compression level won't exceed this.
    /// Defaults to 9 (avoids CPU-intensive levels 10+).
    #[serde(default = "default_max_compression")]
    pub max_compression_level: u32,
}

impl Default for TransferConfig {
    fn default() -> Self {
        Self {
            retrieval_staging_dir: None,
            retrieval_min_free_bytes: 0,
            compression_level: 3,
            exclude_patterns: default_excludes(),
            ssh_server_alive_interval_secs: None,
            ssh_control_persist_secs: None,
            remote_base: default_remote_base(),
            sync_timeout_ms: None,
            source_sync_silence_timeout_secs: default_source_sync_silence_timeout_secs(),
            retry: RetryConfig::default(),
            verify_artifacts: false,
            verify_max_size_bytes: default_verify_max_size(),
            // Transfer optimization (bd-3hho)
            max_transfer_mb: None,
            max_transfer_time_ms: None,
            bwlimit_kbps: None,
            rsync_bin: None,
            estimated_bandwidth_bps: None,
            // Adaptive compression (bd-243w)
            adaptive_compression: false,
            min_compression_level: default_min_compression(),
            max_compression_level: default_max_compression(),
        }
    }
}

impl TransferConfig {
    /// Smallest accepted explicit source-sync timeout (one second).
    pub const MIN_SYNC_TIMEOUT_MS: u64 = 1_000;
    /// Largest accepted explicit or derived source-sync timeout (one hour).
    pub const MAX_SYNC_TIMEOUT_MS: u64 = 3_600_000;
    /// Fixed floor used by the payload-aware default policy.
    pub const DEFAULT_SYNC_TIMEOUT_FLOOR_MS: u64 = 30_000;
    const DEFAULT_SYNC_BYTES_PER_SECOND: u64 = 1024 * 1024;

    /// Whether an explicit source-sync timeout is within the supported range.
    #[must_use]
    pub fn valid_sync_timeout_ms(value: u64) -> bool {
        (Self::MIN_SYNC_TIMEOUT_MS..=Self::MAX_SYNC_TIMEOUT_MS).contains(&value)
    }

    /// Resolve the per-attempt source-sync timeout for an exact payload size.
    ///
    /// Explicit configuration wins. Otherwise the default grants 30 seconds of
    /// setup time plus one second per MiB at a conservative 1 MiB/s floor. This
    /// prevents a cold, large pinned tree from inheriting the old fixed 30-second
    /// ceiling while still bounding every individual attempt.
    #[must_use]
    pub fn sync_timeout_for_payload(&self, payload_bytes: u64) -> std::time::Duration {
        let timeout_ms = self
            .sync_timeout_ms
            .filter(|value| Self::valid_sync_timeout_ms(*value))
            .unwrap_or_else(|| {
                let transfer_seconds = payload_bytes
                    .saturating_add(Self::DEFAULT_SYNC_BYTES_PER_SECOND - 1)
                    / Self::DEFAULT_SYNC_BYTES_PER_SECOND;
                Self::DEFAULT_SYNC_TIMEOUT_FLOOR_MS
                    .saturating_add(transfer_seconds.saturating_mul(1_000))
                    .min(Self::MAX_SYNC_TIMEOUT_MS)
            });
        std::time::Duration::from_millis(timeout_ms)
    }

    /// Select compression level based on estimated transfer size (bd-243w).
    ///
    /// When adaptive compression is enabled, selects an appropriate level
    /// based on payload size:
    /// - < 10MB: level 1 (fast, minimal compression)
    /// - 10-200MB: level 3 (balanced)
    /// - > 200MB: level 7 (better compression)
    ///
    /// The result is clamped between `min_compression_level` and `max_compression_level`.
    ///
    /// When adaptive compression is disabled, returns the fixed `compression_level`.
    pub fn select_compression_level(&self, estimated_bytes: Option<u64>) -> u32 {
        if !self.adaptive_compression {
            return self.compression_level;
        }

        let Some(bytes) = estimated_bytes else {
            // No estimate available, use default
            return self.compression_level;
        };

        // Size thresholds
        const SMALL_THRESHOLD: u64 = 10_000_000; // 10 MB (decimal)
        const LARGE_THRESHOLD: u64 = 200_000_000; // 200 MB (decimal)

        let level = if bytes < SMALL_THRESHOLD {
            1 // Fast compression for small payloads
        } else if bytes < LARGE_THRESHOLD {
            3 // Balanced for medium payloads
        } else {
            7 // Better compression for large payloads
        };

        // Clamp to configured bounds
        level.clamp(self.min_compression_level, self.max_compression_level)
    }
}

// =============================================================================
// Execution Configuration (bd-785w)
// =============================================================================

/// Default allowlist of commands that can be executed remotely.
/// Aligned with the classifier's supported CompilationKind values.
/// Command bases of every compilation kind rch knows how to offload.
///
/// This is the built-in `execution.allowlist`. It is also what a persisted,
/// explicit allowlist is compared against: a config file written before a tool
/// family was added (Go, Nix, TypeScript, ...) silently keeps that family local,
/// and the only symptom is a `[RCH] local (... not in execution.allowlist)` line
/// (issue #51). `ExecutionConfig::missing_builtin_bases` surfaces the gap.
#[rustfmt::skip]
pub const BUILTIN_EXECUTION_BASES: &[&str] = &[
    // Rust
    "cargo",
    "rustc",
    // cargo-nextest
    "nextest",
    // C/C++
    "gcc",
    "g++",
    "clang",
    "clang++",
    "cc",
    "c++",
    // Build systems
    "make",
    "cmake",
    "ninja",
    "meson",
    // Bun
    "bun",
    // Nix (covers both `nix build` and the legacy `nix-build`)
    "nix",
    // Go
    "go",
    // TypeScript (covers both `tsc` and `npx tsc`; command_base is "tsc")
    "tsc",
];

fn default_execution_allowlist() -> Vec<String> {
    BUILTIN_EXECUTION_BASES
        .iter()
        .map(|base| (*base).to_string())
        .collect()
}

/// Execution allowlist configuration for remote builds (bd-785w).
///
/// Controls which command base names are permitted for remote execution.
/// Commands not in the allowlist will fail-open to local execution.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ExecutionConfig {
    /// Allowlist of command base names permitted for remote execution.
    ///
    /// Examples: "cargo", "rustc", "gcc", "bun"
    ///
    /// If a classified compilation command's base name is not in this list,
    /// RCH will fail-open to local execution with a clear reason.
    ///
    /// An empty allowlist disables all remote execution (local only).
    #[serde(default = "default_execution_allowlist")]
    pub allowlist: Vec<String>,
}

impl Default for ExecutionConfig {
    fn default() -> Self {
        Self {
            allowlist: default_execution_allowlist(),
        }
    }
}

impl ExecutionConfig {
    /// Check if a command base name is allowed for remote execution.
    pub fn is_allowed(&self, command_base: &str) -> bool {
        self.allowlist
            .iter()
            .any(|allowed| allowed.eq_ignore_ascii_case(command_base))
    }

    /// Built-in compilation tool bases that this allowlist does NOT admit.
    ///
    /// Empty for the default allowlist. Non-empty means some command family rch
    /// classifies as compilation will always fall open to local execution on
    /// this box — usually because the config file predates that family.
    pub fn missing_builtin_bases(&self) -> Vec<&'static str> {
        BUILTIN_EXECUTION_BASES
            .iter()
            .copied()
            .filter(|base| !self.is_allowed(base))
            .collect()
    }
}

// =============================================================================
// Retry Configuration (bd-x1ek)
// =============================================================================

fn default_retry_max_attempts() -> u32 {
    3
}

fn default_retry_base_delay_ms() -> u64 {
    100
}

fn default_retry_max_delay_ms() -> u64 {
    5000
}

fn default_retry_jitter_factor() -> f64 {
    0.1
}

fn default_retry_total_timeout_ms() -> u64 {
    30000
}

/// Configuration for retry behavior on transient errors.
///
/// Used by rsync and SSH operations to recover from transient network failures
/// without blocking too long or retrying non-recoverable errors.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct RetryConfig {
    /// Maximum number of retry attempts (1 = no retries, just initial attempt).
    #[serde(default = "default_retry_max_attempts")]
    pub max_attempts: u32,

    /// Initial delay between retries in milliseconds.
    #[serde(default = "default_retry_base_delay_ms")]
    pub base_delay_ms: u64,

    /// Maximum delay between retries in milliseconds (caps exponential backoff).
    #[serde(default = "default_retry_max_delay_ms")]
    pub max_delay_ms: u64,

    /// Jitter factor (0.0 - 1.0) to randomize delay.
    /// A value of 0.1 means ±10% randomization.
    #[serde(default = "default_retry_jitter_factor")]
    pub jitter_factor: f64,

    /// Total timeout for all retry attempts in milliseconds.
    /// After this, retries stop even if max_attempts not reached.
    #[serde(default = "default_retry_total_timeout_ms")]
    pub total_timeout_ms: u64,
}

impl Default for RetryConfig {
    fn default() -> Self {
        Self {
            max_attempts: default_retry_max_attempts(),
            base_delay_ms: default_retry_base_delay_ms(),
            max_delay_ms: default_retry_max_delay_ms(),
            jitter_factor: default_retry_jitter_factor(),
            total_timeout_ms: default_retry_total_timeout_ms(),
        }
    }
}

impl RetryConfig {
    /// Create a config that disables retries (single attempt only).
    pub fn no_retry() -> Self {
        Self {
            max_attempts: 1,
            ..Default::default()
        }
    }

    /// Calculate delay for a given attempt number (0-indexed).
    ///
    /// Uses exponential backoff with optional jitter.
    pub fn delay_for_attempt(&self, attempt: u32) -> std::time::Duration {
        if attempt == 0 {
            return std::time::Duration::ZERO;
        }

        // Exponential backoff: base * 2^attempt
        let base = self.base_delay_ms as f64;
        let delay = base * (2.0_f64.powi(attempt as i32 - 1));
        let capped = delay.min(self.max_delay_ms as f64);

        // Apply jitter: delay ± (delay * jitter_factor)
        let jittered = if self.jitter_factor > 0.0 {
            let jitter_range = capped * self.jitter_factor;
            let jitter = (rand::rng().random::<f64>() * 2.0 - 1.0) * jitter_range;
            (capped + jitter).max(0.0)
        } else {
            capped
        };

        std::time::Duration::from_millis(jittered as u64)
    }

    /// Check if we should retry given elapsed time since start.
    pub fn should_retry(&self, attempt: u32, elapsed: std::time::Duration) -> bool {
        if attempt >= self.max_attempts {
            return false;
        }
        elapsed.as_millis() < self.total_timeout_ms as u128
    }
}

/// Default source-sync silence timeout (issue #59): two minutes without a
/// single byte of rsync output means a dead channel or wedged rsync, not a
/// large transfer (progress2 refreshes its line continuously while bytes
/// flow, and each refresh counts as forward progress).
fn default_source_sync_silence_timeout_secs() -> u64 {
    120
}

/// Default remote base path for project transfers.
///
/// Rooted under `/data/tmp` (the fleet-standard managed ephemeral-build zone)
/// rather than `/tmp`: on the Linux workers `/tmp` is a RAM-backed tmpfs, so
/// dispatched build artifacts placed there consume RAM and evade the disk-GC
/// (sbh / datatmp-janitor) that watches `/data/tmp`. Keeping the remote base in
/// the managed zone lets those reapers reclaim orphaned build trees. Must agree
/// with the rchd cache-cleanup and reclaim `remote_base` defaults.
pub fn default_remote_base() -> String {
    "/data/tmp/rch".to_string()
}

/// Validate and normalize a remote base path.
/// Returns an error if the path is not absolute or contains traversal sequences.
pub fn validate_remote_base(path: &str) -> Result<String, String> {
    // Expand tilde for user convenience
    let expanded = shellexpand::tilde(path).into_owned();

    // Check for absolute path
    if !expanded.starts_with('/') {
        return Err(format!(
            "remote_base must be an absolute path, got: {}",
            path
        ));
    }

    // Check for path traversal attempts
    if expanded.contains("..") {
        return Err(format!(
            "remote_base must not contain path traversal (..): {}",
            path
        ));
    }

    // Normalize: remove trailing slashes
    let normalized = expanded.trim_end_matches('/').to_string();

    // Safety checks
    if normalized.is_empty() || normalized == "/" {
        return Err("remote_base cannot be the root directory (safety restriction)".to_string());
    }

    // Enforce at least 2 levels deep (e.g. /tmp/rch, /home/user/rch)
    // /tmp -> depth 1 (unsafe for recursive delete)
    // /tmp/rch -> depth 2 (safe)
    let components: Vec<&str> = normalized.split('/').filter(|c| !c.is_empty()).collect();
    if components.len() < 2 {
        return Err(format!(
            "remote_base must be at least 2 levels deep (e.g. /tmp/rch), got: {}",
            normalized
        ));
    }

    Ok(normalized)
}

fn default_circuit_failure_threshold() -> u32 {
    3
}

fn default_circuit_success_threshold() -> u32 {
    2
}

fn default_circuit_error_rate_threshold() -> f64 {
    0.5
}

fn default_circuit_window_secs() -> u64 {
    60
}

fn default_circuit_open_cooldown_secs() -> u64 {
    30
}

fn default_circuit_half_open_max_probes() -> u32 {
    1
}

fn default_circuit_command_failure_threshold() -> u32 {
    2
}

fn default_circuit_open_cooldown_max_secs() -> u64 {
    // 30s base doubled four times: 30, 60, 120, 240, 480.
    480
}

fn default_true() -> bool {
    true
}

fn default_log_level() -> String {
    "info".to_string()
}

pub fn default_socket_path() -> String {
    // Prefer XDG_RUNTIME_DIR if available (per-user runtime directory).
    if let Ok(runtime_dir) = std::env::var("XDG_RUNTIME_DIR")
        && !runtime_dir.trim().is_empty()
    {
        let path = PathBuf::from(runtime_dir).join("rch.sock");
        return path.to_string_lossy().to_string();
    }

    // Fallback to ~/.cache/rch/rch.sock (persistent across reboots).
    if let Some(cache_dir) = dirs::cache_dir() {
        let rch_cache = cache_dir.join("rch");
        let _ = std::fs::create_dir_all(&rch_cache);
        return rch_cache.join("rch.sock").to_string_lossy().to_string();
    }

    // Last resort: /tmp/rch.sock
    "/tmp/rch.sock".to_string()
}

#[allow(dead_code)]
fn default_output_visibility() -> OutputVisibility {
    OutputVisibility::None
}

fn default_confidence() -> f64 {
    0.85
}

fn default_min_local_time() -> u64 {
    2000
}

/// Default speedup threshold: 1.2 (require 20% speedup to offload).
/// This is conservative to avoid offloading builds that would be only
/// marginally faster remotely when accounting for transfer overhead.
fn default_speedup_threshold() -> f64 {
    1.2
}

fn default_compression() -> u32 {
    3
}

/// Default maximum file size for artifact verification (100MB).
fn default_verify_max_size() -> u64 {
    100 * 1024 * 1024 // 100MB
}

/// Default minimum compression level for adaptive mode.
fn default_min_compression() -> u32 {
    1
}

/// Default maximum compression level for adaptive mode.
fn default_max_compression() -> u32 {
    9
}

fn default_excludes() -> Vec<String> {
    vec![
        // Rust build artifacts
        "target/".to_string(),
        "*.rlib".to_string(),
        "*.rmeta".to_string(),
        // Git metadata is large, often permission-sensitive on workers, and
        // a partial `.git/` tree is not reliable for build scripts anyway.
        ".git/".to_string(),
        // Node.js / Bun dependencies (massive, reinstalled on worker)
        "node_modules/".to_string(),
        // Bun cache and runtime files
        ".bun/".to_string(),
        // npm / pnpm caches
        ".npm/".to_string(),
        ".pnpm-store/".to_string(),
        // Common build output directories
        "dist/".to_string(),
        "build/".to_string(),
        // Framework-specific caches
        ".next/".to_string(),
        ".nuxt/".to_string(),
        ".turbo/".to_string(),
        ".parcel-cache/".to_string(),
        // Agent/session state, not ordinary compilation inputs (issue #109).
        // Directory-only basenames cover nested copies, not similarly named
        // files or directories. .beads/ also covers .beads/.br_recovery/.
        ".beads/".to_string(),
        ".ntm/".to_string(),
        ".beads-recovery/".to_string(),
        // Do not exclude scratch/, data/, artifacts/ or project configuration:
        // they may contain required build inputs. .rchignore adds project-local
        // exclusions; an explicit exclude_patterns list replaces these defaults.
        // Local diagnostic runs may contain host-specific symlinks and are not build inputs.
        ".doctor/".to_string(),
        // Coverage reports (generated during tests)
        "coverage/".to_string(),
        ".nyc_output/".to_string(),
        // Credentials and secrets (must never be transferred to workers)
        ".cargo/credentials".to_string(),
        ".cargo/credentials.toml".to_string(),
        ".env".to_string(),
        ".env.*".to_string(),
        "*.pem".to_string(),
        "*.key".to_string(),
        "credentials.json".to_string(),
        "secrets.json".to_string(),
        "secrets.yaml".to_string(),
        "secrets.yml".to_string(),
        "secrets.toml".to_string(),
        "secrets.env".to_string(),
        "secrets.txt".to_string(),
        "secrets.conf".to_string(),
        "secrets.cfg".to_string(),
        "secrets.properties".to_string(),
        "secrets.xml".to_string(),
        // The classes above were already covered. These are the ones
        // RABS's E027 source-capture seed set denies
        // (rabs-sandbox/src/source_capture.rs) that this list did not,
        // so the two halves of the product now agree about what a
        // secret is. Two components disagreeing on that is the kind of
        // gap nobody notices until one of them is the path that ran.
        //
        // rsync patterns without a `/` match a basename at any depth,
        // so these cover nested copies too — a key checked into a
        // fixtures directory is excluded exactly like one at the root.
        //
        // SSH private keys. The public halves (.pub) are deliberately
        // not excluded: they are not secret and some builds read them.
        "id_rsa".to_string(),
        "id_ed25519".to_string(),
        "id_ecdsa".to_string(),
        "id_dsa".to_string(),
        // Key/certificate container formats (*.pem and *.key are above).
        "*.p12".to_string(),
        "*.pfx".to_string(),
        "*.keystore".to_string(),
        "*.jks".to_string(),
        // Machine credentials and cloud/container config that carries
        // long-lived tokens.
        ".netrc".to_string(),
        "_netrc".to_string(),
        ".aws/credentials".to_string(),
        ".docker/config.json".to_string(),
        // Package-registry auth tokens. These sit in project roots far
        // more often than the key formats above and are the likeliest
        // real leak in a normal repo.
        ".npmrc".to_string(),
        ".pypirc".to_string(),
        // GnuPG private keyring.
        "secring.*".to_string(),
    ]
}

// ============================================================================
// Build History Types
// ============================================================================

/// Location where a build was executed.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
#[serde(rename_all = "snake_case")]
pub enum BuildLocation {
    /// Build executed locally.
    #[default]
    Local,
    /// Build executed on a remote worker.
    Remote,
}

/// Worker health snapshot captured when a cancellation completes.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct BuildCancellationWorkerHealth {
    /// Worker status at cancellation completion.
    pub status: String,
    /// Worker speed score at cancellation completion.
    pub speed_score: f64,
    /// Slots currently used on the worker.
    pub used_slots: u32,
    /// Slots currently available on the worker.
    pub available_slots: u32,
    /// Normalized pressure state.
    pub pressure_state: String,
    /// Stable pressure reason code.
    pub pressure_reason_code: String,
}

/// Structured cancellation metadata attached to cancelled build records.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct BuildCancellationMetadata {
    /// Stable cancellation operation identifier.
    pub operation_id: String,
    /// Cancellation origin (for example `user`, `timeout`).
    pub origin: String,
    /// Stable reason code for cancellation.
    pub reason_code: String,
    /// Ordered cancellation state path.
    #[serde(default)]
    pub decision_path: Vec<String>,
    /// Highest escalation stage reached (`term`, `remote_kill`, `sigkill`).
    pub escalation_stage: String,
    /// Number of escalation transitions taken.
    pub escalation_count: u32,
    /// Whether remote kill was attempted.
    pub remote_kill_attempted: bool,
    /// Whether cleanup completed successfully.
    pub cleanup_ok: bool,
    /// Whether active build history was successfully claimed/cancelled.
    pub history_cancelled: bool,
    /// Final cancellation state.
    pub final_state: String,
    /// Worker health snapshot at completion.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub worker_health: Option<BuildCancellationWorkerHealth>,
}

/// Record of a completed build.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct BuildRecord {
    /// Unique build identifier.
    pub id: u64,
    /// When the build started (ISO 8601 timestamp).
    pub started_at: String,
    /// When the build completed (ISO 8601 timestamp).
    pub completed_at: String,
    /// Project identifier (usually directory name or hash).
    pub project_id: String,
    /// Worker that executed the build (None if local).
    pub worker_id: Option<String>,
    /// Full command executed.
    pub command: String,
    /// Exit code (0 = success).
    pub exit_code: i32,
    /// Duration in milliseconds.
    pub duration_ms: u64,
    /// Build location (local or remote).
    pub location: BuildLocation,
    /// Bytes transferred (if remote).
    pub bytes_transferred: Option<u64>,
    /// Optional timing breakdown for the pipeline.
    #[serde(default)]
    pub timing: Option<CommandTimingBreakdown>,
    /// Structured cancellation metadata for cancelled builds.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub cancellation: Option<BuildCancellationMetadata>,
}

/// Input payload for recording a completed build.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct BuildRecordInput {
    /// When the build started (ISO 8601 timestamp).
    pub started_at: String,
    /// When the build completed (ISO 8601 timestamp).
    pub completed_at: String,
    /// Project identifier (usually directory name or hash).
    pub project_id: String,
    /// Worker that executed the build (None if local).
    pub worker_id: Option<String>,
    /// Full command executed.
    pub command: String,
    /// Exit code (0 = success).
    pub exit_code: i32,
    /// Duration in milliseconds.
    pub duration_ms: u64,
    /// Build location (local or remote).
    pub location: BuildLocation,
    /// Bytes transferred (if remote).
    pub bytes_transferred: Option<u64>,
    /// Optional timing breakdown for the pipeline.
    #[serde(default)]
    pub timing: Option<CommandTimingBreakdown>,
    /// Structured cancellation metadata for cancelled builds.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub cancellation: Option<BuildCancellationMetadata>,
}

impl BuildRecordInput {
    /// Convert input payload into a build record with a supplied ID.
    pub fn into_record(self, id: u64) -> BuildRecord {
        BuildRecord {
            id,
            started_at: self.started_at,
            completed_at: self.completed_at,
            project_id: self.project_id,
            worker_id: self.worker_id,
            command: self.command,
            exit_code: self.exit_code,
            duration_ms: self.duration_ms,
            location: self.location,
            bytes_transferred: self.bytes_transferred,
            timing: self.timing,
            cancellation: self.cancellation,
        }
    }
}

/// Aggregate terminal-command statistics for build history.
///
/// These counts do not infer whether compilation completed before a failing
/// test or interrupted command. The denominator includes every stored record.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct BuildStats {
    /// Total number of builds in history.
    pub total_builds: usize,
    /// Number of successfully completed commands (exit_code == 0).
    pub success_count: usize,
    /// Number of unsuccessful commands, including interruption (exit_code != 0).
    pub failure_count: usize,
    /// Number of remote builds.
    pub remote_count: usize,
    /// Number of local builds.
    pub local_count: usize,
    /// Average build duration in milliseconds.
    pub avg_duration_ms: u64,
}

/// Saved time statistics from remote builds.
///
/// Tracks estimated time savings from offloading builds to remote workers.
/// Estimates are only produced from an observed local baseline; without one the
/// stats carry measured remote durations with `estimate_basis = "none"` and
/// zero estimated savings rather than a fabricated speedup factor.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct SavedTimeStats {
    /// Total time spent on remote builds (milliseconds).
    pub total_remote_duration_ms: u64,
    /// Estimated total time if those builds ran locally (milliseconds).
    /// Zero when no local baseline exists.
    pub estimated_local_duration_ms: u64,
    /// Time saved by offloading (milliseconds).
    /// Computed as max(0, estimated_local - remote); zero without a baseline.
    pub time_saved_ms: u64,
    /// Number of remote builds included in calculation.
    pub builds_counted: usize,
    /// Average speedup factor (local_estimate / remote_duration).
    /// A value of 2.0 means remote builds are ~2x faster than local.
    /// Zero when the estimate basis is "none".
    pub avg_speedup: f64,
    /// Total time saved today (milliseconds).
    pub today_saved_ms: u64,
    /// Total time saved this week (milliseconds).
    pub week_saved_ms: u64,
    /// How the local-time estimate was derived: "observed_local_mean" when
    /// successful local builds back the baseline, "none" otherwise.
    #[serde(default)]
    pub estimate_basis: String,
    /// Number of successful local builds backing the baseline (0 for "none").
    #[serde(default)]
    pub local_baseline_builds: usize,
}

// ============================================================================
// Compilation Timing and Metrics
// ============================================================================

use std::collections::VecDeque;
use std::time::{Duration, Instant};

/// Breakdown of timing for each compilation phase.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct CompilationTimingBreakdown {
    /// Time to sync source files to worker.
    #[serde(with = "duration_millis")]
    pub rsync_up: Duration,
    /// Time for cargo build on worker.
    #[serde(with = "duration_millis")]
    pub remote_build: Duration,
    /// Time to sync artifacts back.
    #[serde(with = "duration_millis")]
    pub rsync_down: Duration,
    /// Total end-to-end latency.
    #[serde(with = "duration_millis")]
    pub total: Duration,
}

/// Timing breakdown for a single command pipeline.
///
/// Uses optional durations so missing phases serialize as `null`.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct CommandTimingBreakdown {
    /// Time spent classifying the command.
    #[serde(with = "option_duration_millis")]
    pub classify: Option<Duration>,
    /// Time spent selecting a worker.
    #[serde(with = "option_duration_millis")]
    pub select: Option<Duration>,
    /// Time spent syncing source files to the worker.
    #[serde(with = "option_duration_millis")]
    pub sync_up: Option<Duration>,
    /// Time spent executing the remote command.
    #[serde(with = "option_duration_millis")]
    pub exec: Option<Duration>,
    /// Time spent syncing artifacts back.
    #[serde(with = "option_duration_millis")]
    pub sync_down: Option<Duration>,
    /// Time spent on cleanup (slot release, bookkeeping).
    #[serde(with = "option_duration_millis")]
    pub cleanup: Option<Duration>,
    /// Total end-to-end time for the pipeline.
    #[serde(with = "option_duration_millis")]
    pub total: Option<Duration>,
}

mod duration_millis {
    use serde::{Deserialize, Deserializer, Serialize, Serializer};
    use std::time::Duration;

    pub fn serialize<S>(duration: &Duration, serializer: S) -> Result<S::Ok, S::Error>
    where
        S: Serializer,
    {
        duration.as_millis().serialize(serializer)
    }

    pub fn deserialize<'de, D>(deserializer: D) -> Result<Duration, D::Error>
    where
        D: Deserializer<'de>,
    {
        let millis = u64::deserialize(deserializer)?;
        Ok(Duration::from_millis(millis))
    }
}

/// Comprehensive metrics for a single compilation.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CompilationMetrics {
    /// Project identifier (usually directory name or hash).
    pub project_id: String,
    /// Worker that performed the compilation.
    pub worker_id: String,
    /// When the compilation occurred.
    pub timestamp: chrono::DateTime<chrono::Utc>,
    /// Timing breakdown for each phase.
    pub timing: CompilationTimingBreakdown,
    /// Local build time for comparison (if available).
    #[serde(with = "option_duration_millis")]
    pub local_build_time: Option<Duration>,
    /// Speedup ratio (local_time / remote_time).
    pub speedup: Option<f64>,
    /// Number of files synced to worker.
    pub files_synced: u64,
    /// Total bytes transferred.
    pub bytes_transferred: u64,
    /// Exit code from compilation.
    pub exit_code: i32,
    /// Whether compilation succeeded.
    pub success: bool,
}

mod option_duration_millis {
    use serde::{Deserialize, Deserializer, Serialize, Serializer};
    use std::time::Duration;

    pub fn serialize<S>(duration: &Option<Duration>, serializer: S) -> Result<S::Ok, S::Error>
    where
        S: Serializer,
    {
        match duration {
            Some(d) => Some(d.as_millis() as u64).serialize(serializer),
            None => Option::<u64>::None.serialize(serializer),
        }
    }

    pub fn deserialize<'de, D>(deserializer: D) -> Result<Option<Duration>, D::Error>
    where
        D: Deserializer<'de>,
    {
        let opt: Option<u64> = Option::deserialize(deserializer)?;
        Ok(opt.map(Duration::from_millis))
    }
}

impl CompilationMetrics {
    /// Calculate speedup based on local vs remote build time.
    pub fn calculate_speedup(&mut self) {
        if let Some(local) = self.local_build_time
            && self.timing.total.as_millis() > 0
        {
            self.speedup = Some(local.as_secs_f64() / self.timing.total.as_secs_f64());
        }
    }

    /// Check if remote compilation was beneficial (faster than local).
    pub fn is_beneficial(&self) -> bool {
        self.speedup.map(|s| s > 1.0).unwrap_or(false)
    }
}

impl Default for CompilationMetrics {
    fn default() -> Self {
        Self {
            project_id: String::new(),
            worker_id: String::new(),
            timestamp: chrono::Utc::now(),
            timing: CompilationTimingBreakdown::default(),
            local_build_time: None,
            speedup: None,
            files_synced: 0,
            bytes_transferred: 0,
            exit_code: 0,
            success: true,
        }
    }
}

/// Timer for tracking compilation phases.
///
/// Use this to measure the duration of each phase of remote compilation.
///
/// # Example
/// ```ignore
/// let mut timer = CompilationTimer::new("my-project", "worker-1");
/// // ... rsync source files ...
/// timer.end_rsync_up();
/// // ... run cargo build ...
/// timer.end_remote_build();
/// // ... rsync artifacts back ...
/// timer.end_rsync_down();
/// let metrics = timer.finish(0, 100, 1_000_000);
/// ```
#[derive(Debug)]
pub struct CompilationTimer {
    project_id: String,
    worker_id: String,
    start: Instant,
    phase_start: Instant,
    rsync_up: Option<Duration>,
    remote_build: Option<Duration>,
    rsync_down: Option<Duration>,
}

impl CompilationTimer {
    /// Create a new timer for a compilation.
    pub fn new(project_id: &str, worker_id: &str) -> Self {
        let now = Instant::now();
        Self {
            project_id: project_id.to_string(),
            worker_id: worker_id.to_string(),
            start: now,
            phase_start: now,
            rsync_up: None,
            remote_build: None,
            rsync_down: None,
        }
    }

    /// Mark the end of the rsync upload phase.
    pub fn end_rsync_up(&mut self) {
        self.rsync_up = Some(self.phase_start.elapsed());
        self.phase_start = Instant::now();
        tracing::info!(
            rsync_up_ms = %self.rsync_up.unwrap().as_millis(),
            "TIMING: rsync_up completed"
        );
    }

    /// Mark the end of the remote build phase.
    pub fn end_remote_build(&mut self) {
        self.remote_build = Some(self.phase_start.elapsed());
        self.phase_start = Instant::now();
        tracing::info!(
            remote_build_ms = %self.remote_build.unwrap().as_millis(),
            "TIMING: remote_build completed"
        );
    }

    /// Mark the end of the rsync download phase.
    pub fn end_rsync_down(&mut self) {
        self.rsync_down = Some(self.phase_start.elapsed());
        tracing::info!(
            rsync_down_ms = %self.rsync_down.unwrap().as_millis(),
            "TIMING: rsync_down completed"
        );
    }

    /// Finish timing and produce metrics.
    pub fn finish(self, exit_code: i32, files: u64, bytes: u64) -> CompilationMetrics {
        let total = self.start.elapsed();
        tracing::info!(
            total_ms = %total.as_millis(),
            exit_code = %exit_code,
            "TIMING: compilation completed"
        );

        CompilationMetrics {
            project_id: self.project_id,
            worker_id: self.worker_id,
            timestamp: chrono::Utc::now(),
            timing: CompilationTimingBreakdown {
                rsync_up: self.rsync_up.unwrap_or_default(),
                remote_build: self.remote_build.unwrap_or_default(),
                rsync_down: self.rsync_down.unwrap_or_default(),
                total,
            },
            local_build_time: None,
            speedup: None,
            files_synced: files,
            bytes_transferred: bytes,
            exit_code,
            success: exit_code == 0,
        }
    }
}

/// Aggregator for compilation metrics with statistics.
#[derive(Debug)]
pub struct MetricsAggregator {
    history: VecDeque<CompilationMetrics>,
    max_history: usize,
}

impl MetricsAggregator {
    /// Create a new aggregator with a maximum history size.
    pub fn new(max_history: usize) -> Self {
        Self {
            history: VecDeque::with_capacity(max_history),
            max_history,
        }
    }

    /// Record a new compilation's metrics.
    pub fn record(&mut self, metrics: CompilationMetrics) {
        if self.history.len() >= self.max_history {
            self.history.pop_front();
        }
        self.history.push_back(metrics);
    }

    /// Get the average speedup across all recorded compilations.
    pub fn average_speedup(&self) -> Option<f64> {
        let speedups: Vec<f64> = self.history.iter().filter_map(|m| m.speedup).collect();

        if speedups.is_empty() {
            None
        } else {
            Some(speedups.iter().sum::<f64>() / speedups.len() as f64)
        }
    }

    /// Get the p50 (median) total compilation time.
    pub fn p50_total_time(&self) -> Option<Duration> {
        self.percentile_total_time(0.50)
    }

    /// Get the p95 total compilation time.
    pub fn p95_total_time(&self) -> Option<Duration> {
        self.percentile_total_time(0.95)
    }

    /// Get the p99 total compilation time.
    pub fn p99_total_time(&self) -> Option<Duration> {
        self.percentile_total_time(0.99)
    }

    /// Get a percentile of total compilation times.
    ///
    /// Uses the "inclusive" percentile method (matching numpy.percentile, Excel PERCENTILE.INC):
    /// index = (n - 1) * percentile, where n is the number of samples.
    fn percentile_total_time(&self, percentile: f64) -> Option<Duration> {
        let mut times: Vec<_> = self.history.iter().map(|m| m.timing.total).collect();

        if times.is_empty() {
            return None;
        }

        times.sort();
        // Use inclusive percentile: index = (n - 1) * p
        // For p50 with 100 items: (99 * 0.50) = 49.5 -> 49 (true median position)
        // For p99 with 100 items: (99 * 0.99) = 98.01 -> 98
        let idx = (((times.len() - 1) as f64 * percentile).round() as usize).min(times.len() - 1);
        Some(times[idx])
    }

    /// Get the success rate as a percentage.
    pub fn success_rate(&self) -> f64 {
        if self.history.is_empty() {
            return 100.0;
        }
        let successes = self.history.iter().filter(|m| m.success).count();
        (successes as f64 / self.history.len() as f64) * 100.0
    }

    /// Get the number of recorded compilations.
    pub fn count(&self) -> usize {
        self.history.len()
    }

    /// Get all recorded metrics.
    pub fn metrics(&self) -> &VecDeque<CompilationMetrics> {
        &self.history
    }

    /// Clear all recorded metrics.
    pub fn clear(&mut self) {
        self.history.clear();
    }
}

impl Default for MetricsAggregator {
    fn default() -> Self {
        Self::new(1000)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::test_guard;

    #[test]
    fn test_declared_os_reads_reserved_tag() {
        assert_eq!(
            declared_os(&["rust".to_string(), "os:windows".to_string()]),
            Some("windows".to_string())
        );
        assert_eq!(
            declared_os(&["os:WINDOWS".to_string()]),
            Some("windows".to_string())
        );
    }

    #[test]
    fn test_declared_os_absent_for_ordinary_tags() {
        // Every historical worker: no reserved tag, so no exclusivity and the
        // worker stays a candidate for any job.
        assert_eq!(declared_os(&[]), None);
        assert_eq!(
            declared_os(&["rust".to_string(), "bun".to_string(), "go".to_string()]),
            None
        );
        // A malformed declaration must not silently fence the worker off.
        assert_eq!(declared_os(&["os:".to_string()]), None);
        assert_eq!(declared_os(&["os:  ".to_string()]), None);
    }

    #[test]
    fn test_os_tag_roundtrips_through_declared_os() {
        for os in ["windows", "Darwin", " linux "] {
            let tag = os_tag(os);
            assert_eq!(
                declared_os(std::slice::from_ref(&tag)).as_deref(),
                Some(os.trim().to_ascii_lowercase().as_str())
            );
        }
    }

    /// Verdict-webhook config must NEVER carry an inline secret. Operators
    /// reference secrets by env-var name (`*_env`); the resolved value is
    /// read at dispatch time and never persisted to the cache or config TOML.
    /// This guards against a future careless `pub secret: String` addition.
    #[test]
    fn test_no_inline_secret_fields() {
        let endpoint = DoctorWebhookEndpoint {
            name: "ops".to_string(),
            url: "https://example.com/hook".to_string(),
            format: DoctorWebhookFormat::Pagerduty,
            on_transitions: vec!["any_to_failing".to_string()],
            retry_max: 3,
            retry_backoff_ms: 200,
            timeout_ms: 5000,
            enabled: true,
            routing_key_env: Some("RCH_PAGERDUTY_ROUTING_KEY".to_string()),
            bearer_token_env: Some("RCH_WEBHOOK_BEARER".to_string()),
            signing_secret_env: Some("RCH_WEBHOOK_HMAC".to_string()),
        };
        // Serialize to a JSON object and inspect the *keys*. The forbidden
        // field names must be entirely absent; only `*_env` references survive.
        let value = serde_json::to_value(&endpoint).unwrap();
        let obj = value.as_object().expect("endpoint serializes to an object");
        for forbidden in ["secret", "routing_key", "bearer_token", "password", "token"] {
            assert!(
                !obj.contains_key(forbidden),
                "endpoint config must not expose an inline `{forbidden}` field"
            );
        }
        assert!(obj.contains_key("routing_key_env"));
        assert!(obj.contains_key("bearer_token_env"));
        assert!(obj.contains_key("signing_secret_env"));
    }

    #[test]
    fn test_webhook_format_toml_spelling() {
        // The config TOML uses snake_case spellings. TOML documents are tables
        // at the root, so a bare enum cannot be (de)serialized standalone —
        // verify the wire spelling through a `format = "..."` table field, the
        // shape it actually appears in within the config.
        #[derive(serde::Serialize, serde::Deserialize, PartialEq, Debug)]
        struct FormatWrap {
            format: DoctorWebhookFormat,
        }
        let rendered = toml::to_string(&FormatWrap {
            format: DoctorWebhookFormat::GenericJson,
        })
        .unwrap();
        assert!(
            rendered.contains("format = \"generic_json\""),
            "unexpected TOML: {rendered}"
        );
        let slack: FormatWrap = toml::from_str("format = \"slack\"").unwrap();
        assert_eq!(slack.format, DoctorWebhookFormat::Slack);
        let pd: FormatWrap = toml::from_str("format = \"pagerduty\"").unwrap();
        assert_eq!(pd.format, DoctorWebhookFormat::Pagerduty);
    }

    #[test]
    fn test_doctor_webhooks_defaults() {
        let cfg = DoctorWebhooksConfig::default();
        assert!(cfg.endpoints.is_empty());
        assert_eq!(cfg.max_inflight, 16);
        // RchConfig wires the section in with serde(default), so an empty
        // config still yields an empty (never-firing) webhook surface.
        let rch: RchConfig = toml::from_str("").unwrap();
        assert!(rch.doctor.webhooks.endpoints.is_empty());
    }

    #[test]
    fn test_doctor_webhook_endpoint_parses_from_toml() {
        let toml_src = r#"
[doctor.webhooks]
max_inflight = 8

[[doctor.webhooks.endpoints]]
name = "ops-slack"
url = "https://hooks.slack.com/services/T0/B0/x"
format = "slack"
on_transitions = ["healthy_to_degraded", "any_to_failing"]
retry_max = 2
"#;
        let cfg: RchConfig = toml::from_str(toml_src).unwrap();
        assert_eq!(cfg.doctor.webhooks.max_inflight, 8);
        assert_eq!(cfg.doctor.webhooks.endpoints.len(), 1);
        let ep = &cfg.doctor.webhooks.endpoints[0];
        assert_eq!(ep.name, "ops-slack");
        assert_eq!(ep.format, DoctorWebhookFormat::Slack);
        assert_eq!(ep.on_transitions.len(), 2);
        assert_eq!(ep.retry_max, 2);
        // Defaults filled in for omitted fields.
        assert_eq!(ep.retry_backoff_ms, 200);
        assert_eq!(ep.timeout_ms, 5000);
        assert!(ep.enabled);
    }

    fn assert_golden_json(
        actual: serde_json::Value,
        expected: &serde_json::Value,
        fixture_path: &str,
    ) {
        if &actual == expected {
            return;
        }

        let actual_json = serde_json::to_string_pretty(&actual).expect("actual JSON renders");
        let expected_json = serde_json::to_string_pretty(expected).expect("expected JSON renders");
        assert_eq!(
            &actual,
            expected,
            "golden mismatch in {fixture_path}\n\
             expected_blake3={}\n\
             actual_blake3={}\n\
             bless path: review the semantic diff, update the fixture expected_* field, \
             and record both hashes in the Beads closeout.\n\
             expected:\n{expected_json}\n\
             actual:\n{actual_json}",
            blake3::hash(expected_json.as_bytes()),
            blake3::hash(actual_json.as_bytes())
        );
    }

    #[test]
    fn test_circuit_state_default() {
        let _guard = test_guard!();
        assert_eq!(CircuitState::default(), CircuitState::Closed);
    }

    #[test]
    fn test_circuit_config_defaults() {
        let _guard = test_guard!();
        let config = CircuitBreakerConfig::default();
        assert_eq!(config.failure_threshold, 3);
        assert_eq!(config.success_threshold, 2);
        assert_eq!(config.error_rate_threshold, 0.5);
        assert_eq!(config.window_secs, 60);
        assert_eq!(config.open_cooldown_secs, 30);
        assert_eq!(config.half_open_max_probes, 1);
    }

    #[test]
    fn test_rch_config_has_circuit_defaults() {
        let _guard = test_guard!();
        let config = RchConfig::default();
        assert_eq!(config.circuit.failure_threshold, 3);
    }

    #[test]
    fn test_circuit_config_serde_roundtrip() {
        let _guard = test_guard!();
        let config = CircuitBreakerConfig {
            failure_threshold: 5,
            success_threshold: 3,
            error_rate_threshold: 0.75,
            window_secs: 120,
            open_cooldown_secs: 45,
            half_open_max_probes: 2,
            command_failure_threshold: 4,
            open_cooldown_max_secs: 900,
        };
        let json = serde_json::to_string(&config).unwrap();
        let parsed: CircuitBreakerConfig = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.command_failure_threshold, 4);
        assert_eq!(parsed.open_cooldown_max_secs, 900);
        assert_eq!(parsed.failure_threshold, 5);
        assert_eq!(parsed.success_threshold, 3);
        assert_eq!(parsed.error_rate_threshold, 0.75);
        assert_eq!(parsed.window_secs, 120);
        assert_eq!(parsed.open_cooldown_secs, 45);
        assert_eq!(parsed.half_open_max_probes, 2);
    }

    // ========================================================================
    // WorkerCapabilities Tests
    // ========================================================================

    #[test]
    fn test_worker_capabilities_default() {
        let _guard = test_guard!();
        let caps = WorkerCapabilities::default();
        assert!(caps.rustc_version.is_none());
        assert!(caps.rustup_toolchains.is_empty());
        assert!(caps.rustup_components.is_empty());
        assert!(caps.bun_version.is_none());
        assert!(caps.node_version.is_none());
        assert!(caps.npm_version.is_none());
        assert!(caps.projects_root_ok.is_none());
        assert!(caps.projects_root_issue.is_none());
        assert!(caps.projects_root_checked_at_unix_ms.is_none());
    }

    #[test]
    fn test_worker_capabilities_new() {
        let _guard = test_guard!();
        let caps = WorkerCapabilities::new();
        assert!(caps.rustc_version.is_none());
        assert!(!caps.has_rust());
        assert!(!caps.has_bun());
        assert!(!caps.has_node());
        assert!(caps.is_topology_healthy().is_none());
    }

    #[test]
    fn test_worker_capabilities_has_rust() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        assert!(!caps.has_rust());

        caps.rustc_version = Some("rustc 1.76.0 (07dca489a 2024-02-04)".to_string());
        assert!(caps.has_rust());
    }

    #[test]
    fn test_worker_capabilities_components_are_toolchain_scoped_and_exact() {
        let _guard = test_guard!();
        let caps = WorkerCapabilities {
            rustup_components: vec![
                "stable-x86_64-unknown-linux-gnu:clippy".to_string(),
                "nightly-2026-07-05-x86_64-unknown-linux-gnu:rustfmt".to_string(),
                "beta-x86_64-unknown-linux-gnu:clippy-preview".to_string(),
            ],
            ..Default::default()
        };

        assert!(caps.has_rustup_component("stable", "clippy"));
        assert!(caps.has_rustup_component("nightly-2026-07-05", "rustfmt"));
        assert!(!caps.has_rustup_component("nightly-2026-07-05", "clippy"));
        assert!(!caps.has_rustup_component("beta", "clippy"));
        assert!(!caps.has_rustup_component("nightly-2026-07-0", "rustfmt"));
    }

    #[test]
    fn test_worker_capabilities_has_bun() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        assert!(!caps.has_bun());

        caps.bun_version = Some("1.0.25".to_string());
        assert!(caps.has_bun());
    }

    #[test]
    fn test_worker_capabilities_has_nix() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        assert!(!caps.has_nix());

        caps.nix_version = Some("nix (Nix) 2.24.9".to_string());
        assert!(caps.has_nix());
    }

    #[test]
    fn test_worker_capabilities_has_zig_requires_both_halves() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        assert!(!caps.has_zig());

        // `zig` alone cannot run `cargo zigbuild` — cargo would report
        // `no such command: 'zigbuild'`.
        caps.zig_version = Some("0.14.1".to_string());
        assert!(!caps.has_zig());

        // cargo-zigbuild alone cannot link — it shells out to `zig cc`.
        caps.zig_version = None;
        caps.cargo_zigbuild_version = Some("cargo-zigbuild 0.23.0".to_string());
        assert!(!caps.has_zig());

        caps.zig_version = Some("0.14.1".to_string());
        assert!(caps.has_zig());
    }

    #[test]
    fn test_nix_build_uses_generous_timeout() {
        let _guard = test_guard!();
        let cfg = CompilationConfig::default();
        // Nix builds get the long test-timeout bucket, not the short build timeout.
        assert_eq!(
            cfg.timeout_for_kind(Some(CompilationKind::NixBuild)),
            std::time::Duration::from_secs(cfg.test_timeout_sec)
        );
    }

    #[test]
    fn test_job_uses_build_timeout_bucket() {
        let _guard = test_guard!();
        // Jobs (explicit `rch exec --job`) reuse the standard build-timeout
        // bucket, so `RCH_BUILD_TIMEOUT_SEC` governs them without new config.
        let cfg = CompilationConfig::default();
        assert_eq!(
            cfg.timeout_for_kind(Some(CompilationKind::Job)),
            std::time::Duration::from_secs(cfg.build_timeout_sec)
        );
    }

    #[test]
    fn test_layer0_disabled_by_default_and_env_gated() {
        let _guard = test_guard!();
        let cfg = Layer0Config::default();
        assert!(!cfg.enabled);
        // Master switch off ⇒ nothing injects even with knobs on.
        let knobs_on = Layer0Config {
            enabled: false,
            release_lto_thin: true,
            release_codegen_units_1: true,
            debug_line_tables_only: true,
        };
        assert!(cfg.active_env().is_empty());
        assert!(knobs_on.active_env().is_empty());
    }

    #[test]
    fn test_layer0_knobs_emit_stable_env_pairs() {
        let _guard = test_guard!();
        let lto_only = Layer0Config {
            enabled: true,
            release_lto_thin: true,
            ..Layer0Config::default()
        };
        assert_eq!(
            lto_only.active_env(),
            vec![("CARGO_PROFILE_RELEASE_LTO", "thin")]
        );
        let debug_only = Layer0Config {
            enabled: true,
            debug_line_tables_only: true,
            ..Layer0Config::default()
        };
        assert_eq!(
            debug_only.active_env(),
            vec![
                ("CARGO_PROFILE_DEV_DEBUG", "line-tables-only"),
                ("CARGO_PROFILE_TEST_DEBUG", "line-tables-only"),
            ]
        );
        let all = Layer0Config {
            enabled: true,
            release_lto_thin: true,
            release_codegen_units_1: true,
            debug_line_tables_only: true,
        };
        assert_eq!(
            all.active_env(),
            vec![
                ("CARGO_PROFILE_RELEASE_LTO", "thin"),
                ("CARGO_PROFILE_RELEASE_CODEGEN_UNITS", "1"),
                ("CARGO_PROFILE_DEV_DEBUG", "line-tables-only"),
                ("CARGO_PROFILE_TEST_DEBUG", "line-tables-only"),
            ],
            "emission order must stay stable for deterministic baselines"
        );
    }

    #[test]
    fn test_layer0_pack_version_is_recorded() {
        let _guard = test_guard!();
        // Baselines record the pack version; it must exist and be non-empty.
        assert!(!LAYER0_PACK_VERSION.is_empty());
    }
    #[test]
    fn test_default_allowlist_includes_nix() {
        let _guard = test_guard!();
        let cfg = ExecutionConfig::default();
        assert!(cfg.is_allowed("nix"));
    }

    #[test]
    fn test_worker_capabilities_has_node() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        assert!(!caps.has_node());

        caps.node_version = Some("v20.11.0".to_string());
        assert!(caps.has_node());
    }

    #[test]
    fn test_worker_capabilities_multiple_runtimes() {
        let _guard = test_guard!();
        let caps = WorkerCapabilities {
            rustc_version: Some("rustc 1.76.0".to_string()),
            bun_version: Some("1.0.25".to_string()),
            node_version: Some("v20.11.0".to_string()),
            npm_version: Some("10.2.4".to_string()),
            ..Default::default()
        };

        assert!(caps.has_rust());
        assert!(caps.has_bun());
        assert!(caps.has_node());
    }

    #[test]
    fn test_worker_capabilities_serialization_empty() {
        let _guard = test_guard!();
        let caps = WorkerCapabilities::new();
        let json = serde_json::to_string(&caps).unwrap();
        // Empty capabilities should serialize to {}
        assert_eq!(json, "{}");

        // And deserialize back correctly
        let parsed: WorkerCapabilities = serde_json::from_str(&json).unwrap();
        assert!(!parsed.has_rust());
        assert!(!parsed.has_bun());
        assert!(!parsed.has_node());
    }

    #[test]
    fn test_worker_capabilities_serialization_with_versions() {
        let _guard = test_guard!();
        let caps = WorkerCapabilities {
            rustc_version: Some("rustc 1.76.0-nightly".to_string()),
            bun_version: Some("1.0.25".to_string()),
            node_version: None,
            npm_version: None,
            ..Default::default()
        };

        let json = serde_json::to_string(&caps).unwrap();
        assert!(json.contains("rustc_version"));
        assert!(json.contains("rustc 1.76.0-nightly"));
        assert!(json.contains("bun_version"));
        assert!(json.contains("1.0.25"));
        // None fields should be skipped
        assert!(!json.contains("node_version"));
        assert!(!json.contains("npm_version"));

        let parsed: WorkerCapabilities = serde_json::from_str(&json).unwrap();
        assert!(parsed.has_rust());
        assert!(parsed.has_bun());
        assert!(!parsed.has_node());
    }

    #[test]
    fn test_worker_capabilities_deserialization_partial() {
        let _guard = test_guard!();
        // Deserialize JSON with only some fields
        let json = r#"{"bun_version": "1.0.0"}"#;
        let caps: WorkerCapabilities = serde_json::from_str(json).unwrap();

        assert!(!caps.has_rust());
        assert!(caps.has_bun());
        assert!(!caps.has_node());
        assert_eq!(caps.bun_version, Some("1.0.0".to_string()));
    }

    #[test]
    fn test_worker_capabilities_clone() {
        let _guard = test_guard!();
        let caps = WorkerCapabilities {
            rustc_version: Some("1.76.0".to_string()),
            bun_version: None,
            node_version: Some("v20".to_string()),
            npm_version: None,
            projects_root_ok: Some(false),
            projects_root_issue: Some("alias_missing".to_string()),
            projects_root_checked_at_unix_ms: Some(123),
            ..Default::default()
        };

        let cloned = caps.clone();
        assert_eq!(cloned.rustc_version, caps.rustc_version);
        assert_eq!(cloned.bun_version, caps.bun_version);
        assert_eq!(cloned.node_version, caps.node_version);
        assert_eq!(cloned.npm_version, caps.npm_version);
        assert_eq!(cloned.projects_root_ok, caps.projects_root_ok);
        assert_eq!(cloned.projects_root_issue, caps.projects_root_issue);
        assert_eq!(
            cloned.projects_root_checked_at_unix_ms,
            caps.projects_root_checked_at_unix_ms
        );
    }

    #[test]
    fn test_worker_capabilities_topology_health_helper() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        assert_eq!(caps.is_topology_healthy(), None);

        caps.projects_root_ok = Some(true);
        assert_eq!(caps.is_topology_healthy(), Some(true));

        caps.projects_root_ok = Some(false);
        assert_eq!(caps.is_topology_healthy(), Some(false));
    }

    #[test]
    fn test_selection_reason_serialization() {
        let _guard = test_guard!();
        // Test that all variants serialize to snake_case
        assert_eq!(
            serde_json::to_string(&SelectionReason::Success).unwrap(),
            "\"success\""
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::NoWorkersConfigured).unwrap(),
            "\"no_workers_configured\""
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::AllWorkersUnreachable).unwrap(),
            "\"all_workers_unreachable\""
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::AllCircuitsOpen).unwrap(),
            "\"all_circuits_open\""
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::AllWorkersBusy).unwrap(),
            "\"all_workers_busy\""
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::NoWorkersPassedHealth).unwrap(),
            "\"no_workers_passed_health\""
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::AllWorkersFailedPreflight).unwrap(),
            "\"all_workers_failed_preflight\""
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::AllWorkersFailedConvergence).unwrap(),
            "\"all_workers_failed_convergence\""
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::NoAdmissibleWorkers(
                "critical_pressure=1,insufficient_slots=1".to_string()
            ))
            .unwrap(),
            "{\"no_admissible_workers\":\"critical_pressure=1,insufficient_slots=1\"}"
        );
        assert_eq!(
            serde_json::to_string(&SelectionReason::NoMatchingWorkers).unwrap(),
            "\"no_matching_workers\""
        );
    }

    #[test]
    fn test_selection_reason_with_error() {
        let _guard = test_guard!();
        let reason = SelectionReason::SelectionError("test error".to_string());
        let json = serde_json::to_string(&reason).unwrap();
        assert!(json.contains("selection_error"));
        assert!(json.contains("test error"));
    }

    #[test]
    fn test_selection_reason_deserialization() {
        let _guard = test_guard!();
        assert_eq!(
            serde_json::from_str::<SelectionReason>("\"success\"").unwrap(),
            SelectionReason::Success
        );
        assert_eq!(
            serde_json::from_str::<SelectionReason>("\"all_workers_busy\"").unwrap(),
            SelectionReason::AllWorkersBusy
        );
        assert_eq!(
            serde_json::from_str::<SelectionReason>("\"no_workers_passed_health\"").unwrap(),
            SelectionReason::NoWorkersPassedHealth
        );
        assert_eq!(
            serde_json::from_str::<SelectionReason>("\"all_workers_failed_preflight\"").unwrap(),
            SelectionReason::AllWorkersFailedPreflight
        );
        assert_eq!(
            serde_json::from_str::<SelectionReason>("\"all_workers_failed_convergence\"").unwrap(),
            SelectionReason::AllWorkersFailedConvergence
        );
        assert_eq!(
            serde_json::from_str::<SelectionReason>(
                "{\"no_admissible_workers\":\"critical_pressure=1\"}"
            )
            .unwrap(),
            SelectionReason::NoAdmissibleWorkers("critical_pressure=1".to_string())
        );
    }

    #[test]
    fn test_selection_reason_display() {
        let _guard = test_guard!();
        assert_eq!(
            SelectionReason::Success.to_string(),
            "worker assigned successfully"
        );
        assert_eq!(
            SelectionReason::NoWorkersConfigured.to_string(),
            "no workers configured"
        );
        assert_eq!(
            SelectionReason::AllWorkersUnreachable.to_string(),
            "all workers unreachable"
        );
        assert_eq!(
            SelectionReason::AllWorkersBusy.to_string(),
            "all workers at capacity"
        );
        assert_eq!(
            SelectionReason::NoWorkersPassedHealth.to_string(),
            "no workers passed health thresholds"
        );
        assert_eq!(
            SelectionReason::AllWorkersFailedPreflight.to_string(),
            "all workers failed preflight checks"
        );
        assert_eq!(
            SelectionReason::AllWorkersFailedConvergence.to_string(),
            "all workers failed repo convergence checks"
        );
        assert_eq!(
            SelectionReason::NoAdmissibleWorkers(
                "critical_pressure=1,insufficient_slots=1".to_string()
            )
            .to_string(),
            "no admissible workers: critical_pressure=1,insufficient_slots=1"
        );
        assert_eq!(
            SelectionReason::SelectionError("oops".to_string()).to_string(),
            "selection error: oops"
        );
    }

    #[test]
    fn test_selection_response_with_worker() {
        let _guard = test_guard!();
        let response = SelectionResponse {
            worker: Some(SelectedWorker {
                id: WorkerId::new("test"),
                host: "localhost".to_string(),
                user: "user".to_string(),
                identity_file: "~/.ssh/id_rsa".to_string(),
                slots_available: 8,
                speed_score: 75.0,
                declared_os: None,
            }),
            reason: SelectionReason::Success,
            build_id: None,
            diagnostics: None,
        };

        let json = serde_json::to_string(&response).unwrap();
        assert!(json.contains("\"reason\":\"success\""));
        assert!(json.contains("\"id\":\"test\""));
    }

    #[test]
    fn test_selection_response_without_worker() {
        let _guard = test_guard!();
        let response = SelectionResponse {
            worker: None,
            reason: SelectionReason::AllWorkersBusy,
            build_id: None,
            diagnostics: None,
        };

        let json = serde_json::to_string(&response).unwrap();
        assert!(json.contains("\"worker\":null"));
        assert!(json.contains("\"reason\":\"all_workers_busy\""));
    }

    #[test]
    fn golden_selection_response_with_no_workers_passed_health_diagnostics() {
        let _guard = test_guard!();
        const FIXTURE_PATH: &str =
            "../../tests/goldens/ft_4tp7g/selection_diagnostics_no_workers_passed_health.json";
        let fixture: serde_json::Value = serde_json::from_str(include_str!(
            "../../tests/goldens/ft_4tp7g/selection_diagnostics_no_workers_passed_health.json"
        ))
        .expect("selection diagnostics golden fixture parses");
        assert_eq!(
            fixture["schema_version"],
            "rch.golden.selection_diagnostics.v1"
        );

        let response = SelectionResponse {
            worker: None,
            reason: SelectionReason::NoWorkersPassedHealth,
            build_id: None,
            diagnostics: Some(SelectionDiagnostics {
                required_runtime: RequiredRuntime::Rust,
                estimated_cores: 2,
                min_success_rate: 0.8,
                fallback_min_success_rate: 0.5,
                active_project_exclusion_count: 0,
                workers: vec![WorkerSelectionDiagnostic {
                    worker_id: WorkerId::new("fixture-worker"),
                    status: "healthy".to_string(),
                    circuit_state: "closed".to_string(),
                    pressure_state: "healthy".to_string(),
                    pressure_reason_code: "pressure.ok".to_string(),
                    success_rate: Some(0.2),
                    min_success_rate: 0.8,
                    fallback_min_success_rate: 0.5,
                    required_runtime: RequiredRuntime::Rust,
                    runtime_available: true,
                    available_slots: 8,
                    total_slots: 8,
                    estimated_cores: 2,
                    active_project_excluded: false,
                    final_decision: WorkerSelectionDiagnosticDecision::Deny,
                    final_reason: "success_rate 0.20 < fallback 0.50".to_string(),
                    reason_codes: vec![
                        "health.below_min_success_rate".to_string(),
                        "health.below_fallback_min_success_rate".to_string(),
                    ],
                }],
            }),
        };

        assert_golden_json(
            serde_json::to_value(&response).expect("response serializes"),
            &fixture["expected_response"],
            FIXTURE_PATH,
        );
    }

    #[test]
    fn test_selection_response_roundtrip() {
        let _guard = test_guard!();
        let original = SelectionResponse {
            worker: Some(SelectedWorker {
                id: WorkerId::new("worker-1"),
                host: "192.168.1.100".to_string(),
                user: "ubuntu".to_string(),
                identity_file: "/path/to/key".to_string(),
                slots_available: 16,
                speed_score: 90.5,
                declared_os: None,
            }),
            reason: SelectionReason::Success,
            build_id: None,
            diagnostics: None,
        };

        let json = serde_json::to_string(&original).unwrap();
        let parsed: SelectionResponse = serde_json::from_str(&json).unwrap();

        assert!(parsed.worker.is_some());
        let worker = parsed.worker.unwrap();
        assert_eq!(worker.id.as_str(), "worker-1");
        assert_eq!(worker.host, "192.168.1.100");
        assert_eq!(worker.slots_available, 16);
        assert_eq!(parsed.reason, SelectionReason::Success);
    }

    #[test]
    fn test_build_heartbeat_phase_roundtrip() {
        let _guard = test_guard!();
        let json = serde_json::to_string(&BuildHeartbeatPhase::SyncUp).unwrap();
        assert_eq!(json, "\"sync_up\"");
        let parsed: BuildHeartbeatPhase = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed, BuildHeartbeatPhase::SyncUp);
    }

    #[test]
    fn test_build_heartbeat_request_roundtrip() {
        let _guard = test_guard!();
        let request = BuildHeartbeatRequest {
            build_id: 42,
            worker_id: WorkerId::new("worker-a"),
            hook_pid: Some(12345),
            local_wrapper_id: Some("wrapper-abc123".to_string()),
            remote_pgid_file: Some("/tmp/rch/project/hash/.rch-run/42.pgid".to_string()),
            phase: BuildHeartbeatPhase::Execute,
            detail: Some("Compiling crates".to_string()),
            progress_counter: Some(7),
            progress_percent: Some(42.5),
        };

        let json = serde_json::to_string(&request).unwrap();
        let parsed: BuildHeartbeatRequest = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.build_id, 42);
        assert_eq!(parsed.worker_id.as_str(), "worker-a");
        assert_eq!(
            parsed.remote_pgid_file,
            Some("/tmp/rch/project/hash/.rch-run/42.pgid".to_string())
        );
        assert_eq!(parsed.phase, BuildHeartbeatPhase::Execute);
        assert_eq!(parsed.progress_counter, Some(7));
        assert_eq!(parsed.progress_percent, Some(42.5));
    }

    // CircuitStats tests

    #[test]
    fn test_circuit_stats_new() {
        let _guard = test_guard!();
        let stats = CircuitStats::new();
        assert_eq!(stats.state(), CircuitState::Closed);
        assert_eq!(stats.consecutive_failures(), 0);
        assert_eq!(stats.error_rate(), 0.0);
        assert!(stats.opened_at().is_none());
    }

    #[test]
    fn test_circuit_stats_record_success() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        stats.record_success();
        stats.record_success();
        assert_eq!(stats.error_rate(), 0.0);
        assert_eq!(stats.consecutive_failures(), 0);
    }

    #[test]
    fn test_circuit_stats_record_failure() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        stats.record_failure();
        stats.record_failure();
        assert_eq!(stats.consecutive_failures(), 2);
        assert_eq!(stats.error_rate(), 1.0); // 2 failures, 0 successes
    }

    #[test]
    fn test_circuit_stats_error_rate() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        stats.record_success();
        stats.record_success();
        stats.record_failure();
        // 1 failure / 3 total = 0.333...
        assert!((stats.error_rate() - 0.333).abs() < 0.01);
    }

    #[test]
    fn test_circuit_stats_should_open_consecutive_failures() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        let config = CircuitBreakerConfig::default(); // threshold = 3

        stats.record_failure();
        assert!(!stats.should_open(&config));

        stats.record_failure();
        assert!(!stats.should_open(&config));

        stats.record_failure();
        assert!(stats.should_open(&config)); // 3 consecutive failures
    }

    #[test]
    fn test_circuit_stats_should_open_error_rate() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        let config = CircuitBreakerConfig {
            error_rate_threshold: 0.5,
            ..Default::default()
        };

        // Need at least 5 samples
        stats.record_success();
        stats.record_success();
        stats.record_failure();
        stats.record_failure();
        assert!(!stats.should_open(&config)); // Only 4 samples

        stats.record_failure(); // 5 samples: 2 success, 3 failures = 60% error rate
        assert!(stats.should_open(&config));
    }

    #[test]
    fn test_circuit_close_clears_recent_results_no_flap() {
        let _guard = test_guard!();
        // Regression: error_rate() is computed over recent_results, so close()
        // must clear that history or a recovered circuit immediately re-opens
        // on stale pre-failure samples (flapping).
        let mut stats = CircuitStats::new();
        let config = CircuitBreakerConfig {
            error_rate_threshold: 0.5,
            ..Default::default()
        };
        for _ in 0..6 {
            stats.record_failure();
        }
        assert!(stats.should_open(&config));
        stats.open();
        // Recover via half-open probes, then close.
        stats.half_open();
        stats.record_success();
        stats.record_success();
        stats.close();
        // Error rate must be reset and the circuit must NOT immediately re-open.
        assert_eq!(stats.error_rate(), 0.0);
        assert!(!stats.should_open(&config));
    }

    #[test]
    fn test_apply_health_outcome_opens_on_failure_threshold() {
        let _guard = test_guard!();
        // failure_threshold=3: three unhealthy outcomes open the circuit.
        let config = CircuitBreakerConfig {
            failure_threshold: 3,
            open_cooldown_secs: 300,
            ..Default::default()
        };
        let mut stats = CircuitStats::new();
        assert_eq!(
            stats.apply_health_outcome(false, false, &config),
            CircuitState::Closed
        );
        assert_eq!(
            stats.apply_health_outcome(false, false, &config),
            CircuitState::Closed
        );
        assert_eq!(
            stats.apply_health_outcome(false, false, &config),
            CircuitState::Open
        );
    }

    #[test]
    fn test_apply_health_outcome_half_open_two_probe_recovery() {
        let _guard = test_guard!();
        // success_threshold=2: recovery from Open needs exactly two clean probes.
        // open_cooldown_secs=0 makes should_half_open() true immediately, so the
        // engine's trailing nudge leaves an opened circuit already in HalfOpen —
        // exactly the state a real daemon would be in on the next poll once the
        // cooldown had elapsed. From HalfOpen the engine credits successes.
        let config = CircuitBreakerConfig {
            failure_threshold: 3,
            success_threshold: 2,
            open_cooldown_secs: 0,
            ..Default::default()
        };
        let mut stats = CircuitStats::new();
        for _ in 0..3 {
            stats.apply_health_outcome(false, false, &config);
        }
        // With a 0s cooldown the engine nudges Open -> HalfOpen immediately.
        assert_eq!(stats.state(), CircuitState::HalfOpen);

        // First healthy probe in half-open: one success credited, not yet closed.
        let after_first = stats.apply_health_outcome(true, false, &config);
        assert_eq!(after_first, CircuitState::HalfOpen);

        // Second healthy probe: threshold met -> Closed.
        let after_second = stats.apply_health_outcome(true, false, &config);
        assert_eq!(after_second, CircuitState::Closed);
    }

    #[test]
    fn test_apply_health_outcome_transient_half_open_does_not_reopen() {
        let _guard = test_guard!();
        // A transient failure while half-open must NOT reset the recovery streak
        // — a network blip can't undo an in-progress recovery.
        let config = CircuitBreakerConfig {
            failure_threshold: 3,
            success_threshold: 2,
            open_cooldown_secs: 0,
            ..Default::default()
        };
        let mut stats = CircuitStats::new();
        for _ in 0..3 {
            stats.apply_health_outcome(false, false, &config);
        }
        // 0s cooldown -> already nudged to HalfOpen.
        assert_eq!(stats.state(), CircuitState::HalfOpen);

        // One clean probe (1 of 2 needed).
        assert_eq!(
            stats.apply_health_outcome(true, false, &config),
            CircuitState::HalfOpen
        );

        // Transient failure in half-open: stays HalfOpen (does NOT reopen the
        // circuit or restart the cooldown). The key contrast with a real failure
        // is that the circuit does not drop back to Open — probing continues in
        // place, so recovery resumes immediately on the next clean probes rather
        // than waiting out a fresh cooldown.
        let after_transient = stats.apply_health_outcome(false, true, &config);
        assert_eq!(after_transient, CircuitState::HalfOpen);

        // Recovery still completes from half-open with the required clean probes.
        assert_eq!(
            stats.apply_health_outcome(true, false, &config),
            CircuitState::HalfOpen
        );
        assert_eq!(
            stats.apply_health_outcome(true, false, &config),
            CircuitState::Closed
        );
    }

    #[test]
    fn test_apply_health_outcome_real_vs_transient_half_open_failure() {
        let _guard = test_guard!();
        // With a positive cooldown the difference is directly observable: a REAL
        // half-open failure reopens the circuit (state -> Open, must wait out the
        // cooldown again), whereas a TRANSIENT half-open failure keeps it in
        // HalfOpen so recovery continues without a fresh cooldown.
        //
        // Manually drive to HalfOpen (open() + half_open()) so the positive
        // cooldown doesn't force a real wallclock wait to reach half-open.
        let config = CircuitBreakerConfig {
            failure_threshold: 3,
            success_threshold: 2,
            open_cooldown_secs: 300,
            ..Default::default()
        };

        // Real failure path.
        let mut real = CircuitStats::new();
        real.open();
        real.half_open();
        assert_eq!(real.state(), CircuitState::HalfOpen);
        let after_real = real.apply_health_outcome(false, false, &config);
        assert_eq!(
            after_real,
            CircuitState::Open,
            "a real half-open failure must reopen the circuit"
        );

        // Transient failure path (same starting state).
        let mut transient = CircuitStats::new();
        transient.open();
        transient.half_open();
        assert_eq!(transient.state(), CircuitState::HalfOpen);
        let after_transient = transient.apply_health_outcome(false, true, &config);
        assert_eq!(
            after_transient,
            CircuitState::HalfOpen,
            "a transient half-open failure must NOT reopen the circuit"
        );
    }

    /// Issue #48 regression: a worker whose liveness probe (`echo` over a warm
    /// session) keeps succeeding while every real command times out must NOT
    /// be re-admitted. Liveness successes alone cannot close a half-open
    /// circuit while command-class evidence is outstanding.
    #[test]
    fn test_command_failure_blocks_half_open_close_on_liveness_successes() {
        let _guard = test_guard!();
        let config = CircuitBreakerConfig {
            failure_threshold: 3,
            success_threshold: 2,
            open_cooldown_secs: 0,
            command_failure_threshold: 2,
            ..Default::default()
        };
        let mut stats = CircuitStats::new();
        stats.open();
        stats.half_open();

        // The half-open trial: a real command hangs on the worker.
        assert_eq!(
            stats.apply_command_outcome(false, &config),
            CircuitState::HalfOpen,
            "cooldown 0: reopen then immediately half-open again"
        );
        assert_eq!(stats.consecutive_command_failures(), 1);

        // Liveness keeps answering — many times over.
        for _ in 0..10 {
            assert_eq!(
                stats.apply_health_outcome(true, false, &config),
                CircuitState::HalfOpen,
                "liveness successes must not close the circuit while a command failure is outstanding"
            );
        }

        // Only a command-class success clears the evidence and recovers.
        assert_eq!(
            stats.apply_command_outcome(true, &config),
            CircuitState::Closed
        );
        assert_eq!(stats.consecutive_command_failures(), 0);
        assert_eq!(stats.reopen_streak(), 0);
    }

    /// Issue #48: interleaved liveness successes reset `failure_threshold`, so
    /// command failures need their own threshold to open a Closed circuit.
    #[test]
    fn test_command_failures_open_closed_circuit_despite_liveness_successes() {
        let _guard = test_guard!();
        let config = CircuitBreakerConfig {
            failure_threshold: 3,
            command_failure_threshold: 2,
            open_cooldown_secs: 300,
            ..Default::default()
        };
        let mut stats = CircuitStats::new();

        assert_eq!(
            stats.apply_command_outcome(false, &config),
            CircuitState::Closed
        );
        // Two liveness successes in between: consecutive_failures is reset...
        stats.apply_health_outcome(true, false, &config);
        stats.apply_health_outcome(true, false, &config);
        assert_eq!(stats.consecutive_failures(), 0);
        // ...but the command-failure count is not, and the second one opens.
        assert_eq!(
            stats.apply_command_outcome(false, &config),
            CircuitState::Open
        );
        assert_eq!(stats.reopen_streak(), 1);
    }

    /// A command failure in HalfOpen always reopens: "connected but the
    /// command never completed" is exactly what the trial exists to detect,
    /// so there is no transient exemption on this channel.
    #[test]
    fn test_command_failure_in_half_open_always_reopens() {
        let _guard = test_guard!();
        let config = CircuitBreakerConfig {
            open_cooldown_secs: 300,
            ..Default::default()
        };
        let mut stats = CircuitStats::new();
        stats.open();
        stats.half_open();
        assert_eq!(
            stats.apply_command_outcome(false, &config),
            CircuitState::Open
        );
        assert_eq!(stats.reopen_streak(), 2);
    }

    /// Repeated Open -> HalfOpen -> Open cycles lengthen the cooldown
    /// (exponential, capped) so a flapping worker converges to Open instead of
    /// being re-admitted on every base cooldown.
    #[test]
    fn test_reopen_streak_backs_off_cooldown_until_cap() {
        let _guard = test_guard!();
        let config = CircuitBreakerConfig {
            open_cooldown_secs: 30,
            open_cooldown_max_secs: 100,
            ..Default::default()
        };
        assert_eq!(config.cooldown_ms_for_streak(0), 30_000);
        assert_eq!(config.cooldown_ms_for_streak(1), 30_000);
        assert_eq!(config.cooldown_ms_for_streak(2), 60_000);
        assert_eq!(config.cooldown_ms_for_streak(3), 100_000, "capped");
        assert_eq!(config.cooldown_ms_for_streak(40), 100_000, "no overflow");

        // A cap below the base never shortens the base cooldown.
        let inverted = CircuitBreakerConfig {
            open_cooldown_secs: 30,
            open_cooldown_max_secs: 5,
            ..Default::default()
        };
        assert_eq!(inverted.cooldown_ms_for_streak(3), 30_000);

        // The streak is observable through recovery_remaining_secs: after two
        // opens the remaining cooldown exceeds the base.
        let mut stats = CircuitStats::new();
        stats.open();
        stats.half_open();
        stats.open();
        assert_eq!(stats.reopen_streak(), 2);
        let remaining = stats.recovery_remaining_secs(&config).unwrap();
        assert!(
            remaining > 30 && remaining <= 60,
            "second open should wait ~60s, got {remaining}"
        );

        // Closing resets the streak, so the next open starts from the base.
        stats.close();
        assert_eq!(stats.reopen_streak(), 0);
        stats.open();
        assert!(stats.recovery_remaining_secs(&config).unwrap() <= 30);
    }

    #[test]
    fn test_circuit_stats_success_resets_consecutive_failures() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        stats.record_failure();
        stats.record_failure();
        assert_eq!(stats.consecutive_failures(), 2);

        stats.record_success();
        assert_eq!(stats.consecutive_failures(), 0);
    }

    #[test]
    fn test_circuit_stats_open_transition() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        stats.open();

        assert_eq!(stats.state(), CircuitState::Open);
        assert!(stats.opened_at().is_some());
    }

    #[test]
    fn test_circuit_stats_half_open_transition() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        stats.open();
        stats.half_open();

        assert_eq!(stats.state(), CircuitState::HalfOpen);
    }

    #[test]
    fn test_circuit_stats_close_transition() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        stats.open();
        stats.half_open();
        stats.close();

        assert_eq!(stats.state(), CircuitState::Closed);
        assert!(stats.opened_at().is_none());
        assert_eq!(stats.consecutive_failures(), 0);
    }

    #[test]
    fn test_circuit_stats_should_close() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        let config = CircuitBreakerConfig {
            success_threshold: 2,
            ..Default::default()
        };

        stats.open();
        stats.half_open();

        assert!(!stats.should_close(&config)); // 0 successes

        stats.record_success();
        assert!(!stats.should_close(&config)); // 1 success

        stats.record_success();
        assert!(stats.should_close(&config)); // 2 successes
    }

    #[test]
    fn test_circuit_stats_can_probe() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        let config = CircuitBreakerConfig {
            half_open_max_probes: 1,
            ..Default::default()
        };

        // Can't probe when closed
        assert!(!stats.can_probe(&config));

        stats.open();
        // Can't probe when open
        assert!(!stats.can_probe(&config));

        stats.half_open();
        // Can probe when half-open
        assert!(stats.can_probe(&config));

        // Start a probe
        assert!(stats.start_probe(&config));
        // Can't start another probe
        assert!(!stats.can_probe(&config));
        assert!(!stats.start_probe(&config));
    }

    #[test]
    fn test_circuit_stats_probe_completion() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        let config = CircuitBreakerConfig {
            half_open_max_probes: 1,
            ..Default::default()
        };

        stats.open();
        stats.half_open();
        stats.start_probe(&config);

        // Probe completes with success
        stats.record_success();

        // Can start another probe
        assert!(stats.can_probe(&config));
    }

    #[test]
    fn test_circuit_stats_failure_in_half_open() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        let config = CircuitBreakerConfig {
            success_threshold: 2,
            ..Default::default()
        };

        stats.open();
        stats.half_open();

        stats.record_success();
        assert_eq!(stats.consecutive_failures(), 0);

        // Failure resets consecutive successes
        stats.record_failure();
        assert!(!stats.should_close(&config));

        // Need 2 more successes
        stats.record_success();
        stats.record_success();
        assert!(stats.should_close(&config));
    }

    #[test]
    fn test_circuit_stats_reset_window() {
        let _guard = test_guard!();
        let mut stats = CircuitStats::new();
        stats.record_success();
        stats.record_failure();

        assert!(stats.error_rate() > 0.0);

        stats.reset_window();
        assert_eq!(stats.error_rate(), 0.0);
    }

    #[test]
    fn test_circuit_state_transitions_are_deterministic() {
        let _guard = test_guard!();
        let config = CircuitBreakerConfig::default();
        let mut stats = CircuitStats::new();

        // Start closed
        assert_eq!(stats.state(), CircuitState::Closed);

        // Cause failures to open
        for _ in 0..3 {
            stats.record_failure();
        }
        assert!(stats.should_open(&config));
        stats.open();
        assert_eq!(stats.state(), CircuitState::Open);

        // Transition to half-open after cooldown would be checked
        stats.half_open();
        assert_eq!(stats.state(), CircuitState::HalfOpen);

        // Successes to close
        for _ in 0..2 {
            stats.record_success();
        }
        assert!(stats.should_close(&config));
        stats.close();
        assert_eq!(stats.state(), CircuitState::Closed);
    }

    // ========================================================================
    // Compilation Timing Tests
    // ========================================================================

    fn make_test_metrics(
        speedup: Option<f64>,
        total_secs: u64,
        success: bool,
    ) -> CompilationMetrics {
        CompilationMetrics {
            project_id: "test-project".to_string(),
            worker_id: "worker-1".to_string(),
            timestamp: chrono::Utc::now(),
            timing: CompilationTimingBreakdown {
                rsync_up: Duration::from_millis(100),
                remote_build: Duration::from_millis(800),
                rsync_down: Duration::from_millis(100),
                total: Duration::from_secs(total_secs),
            },
            local_build_time: speedup.map(|s| Duration::from_secs_f64(total_secs as f64 * s)),
            speedup,
            files_synced: 100,
            bytes_transferred: 1_000_000,
            exit_code: if success { 0 } else { 1 },
            success,
        }
    }

    #[test]
    fn test_compilation_timing_breakdown_default() {
        let _guard = test_guard!();
        let timing = CompilationTimingBreakdown::default();
        assert_eq!(timing.rsync_up, Duration::ZERO);
        assert_eq!(timing.remote_build, Duration::ZERO);
        assert_eq!(timing.rsync_down, Duration::ZERO);
        assert_eq!(timing.total, Duration::ZERO);
    }

    #[test]
    fn test_compilation_timing_breakdown_serialization() {
        let _guard = test_guard!();
        let timing = CompilationTimingBreakdown {
            rsync_up: Duration::from_millis(100),
            remote_build: Duration::from_millis(2000),
            rsync_down: Duration::from_millis(50),
            total: Duration::from_millis(2150),
        };

        let json = serde_json::to_string(&timing).unwrap();
        assert!(json.contains("100")); // rsync_up
        assert!(json.contains("2000")); // remote_build
        assert!(json.contains("50")); // rsync_down
        assert!(json.contains("2150")); // total

        let parsed: CompilationTimingBreakdown = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.rsync_up, timing.rsync_up);
        assert_eq!(parsed.remote_build, timing.remote_build);
        assert_eq!(parsed.rsync_down, timing.rsync_down);
        assert_eq!(parsed.total, timing.total);
    }

    #[test]
    fn test_command_timing_breakdown_serialization() {
        let _guard = test_guard!();
        let timing = CommandTimingBreakdown {
            classify: Some(Duration::from_millis(2)),
            select: None,
            sync_up: Some(Duration::from_millis(120)),
            exec: Some(Duration::from_millis(900)),
            sync_down: None,
            cleanup: Some(Duration::from_millis(5)),
            total: Some(Duration::from_millis(1027)),
        };

        let json = serde_json::to_string(&timing).unwrap();
        assert!(json.contains("\"classify\":2"));
        assert!(json.contains("\"select\":null"));
        assert!(json.contains("\"sync_up\":120"));
        assert!(json.contains("\"exec\":900"));
        assert!(json.contains("\"sync_down\":null"));
        assert!(json.contains("\"cleanup\":5"));
        assert!(json.contains("\"total\":1027"));

        let parsed: CommandTimingBreakdown = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.classify, timing.classify);
        assert_eq!(parsed.select, timing.select);
        assert_eq!(parsed.sync_up, timing.sync_up);
        assert_eq!(parsed.exec, timing.exec);
        assert_eq!(parsed.sync_down, timing.sync_down);
        assert_eq!(parsed.cleanup, timing.cleanup);
        assert_eq!(parsed.total, timing.total);
    }

    #[test]
    fn test_compilation_metrics_calculate_speedup() {
        let _guard = test_guard!();
        let mut metrics = CompilationMetrics {
            timing: CompilationTimingBreakdown {
                total: Duration::from_secs(10),
                ..Default::default()
            },
            local_build_time: Some(Duration::from_secs(30)),
            ..Default::default()
        };

        metrics.calculate_speedup();

        assert!(metrics.speedup.is_some());
        let speedup = metrics.speedup.unwrap();
        assert!(
            (speedup - 3.0).abs() < 0.01,
            "Expected 3.0x speedup, got {}",
            speedup
        );
    }

    #[test]
    fn test_compilation_metrics_is_beneficial() {
        let _guard = test_guard!();
        let mut metrics = CompilationMetrics::default();

        // No speedup calculated
        assert!(!metrics.is_beneficial());

        // Speedup > 1 is beneficial
        metrics.speedup = Some(1.5);
        assert!(metrics.is_beneficial());

        // Speedup < 1 is not beneficial
        metrics.speedup = Some(0.8);
        assert!(!metrics.is_beneficial());

        // Speedup = 1 is not beneficial (exactly same speed)
        metrics.speedup = Some(1.0);
        assert!(!metrics.is_beneficial());
    }

    #[test]
    fn test_compilation_metrics_serialization() {
        let _guard = test_guard!();
        let metrics = make_test_metrics(Some(2.5), 10, true);
        let json = serde_json::to_string(&metrics).unwrap();

        assert!(json.contains("test-project"));
        assert!(json.contains("worker-1"));
        assert!(json.contains("2.5")); // speedup

        let parsed: CompilationMetrics = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.project_id, "test-project");
        assert_eq!(parsed.worker_id, "worker-1");
        assert!(parsed.success);
    }

    #[test]
    fn test_compilation_timer_new() {
        let _guard = test_guard!();
        let timer = CompilationTimer::new("my-project", "my-worker");
        assert_eq!(timer.project_id, "my-project");
        assert_eq!(timer.worker_id, "my-worker");
        assert!(timer.rsync_up.is_none());
        assert!(timer.remote_build.is_none());
        assert!(timer.rsync_down.is_none());
    }

    #[test]
    fn test_compilation_timer_phases() {
        let _guard = test_guard!();
        let mut timer = CompilationTimer::new("test", "worker");

        // Simulate rsync up
        std::thread::sleep(Duration::from_millis(5));
        timer.end_rsync_up();
        assert!(timer.rsync_up.is_some());
        assert!(timer.rsync_up.unwrap() >= Duration::from_millis(5));

        // Simulate remote build
        std::thread::sleep(Duration::from_millis(10));
        timer.end_remote_build();
        assert!(timer.remote_build.is_some());
        assert!(timer.remote_build.unwrap() >= Duration::from_millis(10));

        // Simulate rsync down
        std::thread::sleep(Duration::from_millis(5));
        timer.end_rsync_down();
        assert!(timer.rsync_down.is_some());
        assert!(timer.rsync_down.unwrap() >= Duration::from_millis(5));

        // Finish
        let metrics = timer.finish(0, 50, 500_000);
        assert_eq!(metrics.project_id, "test");
        assert_eq!(metrics.worker_id, "worker");
        assert!(metrics.success);
        assert!(metrics.timing.total >= Duration::from_millis(20));
        assert_eq!(metrics.files_synced, 50);
        assert_eq!(metrics.bytes_transferred, 500_000);
    }

    #[test]
    fn test_metrics_aggregator_new() {
        let _guard = test_guard!();
        let agg = MetricsAggregator::new(100);
        assert_eq!(agg.count(), 0);
        assert!(agg.average_speedup().is_none());
    }

    #[test]
    fn test_metrics_aggregator_record() {
        let _guard = test_guard!();
        let mut agg = MetricsAggregator::new(3);

        // Record first metric
        agg.record(make_test_metrics(Some(2.0), 10, true));
        assert_eq!(agg.count(), 1);

        // Record up to max
        agg.record(make_test_metrics(Some(3.0), 20, true));
        agg.record(make_test_metrics(Some(4.0), 30, true));
        assert_eq!(agg.count(), 3);

        // Exceed max should drop oldest
        agg.record(make_test_metrics(Some(5.0), 40, true));
        assert_eq!(agg.count(), 3);
    }

    #[test]
    fn test_metrics_aggregator_average_speedup() {
        let _guard = test_guard!();
        let mut agg = MetricsAggregator::new(100);

        // Empty case
        assert!(agg.average_speedup().is_none());

        // Add metrics with speedups 1, 2, 3, 4, 5 -> average = 3
        for i in 1..=5 {
            agg.record(make_test_metrics(Some(i as f64), 10, true));
        }

        let avg = agg.average_speedup().unwrap();
        assert!((avg - 3.0).abs() < 0.01, "Expected 3.0, got {}", avg);
    }

    #[test]
    fn test_metrics_aggregator_average_speedup_with_none() {
        let _guard = test_guard!();
        let mut agg = MetricsAggregator::new(100);

        // Mix of Some and None speedups
        agg.record(make_test_metrics(Some(2.0), 10, true));
        agg.record(make_test_metrics(None, 10, true)); // No speedup
        agg.record(make_test_metrics(Some(4.0), 10, true));

        // Average should only consider metrics with speedup
        let avg = agg.average_speedup().unwrap();
        assert!((avg - 3.0).abs() < 0.01, "Expected 3.0, got {}", avg);
    }

    #[test]
    fn test_metrics_aggregator_percentiles() {
        let _guard = test_guard!();
        let mut agg = MetricsAggregator::new(100);

        // Add metrics with total times 1s, 2s, 3s, ..., 10s
        for i in 1..=10 {
            agg.record(make_test_metrics(Some(1.0), i, true));
        }

        // p50 (median) should be around 5s
        let p50 = agg.p50_total_time().unwrap();
        assert!(p50 >= Duration::from_secs(5) && p50 <= Duration::from_secs(6));

        // p95 should be around 9s or 10s
        let p95 = agg.p95_total_time().unwrap();
        assert!(p95 >= Duration::from_secs(9) && p95 <= Duration::from_secs(10));

        // p99 should be 10s
        let p99 = agg.p99_total_time().unwrap();
        assert!(p99 >= Duration::from_secs(9));
    }

    #[test]
    fn test_metrics_aggregator_success_rate() {
        let _guard = test_guard!();
        let mut agg = MetricsAggregator::new(100);

        // Empty case
        assert_eq!(agg.success_rate(), 100.0);

        // All successful
        agg.record(make_test_metrics(Some(1.0), 10, true));
        agg.record(make_test_metrics(Some(1.0), 10, true));
        assert_eq!(agg.success_rate(), 100.0);

        // Add a failure (2 success, 1 failure = 66.67%)
        agg.record(make_test_metrics(Some(1.0), 10, false));
        let rate = agg.success_rate();
        assert!((rate - 66.67).abs() < 1.0, "Expected ~66.67%, got {}", rate);
    }

    #[test]
    fn test_metrics_aggregator_clear() {
        let _guard = test_guard!();
        let mut agg = MetricsAggregator::new(100);
        agg.record(make_test_metrics(Some(1.0), 10, true));
        agg.record(make_test_metrics(Some(2.0), 20, true));

        assert_eq!(agg.count(), 2);

        agg.clear();
        assert_eq!(agg.count(), 0);
        assert!(agg.average_speedup().is_none());
    }

    // CompilationConfig timeout tests

    #[test]
    fn test_remote_build_jobs_parse_display_and_serde() {
        let _guard = test_guard!();
        assert_eq!(RemoteBuildJobs::default(), RemoteBuildJobs::Auto);
        for (spelling, expected) in [
            ("auto", RemoteBuildJobs::Auto),
            (" AUTO ", RemoteBuildJobs::Auto),
            ("off", RemoteBuildJobs::Off),
            ("none", RemoteBuildJobs::Off),
            ("4", RemoteBuildJobs::Fixed(4)),
        ] {
            assert_eq!(
                RemoteBuildJobs::parse(spelling).unwrap(),
                expected,
                "{spelling}"
            );
        }
        assert!(RemoteBuildJobs::parse("0").is_err(), "cargo rejects -j 0");
        assert!(RemoteBuildJobs::parse("many").is_err());
        assert_eq!(RemoteBuildJobs::Fixed(4).to_string(), "4");
        assert_eq!(RemoteBuildJobs::Off.to_string(), "off");

        // TOML accepts both the string spellings and a bare integer, and an
        // absent key keeps the default.
        #[derive(Deserialize)]
        struct Wrap {
            compilation: CompilationConfig,
        }
        let parsed: Wrap = toml::from_str("[compilation]\nremote_build_jobs = \"off\"\n").unwrap();
        assert_eq!(parsed.compilation.remote_build_jobs, RemoteBuildJobs::Off);
        let parsed: Wrap = toml::from_str("[compilation]\nremote_build_jobs = 6\n").unwrap();
        assert_eq!(
            parsed.compilation.remote_build_jobs,
            RemoteBuildJobs::Fixed(6)
        );
        let parsed: Wrap = toml::from_str("[compilation]\nbuild_slots = 4\n").unwrap();
        assert_eq!(parsed.compilation.remote_build_jobs, RemoteBuildJobs::Auto);
        assert!(toml::from_str::<Wrap>("[compilation]\nremote_build_jobs = 0\n").is_err());

        // Serialization round-trips through the same spellings.
        let json = serde_json::to_string(&RemoteBuildJobs::Fixed(3)).unwrap();
        assert_eq!(json, "3");
        assert_eq!(
            serde_json::to_string(&RemoteBuildJobs::Auto).unwrap(),
            "\"auto\""
        );
        let back: RemoteBuildJobs = serde_json::from_str("\"auto\"").unwrap();
        assert_eq!(back, RemoteBuildJobs::Auto);
    }

    #[test]
    fn test_compilation_config_default_timeouts() {
        let _guard = test_guard!();
        let config = CompilationConfig::default();
        // Default build timeout: 5 minutes
        assert_eq!(config.build_timeout_sec, 300);
        // Default test timeout: 30 minutes
        assert_eq!(config.test_timeout_sec, 1800);
    }

    #[test]
    fn test_compilation_config_timeout_for_test_kinds() {
        let _guard = test_guard!();
        let config = CompilationConfig::default();

        // Test commands should get test_timeout_sec
        assert_eq!(
            config.timeout_for_kind(Some(crate::CompilationKind::CargoTest)),
            std::time::Duration::from_secs(1800)
        );
        assert_eq!(
            config.timeout_for_kind(Some(crate::CompilationKind::CargoNextest)),
            std::time::Duration::from_secs(1800)
        );
        assert_eq!(
            config.timeout_for_kind(Some(crate::CompilationKind::BunTest)),
            std::time::Duration::from_secs(config.bun_timeout_sec)
        );
    }

    #[test]
    fn test_compilation_config_timeout_for_build_kinds() {
        let _guard = test_guard!();
        let config = CompilationConfig::default();

        // Build/check commands should get build_timeout_sec
        assert_eq!(
            config.timeout_for_kind(Some(crate::CompilationKind::CargoBuild)),
            std::time::Duration::from_secs(300)
        );
        assert_eq!(
            config.timeout_for_kind(Some(crate::CompilationKind::CargoCheck)),
            std::time::Duration::from_secs(300)
        );
        assert_eq!(
            config.timeout_for_kind(Some(crate::CompilationKind::CargoClippy)),
            std::time::Duration::from_secs(300)
        );
        assert_eq!(
            config.timeout_for_kind(None),
            std::time::Duration::from_secs(300)
        );
    }

    #[test]
    fn test_compilation_config_custom_timeouts() {
        let _guard = test_guard!();
        let config = CompilationConfig {
            build_timeout_sec: 600, // 10 minutes
            test_timeout_sec: 3600, // 1 hour
            ..Default::default()
        };

        assert_eq!(
            config.timeout_for_kind(Some(crate::CompilationKind::CargoBuild)),
            std::time::Duration::from_secs(600)
        );
        assert_eq!(
            config.timeout_for_kind(Some(crate::CompilationKind::CargoTest)),
            std::time::Duration::from_secs(3600)
        );
    }

    #[test]
    fn test_compilation_config_speedup_threshold_default() {
        let _guard = test_guard!();
        let config = CompilationConfig::default();
        // Default speedup threshold: 1.2 (20% faster required)
        assert!((config.remote_speedup_threshold - 1.2).abs() < 0.001);
    }

    #[test]
    fn test_compilation_config_speedup_threshold_custom() {
        let _guard = test_guard!();
        let config = CompilationConfig {
            remote_speedup_threshold: 1.5, // Require 50% faster
            ..Default::default()
        };
        assert!((config.remote_speedup_threshold - 1.5).abs() < 0.001);
    }

    #[test]
    fn test_compilation_config_speedup_threshold_no_minimum() {
        let _guard = test_guard!();
        // Setting to 1.0 means "always offload when other criteria met"
        let config = CompilationConfig {
            remote_speedup_threshold: 1.0,
            ..Default::default()
        };
        assert!((config.remote_speedup_threshold - 1.0).abs() < 0.001);
    }

    #[test]
    fn test_compilation_config_min_local_time_ms_default() {
        let _guard = test_guard!();
        let config = CompilationConfig::default();
        // Default min_local_time_ms: 2000ms
        assert_eq!(config.min_local_time_ms, 2000);
    }

    #[test]
    fn test_compilation_config_serde_roundtrip_speedup_threshold() {
        let _guard = test_guard!();
        let config = CompilationConfig {
            remote_speedup_threshold: 2.5,
            min_local_time_ms: 5000,
            ..Default::default()
        };
        let json = serde_json::to_string(&config).unwrap();
        let parsed: CompilationConfig = serde_json::from_str(&json).unwrap();
        assert!((parsed.remote_speedup_threshold - 2.5).abs() < 0.001);
        assert_eq!(parsed.min_local_time_ms, 5000);
    }

    // ========================================================================
    // validate_remote_base Tests
    // ========================================================================

    #[test]
    fn test_validate_remote_base_absolute_path() {
        let _guard = test_guard!();
        assert_eq!(validate_remote_base("/tmp/rch").unwrap(), "/tmp/rch");
        assert_eq!(
            validate_remote_base("/var/rch-builds").unwrap(),
            "/var/rch-builds"
        );
        assert_eq!(
            validate_remote_base("/home/builder/.rch").unwrap(),
            "/home/builder/.rch"
        );
    }

    #[test]
    fn test_validate_remote_base_tilde_expansion() {
        let _guard = test_guard!();
        // Tilde expansion should work
        let result = validate_remote_base("~/rch");
        assert!(result.is_ok());
        let path = result.unwrap();
        assert!(
            path.starts_with('/'),
            "Path should be absolute after expansion: {}",
            path
        );
        assert!(!path.contains('~'), "Tilde should be expanded: {}", path);
    }

    #[test]
    fn test_validate_remote_base_rejects_relative_path() {
        let _guard = test_guard!();
        let result = validate_remote_base("tmp/rch");
        assert!(result.is_err());
        assert!(result.unwrap_err().contains("absolute path"));
    }

    #[test]
    fn test_validate_remote_base_rejects_path_traversal() {
        let _guard = test_guard!();
        let result = validate_remote_base("/tmp/../etc/rch");
        assert!(result.is_err());
        assert!(result.unwrap_err().contains("path traversal"));

        let result = validate_remote_base("/tmp/rch/../other");
        assert!(result.is_err());
        assert!(result.unwrap_err().contains("path traversal"));
    }

    #[test]
    fn test_validate_remote_base_normalizes_trailing_slash() {
        let _guard = test_guard!();
        assert_eq!(validate_remote_base("/tmp/rch/").unwrap(), "/tmp/rch");
        assert_eq!(validate_remote_base("/tmp/rch///").unwrap(), "/tmp/rch");
    }

    #[test]
    fn test_validate_remote_base_root_path() {
        let _guard = test_guard!();
        // Root path should be rejected
        let result = validate_remote_base("/");
        assert!(result.is_err());
        assert!(result.unwrap_err().contains("root directory"));
    }

    #[test]
    fn test_validate_remote_base_rejects_top_level() {
        let _guard = test_guard!();
        // Top-level directories should be rejected (depth 1)
        let result = validate_remote_base("/tmp");
        assert!(result.is_err());
        assert!(result.unwrap_err().contains("at least 2 levels deep"));

        let result = validate_remote_base("/home");
        assert!(result.is_err());

        // Depth 2 should be allowed
        assert_eq!(validate_remote_base("/tmp/rch").unwrap(), "/tmp/rch");
    }

    #[test]
    fn test_default_remote_base() {
        let _guard = test_guard!();
        assert_eq!(default_remote_base(), "/data/tmp/rch");
    }

    #[test]
    fn test_transfer_config_default_has_remote_base() {
        let _guard = test_guard!();
        let config = TransferConfig::default();
        assert_eq!(config.remote_base, "/data/tmp/rch");
    }

    #[test]
    fn test_source_sync_timeout_validation_boundaries() {
        let _guard = test_guard!();
        assert!(!TransferConfig::valid_sync_timeout_ms(0));
        assert!(!TransferConfig::valid_sync_timeout_ms(999));
        assert!(TransferConfig::valid_sync_timeout_ms(1_000));
        assert!(TransferConfig::valid_sync_timeout_ms(3_600_000));
        assert!(!TransferConfig::valid_sync_timeout_ms(3_600_001));
    }

    #[test]
    fn test_source_sync_timeout_is_explicit_or_payload_aware() {
        let _guard = test_guard!();
        let default = TransferConfig::default();
        assert_eq!(
            default.sync_timeout_for_payload(0),
            std::time::Duration::from_millis(30_000)
        );
        assert_eq!(
            default.sync_timeout_for_payload(10 * 1024 * 1024),
            std::time::Duration::from_millis(40_000)
        );
        assert_eq!(
            default.sync_timeout_for_payload(u64::MAX),
            std::time::Duration::from_millis(3_600_000)
        );

        let explicit = TransferConfig {
            sync_timeout_ms: Some(120_000),
            ..TransferConfig::default()
        };
        assert_eq!(
            explicit.sync_timeout_for_payload(u64::MAX),
            std::time::Duration::from_millis(120_000)
        );
    }

    // ========================================================================
    // RetryConfig Tests (bd-x1ek)
    // ========================================================================

    #[test]
    fn test_retry_config_default() {
        let _guard = test_guard!();
        let config = RetryConfig::default();
        assert_eq!(config.max_attempts, 3);
        assert_eq!(config.base_delay_ms, 100);
        assert_eq!(config.max_delay_ms, 5000);
        assert_eq!(config.jitter_factor, 0.1);
        assert_eq!(config.total_timeout_ms, 30000);
    }

    #[test]
    fn test_retry_config_no_retry() {
        let _guard = test_guard!();
        let config = RetryConfig::no_retry();
        assert_eq!(config.max_attempts, 1);
    }

    #[test]
    fn test_retry_config_delay_for_attempt_zero() {
        let _guard = test_guard!();
        let config = RetryConfig::default();
        assert_eq!(config.delay_for_attempt(0), std::time::Duration::ZERO);
    }

    #[test]
    fn test_retry_config_delay_for_attempt_exponential() {
        let _guard = test_guard!();
        let config = RetryConfig {
            base_delay_ms: 100,
            max_delay_ms: 10000,
            jitter_factor: 0.0, // No jitter for deterministic test
            ..Default::default()
        };

        // attempt 1: 100 * 2^0 = 100ms
        assert_eq!(
            config.delay_for_attempt(1),
            std::time::Duration::from_millis(100)
        );
        // attempt 2: 100 * 2^1 = 200ms
        assert_eq!(
            config.delay_for_attempt(2),
            std::time::Duration::from_millis(200)
        );
        // attempt 3: 100 * 2^2 = 400ms
        assert_eq!(
            config.delay_for_attempt(3),
            std::time::Duration::from_millis(400)
        );
    }

    #[test]
    fn test_retry_config_delay_capped_at_max() {
        let _guard = test_guard!();
        let config = RetryConfig {
            base_delay_ms: 1000,
            max_delay_ms: 2000,
            jitter_factor: 0.0,
            ..Default::default()
        };

        // attempt 2: 1000 * 2^1 = 2000ms (at cap)
        assert_eq!(
            config.delay_for_attempt(2),
            std::time::Duration::from_millis(2000)
        );
        // attempt 3: 1000 * 2^2 = 4000ms, capped to 2000ms
        assert_eq!(
            config.delay_for_attempt(3),
            std::time::Duration::from_millis(2000)
        );
    }

    #[test]
    fn test_retry_config_random_jitter_stays_within_bounds() {
        let _guard = test_guard!();
        let config = RetryConfig {
            base_delay_ms: 100,
            max_delay_ms: 200,
            jitter_factor: 0.25,
            ..Default::default()
        };

        // Jitter applies after the exponential delay is capped. Exercise the
        // real RNG without requiring any particular random value or sequence.
        for (attempt, minimum_ms, maximum_ms) in [(1, 75, 125), (3, 150, 250)] {
            for _ in 0..128 {
                let delay = config.delay_for_attempt(attempt);
                assert!(delay >= std::time::Duration::from_millis(minimum_ms));
                assert!(delay <= std::time::Duration::from_millis(maximum_ms));
            }
        }
    }

    #[test]
    fn test_retry_config_should_retry() {
        let _guard = test_guard!();
        let config = RetryConfig {
            max_attempts: 3,
            total_timeout_ms: 1000,
            ..Default::default()
        };

        // Before max attempts and within timeout
        assert!(config.should_retry(0, std::time::Duration::from_millis(0)));
        assert!(config.should_retry(1, std::time::Duration::from_millis(500)));
        assert!(config.should_retry(2, std::time::Duration::from_millis(900)));

        // At max attempts
        assert!(!config.should_retry(3, std::time::Duration::from_millis(0)));

        // Past timeout
        assert!(!config.should_retry(1, std::time::Duration::from_millis(1001)));
    }

    #[test]
    fn test_retry_config_serde_roundtrip() {
        let _guard = test_guard!();
        let config = RetryConfig {
            max_attempts: 5,
            base_delay_ms: 200,
            max_delay_ms: 8000,
            jitter_factor: 0.2,
            total_timeout_ms: 60000,
        };
        let json = serde_json::to_string(&config).unwrap();
        let parsed: RetryConfig = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.max_attempts, 5);
        assert_eq!(parsed.base_delay_ms, 200);
        assert_eq!(parsed.max_delay_ms, 8000);
        assert_eq!(parsed.jitter_factor, 0.2);
        assert_eq!(parsed.total_timeout_ms, 60000);
    }

    #[test]
    fn test_transfer_config_includes_retry() {
        let _guard = test_guard!();
        let config = TransferConfig::default();
        assert_eq!(config.retry.max_attempts, 3);
    }

    // =========================================================================
    // ExecutionConfig Tests (bd-785w)
    // =========================================================================

    #[test]
    fn test_execution_config_default_allowlist() {
        let _guard = test_guard!();
        let config = ExecutionConfig::default();
        // All supported compilers should be in the default allowlist
        assert!(config.is_allowed("cargo"));
        assert!(config.is_allowed("rustc"));
        assert!(config.is_allowed("gcc"));
        assert!(config.is_allowed("g++"));
        assert!(config.is_allowed("clang"));
        assert!(config.is_allowed("clang++"));
        assert!(config.is_allowed("make"));
        assert!(config.is_allowed("cmake"));
        assert!(config.is_allowed("ninja"));
        assert!(config.is_allowed("meson"));
        assert!(config.is_allowed("bun"));
        assert!(config.is_allowed("nextest"));
        assert!(config.is_allowed("cc"));
        assert!(config.is_allowed("c++"));
    }

    #[test]
    fn test_execution_config_is_allowed_case_insensitive() {
        let _guard = test_guard!();
        let config = ExecutionConfig::default();
        // Should be case-insensitive
        assert!(config.is_allowed("CARGO"));
        assert!(config.is_allowed("Cargo"));
        assert!(config.is_allowed("GCC"));
        assert!(config.is_allowed("Gcc"));
    }

    #[test]
    fn test_execution_config_is_not_allowed() {
        let _guard = test_guard!();
        let config = ExecutionConfig::default();
        // Unknown commands should not be allowed
        assert!(!config.is_allowed("python"));
        assert!(!config.is_allowed("npm"));
        assert!(!config.is_allowed(""));
        // `go` and `tsc` ARE allowed now that Go/TypeScript are offload targets.
        // If they were missing from the allowlist the hook would fail open to
        // LOCAL execution — classification would look correct while every build
        // silently ran on the orchestrator.
        assert!(config.is_allowed("go"));
        assert!(config.is_allowed("tsc"));
    }

    #[test]
    fn test_execution_config_empty_allowlist() {
        let _guard = test_guard!();
        let config = ExecutionConfig { allowlist: vec![] };
        // Empty allowlist should block everything
        assert!(!config.is_allowed("cargo"));
        assert!(!config.is_allowed("gcc"));
    }

    #[test]
    fn test_execution_config_custom_allowlist() {
        let _guard = test_guard!();
        let config = ExecutionConfig {
            allowlist: vec!["cargo".to_string(), "custom_tool".to_string()],
        };
        assert!(config.is_allowed("cargo"));
        assert!(config.is_allowed("custom_tool"));
        assert!(!config.is_allowed("gcc"));
    }

    #[test]
    fn test_execution_config_serde() {
        let _guard = test_guard!();
        let config = ExecutionConfig {
            allowlist: vec!["cargo".to_string(), "rustc".to_string()],
        };
        let json = serde_json::to_string(&config).unwrap();
        let parsed: ExecutionConfig = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.allowlist, config.allowlist);
    }

    #[test]
    fn test_rch_config_includes_execution() {
        let _guard = test_guard!();
        let config = RchConfig::default();
        // RchConfig should include ExecutionConfig with default allowlist
        assert!(config.execution.is_allowed("cargo"));
    }

    // ========================================================================
    // Adaptive Compression Tests (bd-243w)
    // ========================================================================

    #[test]
    fn test_adaptive_compression_disabled_by_default() {
        let _guard = test_guard!();
        let config = TransferConfig::default();
        assert!(!config.adaptive_compression);
        // Should use fixed level when disabled
        assert_eq!(config.select_compression_level(Some(1_000_000)), 3);
        assert_eq!(config.select_compression_level(Some(500_000_000)), 3);
    }

    #[test]
    fn test_adaptive_compression_small_payload() {
        let _guard = test_guard!();
        let config = TransferConfig {
            adaptive_compression: true,
            ..Default::default()
        };
        // < 10MB should use level 1
        assert_eq!(config.select_compression_level(Some(0)), 1);
        assert_eq!(config.select_compression_level(Some(1_000_000)), 1); // 1MB
        assert_eq!(config.select_compression_level(Some(9_999_999)), 1); // ~10MB
    }

    #[test]
    fn test_adaptive_compression_medium_payload() {
        let _guard = test_guard!();
        let config = TransferConfig {
            adaptive_compression: true,
            ..Default::default()
        };
        // 10-200MB should use level 3
        assert_eq!(config.select_compression_level(Some(10_000_000)), 3); // 10MB
        assert_eq!(config.select_compression_level(Some(100_000_000)), 3); // 100MB
        assert_eq!(config.select_compression_level(Some(199_999_999)), 3); // ~200MB
    }

    #[test]
    fn test_adaptive_compression_large_payload() {
        let _guard = test_guard!();
        let config = TransferConfig {
            adaptive_compression: true,
            ..Default::default()
        };
        // > 200MB should use level 7
        assert_eq!(config.select_compression_level(Some(200_000_000)), 7); // 200MB
        assert_eq!(config.select_compression_level(Some(500_000_000)), 7); // 500MB
        assert_eq!(config.select_compression_level(Some(1_000_000_000)), 7); // 1GB
    }

    #[test]
    fn test_adaptive_compression_no_estimate() {
        let _guard = test_guard!();
        let config = TransferConfig {
            adaptive_compression: true,
            compression_level: 5,
            ..Default::default()
        };
        // No estimate should fallback to default level
        assert_eq!(config.select_compression_level(None), 5);
    }

    #[test]
    fn test_adaptive_compression_respects_min_level() {
        let _guard = test_guard!();
        let config = TransferConfig {
            adaptive_compression: true,
            min_compression_level: 3,
            ..Default::default()
        };
        // Small payload would want level 1, but min is 3
        assert_eq!(config.select_compression_level(Some(1_000_000)), 3);
    }

    #[test]
    fn test_adaptive_compression_respects_max_level() {
        let _guard = test_guard!();
        let config = TransferConfig {
            adaptive_compression: true,
            max_compression_level: 5,
            ..Default::default()
        };
        // Large payload would want level 7, but max is 5
        assert_eq!(config.select_compression_level(Some(500_000_000)), 5);
    }

    #[test]
    fn test_adaptive_compression_serde_roundtrip() {
        let _guard = test_guard!();
        let config = TransferConfig {
            adaptive_compression: true,
            min_compression_level: 2,
            max_compression_level: 8,
            ..Default::default()
        };
        let json = serde_json::to_string(&config).unwrap();
        let parsed: TransferConfig = serde_json::from_str(&json).unwrap();
        assert!(parsed.adaptive_compression);
        assert_eq!(parsed.min_compression_level, 2);
        assert_eq!(parsed.max_compression_level, 8);
    }

    // -------------------------------------------------------------------------
    // Self-Healing Config Tests (bd-59kg)
    // -------------------------------------------------------------------------

    #[test]
    fn test_self_healing_config_defaults() {
        let _guard = test_guard!();
        // TEST START: SelfHealingConfig::default() returns expected values
        let config = SelfHealingConfig::default();

        // Both self-healing behaviors enabled by default
        assert!(
            config.hook_starts_daemon,
            "hook_starts_daemon should be true by default"
        );
        assert!(
            config.daemon_installs_hooks,
            "daemon_installs_hooks should be true by default"
        );

        // Timing defaults
        assert_eq!(
            config.auto_start_cooldown_secs, 30,
            "auto_start_cooldown_secs should default to 30"
        );
        assert_eq!(
            config.auto_start_timeout_secs, 3,
            "auto_start_timeout_secs should default to 3"
        );
        // TEST PASS: SelfHealingConfig defaults
    }

    #[test]
    fn test_self_healing_config_serde_full() {
        let _guard = test_guard!();
        // TEST START: Full SelfHealingConfig serialization/deserialization
        let config = SelfHealingConfig {
            hook_starts_daemon: false,
            daemon_installs_hooks: false,
            auto_start_cooldown_secs: 60,
            auto_start_timeout_secs: 10,
            self_healing_log_level: SelfHealingLogLevel::Debug,
        };

        let json = serde_json::to_string(&config).unwrap();
        assert!(json.contains("\"hook_starts_daemon\":false"));
        assert!(json.contains("\"daemon_installs_hooks\":false"));
        assert!(json.contains("\"auto_start_cooldown_secs\":60"));
        assert!(json.contains("\"auto_start_timeout_secs\":10"));
        assert!(json.contains("\"self_healing_log_level\":\"debug\""));

        let parsed: SelfHealingConfig = serde_json::from_str(&json).unwrap();
        assert!(!parsed.hook_starts_daemon);
        assert!(!parsed.daemon_installs_hooks);
        assert_eq!(parsed.auto_start_cooldown_secs, 60);
        assert_eq!(parsed.auto_start_timeout_secs, 10);
        assert_eq!(parsed.self_healing_log_level, SelfHealingLogLevel::Debug);
        // TEST PASS: Full SelfHealingConfig serde
    }

    // br-4zf3p: Tests for the self_healing_log_level field.

    #[test]
    fn test_self_healing_log_level_default_is_info() {
        // TEST START: SelfHealingLogLevel::default() == Info
        let level = SelfHealingLogLevel::default();
        assert_eq!(level, SelfHealingLogLevel::Info);
        // Also verify the SelfHealingConfig default surfaces it.
        let config = SelfHealingConfig::default();
        assert_eq!(
            config.self_healing_log_level,
            SelfHealingLogLevel::Info,
            "config default should be Info"
        );
        // TEST PASS
    }

    #[test]
    fn test_self_healing_log_level_serializes_lowercase() {
        // TEST START: serde uses lowercase variants
        let json = serde_json::to_string(&SelfHealingLogLevel::Warn).unwrap();
        assert_eq!(json, "\"warn\"");
        let parsed: SelfHealingLogLevel = serde_json::from_str("\"error\"").unwrap();
        assert_eq!(parsed, SelfHealingLogLevel::Error);
        // TEST PASS
    }

    #[test]
    fn test_self_healing_log_level_from_env_str_accepts_known() {
        // TEST START: from_env_str parses each level (case-insensitive) and aliases
        assert_eq!(
            SelfHealingLogLevel::from_env_str("debug"),
            Some(SelfHealingLogLevel::Debug)
        );
        assert_eq!(
            SelfHealingLogLevel::from_env_str("DEBUG"),
            Some(SelfHealingLogLevel::Debug)
        );
        assert_eq!(
            SelfHealingLogLevel::from_env_str("info"),
            Some(SelfHealingLogLevel::Info)
        );
        assert_eq!(
            SelfHealingLogLevel::from_env_str("warn"),
            Some(SelfHealingLogLevel::Warn)
        );
        assert_eq!(
            SelfHealingLogLevel::from_env_str("warning"),
            Some(SelfHealingLogLevel::Warn),
            "warning should map to Warn"
        );
        assert_eq!(
            SelfHealingLogLevel::from_env_str("error"),
            Some(SelfHealingLogLevel::Error)
        );
        // Whitespace tolerance
        assert_eq!(
            SelfHealingLogLevel::from_env_str("  debug  "),
            Some(SelfHealingLogLevel::Debug)
        );
        // TEST PASS
    }

    #[test]
    fn test_self_healing_log_level_from_env_str_rejects_unknown() {
        // TEST START: unknown values return None (caller falls back to default)
        assert!(SelfHealingLogLevel::from_env_str("banana").is_none());
        assert!(SelfHealingLogLevel::from_env_str("").is_none());
        assert!(SelfHealingLogLevel::from_env_str("trace").is_none());
        // TEST PASS
    }

    #[test]
    fn test_self_healing_config_serde_partial_uses_defaults() {
        let _guard = test_guard!();
        // TEST START: Partial TOML/JSON uses defaults for missing fields
        let json = r#"{"hook_starts_daemon": false}"#;
        let config: SelfHealingConfig = serde_json::from_str(json).unwrap();

        assert!(
            !config.hook_starts_daemon,
            "Explicit false should be parsed"
        );
        assert!(
            config.daemon_installs_hooks,
            "Missing field should use default (true)"
        );
        assert_eq!(
            config.auto_start_cooldown_secs, 30,
            "Missing field should use default (30)"
        );
        assert_eq!(
            config.auto_start_timeout_secs, 3,
            "Missing field should use default (3)"
        );
        // TEST PASS: Partial SelfHealingConfig uses defaults
    }

    #[test]
    fn test_self_healing_config_toml_with_alias() {
        let _guard = test_guard!();
        // TEST START: daemon_start_timeout alias works for auto_start_timeout_secs
        let toml_str = r#"
            hook_starts_daemon = true
            daemon_installs_hooks = false
            auto_start_cooldown_secs = 45
            daemon_start_timeout = 7
        "#;

        let config: SelfHealingConfig = toml::from_str(toml_str).unwrap();
        assert!(config.hook_starts_daemon);
        assert!(!config.daemon_installs_hooks);
        assert_eq!(config.auto_start_cooldown_secs, 45);
        assert_eq!(
            config.auto_start_timeout_secs, 7,
            "daemon_start_timeout alias should set auto_start_timeout_secs"
        );
        // TEST PASS: TOML alias works
    }

    #[allow(unsafe_code)]
    mod self_healing_env_override_tests {
        use super::*;
        use crate::config::env_test_lock;

        fn env_guard() -> std::sync::MutexGuard<'static, ()> {
            env_test_lock()
        }

        fn set_env(key: &str, value: &str) {
            // SAFETY: Tests are serialized with env_guard().
            unsafe { std::env::set_var(key, value) };
        }

        fn remove_env(key: &str) {
            // SAFETY: Tests are serialized with env_guard().
            unsafe { std::env::remove_var(key) };
        }

        struct EnvVarGuard {
            key: &'static str,
            old: Option<String>,
        }

        impl EnvVarGuard {
            fn set(key: &'static str, value: &str) -> Self {
                let old = std::env::var(key).ok();
                set_env(key, value);
                Self { key, old }
            }
        }

        impl Drop for EnvVarGuard {
            fn drop(&mut self) {
                if let Some(old) = &self.old {
                    set_env(self.key, old);
                } else {
                    remove_env(self.key);
                }
            }
        }

        #[test]
        fn test_self_healing_config_with_env_overrides_master_disable() {
            let _guard = test_guard!();
            let _guard = env_guard();

            let _no_self_healing = EnvVarGuard::set("RCH_NO_SELF_HEALING", "1");
            let _hook_starts_daemon = EnvVarGuard::set("RCH_HOOK_STARTS_DAEMON", "1");
            let _daemon_installs_hooks = EnvVarGuard::set("RCH_DAEMON_INSTALLS_HOOKS", "1");
            let _cooldown = EnvVarGuard::set("RCH_AUTO_START_COOLDOWN_SECS", "99");
            let _timeout = EnvVarGuard::set("RCH_AUTO_START_TIMEOUT_SECS", "99");

            let config = SelfHealingConfig::default().with_env_overrides();
            assert!(
                !config.hook_starts_daemon,
                "RCH_NO_SELF_HEALING should disable hook auto-start"
            );
            assert!(
                !config.daemon_installs_hooks,
                "RCH_NO_SELF_HEALING should disable daemon hook installation"
            );

            // Master disable returns early; numeric overrides should not apply.
            assert_eq!(config.auto_start_cooldown_secs, 30);
            assert_eq!(config.auto_start_timeout_secs, 3);
        }

        #[test]
        fn test_self_healing_config_with_env_overrides_toggles_and_numbers() {
            let _guard = test_guard!();
            let _guard = env_guard();

            let _hook_starts_daemon = EnvVarGuard::set("RCH_HOOK_STARTS_DAEMON", "0");
            let _daemon_installs_hooks = EnvVarGuard::set("RCH_DAEMON_INSTALLS_HOOKS", "false");
            let _cooldown = EnvVarGuard::set("RCH_AUTO_START_COOLDOWN_SECS", "45");
            let _timeout = EnvVarGuard::set("RCH_AUTO_START_TIMEOUT_SECS", "7");

            let config = SelfHealingConfig::default().with_env_overrides();
            assert!(!config.hook_starts_daemon);
            assert!(!config.daemon_installs_hooks);
            assert_eq!(config.auto_start_cooldown_secs, 45);
            assert_eq!(config.auto_start_timeout_secs, 7);
        }

        #[test]
        fn test_self_healing_config_with_env_overrides_invalid_numbers_ignored() {
            let _guard = test_guard!();
            let _guard = env_guard();

            let _cooldown = EnvVarGuard::set("RCH_AUTO_START_COOLDOWN_SECS", "not-a-number");
            let _timeout = EnvVarGuard::set("RCH_AUTO_START_TIMEOUT_SECS", "nope");

            let config = SelfHealingConfig::default().with_env_overrides();
            assert_eq!(config.auto_start_cooldown_secs, 30);
            assert_eq!(config.auto_start_timeout_secs, 3);
        }
    }

    // -------------------------------------------------------------------------
    // Preflight Health Guard Tests (bd-3eaa)
    // -------------------------------------------------------------------------

    #[test]
    fn test_load_per_core_calculation() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        // No metrics -> None
        assert!(caps.load_per_core().is_none());

        // Only load, no cpus -> None
        caps.load_avg_1 = Some(4.0);
        assert!(caps.load_per_core().is_none());

        // Both available
        caps.num_cpus = Some(4);
        assert_eq!(caps.load_per_core(), Some(1.0)); // 4.0 / 4 = 1.0

        // High load
        caps.load_avg_1 = Some(16.0);
        assert_eq!(caps.load_per_core(), Some(4.0)); // 16.0 / 4 = 4.0
    }

    #[test]
    fn test_is_high_load() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        caps.load_avg_1 = Some(8.0);
        caps.num_cpus = Some(4);
        // load_per_core = 2.0

        // Below threshold
        assert_eq!(caps.is_high_load(3.0), Some(false));
        // At threshold
        assert_eq!(caps.is_high_load(2.0), Some(false)); // equal is OK
        // Above threshold
        assert_eq!(caps.is_high_load(1.5), Some(true));

        // No metrics -> None (fail-open)
        caps.load_avg_1 = None;
        assert!(caps.is_high_load(1.0).is_none());
    }

    #[test]
    fn test_is_low_disk() {
        let _guard = test_guard!();
        let mut caps = WorkerCapabilities::new();
        caps.disk_free_gb = Some(15.0);

        // Above threshold
        assert_eq!(caps.is_low_disk(10.0), Some(false));
        // At threshold
        assert_eq!(caps.is_low_disk(15.0), Some(false)); // equal is OK
        // Below threshold
        assert_eq!(caps.is_low_disk(20.0), Some(true));

        // No metrics -> None (fail-open)
        caps.disk_free_gb = None;
        assert!(caps.is_low_disk(10.0).is_none());
    }

    #[test]
    fn low_disk_judges_the_build_disk_not_a_small_tmpfs() {
        let _guard = test_guard!();
        // GH #78: a 16 GB tmpfs /tmp is the tightest mount by ratio, but the
        // build trees live on a roomy data disk.
        let mut caps = WorkerCapabilities::new();
        caps.disk_free_gb = Some(8.0);
        caps.disk_total_gb = Some(15.5);
        caps.build_disk_free_gb = Some(1700.0);
        caps.build_disk_total_gb = Some(1900.0);
        assert_eq!(caps.build_disk_gb(), (Some(1700.0), Some(1900.0)));
        assert_eq!(caps.is_low_disk(10.0), Some(false));

        // A full build disk is low even when /tmp is roomy.
        caps.build_disk_free_gb = Some(4.0);
        assert_eq!(caps.is_low_disk(10.0), Some(true));

        // An older rch-wkr without the build sample keeps the old behaviour.
        caps.build_disk_free_gb = None;
        caps.build_disk_total_gb = None;
        assert_eq!(caps.build_disk_gb(), (Some(8.0), Some(15.5)));
        assert_eq!(caps.is_low_disk(10.0), Some(true));
    }

    #[test]
    fn build_disk_sample_is_optional_on_the_wire() {
        let _guard = test_guard!();
        let old: WorkerCapabilities =
            serde_json::from_str(r#"{"disk_free_gb":13.9,"disk_total_gb":15.5}"#).unwrap();
        assert_eq!(old.build_disk_free_gb, None);
        assert_eq!(old.build_disk_gb(), (Some(13.9), Some(15.5)));

        let new: WorkerCapabilities = serde_json::from_str(
            r#"{"disk_free_gb":13.9,"disk_total_gb":15.5,"build_disk_free_gb":1700.0,"build_disk_total_gb":1900.0}"#,
        )
        .unwrap();
        assert_eq!(new.build_disk_gb(), (Some(1700.0), Some(1900.0)));
        let json = serde_json::to_string(&new).unwrap();
        assert!(json.contains("\"build_disk_free_gb\":1700.0"), "{json}");
    }

    #[test]
    fn test_selection_config_preflight_defaults() {
        let _guard = test_guard!();
        let config = SelectionConfig::default();
        // Default thresholds
        assert_eq!(config.max_load_per_core, Some(2.0));
        assert_eq!(config.min_free_gb, Some(10.0));
        assert_eq!(config.disk_gb_per_slot, 10.0);
    }

    #[test]
    fn test_selection_config_preflight_serde() {
        let _guard = test_guard!();
        let config = SelectionConfig {
            max_load_per_core: Some(3.5),
            min_free_gb: Some(25.0),
            disk_gb_per_slot: 0.83,
            ..Default::default()
        };
        let json = serde_json::to_string(&config).unwrap();
        let parsed: SelectionConfig = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.max_load_per_core, Some(3.5));
        assert_eq!(parsed.min_free_gb, Some(25.0));
        assert_eq!(parsed.disk_gb_per_slot, 0.83);
    }

    #[test]
    fn test_selection_disk_slot_budget_validation() {
        let _guard = test_guard!();
        for value in ["0.0", "-1.0", "nan", "inf", "-inf"] {
            let result = toml::from_str::<SelectionConfig>(&format!("disk_gb_per_slot = {value}"));
            assert!(result.is_err(), "invalid disk slot budget {value}");
        }
        let default: SelectionConfig = toml::from_str("").unwrap();
        assert_eq!(default.disk_gb_per_slot, 10.0);
    }

    #[test]
    fn test_selection_disk_weight_validation_and_roundtrip() {
        for value in ["-0.1", "1.1", "nan", "inf", "-inf"] {
            assert!(toml::from_str::<SelectionWeightConfig>(&format!("disk = {value}")).is_err());
        }
        for value in [0.0, 0.2, 1.0] {
            let weights: SelectionWeightConfig =
                toml::from_str(&format!("disk = {value}")).unwrap();
            assert_eq!(weights.disk, value);
            let roundtrip: SelectionWeightConfig =
                serde_json::from_str(&serde_json::to_string(&weights).unwrap()).unwrap();
            assert_eq!(roundtrip.disk, value);
        }
        let defaults: SelectionWeightConfig = toml::from_str("").unwrap();
        assert_eq!(defaults.disk, 0.2);
    }

    #[test]
    fn test_worker_capabilities_health_metrics_serde() {
        let _guard = test_guard!();
        let caps = WorkerCapabilities {
            num_cpus: Some(8),
            load_avg_1: Some(1.5),
            load_avg_5: Some(2.0),
            load_avg_15: Some(1.8),
            disk_free_gb: Some(50.5),
            disk_total_gb: Some(100.0),
            ..Default::default()
        };
        let json = serde_json::to_string(&caps).unwrap();
        let parsed: WorkerCapabilities = serde_json::from_str(&json).unwrap();
        assert_eq!(parsed.num_cpus, Some(8));
        assert_eq!(parsed.load_avg_1, Some(1.5));
        assert_eq!(parsed.disk_free_gb, Some(50.5));
    }

    #[test]
    fn microarch_level_classifies_ivy_bridge_as_v2() {
        // Abbreviated real flag set from ovh-b (Xeon E3-1245 V2, Ivy Bridge):
        // AVX present, but no avx2/bmi/fma — the bd-68hon SIGILL class.
        let ivy = "fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov \
                   pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx rdtscp lm \
                   constant_tsc pni pclmulqdq ssse3 cx16 sse4_1 sse4_2 x2apic \
                   popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm";
        assert_eq!(x86_64_microarch_level_from_flags(ivy), 2);
    }

    #[test]
    fn microarch_level_classifies_v3_and_v4() {
        let v3 = "cx16 lahf_lm popcnt sse4_1 sse4_2 ssse3 abm avx avx2 bmi1 bmi2 \
                  f16c fma movbe xsave";
        assert_eq!(x86_64_microarch_level_from_flags(v3), 3);

        let v4 = format!("{v3} avx512f avx512bw avx512cd avx512dq avx512vl");
        assert_eq!(x86_64_microarch_level_from_flags(&v4), 4);
    }

    #[test]
    fn microarch_level_v1_for_ancient_or_empty_flags() {
        assert_eq!(x86_64_microarch_level_from_flags(""), 1);
        assert_eq!(x86_64_microarch_level_from_flags("fpu mmx sse sse2"), 1);
    }

    #[test]
    fn pre_v3_predicate_only_fires_on_affirmative_evidence() {
        let mut caps = WorkerCapabilities::new();
        assert!(!caps.is_pre_v3_x86(), "unknown must not deprioritize");
        caps.cpu_microarch_level = Some(2);
        assert!(caps.is_pre_v3_x86());
        caps.cpu_microarch_level = Some(3);
        assert!(!caps.is_pre_v3_x86());
    }
}