cpm-planner 0.2.2

//! CPM Algorithm Implementation
//!
//! Implements the Critical Path Method (CPM) algorithm:
//! 1. Forward pass: Calculate earliest start/finish times (ES/EF)
//! 2. Backward pass: Calculate latest start/finish times (LS/LF)
//! 3. Float calculation: slack = LS - ES
//! 4. Critical path: `critical_path` is one longest prerequisite chain in
//!    execution order; `critical_ids` lists every zero-float (critical) task.
//! 5. Batch identification: Group tasks by earliest start time
//! 6. Bottleneck analysis: Identify tasks that block the most work

use crate::task::{Bottleneck, CriticalPathResult, Task, TaskBatch};
use std::collections::{HashMap, HashSet, VecDeque};

/// Version of the CPM kernel's output semantics. Bump whenever a change
/// alters any field of [`CriticalPathResult`] for the same input, so stored
/// plans are recomputed on open (see `plan_store`). Bump also when
/// `EffortEstimator` defaults change, since stored plans are recomputed with
/// the current estimator.
///
/// - v2: synthetic `__start__`/`__finish__` endpoints
/// - v3: calendar `duration_hours` and per-edge `lag_hours`
pub const CPM_VERSION: i64 = 3;

const TIGHT_EPS: f32 = 1e-3;

/// CPM Algorithm calculator
pub struct CpmAlgorithm;

impl CpmAlgorithm {
    /// Calculate critical path and all related metrics
    ///
    /// # Arguments
    /// * `tasks` - Mutable slice of tasks to analyze
    ///
    /// # Returns
    /// `CriticalPathResult` with all calculated metrics
    pub fn calculate(tasks: &mut [Task]) -> CriticalPathResult {
        if tasks.is_empty() {
            return CriticalPathResult::default();
        }

        // Duplicate ids make the id-keyed passes ill-defined (and the
        // forward pass could requeue a shared id forever). Bail out with
        // every task unscheduled so callers reject the result.
        let mut ids: HashSet<&str> = HashSet::with_capacity(tasks.len());
        if !tasks.iter().all(|t| ids.insert(t.id.as_str())) {
            let mut unscheduled: Vec<String> = tasks.iter().map(|t| t.id.clone()).collect();
            unscheduled.sort_unstable();
            unscheduled.dedup();
            return CriticalPathResult {
                total_tasks: tasks.len(),
                tasks: tasks.to_vec(),
                unscheduled,
                ..CriticalPathResult::default()
            };
        }

        // Build dependency graphs
        let (successors, _predecessors) = Self::build_dependency_graphs(tasks);

        // Forward pass: calculate ES/EF. This is the authoritative cycle
        // detector: an id here is a task whose dependencies could not be
        // topologically ordered.
        let (mut unscheduled, order) = Self::forward_pass(tasks);

        // Backward pass: calculate LS/LF. It returns any task whose
        // latest-finish never relaxed off the sentinel. In a well-formed
        // (acyclic) graph that set is always empty, so the only way it is
        // non-empty is a cycle the forward pass already flagged.
        //
        // When the forward pass already identified the unschedulable set we
        // treat it as authoritative and do NOT widen it with the backward
        // signal: backward also flags acyclic nodes that merely sit upstream
        // of a cycle (their LF can't be finalised), and those were correctly
        // scheduled by the forward pass. We only fold the backward findings
        // in as a defence-in-depth net for the "forward saw nothing wrong but
        // a task still never relaxed" case, which should not occur.
        let backward_unscheduled = Self::backward_pass(tasks, &successors, &order);
        if unscheduled.is_empty() {
            unscheduled = backward_unscheduled;
        }
        unscheduled.sort_unstable();
        unscheduled.dedup();

        // Calculate float and identify critical path
        Self::calculate_float(tasks);

        // Identify parallel batches
        let batches = Self::identify_parallel_batches(tasks);

        // Identify bottlenecks
        let bottlenecks = Self::identify_bottlenecks(tasks);

        // Build result
        Self::build_result(tasks, batches, bottlenecks, unscheduled)
    }

    /// Build forward (successors) and reverse (predecessors) dependency graphs
    fn build_dependency_graphs(
        tasks: &[Task],
    ) -> (HashMap<String, Vec<String>>, HashMap<String, Vec<String>>) {
        let mut successors: HashMap<String, Vec<String>> = HashMap::new();
        let mut predecessors: HashMap<String, Vec<String>> = HashMap::new();

        // Initialize empty lists for all tasks
        for task in tasks {
            successors.entry(task.id.clone()).or_default();
            predecessors.entry(task.id.clone()).or_default();
        }

        // Build the graphs from dependencies
        for task in tasks {
            for dep in &task.dependencies {
                // task depends on dep, so:
                // dep -> task (dep is predecessor of task)
                // task is successor of dep
                successors
                    .entry(dep.clone())
                    .or_default()
                    .push(task.id.clone());
                predecessors
                    .entry(task.id.clone())
                    .or_default()
                    .push(dep.clone());
            }
        }

        (successors, predecessors)
    }

    /// Forward pass: Calculate earliest start (ES) and earliest finish (EF)
    ///
    /// ES = max(EF of each predecessor + edge lag), or 0 if no predecessors
    /// EF = ES + scheduled length (`duration_hours` when set, else effort)
    ///
    /// Returns the ids of any tasks that could not be scheduled because
    /// they (or their predecessors) sit on a dependency cycle (an empty list
    /// means the whole graph was schedulable), plus the task indices in the
    /// order they were scheduled. Ids are unique here (`calculate` bails out
    /// on duplicates first).
    fn forward_pass(tasks: &mut [Task]) -> (Vec<String>, Vec<usize>) {
        // Passes only ever raise ES, so clear stale values from earlier runs.
        for task in tasks.iter_mut() {
            task.earliest_start = 0.0;
            task.earliest_finish = 0.0;
        }
        let task_count = tasks.len();
        let ids: Vec<String> = tasks.iter().map(|t| t.id.clone()).collect();
        let task_map: HashMap<&str, usize> = ids
            .iter()
            .enumerate()
            .map(|(i, id)| (id.as_str(), i))
            .collect();

        // Dependents of each task, each listed once, in task order; and the
        // number of distinct dependencies (unknown ids included, so a task
        // depending on a missing id is never scheduled).
        let mut dependents: Vec<Vec<usize>> = vec![Vec::new(); task_count];
        let mut in_degree: Vec<usize> = Vec::with_capacity(task_count);
        for (i, task) in tasks.iter().enumerate() {
            let distinct: HashSet<&str> = task.dependencies.iter().map(String::as_str).collect();
            in_degree.push(distinct.len());
            for dep in distinct {
                if let Some(&d) = task_map.get(dep) {
                    dependents[d].push(i);
                }
            }
        }
        for list in &mut dependents {
            list.sort_unstable();
        }

        // Kahn's algorithm, starting with tasks that have no dependencies.
        let mut queue: VecDeque<usize> = (0..task_count)
            .filter(|&i| tasks[i].dependencies.is_empty())
            .collect();
        for &i in &queue {
            tasks[i].earliest_finish = tasks[i].effort_hours;
        }

        let mut order: Vec<usize> = Vec::with_capacity(task_count);
        while let Some(current) = queue.pop_front() {
            order.push(current);
            let current_ef = tasks[current].earliest_finish;
            for &i in &dependents[current] {
                let task = &mut tasks[i];
                // Update ES if this predecessor (plus edge lag) is later.
                let lag = task
                    .lag_by_dependency
                    .get(&ids[current])
                    .copied()
                    .unwrap_or(0.0);
                if current_ef + lag > task.earliest_start {
                    task.earliest_start = current_ef + lag;
                    task.earliest_finish = task.earliest_start + task.effort_hours;
                }
                in_degree[i] -= 1;
                if in_degree[i] == 0 {
                    // All predecessors processed: lock in the final EF.
                    task.earliest_finish = task.earliest_start + task.effort_hours;
                    queue.push_back(i);
                }
            }
        }

        // Tasks never drained from the queue sit on (or downstream of) a
        // cycle, or depend on an unknown id: report them so callers reject
        // the otherwise confidently-wrong result.
        if order.len() < task_count {
            let mut scheduled = vec![false; task_count];
            for &i in &order {
                scheduled[i] = true;
            }
            let mut unscheduled: Vec<String> = (0..task_count)
                .filter(|&i| !scheduled[i])
                .map(|i| tasks[i].id.clone())
                .collect();
            unscheduled.sort_unstable();
            tracing::warn!(
                unscheduled = unscheduled.len(),
                "CPM forward pass could not schedule all tasks (possible dependency cycle)"
            );
            return (unscheduled, order);
        }

        (Vec::new(), order)
    }

    /// Backward pass: Calculate latest start (LS) and latest finish (LF)
    ///
    /// LF = min(LS of each successor - edge lag), or `project_end` if no
    /// successors
    /// LS = LF - scheduled length
    ///
    /// Returns the ids of any tasks whose `latest_finish` never relaxed off
    /// the `f32::MAX` sentinel. In a well-formed graph this is empty; a
    /// non-empty return means those tasks are unschedulable (the same cycle
    /// signal the forward pass surfaces) and the result must not be trusted.
    ///
    /// Tasks are relaxed in reverse `order` (the forward pass's topological
    /// order), then any task it never scheduled, so an acyclic graph settles
    /// in one sweep plus a confirming one.
    fn backward_pass(
        tasks: &mut [Task],
        successors: &HashMap<String, Vec<String>>,
        order: &[usize],
    ) -> Vec<String> {
        // Find project duration (max EF)
        let project_duration = tasks
            .iter()
            .map(|t| t.earliest_finish)
            .fold(0.0_f32, f32::max);

        // Create a map for quick lookup
        let task_map: HashMap<String, usize> = tasks
            .iter()
            .enumerate()
            .map(|(i, t)| (t.id.clone(), i))
            .collect();

        // Initialize LF/LS for ending tasks (no successors)
        for task in tasks.iter_mut() {
            let has_successors = successors.get(&task.id).is_some_and(|s| !s.is_empty());
            if has_successors {
                // Initialize to max values for later minimization
                task.latest_finish = f32::MAX;
                task.latest_start = f32::MAX;
            } else {
                task.latest_finish = project_duration;
                task.latest_start = project_duration - task.effort_hours;
            }
        }

        // Reverse topological order processing
        // We iterate until no changes (simpler than proper reverse topo sort)
        let mut visit: Vec<usize> = order.iter().rev().copied().collect();
        if visit.len() < tasks.len() {
            let mut in_order = vec![false; tasks.len()];
            for &i in order {
                in_order[i] = true;
            }
            visit.extend((0..tasks.len()).filter(|&i| !in_order[i]));
        }
        let mut changed = true;
        let mut iterations = 0;
        let max_iterations = tasks.len() * 2;

        while changed && iterations < max_iterations {
            changed = false;
            iterations += 1;

            for &i in &visit {
                let task_id = tasks[i].id.clone();
                let task_effort = tasks[i].effort_hours;

                // Find minimum LS of all successors
                if let Some(succ_ids) = successors.get(&task_id)
                    && !succ_ids.is_empty()
                {
                    let min_succ_ls = succ_ids
                        .iter()
                        .filter_map(|sid| {
                            task_map.get(sid).map(|&idx| {
                                let lag = tasks[idx]
                                    .lag_by_dependency
                                    .get(&task_id)
                                    .copied()
                                    .unwrap_or(0.0);
                                if tasks[idx].latest_start < f32::MAX {
                                    tasks[idx].latest_start - lag
                                } else {
                                    f32::MAX
                                }
                            })
                        })
                        .filter(|&ls| ls < f32::MAX)
                        .fold(f32::MAX, f32::min);

                    if min_succ_ls < f32::MAX && min_succ_ls < tasks[i].latest_finish {
                        tasks[i].latest_finish = min_succ_ls;
                        tasks[i].latest_start = min_succ_ls - task_effort;
                        changed = true;
                    }
                }
            }
        }

        // Any task still pinned at the MAX sentinel never had its LF/LS
        // relaxed — in a well-formed DAG that cannot happen, so treat it as
        // the same unschedulable signal the forward pass raises rather than
        // silently clamping to project_duration and emitting a plausible-
        // looking (but wrong) plan. We still clamp the numeric fields so the
        // struct holds finite values, but the returned ids let callers
        // reject the result.
        let mut unscheduled: Vec<String> = Vec::new();
        for task in tasks.iter_mut() {
            if task.latest_finish >= f32::MAX - 1.0 {
                task.latest_finish = project_duration;
                task.latest_start = project_duration - task.effort_hours;
                unscheduled.push(task.id.clone());
            }
        }
        unscheduled.sort_unstable();
        unscheduled
    }

    /// Calculate float (slack) for each task and mark critical path
    fn calculate_float(tasks: &mut [Task]) {
        for task in tasks.iter_mut() {
            task.float = task.latest_start - task.earliest_start;
            // Critical if float is essentially zero (within tolerance)
            task.is_critical = task.float.abs() < 0.001;
        }
    }

    /// Identify parallel batches by grouping tasks with same earliest start
    fn identify_parallel_batches(tasks: &[Task]) -> Vec<TaskBatch> {
        // Group tasks by earliest start time (rounded to avoid float issues)
        let mut batches_map: HashMap<i32, Vec<&Task>> = HashMap::new();

        for task in tasks {
            // Round to nearest 0.1 hour for batching
            let es_key = (task.earliest_start * 10.0).round() as i32;
            batches_map.entry(es_key).or_default().push(task);
        }

        // Convert to sorted Vec<TaskBatch>
        let mut sorted_keys: Vec<i32> = batches_map.keys().copied().collect();
        sorted_keys.sort_unstable();

        let mut batches = Vec::new();
        let mut batch_num = 0;

        for es_key in sorted_keys {
            if let Some(batch_tasks) = batches_map.get(&es_key) {
                let total_effort: f32 = batch_tasks.iter().map(|t| t.effort_hours).sum();
                let duration = batch_tasks
                    .iter()
                    .map(|t| t.effort_hours)
                    .fold(0.0_f32, f32::max);
                let start_time = es_key as f32 / 10.0;

                batches.push(TaskBatch {
                    id: format!("Batch-{batch_num}"),
                    tasks: batch_tasks.iter().map(|t| t.id.clone()).collect(),
                    total_effort_hours: total_effort,
                    duration_hours: duration,
                    start_time,
                });
                batch_num += 1;
            }
        }

        batches
    }

    /// Identify bottleneck tasks based on transitive impact. Each task's
    /// transitive successors are found by an index-based walk with a
    /// visit stamp, so no per-task sets are allocated.
    fn identify_bottlenecks(tasks: &[Task]) -> Vec<Bottleneck> {
        let task_map: HashMap<&str, usize> = tasks
            .iter()
            .enumerate()
            .map(|(i, t)| (t.id.as_str(), i))
            .collect();
        let mut dependents: Vec<Vec<usize>> = vec![Vec::new(); tasks.len()];
        for (i, task) in tasks.iter().enumerate() {
            for dep in &task.dependencies {
                if let Some(&d) = task_map.get(dep.as_str()) {
                    dependents[d].push(i);
                }
            }
        }

        let mut bottlenecks = Vec::new();
        let mut stamp: Vec<usize> = vec![usize::MAX; tasks.len()];
        let mut stack: Vec<usize> = Vec::new();

        for (root, task) in tasks.iter().enumerate() {
            // Count transitively blocked tasks and their hours.
            let mut blocks_count = 0;
            let mut blocked_hours = 0.0_f32;
            stack.clear();
            stack.push(root);
            while let Some(current) = stack.pop() {
                for &succ in &dependents[current] {
                    if stamp[succ] != root {
                        stamp[succ] = root;
                        blocks_count += 1;
                        blocked_hours += tasks[succ].effort_hours;
                        stack.push(succ);
                    }
                }
            }

            if blocks_count == 0 {
                continue;
            }

            // Calculate ROI
            let roi = if task.effort_hours > 0.0 {
                blocked_hours / task.effort_hours
            } else {
                0.0
            };

            bottlenecks.push(Bottleneck {
                task_id: task.id.clone(),
                task_name: task.name.clone(),
                blocks_count,
                blocked_hours,
                roi,
                effort_hours: task.effort_hours,
            });
        }

        // Sort by ROI descending
        bottlenecks.sort_by(|a, b| {
            b.roi
                .partial_cmp(&a.roi)
                .unwrap_or(std::cmp::Ordering::Equal)
        });

        bottlenecks
    }

    /// Trace one longest chain backwards from the task with the maximum
    /// earliest finish, following predecessors whose EF equals the
    /// successor's ES. Sink ties resolve to the later earliest start, then the
    /// smallest id; predecessor ties resolve to the smallest id. Bounded by the
    /// task count so malformed (cyclic) input cannot loop.
    fn trace_critical_path(tasks: &[Task]) -> Vec<String> {
        let Some(sink) = tasks.iter().max_by(|a, b| {
            a.earliest_finish
                .partial_cmp(&b.earliest_finish)
                .unwrap_or(std::cmp::Ordering::Equal)
                .then_with(|| {
                    a.earliest_start
                        .partial_cmp(&b.earliest_start)
                        .unwrap_or(std::cmp::Ordering::Equal)
                })
                .then_with(|| b.id.cmp(&a.id))
        }) else {
            return Vec::new();
        };
        Self::trace_back(tasks, sink)
    }

    /// Trace the longest chain ending at the task `end_id` (same tie rules as
    /// [`Self::trace_critical_path`]), restricted to that task's ancestors.
    /// Empty when `end_id` is unknown.
    pub(crate) fn trace_path_to(tasks: &[Task], end_id: &str) -> Vec<String> {
        tasks
            .iter()
            .find(|t| t.id == end_id)
            .map_or_else(Vec::new, |end| Self::trace_back(tasks, end))
    }

    /// Walk backwards from `sink` along tight edges (EF + lag == ES),
    /// smallest id on ties. Lengths are scheduled lengths, not effort.
    fn trace_back(tasks: &[Task], sink: &Task) -> Vec<String> {
        let by_id: HashMap<&str, &Task> = tasks.iter().map(|t| (t.id.as_str(), t)).collect();
        let mut path = vec![sink.id.clone()];
        let mut current = sink;
        let mut seen: HashSet<&str> = HashSet::from([sink.id.as_str()]);
        while path.len() <= tasks.len() {
            let next = current
                .dependencies
                .iter()
                .filter_map(|d| by_id.get(d.as_str()).copied())
                .filter(|p| {
                    let lag = current.lag_by_dependency.get(&p.id).copied().unwrap_or(0.0);
                    (p.earliest_finish + lag - current.earliest_start).abs() < TIGHT_EPS
                })
                .filter(|p| !seen.contains(p.id.as_str()))
                .min_by(|a, b| a.id.cmp(&b.id));
            let Some(pred) = next else { break };
            seen.insert(pred.id.as_str());
            path.push(pred.id.clone());
            current = pred;
        }
        path.reverse();
        path
    }

    /// Build the final result
    fn build_result(
        tasks: &[Task],
        batches: Vec<TaskBatch>,
        bottlenecks: Vec<Bottleneck>,
        unscheduled: Vec<String>,
    ) -> CriticalPathResult {
        let critical_path = Self::trace_critical_path(tasks);
        let project_length = tasks
            .iter()
            .map(|t| t.earliest_finish)
            .fold(0.0_f32, f32::max);

        let mut critical_tasks: Vec<&Task> = tasks.iter().filter(|t| t.is_critical).collect();
        critical_tasks.sort_by(|a, b| {
            a.earliest_start
                .partial_cmp(&b.earliest_start)
                .unwrap_or(std::cmp::Ordering::Equal)
                .then_with(|| a.id.cmp(&b.id))
        });
        let critical_ids: Vec<String> = critical_tasks.iter().map(|t| t.id.clone()).collect();

        let total_duration_sequential: f32 = tasks.iter().map(|t| t.effort_hours).sum();
        let optimal_duration_parallel = project_length;

        let speedup_factor = if optimal_duration_parallel > 0.0 {
            total_duration_sequential / optimal_duration_parallel
        } else {
            1.0
        };

        CriticalPathResult {
            total_tasks: tasks.len(),
            critical_path,
            critical_ids,
            critical_path_duration: project_length,
            total_duration_sequential,
            optimal_duration_parallel,
            speedup_factor,
            parallelizable_batches: batches,
            bottlenecks,
            tasks: tasks.to_vec(),
            unscheduled,
        }
    }
}

#[cfg(test)]
#[allow(clippy::float_cmp)]
mod tests {
    use super::*;

    fn make_task(id: &str, effort: f32, deps: Vec<&str>) -> Task {
        Task {
            id: id.to_string(),
            name: format!("Task {id}"),
            effort_hours: effort,
            dependencies: deps.into_iter().map(String::from).collect(),
            ..Default::default()
        }
    }

    #[test]
    fn calculate_terminates_on_duplicate_ids() {
        // The second "a" closes a loop through "b" while the first is a root.
        let mut tasks = vec![
            make_task("a", 1.0, vec![]),
            make_task("a", 1.0, vec!["b"]),
            make_task("b", 1.0, vec!["a"]),
        ];
        let result = CpmAlgorithm::calculate(&mut tasks);
        assert_eq!(result.unscheduled, vec!["a".to_string(), "b".to_string()]);
    }

    #[test]
    fn test_simple_chain() {
        // A -> B -> C (linear dependency)
        let mut tasks = vec![
            make_task("A", 2.0, vec![]),
            make_task("B", 3.0, vec!["A"]),
            make_task("C", 1.0, vec!["B"]),
        ];

        let result = CpmAlgorithm::calculate(&mut tasks);

        // All tasks should be on critical path
        assert_eq!(result.critical_path.len(), 3);
        assert_eq!(result.critical_path_duration, 6.0);
        // No parallelization possible
        assert!((result.speedup_factor - 1.0).abs() < 0.01);
    }

    #[test]
    fn test_parallel_tasks() {
        // A -> B
        // A -> C (B and C can run in parallel)
        let mut tasks = vec![
            make_task("A", 1.0, vec![]),
            make_task("B", 2.0, vec!["A"]),
            make_task("C", 3.0, vec!["A"]),
        ];

        let result = CpmAlgorithm::calculate(&mut tasks);

        // Critical path: A -> C (1 + 3 = 4h)
        assert_eq!(result.critical_path_duration, 4.0);
        // Total sequential: 1 + 2 + 3 = 6h
        assert_eq!(result.total_duration_sequential, 6.0);
        // Speedup should be > 1
        assert!(result.speedup_factor > 1.0);
    }

    #[test]
    fn test_diamond_dependency() {
        //     B
        //   /   \
        // A       D
        //   \   /
        //     C
        let mut tasks = vec![
            make_task("A", 1.0, vec![]),
            make_task("B", 2.0, vec!["A"]),
            make_task("C", 4.0, vec!["A"]),
            make_task("D", 1.0, vec!["B", "C"]),
        ];

        let result = CpmAlgorithm::calculate(&mut tasks);

        // Critical path: A -> C -> D (1 + 4 + 1 = 6h)
        assert_eq!(result.critical_path_duration, 6.0);
        // B should not be on critical path (has float)
        assert!(!result.critical_path.contains(&"B".to_string()));
    }

    #[test]
    fn test_bottleneck_identification() {
        // A blocks B, C, D (high ROI)
        let mut tasks = vec![
            make_task("A", 1.0, vec![]),
            make_task("B", 5.0, vec!["A"]),
            make_task("C", 5.0, vec!["A"]),
            make_task("D", 5.0, vec!["A"]),
        ];

        let result = CpmAlgorithm::calculate(&mut tasks);

        // A should be identified as top bottleneck
        assert!(!result.bottlenecks.is_empty());
        assert_eq!(result.bottlenecks[0].task_id, "A");
        // ROI = 15h blocked / 1h effort = 15
        assert!(result.bottlenecks[0].roi >= 14.0);
    }

    #[test]
    fn test_empty_tasks() {
        let mut tasks: Vec<Task> = vec![];
        let result = CpmAlgorithm::calculate(&mut tasks);
        assert_eq!(result.total_tasks, 0);
        assert!(result.critical_path.is_empty());
    }

    #[test]
    fn test_single_task() {
        let mut tasks = vec![make_task("A", 5.0, vec![])];
        let result = CpmAlgorithm::calculate(&mut tasks);

        assert_eq!(result.total_tasks, 1);
        assert_eq!(result.critical_path_duration, 5.0);
        assert_eq!(result.critical_path, vec!["A".to_string()]);
    }

    #[test]
    fn test_cyclic_graph_is_reported_unscheduled() {
        // A -> B -> C -> A forms a cycle: none of these can be scheduled.
        let mut tasks = vec![
            make_task("A", 1.0, vec!["C"]),
            make_task("B", 1.0, vec!["A"]),
            make_task("C", 1.0, vec!["B"]),
        ];

        let result = CpmAlgorithm::calculate(&mut tasks);

        // The forward pass cannot drain the topo queue, so every member of
        // the cycle must surface in `unscheduled` (sorted) rather than the
        // result silently looking valid.
        assert_eq!(
            result.unscheduled,
            vec!["A".to_string(), "B".to_string(), "C".to_string()]
        );
    }

    #[test]
    fn test_partial_cycle_reports_only_cycle_members() {
        // ROOT is schedulable; X<->Y form a 2-cycle and are not.
        let mut tasks = vec![
            make_task("ROOT", 1.0, vec![]),
            make_task("X", 1.0, vec!["ROOT", "Y"]),
            make_task("Y", 1.0, vec!["X"]),
        ];

        let result = CpmAlgorithm::calculate(&mut tasks);

        assert_eq!(result.unscheduled, vec!["X".to_string(), "Y".to_string()]);
    }

    #[test]
    fn test_acyclic_graph_has_empty_unscheduled() {
        let mut tasks = vec![make_task("A", 2.0, vec![]), make_task("B", 3.0, vec!["A"])];
        let result = CpmAlgorithm::calculate(&mut tasks);
        assert!(result.unscheduled.is_empty());
    }

    #[test]
    fn test_batch_grouping() {
        // Two independent tasks starting at same time
        let mut tasks = vec![make_task("A", 2.0, vec![]), make_task("B", 3.0, vec![])];

        let result = CpmAlgorithm::calculate(&mut tasks);

        // Should be in same batch (start at time 0)
        assert_eq!(result.parallelizable_batches.len(), 1);
        assert_eq!(result.parallelizable_batches[0].tasks.len(), 2);
        // Duration is max (3h), not sum (5h)
        assert_eq!(result.parallelizable_batches[0].duration_hours, 3.0);
        assert_eq!(result.parallelizable_batches[0].total_effort_hours, 5.0);
    }
}