entrenar 0.7.12

//! Golden trace baselines for pruning module (PMAT QA Phase 6)
//!
//! This module defines expected performance characteristics and trace baselines
//! for the pruning operations. These serve as regression tests for performance
//! and can be verified with Renacer tracing when enabled.
//!
//! # Toyota Way: Hansei (Reflection)
//! Golden traces capture expected behavior for continuous improvement.

#![allow(unreachable_pub)]

use serde::{Deserialize, Serialize};

/// Maximum allowed latency for importance computation per layer (milliseconds)
const IMPORTANCE_LATENCY_MAX_MS: u64 = 5000;
/// Maximum allowed syscall/span count during calibration
const CALIBRATION_SYSCALL_BUDGET: u64 = 5000;

/// Performance assertion for a pruning operation.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct PruningAssertion {
    /// Name of the assertion
    pub name: &'static str,
    /// Type of assertion (latency, memory, etc.)
    pub assertion_type: AssertionType,
    /// Maximum allowed value
    pub max_value: u64,
    /// Whether violation causes test failure
    pub fail_on_violation: bool,
    /// Whether this assertion is enabled
    pub enabled: bool,
}

/// Type of performance assertion.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
pub enum AssertionType {
    /// Maximum duration in milliseconds
    Latency,
    /// Maximum memory usage in bytes
    Memory,
    /// Maximum number of spans/calls
    SpanCount,
    /// Anti-pattern detection threshold (0.0-1.0 as percentage * 100)
    AntiPattern,
}

/// Golden trace baseline for a pruning schedule.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ScheduleGoldenTrace {
    /// Schedule type name
    pub schedule_type: &'static str,
    /// Expected steps where pruning should trigger
    pub expected_prune_steps: Vec<usize>,
    /// Expected sparsity at each step (step, sparsity)
    pub expected_sparsity_curve: Vec<(usize, f32)>,
}

/// Golden trace baseline for a pruning configuration.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ConfigGoldenTrace {
    /// Configuration identifier
    pub config_id: &'static str,
    /// Expected calibration requirement
    pub requires_calibration: bool,
    /// Expected validation result (true = valid)
    pub expected_valid: bool,
}

/// Collection of all golden trace baselines for pruning.
pub struct PruningGoldenTraces;

impl PruningGoldenTraces {
    /// Get performance assertions for pruning operations.
    pub fn performance_assertions() -> Vec<PruningAssertion> {
        vec![
            // Importance computation latency
            PruningAssertion {
                name: "pruning_importance_latency",
                assertion_type: AssertionType::Latency,
                max_value: IMPORTANCE_LATENCY_MAX_MS, // <5s per layer
                fail_on_violation: true,
                enabled: true,
            },
            // Mask generation latency
            PruningAssertion {
                name: "pruning_mask_generation_latency",
                assertion_type: AssertionType::Latency,
                max_value: 1000, // <1s
                fail_on_violation: true,
                enabled: true,
            },
            // Memory budget
            PruningAssertion {
                name: "pruning_memory_budget",
                assertion_type: AssertionType::Memory,
                max_value: 2_147_483_648, // 2GB
                fail_on_violation: true,
                enabled: true,
            },
            // Calibration syscall budget
            PruningAssertion {
                name: "calibration_syscall_budget",
                assertion_type: AssertionType::SpanCount,
                max_value: CALIBRATION_SYSCALL_BUDGET,
                fail_on_violation: false, // Warning only
                enabled: true,
            },
            // Redundant computation detection
            PruningAssertion {
                name: "detect_redundant_computation",
                assertion_type: AssertionType::AntiPattern,
                max_value: 70, // 70% threshold
                fail_on_violation: false,
                enabled: true,
            },
            // Memory thrashing detection
            PruningAssertion {
                name: "detect_memory_thrashing",
                assertion_type: AssertionType::AntiPattern,
                max_value: 80, // 80% threshold
                fail_on_violation: false,
                enabled: true,
            },
        ]
    }

    /// Get golden traces for schedule types.
    pub fn schedule_traces() -> Vec<ScheduleGoldenTrace> {
        vec![
            // OneShot schedule at step 1000
            ScheduleGoldenTrace {
                schedule_type: "oneshot_1000",
                expected_prune_steps: vec![1000],
                expected_sparsity_curve: vec![(0, 0.0), (999, 0.0), (1000, 1.0), (1001, 1.0)],
            },
            // Gradual schedule 0-100 with freq=10, sparsity 0.0->0.5
            ScheduleGoldenTrace {
                schedule_type: "gradual_0_100_50pct",
                expected_prune_steps: vec![0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100],
                expected_sparsity_curve: vec![
                    (0, 0.0),
                    (25, 0.125),
                    (50, 0.25),
                    (75, 0.375),
                    (100, 0.5),
                ],
            },
            // Cubic schedule 0-100, sparsity 0.0->0.5
            // Formula: s_t = s_f * (1 - (1 - t/T)^3)
            ScheduleGoldenTrace {
                schedule_type: "cubic_0_100_50pct",
                expected_prune_steps: (0..=100).collect(),
                expected_sparsity_curve: vec![
                    (0, 0.0),
                    (25, 0.2890625), // 0.5 * (1 - 0.75^3) = 0.5 * 0.578125
                    (50, 0.4375),    // 0.5 * (1 - 0.5^3) = 0.5 * 0.875
                    (75, 0.4921875), // 0.5 * (1 - 0.25^3) = 0.5 * 0.984375
                    (100, 0.5),
                ],
            },
        ]
    }

    /// Get golden traces for configurations.
    pub fn config_traces() -> Vec<ConfigGoldenTrace> {
        vec![
            ConfigGoldenTrace {
                config_id: "default",
                requires_calibration: false, // Magnitude doesn't require calibration
                expected_valid: true,
            },
            ConfigGoldenTrace {
                config_id: "wanda_nm24",
                requires_calibration: true,
                expected_valid: true,
            },
            ConfigGoldenTrace {
                config_id: "sparsegpt_unstructured",
                requires_calibration: true,
                expected_valid: true,
            },
            ConfigGoldenTrace {
                config_id: "invalid_nm_n_gte_m",
                requires_calibration: false,
                expected_valid: false,
            },
        ]
    }
}

// =============================================================================
// Tests
// =============================================================================

#[cfg(test)]
mod tests {
    use super::*;
    use crate::prune::{PruneMethod, PruningConfig, PruningSchedule, SparsityPatternConfig};

    // =========================================================================
    // Golden Trace Verification Tests
    // =========================================================================

    #[test]
    fn test_oneshot_schedule_matches_golden_trace() {
        // TEST_ID: GOLD-001
        // Verify OneShot schedule behavior matches golden trace
        let schedule = PruningSchedule::OneShot { step: 1000 };
        let golden = PruningGoldenTraces::schedule_traces()
            .into_iter()
            .find(|t| t.schedule_type == "oneshot_1000")
            .expect("Golden trace not found");

        for (step, expected) in &golden.expected_sparsity_curve {
            let actual = schedule.sparsity_at_step(*step);
            assert!(
                (actual - expected).abs() < 1e-6,
                "GOLD-001 FALSIFIED: OneShot at step {step} expected {expected}, got {actual}"
            );
        }

        // Verify prune steps
        for step in &golden.expected_prune_steps {
            assert!(
                schedule.should_prune_at_step(*step),
                "GOLD-001 FALSIFIED: OneShot should prune at step {step}"
            );
        }
    }

    #[test]
    fn test_gradual_schedule_matches_golden_trace() {
        // TEST_ID: GOLD-002
        // Verify Gradual schedule behavior matches golden trace
        let schedule = PruningSchedule::Gradual {
            start_step: 0,
            end_step: 100,
            initial_sparsity: 0.0,
            final_sparsity: 0.5,
            frequency: 10,
        };
        let golden = PruningGoldenTraces::schedule_traces()
            .into_iter()
            .find(|t| t.schedule_type == "gradual_0_100_50pct")
            .expect("Golden trace not found");

        for (step, expected) in &golden.expected_sparsity_curve {
            let actual = schedule.sparsity_at_step(*step);
            assert!(
                (actual - expected).abs() < 1e-6,
                "GOLD-002 FALSIFIED: Gradual at step {step} expected {expected}, got {actual}"
            );
        }

        // Verify prune steps count matches
        assert_eq!(
            schedule.num_pruning_steps(),
            golden.expected_prune_steps.len(),
            "GOLD-002 FALSIFIED: Gradual num_pruning_steps mismatch"
        );
    }

    #[test]
    fn test_cubic_schedule_matches_golden_trace() {
        // TEST_ID: GOLD-003
        // Verify Cubic schedule behavior matches golden trace
        let schedule = PruningSchedule::Cubic { start_step: 0, end_step: 100, final_sparsity: 0.5 };
        let golden = PruningGoldenTraces::schedule_traces()
            .into_iter()
            .find(|t| t.schedule_type == "cubic_0_100_50pct")
            .expect("Golden trace not found");

        for (step, expected) in &golden.expected_sparsity_curve {
            let actual = schedule.sparsity_at_step(*step);
            assert!(
                (actual - expected).abs() < 1e-6,
                "GOLD-003 FALSIFIED: Cubic at step {step} expected {expected}, got {actual}"
            );
        }
    }

    #[test]
    fn test_config_calibration_matches_golden_trace() {
        // TEST_ID: GOLD-004
        // Verify config calibration requirements match golden traces
        let configs = [
            ("default", PruningConfig::default()),
            (
                "wanda_nm24",
                PruningConfig::new()
                    .with_method(PruneMethod::Wanda)
                    .with_pattern(SparsityPatternConfig::nm_2_4()),
            ),
            (
                "sparsegpt_unstructured",
                PruningConfig::new()
                    .with_method(PruneMethod::SparseGpt)
                    .with_pattern(SparsityPatternConfig::Unstructured),
            ),
        ];

        for (id, config) in &configs {
            let golden = PruningGoldenTraces::config_traces()
                .into_iter()
                .find(|t| t.config_id == *id)
                .expect("Golden trace not found");

            assert_eq!(
                config.requires_calibration(),
                golden.requires_calibration,
                "GOLD-004 FALSIFIED: Config {id} calibration requirement mismatch"
            );
        }
    }

    #[test]
    fn test_config_validation_matches_golden_trace() {
        // TEST_ID: GOLD-005
        // Verify config validation matches golden traces
        let valid_config = PruningConfig::default();
        assert!(
            valid_config.validate().is_ok(),
            "GOLD-005 FALSIFIED: Default config should be valid"
        );

        let invalid_config =
            PruningConfig::new().with_pattern(SparsityPatternConfig::NM { n: 5, m: 4 });
        assert!(
            invalid_config.validate().is_err(),
            "GOLD-005 FALSIFIED: Config with n >= m should be invalid"
        );
    }

    #[test]
    fn test_performance_assertions_defined() {
        // TEST_ID: GOLD-010
        // Verify all expected performance assertions are defined
        let assertions = PruningGoldenTraces::performance_assertions();

        let expected_names = [
            "pruning_importance_latency",
            "pruning_mask_generation_latency",
            "pruning_memory_budget",
            "calibration_syscall_budget",
            "detect_redundant_computation",
            "detect_memory_thrashing",
        ];

        for name in &expected_names {
            assert!(
                assertions.iter().any(|a| a.name == *name),
                "GOLD-010 FALSIFIED: Missing assertion: {name}"
            );
        }
    }

    #[test]
    fn test_latency_assertions_reasonable() {
        // TEST_ID: GOLD-011
        // Verify latency assertions have reasonable values
        let assertions = PruningGoldenTraces::performance_assertions();

        for assertion in &assertions {
            if assertion.assertion_type == AssertionType::Latency {
                assert!(
                    assertion.max_value <= 60000, // Max 60s
                    "GOLD-011 FALSIFIED: Latency {} unreasonably high: {}ms",
                    assertion.name,
                    assertion.max_value
                );
                assert!(
                    assertion.max_value >= 100, // Min 100ms
                    "GOLD-011 FALSIFIED: Latency {} unreasonably low: {}ms",
                    assertion.name,
                    assertion.max_value
                );
            }
        }
    }

    #[test]
    fn test_memory_assertion_reasonable() {
        // TEST_ID: GOLD-012
        // Verify memory assertion has reasonable value
        let assertions = PruningGoldenTraces::performance_assertions();

        let memory_assertion = assertions
            .iter()
            .find(|a| a.assertion_type == AssertionType::Memory)
            .expect("Memory assertion not found");

        // Should be between 100MB and 16GB
        assert!(
            memory_assertion.max_value >= 100_000_000,
            "GOLD-012 FALSIFIED: Memory budget too low: {}",
            memory_assertion.max_value
        );
        assert!(
            memory_assertion.max_value <= 17_179_869_184, // 16GB
            "GOLD-012 FALSIFIED: Memory budget too high: {}",
            memory_assertion.max_value
        );
    }

    #[test]
    fn test_golden_trace_serialization() {
        // TEST_ID: GOLD-020
        // Verify golden traces can be serialized
        let schedule_traces = PruningGoldenTraces::schedule_traces();
        let json = serde_json::to_string(&schedule_traces)
            .expect("GOLD-020 FALSIFIED: Failed to serialize schedule traces");
        assert!(
            json.contains("oneshot_1000"),
            "GOLD-020 FALSIFIED: Serialized trace should contain schedule type"
        );

        let assertions = PruningGoldenTraces::performance_assertions();
        let json = serde_json::to_string(&assertions)
            .expect("GOLD-020 FALSIFIED: Failed to serialize assertions");
        assert!(
            json.contains("pruning_importance_latency"),
            "GOLD-020 FALSIFIED: Serialized assertion should contain name"
        );
    }
}