car-eventlog 0.54.0

//! Harness-level evaluation metrics.
//!
//! Survey "Code as Agent Harness" §5.2.1: end-task success conflates the
//! base model, the harness, the tools, and the environment. To evaluate the
//! *operational substrate itself*, success accuracy must be complemented by
//! measurements of "execution reliability, feedback quality, context
//! sustainability, safety, coordination, and reproducibility." This module
//! derives those dimensions from the deep telemetry the event log now
//! records (token/cost/latency metrics, branch decisions, rejected
//! alternatives, permission decisions — see `car_eventlog`).
//!
//! These are descriptive, not normative: they characterize a trajectory so
//! harness variants can be compared, and an Evolution Agent (§3.5.2) can
//! attribute cost/failure to specific harness components.

use crate::{Event, EventKind, MetricsTotals};
use serde::{Deserialize, Serialize};

/// The six §5.2.1 dimensions, computed from a trajectory's event stream.
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct HarnessMetrics {
    pub trajectory_efficiency: TrajectoryEfficiency,
    pub verification_strength: VerificationStrength,
    pub recovery: Recovery,
    pub state_consistency: StateConsistency,
    pub safety: Safety,
    pub replayability: Replayability,
    /// JSONL lines that failed to parse when computing from a journal tail
    /// (0 when computed from in-memory events). Nonzero means the metrics
    /// were computed over a partial event set — do not compare across
    /// harnesses without accounting for it (neo review).
    pub parse_errors: usize,
    /// Share of graded TASKS the harness completed, when a task-suite runner
    /// measured one (`car-bench-harness --metrics-out`).
    ///
    /// This is **end-task** success, and it is a different quantity from
    /// [`TrajectoryEfficiency::success_rate`], which is attempt-level: a
    /// candidate whose every tool call succeeds while it fails more tasks
    /// scores 1.0 there and lower here. That gap is why the regression gate
    /// guards both.
    ///
    /// It cannot be derived from an event stream — the log knows what ran, not
    /// whether the task was satisfied — so [`compute_harness_metrics`] always
    /// leaves it `None`. Only a runner holding the task suite and its grading
    /// criteria can fill it in.
    ///
    /// **`None` means "not measured", never "zero".** A consumer that defaults
    /// it to 0.0 turns an unmeasured run into a total failure; a consumer that
    /// defaults it to 1.0 turns one into a perfect score. The gate does neither
    /// — with either side absent it simply does not fire.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub task_pass_rate: Option<f64>,
    /// How many tasks [`HarnessMetrics::task_pass_rate`] is a rate **over**.
    ///
    /// A bare rate is not comparable across runs, and the gap is exploitable
    /// rather than merely untidy. A runner drops a task from its denominator
    /// whenever it cannot measure it — and whether it can measure a task
    /// depends on the harness's own toolset, which is a declared, promotable
    /// mutation target (`HarnessComponent::ToolSchema`). Shrink the toolset so
    /// the four file-reading tasks become unmeasurable and they leave the
    /// denominator: exactly the four a file-blind harness would have *failed*.
    /// The scalar goes up, and without this field nothing downstream can see
    /// that the two rates count different things.
    ///
    /// `EvolutionAgent::evaluate` refuses the comparison outright when both
    /// sides carry a denominator and the two disagree.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub task_pass_denominator: Option<usize>,
    /// Tasks the runner ran against but could **not** measure, and therefore
    /// left out of [`HarnessMetrics::task_pass_denominator`].
    ///
    /// Reported for the same reason: a denominator that shrank is a fact, and
    /// the count of what fell out of it is the first thing anyone comparing two
    /// runs needs. Zero and absent are different — absent means the runner does
    /// not track it.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub tasks_unrunnable: Option<usize>,
}

/// (i) Trajectory efficiency — how much work was spent reaching the outcome.
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct TrajectoryEfficiency {
    /// Successful actions plus *failed attempts* (a retried action
    /// contributes one success and N failed attempts).
    pub attempts_total: usize,
    pub actions_succeeded: usize,
    /// Failed *attempts* (`ActionFailed` events), which include retries —
    /// see `success_rate`.
    pub failed_attempts: usize,
    /// Sum of input + output tokens across inference events.
    pub total_tokens: u64,
    pub total_cost_usd: f64,
    /// Summed wall-clock across metered events (ms).
    pub wall_clock_ms: f64,
    /// Number of model calls in the trajectory — one per
    /// [`EventKind::InferenceMetered`] event.
    ///
    /// This is the "model calls" leg of the #813 A/B (pass rate, model calls,
    /// tokens per task): a harness change that cuts `total_tokens` by making
    /// *more* calls has not saved anything, so the two are only interpretable
    /// together. Counted from the metered events rather than from a loop's own
    /// turn counter, so it stays correct if a turn ever issues more than one
    /// call — today the assistant loop issues exactly one, but that is a
    /// property of the loop, not of this metric.
    ///
    /// A call whose provider reported no usage still counts here (the call
    /// happened) while contributing nothing to `total_tokens` — which is why
    /// `total_tokens / model_calls` is not a safe per-call average.
    pub model_calls: usize,
    /// **Attempt-level** success: succeeded / (succeeded + failed_attempts).
    /// Counts every `ActionFailed` including retries, so a
    /// retried-then-succeeded action lowers this — a harness with
    /// aggressive retry (good recovery) scores lower here than one that
    /// gives up. Read it alongside `recovery.retries`. `None` when no
    /// attempts ran.
    pub success_rate: Option<f64>,
}

/// (ii) Verification strength — how much the harness checked before
/// accepting. False-acceptance rate needs an external oracle, so it is left
/// `None` here; what the log *can* show is how often verification rejected
/// or policy blocked an action (the verifier doing work).
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct VerificationStrength {
    pub actions_validated: usize,
    pub actions_rejected: usize,
    pub policy_violations: usize,
    /// Rejections / (validated + rejected) — the share caught **by the
    /// validator** before execution (excludes policy blocks, counted
    /// separately). `None` when nothing was validated.
    pub rejection_rate: Option<f64>,
}

/// (iii) Recovery ability — can the harness diagnose and repair failures?
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct Recovery {
    pub replan_attempts: usize,
    pub replan_rejected: usize,
    pub replan_exhausted: usize,
    /// Action-level retries (`ActionRetrying` events) — the within-action
    /// recovery the attempt-level `success_rate` counts against.
    pub retries: usize,
    /// Branch decisions + rejected alternatives — the size of the search
    /// the harness explored when things went wrong.
    pub branch_decisions: usize,
    pub alternatives_rejected: usize,
}

/// (iv) State consistency — how much the shared state churned and how often
/// the harness had to roll back.
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct StateConsistency {
    pub state_changes: usize,
    pub snapshots: usize,
    pub rollbacks: usize,
}

/// (v) Safety compliance — permission-gate activity (from the §5.2.5 tier
/// gate). High escalation/denial counts mean the harness governed real
/// risk rather than running unsupervised.
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct Safety {
    pub permission_decisions: usize,
    pub escalations: usize,
    pub denials: usize,
    /// EVERY durable human-in-the-loop decision, approvals and rejections
    /// alike.
    pub approvals_recorded: usize,
    /// The subset of [`Self::approvals_recorded`] that were REJECTIONS.
    ///
    /// `ApprovalRecorded` has always carried `approval: approved | rejected`
    /// in its data, and this fold has always thrown that away — so "the
    /// overseer approved 40 things" and "the overseer rejected 40 things"
    /// produced an identical `Safety`, despite being opposite safety facts.
    /// Splitting them makes the disagreement rate (`approvals_rejected /
    /// approvals_recorded`) computable, which is the one signal that
    /// distinguishes a reviewer who is still reading from one who has
    /// started waving things through.
    #[serde(default)]
    pub approvals_rejected: usize,
}

/// (vi) Replayability — what the log carries for reconstruction and audit.
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct Replayability {
    pub total_events: usize,
    /// True when the log carries engine-state delta records (`StateChanged`
    /// / `StateSnapshot`) — i.e. the trajectory's *state effects* can be
    /// replayed. This is distinct from auditability: a pure side-effecting
    /// tool (sends an email, writes an untracked file) succeeds with no
    /// state delta, so this is `false` even though the action is fully
    /// logged. Read it as "captured state effects", not "is auditable"
    /// (neo review — the append-only log is auditable regardless).
    pub state_effects_captured: bool,
}

/// Compute harness-level metrics from a trajectory's events.
///
/// [`HarnessMetrics::task_pass_rate`] — and with it
/// [`HarnessMetrics::task_pass_denominator`] and
/// [`HarnessMetrics::tasks_unrunnable`] — is deliberately left `None`:
/// end-task success is not in the event stream, and inventing a value here
/// would put a fabricated number in front of the regression gate.
pub fn compute_harness_metrics(events: &[Event]) -> HarnessMetrics {
    let mut m = HarnessMetrics::default();
    let totals: MetricsTotals = crate::metrics_totals_of(events);

    let mut succeeded = 0usize;
    let mut failed = 0usize;
    let mut model_calls = 0usize;
    let mut any_state_record = false;

    for ev in events {
        match ev.kind {
            EventKind::InferenceMetered => model_calls += 1,
            EventKind::ActionSucceeded => succeeded += 1,
            EventKind::ActionFailed => failed += 1,
            EventKind::ActionRetrying => m.recovery.retries += 1,
            EventKind::ActionValidated => m.verification_strength.actions_validated += 1,
            EventKind::ActionRejected => m.verification_strength.actions_rejected += 1,
            EventKind::PolicyViolation => m.verification_strength.policy_violations += 1,
            EventKind::ReplanAttempted => m.recovery.replan_attempts += 1,
            EventKind::ReplanRejected => m.recovery.replan_rejected += 1,
            EventKind::ReplanExhausted => m.recovery.replan_exhausted += 1,
            EventKind::BranchDecision => m.recovery.branch_decisions += 1,
            EventKind::AlternativeRejected => m.recovery.alternatives_rejected += 1,
            EventKind::StateChanged => {
                m.state_consistency.state_changes += 1;
                any_state_record = true;
            }
            EventKind::StateSnapshot => {
                m.state_consistency.snapshots += 1;
                any_state_record = true;
            }
            EventKind::StateRollback => m.state_consistency.rollbacks += 1,
            EventKind::PermissionDecision => {
                m.safety.permission_decisions += 1;
                match ev.data.get("gate_decision").and_then(|v| v.as_str()) {
                    Some("needs_approval") => m.safety.escalations += 1,
                    Some("deny") => m.safety.denials += 1,
                    _ => {}
                }
            }
            EventKind::ApprovalRecorded => {
                m.safety.approvals_recorded += 1;
                if ev.data.get("approval").and_then(|v| v.as_str()) == Some("rejected") {
                    m.safety.approvals_rejected += 1;
                }
            }
            _ => {}
        }
    }

    m.trajectory_efficiency = TrajectoryEfficiency {
        attempts_total: succeeded + failed,
        actions_succeeded: succeeded,
        failed_attempts: failed,
        total_tokens: totals.tokens,
        total_cost_usd: totals.cost_usd,
        wall_clock_ms: totals.duration_ms,
        model_calls,
        success_rate: ratio(succeeded, succeeded + failed),
    };
    m.verification_strength.rejection_rate = ratio(
        m.verification_strength.actions_rejected,
        m.verification_strength.actions_validated + m.verification_strength.actions_rejected,
    );
    m.replayability = Replayability {
        total_events: events.len(),
        state_effects_captured: any_state_record,
    };
    m
}

fn ratio(num: usize, denom: usize) -> Option<f64> {
    (denom > 0).then(|| num as f64 / denom as f64)
}

/// Compute harness metrics from a JSONL string of events (one `Event` per
/// line) — the form the FFI passes from a journal tail. Unparseable lines
/// are counted in `parse_errors` (not silently dropped), so a half-corrupt
/// tail is distinguishable from a clean one (neo review).
pub fn compute_from_jsonl(jsonl: &str) -> HarnessMetrics {
    let mut parse_errors = 0usize;
    let events: Vec<Event> = jsonl
        .lines()
        .filter(|l| !l.trim().is_empty())
        .filter_map(|l| match serde_json::from_str::<Event>(l) {
            Ok(ev) => Some(ev),
            Err(_) => {
                parse_errors += 1;
                None
            }
        })
        .collect();
    let mut m = compute_harness_metrics(&events);
    m.parse_errors = parse_errors;
    m
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::{EventLog, Metrics};

    #[test]
    fn efficiency_and_success_rate() {
        let mut log = EventLog::new();
        log.append_metered(
            EventKind::ActionSucceeded,
            Some("a1"),
            None,
            Default::default(),
            Metrics::latency(50.0),
        );
        log.append_metered(
            EventKind::InferenceMetered,
            None,
            None,
            Default::default(),
            Metrics::inference(100, 40, Some(0.01)),
        );
        log.append(
            EventKind::ActionFailed,
            Some("a2"),
            None,
            Default::default(),
        );

        let m = compute_harness_metrics(log.events());
        assert_eq!(m.trajectory_efficiency.attempts_total, 2);
        assert_eq!(m.trajectory_efficiency.actions_succeeded, 1);
        assert_eq!(m.trajectory_efficiency.failed_attempts, 1);
        assert_eq!(m.trajectory_efficiency.total_tokens, 140);
        assert_eq!(m.trajectory_efficiency.wall_clock_ms, 50.0);
        assert_eq!(m.trajectory_efficiency.success_rate, Some(0.5));
    }

    #[test]
    fn retried_then_succeeded_lowers_attempt_success_but_counts_retries() {
        // One logical action: 2 failed attempts + 2 retries + 1 success.
        let mut log = EventLog::new();
        log.append(
            EventKind::ActionFailed,
            Some("a1"),
            None,
            Default::default(),
        );
        log.append(
            EventKind::ActionRetrying,
            Some("a1"),
            None,
            Default::default(),
        );
        log.append(
            EventKind::ActionFailed,
            Some("a1"),
            None,
            Default::default(),
        );
        log.append(
            EventKind::ActionRetrying,
            Some("a1"),
            None,
            Default::default(),
        );
        log.append(
            EventKind::ActionSucceeded,
            Some("a1"),
            None,
            Default::default(),
        );

        let m = compute_harness_metrics(log.events());
        assert_eq!(m.trajectory_efficiency.failed_attempts, 2);
        assert_eq!(m.recovery.retries, 2);
        // Attempt-level success_rate reflects the retries (1/3), and the
        // retries field explains why — the two are read together.
        assert_eq!(m.trajectory_efficiency.success_rate, Some(1.0 / 3.0));
    }

    /// `model_calls` must equal the number of `InferenceMetered` events —
    /// not the number of actions, and not the number of events that happen to
    /// carry token metrics. The #813 A/B reads tokens *per model call*, so a
    /// count that drifted from the metered events would misattribute the whole
    /// comparison.
    #[test]
    fn model_calls_equals_inference_metered_event_count() {
        let mut log = EventLog::new();
        // Three model calls, one of which reported no usage (a provider that
        // cannot count tokens) — it is still a call.
        log.append_metered(
            EventKind::InferenceMetered,
            None,
            None,
            Default::default(),
            Metrics::inference(100, 40, Some(0.01)),
        );
        log.append_metered(
            EventKind::InferenceMetered,
            None,
            None,
            Default::default(),
            Metrics::inference(60, 10, None),
        );
        log.append_metered(
            EventKind::InferenceMetered,
            None,
            None,
            Default::default(),
            Metrics::latency(12.0),
        );
        // Non-inference events must not inflate the count, including a metered
        // action that carries latency.
        log.append_metered(
            EventKind::ActionSucceeded,
            Some("a1"),
            None,
            Default::default(),
            Metrics::latency(50.0),
        );
        log.append(
            EventKind::ActionFailed,
            Some("a2"),
            None,
            Default::default(),
        );

        let m = compute_harness_metrics(log.events());
        let metered = log
            .events()
            .iter()
            .filter(|e| e.kind == EventKind::InferenceMetered)
            .count();
        assert_eq!(metered, 3, "fixture should hold three metered calls");
        assert_eq!(m.trajectory_efficiency.model_calls, metered);
        // The usage-less third call contributes no tokens but still counts.
        assert_eq!(m.trajectory_efficiency.total_tokens, 210);
        assert_eq!(m.trajectory_efficiency.attempts_total, 2);
    }

    /// An empty trajectory has made no model calls — `model_calls` must be 0,
    /// never a default that reads as "one call".
    #[test]
    fn no_inference_events_means_no_model_calls() {
        let mut log = EventLog::new();
        log.append(
            EventKind::ActionSucceeded,
            Some("a1"),
            None,
            Default::default(),
        );
        assert_eq!(
            compute_harness_metrics(log.events())
                .trajectory_efficiency
                .model_calls,
            0
        );
        assert_eq!(
            compute_harness_metrics(&[])
                .trajectory_efficiency
                .model_calls,
            0
        );
    }

    /// `model_calls` must survive the JSONL round trip the FFI and
    /// `car-bench-harness --metrics-out` use, since that is the form the
    /// evolution gate receives as `harness_candidate_metrics`.
    #[test]
    fn model_calls_survives_json_round_trip() {
        let mut log = EventLog::new();
        log.append_metered(
            EventKind::InferenceMetered,
            None,
            None,
            Default::default(),
            Metrics::inference(7, 3, None),
        );
        let jsonl = serde_json::to_string(&log.events()[0]).unwrap();
        let from_jsonl = compute_from_jsonl(&jsonl);
        assert_eq!(from_jsonl.trajectory_efficiency.model_calls, 1);

        let encoded = serde_json::to_string(&from_jsonl).unwrap();
        let decoded: HarnessMetrics = serde_json::from_str(&encoded).unwrap();
        assert_eq!(decoded.trajectory_efficiency.model_calls, 1);
        assert_eq!(decoded.trajectory_efficiency.total_tokens, 10);
    }

    /// `task_pass_rate` is a task-suite measurement, not an event-stream one.
    /// Computing it here would mean fabricating it, so it must come back
    /// `None` — "not measured" — even from a rich trajectory.
    #[test]
    fn task_pass_rate_is_never_derived_from_events() {
        let mut log = EventLog::new();
        log.append_metered(
            EventKind::InferenceMetered,
            None,
            None,
            Default::default(),
            Metrics::inference(100, 40, Some(0.01)),
        );
        log.append(
            EventKind::ActionSucceeded,
            Some("a1"),
            None,
            Default::default(),
        );
        assert_eq!(compute_harness_metrics(log.events()).task_pass_rate, None);
        assert_eq!(compute_harness_metrics(&[]).task_pass_rate, None);
        assert_eq!(compute_from_jsonl("").task_pass_rate, None);

        // Same rule for the denominator that qualifies it. A folded event
        // stream that reported "0 tasks over 0" would be a fabricated
        // denominator, and the gate would then compare it against a real one.
        let m = compute_harness_metrics(log.events());
        assert_eq!(m.task_pass_denominator, None);
        assert_eq!(m.tasks_unrunnable, None);
    }

    /// The denominator has to reach the gate through the same file
    /// `--metrics-out` writes, and a document written before the field existed
    /// must still load — as `None`, which is what makes the gate's
    /// comparability check skip rather than reject.
    #[test]
    fn task_pass_denominator_round_trips_and_is_backward_compatible() {
        let mut m = compute_harness_metrics(&[]);
        m.task_pass_rate = Some(0.75);
        m.task_pass_denominator = Some(12);
        m.tasks_unrunnable = Some(2);

        let encoded = serde_json::to_string(&m).unwrap();
        let decoded: HarnessMetrics = serde_json::from_str(&encoded).unwrap();
        assert_eq!(decoded.task_pass_denominator, Some(12));
        assert_eq!(decoded.tasks_unrunnable, Some(2));

        // Absent in an older document → None, never 0.
        let old = r#"{"task_pass_rate":0.75}"#;
        let decoded: HarnessMetrics = serde_json::from_str(old).unwrap();
        assert_eq!(decoded.task_pass_rate, Some(0.75));
        assert_eq!(decoded.task_pass_denominator, None);
        assert_eq!(decoded.tasks_unrunnable, None);

        // And an unmeasured document carries neither key at all, so a reader
        // cannot mistake a zero for a measurement.
        let bare = serde_json::to_value(compute_harness_metrics(&[])).unwrap();
        assert!(bare.get("task_pass_denominator").is_none());
        assert!(bare.get("tasks_unrunnable").is_none());
    }

    /// A runner that DID measure it must be able to ship the number to the
    /// gate: it has to survive the JSON round trip `--metrics-out` uses, and an
    /// older document without the field must still load (as `None`).
    #[test]
    fn task_pass_rate_round_trips_and_is_backward_compatible() {
        let mut m = compute_harness_metrics(&[]);
        m.task_pass_rate = Some(0.75);
        let encoded = serde_json::to_string(&m).unwrap();
        let decoded: HarnessMetrics = serde_json::from_str(&encoded).unwrap();
        assert_eq!(decoded.task_pass_rate, Some(0.75));

        // A document written before this field existed carries no key at all.
        let older = r#"{"trajectory_efficiency":{"model_calls":3}}"#;
        let loaded: HarnessMetrics = serde_json::from_str(older).unwrap();
        assert_eq!(loaded.task_pass_rate, None);
        assert_eq!(loaded.trajectory_efficiency.model_calls, 3);
    }

    /// Approve and reject are opposite safety facts. The fold used to
    /// increment one counter for both, so a `Safety` could not say which had
    /// happened.
    #[test]
    fn approvals_and_rejections_are_counted_separately() {
        let mut log = EventLog::new();
        for decision in ["approved", "rejected", "rejected"] {
            log.append(
                EventKind::ApprovalRecorded,
                None,
                None,
                [("approval".to_string(), decision.into())].into(),
            );
        }

        let m = compute_harness_metrics(log.events());
        assert_eq!(m.safety.approvals_recorded, 3, "every decision is recorded");
        assert_eq!(
            m.safety.approvals_rejected, 2,
            "two of them were rejections"
        );
    }

    #[test]
    fn recovery_and_safety_counts() {
        let mut log = EventLog::new();
        log.append(EventKind::ReplanAttempted, None, None, Default::default());
        log.append(EventKind::BranchDecision, None, None, Default::default());
        log.append(
            EventKind::AlternativeRejected,
            None,
            None,
            Default::default(),
        );
        log.append(
            EventKind::PermissionDecision,
            None,
            None,
            [("gate_decision".to_string(), "needs_approval".into())].into(),
        );
        log.append(
            EventKind::PermissionDecision,
            None,
            None,
            [("gate_decision".to_string(), "deny".into())].into(),
        );
        log.append(EventKind::ApprovalRecorded, None, None, Default::default());

        let m = compute_harness_metrics(log.events());
        assert_eq!(m.recovery.replan_attempts, 1);
        assert_eq!(m.recovery.branch_decisions, 1);
        assert_eq!(m.recovery.alternatives_rejected, 1);
        assert_eq!(m.safety.permission_decisions, 2);
        assert_eq!(m.safety.escalations, 1);
        assert_eq!(m.safety.denials, 1);
        assert_eq!(m.safety.approvals_recorded, 1);
    }

    #[test]
    fn state_effects_captured_tracks_state_records() {
        // A success with no state record → no captured state effects (but
        // still fully logged/auditable — that's the point of the rename).
        let mut log = EventLog::new();
        log.append(
            EventKind::ActionSucceeded,
            Some("a1"),
            None,
            Default::default(),
        );
        assert!(
            !compute_harness_metrics(log.events())
                .replayability
                .state_effects_captured
        );

        // Add a StateChanged → state effects captured.
        log.append(
            EventKind::StateChanged,
            Some("a1"),
            None,
            Default::default(),
        );
        assert!(
            compute_harness_metrics(log.events())
                .replayability
                .state_effects_captured
        );
    }

    #[test]
    fn empty_trajectory_has_no_success_rate() {
        let m = compute_harness_metrics(&[]);
        assert_eq!(m.trajectory_efficiency.success_rate, None);
        assert_eq!(m.replayability.total_events, 0);
    }

    #[test]
    fn jsonl_counts_parse_errors() {
        let mut log = EventLog::new();
        log.append(
            EventKind::ActionSucceeded,
            Some("a1"),
            None,
            Default::default(),
        );
        let good = serde_json::to_string(&log.events()[0]).unwrap();
        let jsonl = format!("{good}\nnot json\n{{\"kind\":\"bogus\"}}\n");
        let m = compute_from_jsonl(&jsonl);
        assert_eq!(m.trajectory_efficiency.actions_succeeded, 1);
        // "not json" and the unknown-kind line both fail to parse.
        assert_eq!(m.parse_errors, 2);
    }
}