car-multi 0.54.0

//! Turning coordination runs into topology-selection records.
//!
//! [`car_topology`] fits a selector on `(topology, query, utility, tokens)`
//! tuples. `car-multi` is where those tuples come from: every pattern already
//! returns `AgentOutput`s carrying [`TokenAccounting`](crate::types::TokenAccounting),
//! so the measured cost of a coordination run is already in hand — it just was
//! not being kept.
//!
//! This module is the bridge in both directions:
//!
//! - [`execution_record`] folds a finished run into one
//!   [`car_topology::ExecutionRecord`].
//! - [`team_homogeneity`] answers, for a real `Vec<AgentSpec>`, the question
//!   the paper shows decides whether a profile-node message-passing scorer can
//!   rank anything at all.
//!
//! Nothing here calls a model. Collecting records costs LLM calls because
//! *running the coordination* does; the bookkeeping is free.

use car_topology::{
    CoordinationShape, ExecutionRecord, Homogeneity, JournalError, RecordJournal, ScorerAdvice,
    Selection, Topology, TopologyError, TopologySelector,
};

use car_ir::{AgentOutcome, EvidenceKind, OutcomeStatus};

use crate::types::{AgentOutput, AgentSpec};

/// Total measured tokens across a run's outputs, or `None` when no output
/// reported any accounting.
///
/// `None` rather than `0` on purpose. [`TokenAccounting`](crate::types::TokenAccounting)
/// is opt-in — a runner that does not populate it is not reporting a zero-cost
/// run, it is reporting nothing — and a record claiming zero tokens would drag
/// the proxy's cost head toward a topology that merely went unmetered.
pub fn measured_tokens(outputs: &[AgentOutput]) -> Option<u64> {
    let mut total = 0u64;
    let mut any = false;
    for output in outputs {
        if let Some(tokens) = &output.tokens {
            any = true;
            total = total
                .saturating_add(tokens.input_tokens)
                .saturating_add(tokens.output_tokens);
        }
    }
    any.then_some(total)
}

/// Fold a finished coordination run into one execution record.
///
/// `team_size` is the number of agents the *topology* is over, which is not
/// always `outputs.len()`: a swarm's optional synthesizer and a supervisor's
/// reviewer rounds add outputs without adding nodes. Pass the length of the
/// `AgentSpec` list the pattern was built from.
///
/// `utility` is the caller's — CAR does not know whether a coordination run
/// answered the question. Use 1.0/0.0 for a pass/fail check, or a graded score
/// where one exists.
///
/// Returns `None` when the run reported no token accounting, so an unmetered
/// run is skipped rather than logged as free.
pub fn execution_record(
    task_id: impl Into<String>,
    query: Vec<f32>,
    shape: CoordinationShape,
    team_size: usize,
    outputs: &[AgentOutput],
    utility: f32,
) -> Result<Option<ExecutionRecord>, TopologyError> {
    // `Solo` is refused rather than recorded. It folds to the empty adjacency,
    // which is byte-identical to `Swarm`, so a one-agent run logged with
    // `team_size = 4` becomes a *Swarm* record carrying one agent's token cost.
    // Nothing downstream can undo that: `shape_of` cannot recover `Solo` from
    // the matrix, so `select_shape` hands the caller `Swarm`, who then runs
    // four agents against a price learned from one. The paper's model has no
    // room for it either — a solo run is a different `N`, not a different
    // adjacency over the same `N`.
    if shape == CoordinationShape::Solo {
        return Err(TopologyError::BadConfig {
            field: "shape",
            expected: "a shape with a distinguishable adjacency",
            found: "solo — indistinguishable from swarm; record the run under                     its real team size, or not at all"
                .into(),
        });
    }
    let Some(tokens) = measured_tokens(outputs) else {
        return Ok(None);
    };
    let topology = shape.topology(team_size)?;
    Ok(Some(ExecutionRecord::new(
        task_id, query, topology, utility, tokens,
    )))
}

/// The same fold for a run whose topology is not one of the named shapes —
/// a topology a selector chose, for instance.
pub fn execution_record_for(
    task_id: impl Into<String>,
    query: Vec<f32>,
    topology: Topology,
    outputs: &[AgentOutput],
    utility: f32,
) -> Option<ExecutionRecord> {
    let tokens = measured_tokens(outputs)?;
    Some(ExecutionRecord::new(
        task_id, query, topology, utility, tokens,
    ))
}

/// How much of a success claim has to be demonstrated before it counts.
///
/// CAR already draws this line elsewhere — the flagship assistant cross-checks
/// a final summary's operational claims against same-run tool receipts rather
/// than believing the prose. The same reasoning applies here, and harder: a
/// codebook fitted on runs that merely *said* they succeeded is a codebook of
/// whatever the model is most confident about.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum UtilityEvidence {
    /// Take [`OutcomeStatus`] at face value, whatever backs it.
    Reported,
    /// Require at least one evidence item that is not the agent's own
    /// assessment — a tool result, a state change, an external verification, or
    /// a product evaluator. A `Success` backed only by
    /// [`EvidenceKind::SelfAssessment`] yields `None`, not `0.0`: "nobody
    /// checked" is not "it failed", and recording it either way is a fabricated
    /// training label.
    Grounded,
}

/// How to reduce a multi-agent run's outcomes to one utility.
///
/// There is no universal rule, so this is the caller's to state rather than
/// something to infer from the outputs.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum UtilityAggregation {
    /// The last output carrying an outcome is the run's answer.
    ///
    /// Right for the shapes where one agent's output *is* the result —
    /// [`CoordinationShape::Pipeline`] (the last stage is what the caller sees)
    /// and [`CoordinationShape::Supervisor`] (the reviewer has the last word).
    Final,
    /// The fraction of outcome-carrying outputs that succeeded.
    ///
    /// Right for the shapes where agents answer independently and the answer is
    /// a vote or a pick — [`CoordinationShape::Swarm`] and
    /// [`CoordinationShape::Debate`]. Yields the graded utility in `[0, 1]` that
    /// [`ExecutionRecord`] already supports, rather than forcing a binary.
    Consensus,
}

/// Map one agent outcome to a utility in `[0, 1]`, or `None` when it carries no
/// success signal.
///
/// The mapping is dictated by the type rather than chosen:
///
/// | [`OutcomeStatus`] | utility |
/// |---|---|
/// | `Success` | `1.0` |
/// | `PartialSuccess` | `0.5` |
/// | `Failure`, `GiveUp`, `Timeout` | `0.0` |
/// | `Done` | `None` |
///
/// `Done` is `None` because `car-ir` documents it as *"neutral — may or may not
/// have succeeded"*. Scoring a deliberately neutral status as either outcome
/// would invent the one number this crate exists to measure honestly.
pub fn outcome_utility(outcome: &AgentOutcome, evidence: UtilityEvidence) -> Option<f32> {
    if evidence == UtilityEvidence::Grounded
        && !outcome
            .evidence
            .iter()
            .any(|e| e.kind != EvidenceKind::SelfAssessment)
    {
        return None;
    }
    match outcome.status {
        OutcomeStatus::Success => Some(1.0),
        OutcomeStatus::PartialSuccess => Some(0.5),
        OutcomeStatus::Failure | OutcomeStatus::GiveUp | OutcomeStatus::Timeout => Some(0.0),
        OutcomeStatus::Done => None,
    }
}

/// Derive a run's utility from its agents' structured outcomes.
///
/// Returns `None` when no output carries a usable outcome — a runner that
/// reports nothing structured is not reporting success, and the caller should
/// skip the record rather than label it. That is the same discipline
/// [`measured_tokens`] applies to cost, for the same reason: the paper's whole
/// argument is that these two numbers must be *measured*, and a default is not
/// a measurement.
///
/// **This is only as good as the runner.** `AgentOutput::outcome` is populated
/// by the caller's [`AgentRunner`](crate::AgentRunner), not by this crate: every
/// aggregate output the coordination patterns build themselves (a swarm's
/// synthesizer, a map-reduce reducer, a vote's tally) sets `outcome: None` by
/// construction, because the runtime does not own the model and cannot classify
/// what it did not run. A team whose runner never fills the field will record
/// nothing through this path, which is the intended failure — fail closed, not
/// fabricate.
pub fn run_utility(
    outputs: &[AgentOutput],
    aggregation: UtilityAggregation,
    evidence: UtilityEvidence,
) -> Option<f32> {
    let scored: Vec<f32> = outputs
        .iter()
        .filter_map(|o| o.outcome.as_ref())
        .filter_map(|o| outcome_utility(o, evidence))
        .collect();
    if scored.is_empty() {
        return None;
    }
    match aggregation {
        UtilityAggregation::Final => scored.last().copied(),
        UtilityAggregation::Consensus => Some(scored.iter().sum::<f32>() / scored.len() as f32),
    }
}

/// Record a finished coordination run, deriving BOTH measured numbers from the
/// run itself.
///
/// The ergonomic path deliberately cannot fabricate either one: it writes
/// nothing and returns `false` when the runner metered no tokens *or* when no
/// agent reported a usable outcome. Defaulting utility to 1.0 — the obvious
/// shortcut — would make the codebook the set of everything ever run, which is
/// exactly the short list the paper says does not exist.
///
/// Use [`record_run`] instead when the caller has a better utility signal than
/// the agents' self-reported outcomes: a `car-verify` goal verdict, a product
/// evaluator, or a graded score.
pub fn record_run_from_outcomes(
    journal: &mut RecordJournal,
    embedder: &str,
    task_id: impl Into<String>,
    query: Vec<f32>,
    shape: CoordinationShape,
    team_size: usize,
    outputs: &[AgentOutput],
    aggregation: UtilityAggregation,
    evidence: UtilityEvidence,
) -> Result<bool, JournalError> {
    let Some(utility) = run_utility(outputs, aggregation, evidence) else {
        return Ok(false);
    };
    record_run(
        journal, embedder, task_id, query, shape, team_size, outputs, utility,
    )
}

/// Record a finished coordination run to a [`RecordJournal`].
///
/// The one-call form of [`execution_record`] + [`RecordJournal::append`], which
/// is what a daemon call site wants: run the coordination, decide the utility,
/// hand the run here.
///
/// `embedder` names the encoder that produced `query`. It is not optional
/// because the journal's whole guard is that a record can never be folded into
/// a set from a different embedding space — see [`car_topology::journal`].
///
/// Returns `false` when the run reported no token accounting and so was not
/// recorded. That is the honest outcome, not a failure: an unmetered run has no
/// measured cost, and this crate's whole premise is scoring on measured cost.
pub fn record_run(
    journal: &mut RecordJournal,
    embedder: &str,
    task_id: impl Into<String>,
    query: Vec<f32>,
    shape: CoordinationShape,
    team_size: usize,
    outputs: &[AgentOutput],
    utility: f32,
) -> Result<bool, JournalError> {
    let Some(record) = execution_record(task_id, query, shape, team_size, outputs, utility)? else {
        return Ok(false);
    };
    journal.append(embedder, &record)?;
    Ok(true)
}

/// Whether a team's agents are distinguishable, judged on the specs themselves.
///
/// The paper measures homogeneity over profile *embeddings*; for CAR the
/// question is answerable without an encoder, because a team is usually built
/// by cloning one `AgentSpec` N times. Two agents count as distinguishable when
/// their system prompts differ — the prompt is the profile.
///
/// Names are deliberately ignored: `solver_1` … `solver_4` around one identical
/// prompt is the homogeneous case, and letting the name split them would give
/// exactly the false negative this check exists to prevent.
pub fn team_homogeneity(agents: &[AgentSpec]) -> Homogeneity {
    let Some(first) = agents.first() else {
        return Homogeneity::Homogeneous;
    };
    if agents
        .iter()
        .skip(1)
        .any(|a| a.system_prompt != first.system_prompt)
    {
        Homogeneity::Heterogeneous
    } else {
        Homogeneity::Homogeneous
    }
}

/// What a profile-node message-passing scorer could do on this team, phrased
/// for a log line next to a selection.
pub fn scorer_advice(agents: &[AgentSpec]) -> ScorerAdvice {
    match team_homogeneity(agents) {
        Homogeneity::Homogeneous => ScorerAdvice {
            homogeneity: Homogeneity::Homogeneous,
            message_passing_is_adjacency_blind: true,
            reason: format!(
                "all {} agents share one system prompt, so a message-passing scorer over \
                 profile nodes pools identical features and scores every candidate topology \
                 the same; rank on the adjacency itself",
                agents.len()
            ),
        },
        Homogeneity::Heterogeneous => ScorerAdvice {
            homogeneity: Homogeneity::Heterogeneous,
            message_passing_is_adjacency_blind: false,
            reason: "agents carry different system prompts, so profile-node message passing \
                     can in principle distinguish adjacencies"
                .into(),
        },
    }
}

/// Select a coordination shape for a query, falling back to `default_shape`
/// when the selected topology matches no named pattern.
///
/// A learned codebook can hold topologies with no `car-multi` pattern behind
/// them — that is a real outcome of fitting on records, not a bug — and this
/// function is for the caller who can only execute the named patterns. Callers
/// that can execute an arbitrary adjacency should use
/// [`TopologySelector::select`] directly and read [`Selection::topology`].
pub fn select_shape(
    selector: &TopologySelector,
    query: &[f32],
    default_shape: CoordinationShape,
) -> Result<(CoordinationShape, Selection), TopologyError> {
    let selection = selector.select(query)?;
    // Prefer the winner's own shape; otherwise the best-scoring candidate that
    // has one; otherwise the caller's default.
    let shape = selection.shape.or_else(|| {
        let mut named: Vec<_> = selection
            .considered
            .iter()
            .filter(|c| c.shape.is_some())
            .collect();
        named.sort_by(|a, b| {
            b.objective
                .total_cmp(&a.objective)
                .then(a.code.cmp(&b.code))
        });
        named.first().and_then(|c| c.shape)
    });
    Ok((shape.unwrap_or(default_shape), selection))
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::types::TokenAccounting;
    use car_topology::{RecordSet, SelectorConfig};

    fn output(name: &str, tokens: Option<(u64, u64)>) -> AgentOutput {
        AgentOutput {
            name: name.into(),
            answer: "done".into(),
            turns: 1,
            tool_calls: 0,
            duration_ms: 1.0,
            error: None,
            outcome: None,
            tokens: tokens.map(|(i, o)| TokenAccounting::new(i, o, 0.0)),
            tools_used: Vec::new(),
        }
    }

    fn spec(name: &str, prompt: &str) -> AgentSpec {
        AgentSpec::new(name, prompt)
    }

    #[test]
    fn tokens_sum_across_outputs() {
        let outputs = vec![output("a", Some((100, 20))), output("b", Some((50, 30)))];
        assert_eq!(measured_tokens(&outputs), Some(200));
    }

    #[test]
    fn an_unmetered_run_reports_none_not_zero() {
        let outputs = vec![output("a", None), output("b", None)];
        assert_eq!(measured_tokens(&outputs), None);
        assert_eq!(
            execution_record(
                "t",
                vec![0.5],
                CoordinationShape::Pipeline,
                4,
                &outputs,
                1.0
            )
            .unwrap(),
            None
        );
    }

    #[test]
    fn a_partially_metered_run_still_counts() {
        let outputs = vec![output("a", Some((100, 20))), output("b", None)];
        assert_eq!(measured_tokens(&outputs), Some(120));
    }

    #[test]
    fn a_solo_run_is_refused_rather_than_recorded_as_a_swarm() {
        // Solo and Swarm share the empty adjacency, so recording a 1-agent run
        // at team_size 4 would mint a Swarm record priced for one agent — and
        // `shape_of` could never recover the Solo, so a later `select_shape`
        // would hand back Swarm and run four.
        let outputs = vec![output("a", Some((100, 20)))];
        assert!(matches!(
            execution_record("t", vec![0.5], CoordinationShape::Solo, 4, &outputs, 1.0),
            Err(TopologyError::BadConfig { field: "shape", .. })
        ));
        // Swarm itself still records fine.
        assert!(
            execution_record("t", vec![0.5], CoordinationShape::Swarm, 4, &outputs, 1.0)
                .unwrap()
                .is_some()
        );
    }

    #[test]
    fn a_record_carries_the_shapes_topology_over_the_team_size() {
        let outputs = vec![output("a", Some((100, 20)))];
        let record = execution_record("t", vec![0.5], CoordinationShape::Debate, 4, &outputs, 1.0)
            .unwrap()
            .unwrap();
        assert_eq!(record.topology, Topology::complete(4).unwrap());
        assert_eq!(record.tokens, 120);
        assert_eq!(record.task_id, "t");
    }

    #[test]
    fn a_cloned_team_is_homogeneous_whatever_the_names() {
        let team = vec![
            spec("solver_1", "You solve math problems."),
            spec("solver_2", "You solve math problems."),
            spec("solver_3", "You solve math problems."),
        ];
        let advice = scorer_advice(&team);
        assert_eq!(advice.homogeneity, Homogeneity::Homogeneous);
        assert!(advice.message_passing_is_adjacency_blind);
        assert!(advice.reason.contains("share one system prompt"));
    }

    #[test]
    fn distinct_roles_are_heterogeneous() {
        let team = vec![
            spec("researcher", "You gather evidence."),
            spec("verifier", "You check the answer."),
        ];
        assert_eq!(team_homogeneity(&team), Homogeneity::Heterogeneous);
        assert!(!scorer_advice(&team).message_passing_is_adjacency_blind);
    }

    #[test]
    fn an_empty_team_is_homogeneous() {
        assert_eq!(team_homogeneity(&[]), Homogeneity::Homogeneous);
    }

    fn outcome(status: OutcomeStatus, kinds: &[EvidenceKind]) -> AgentOutcome {
        AgentOutcome {
            status,
            summary: String::new(),
            evidence: kinds
                .iter()
                .map(|&kind| car_ir::Evidence {
                    kind,
                    description: String::new(),
                    data: None,
                })
                .collect(),
            metrics: Default::default(),
            timestamp: chrono::Utc::now(),
        }
    }

    fn output_with(
        name: &str,
        tokens: Option<(u64, u64)>,
        status: OutcomeStatus,
        kinds: &[EvidenceKind],
    ) -> AgentOutput {
        let mut o = output(name, tokens);
        o.outcome = Some(outcome(status, kinds));
        o
    }

    #[test]
    fn the_status_to_utility_mapping_is_the_one_the_type_dictates() {
        let ev = &[EvidenceKind::ToolResult];
        for (status, expected) in [
            (OutcomeStatus::Success, Some(1.0)),
            (OutcomeStatus::PartialSuccess, Some(0.5)),
            (OutcomeStatus::Failure, Some(0.0)),
            (OutcomeStatus::GiveUp, Some(0.0)),
            (OutcomeStatus::Timeout, Some(0.0)),
        ] {
            assert_eq!(
                outcome_utility(&outcome(status, ev), UtilityEvidence::Reported),
                expected,
                "{status:?}"
            );
        }
    }

    #[test]
    fn a_neutral_done_carries_no_success_signal_and_is_not_scored() {
        // car-ir documents Done as "neutral -- may or may not have succeeded".
        // Scoring it either way would invent the number this crate measures.
        assert_eq!(
            outcome_utility(
                &outcome(OutcomeStatus::Done, &[EvidenceKind::ToolResult]),
                UtilityEvidence::Reported
            ),
            None
        );
    }

    #[test]
    fn grounded_evidence_rejects_a_success_backed_only_by_self_assessment() {
        let self_only = outcome(OutcomeStatus::Success, &[EvidenceKind::SelfAssessment]);
        assert_eq!(
            outcome_utility(&self_only, UtilityEvidence::Reported),
            Some(1.0)
        );
        assert_eq!(
            outcome_utility(&self_only, UtilityEvidence::Grounded),
            None,
            "unchecked is not failed — it must not be scored 0.0 either"
        );

        let backed = outcome(
            OutcomeStatus::Success,
            &[
                EvidenceKind::SelfAssessment,
                EvidenceKind::ExternalVerification,
            ],
        );
        assert_eq!(
            outcome_utility(&backed, UtilityEvidence::Grounded),
            Some(1.0)
        );
    }

    #[test]
    fn an_outcome_with_no_evidence_at_all_is_ungrounded() {
        let bare = outcome(OutcomeStatus::Success, &[]);
        assert_eq!(outcome_utility(&bare, UtilityEvidence::Reported), Some(1.0));
        assert_eq!(outcome_utility(&bare, UtilityEvidence::Grounded), None);
    }

    #[test]
    fn final_aggregation_takes_the_last_scored_stage() {
        let ev = &[EvidenceKind::ToolResult];
        let outputs = vec![
            output_with("a", Some((10, 10)), OutcomeStatus::Failure, ev),
            output_with("b", Some((10, 10)), OutcomeStatus::Success, ev),
        ];
        assert_eq!(
            run_utility(
                &outputs,
                UtilityAggregation::Final,
                UtilityEvidence::Reported
            ),
            Some(1.0)
        );
    }

    #[test]
    fn final_aggregation_skips_trailing_outputs_that_carry_no_signal() {
        let ev = &[EvidenceKind::ToolResult];
        let outputs = vec![
            output_with("a", Some((10, 10)), OutcomeStatus::Success, ev),
            // A neutral Done and an outcome-less aggregate must not displace it.
            output_with("done", Some((10, 10)), OutcomeStatus::Done, ev),
            output("aggregate", Some((10, 10))),
        ];
        assert_eq!(
            run_utility(
                &outputs,
                UtilityAggregation::Final,
                UtilityEvidence::Reported
            ),
            Some(1.0)
        );
    }

    #[test]
    fn consensus_aggregation_grades_independent_answers() {
        let ev = &[EvidenceKind::ToolResult];
        let outputs = vec![
            output_with("a", Some((10, 10)), OutcomeStatus::Success, ev),
            output_with("b", Some((10, 10)), OutcomeStatus::Success, ev),
            output_with("c", Some((10, 10)), OutcomeStatus::Failure, ev),
            output_with("d", Some((10, 10)), OutcomeStatus::Failure, ev),
        ];
        assert_eq!(
            run_utility(
                &outputs,
                UtilityAggregation::Consensus,
                UtilityEvidence::Reported
            ),
            Some(0.5)
        );
    }

    #[test]
    fn a_run_with_no_structured_outcomes_yields_no_utility() {
        let outputs = vec![output("a", Some((10, 10))), output("b", Some((10, 10)))];
        for aggregation in [UtilityAggregation::Final, UtilityAggregation::Consensus] {
            assert_eq!(
                run_utility(&outputs, aggregation, UtilityEvidence::Reported),
                None
            );
        }
    }

    #[test]
    fn the_derived_path_refuses_to_fabricate_either_measured_number() {
        let dir = tempfile::tempdir().unwrap();
        let path = car_topology::journal_path(dir.path());
        let mut journal = RecordJournal::open(&path).unwrap();
        let ev = &[EvidenceKind::ToolResult];

        let write = |journal: &mut RecordJournal, task: &str, outputs: &[AgentOutput]| {
            record_run_from_outcomes(
                journal,
                "mini-lm",
                task,
                vec![0.5, 0.5],
                CoordinationShape::Debate,
                4,
                outputs,
                UtilityAggregation::Final,
                UtilityEvidence::Grounded,
            )
            .unwrap()
        };

        // Metered and grounded -> recorded.
        assert!(write(
            &mut journal,
            "ok",
            &[output_with(
                "a",
                Some((400, 100)),
                OutcomeStatus::Success,
                ev
            )]
        ));
        // Grounded but unmetered -> no measured cost, so no record.
        assert!(!write(
            &mut journal,
            "unmetered",
            &[output_with("a", None, OutcomeStatus::Success, ev)]
        ));
        // Metered but only self-assessed -> no measured utility, so no record.
        assert!(!write(
            &mut journal,
            "ungrounded",
            &[output_with(
                "a",
                Some((400, 100)),
                OutcomeStatus::Success,
                &[EvidenceKind::SelfAssessment]
            )]
        ));
        // Metered but neutral -> likewise.
        assert!(!write(
            &mut journal,
            "neutral",
            &[output_with("a", Some((400, 100)), OutcomeStatus::Done, ev)]
        ));

        drop(journal);
        let entries = RecordJournal::load(&path).unwrap();
        assert_eq!(entries.len(), 1, "only the fully measured run is recorded");
        assert_eq!(entries[0].record.task_id, "ok");
        assert_eq!(entries[0].record.utility, 1.0);
    }

    #[test]
    fn record_run_writes_a_metered_run_and_skips_an_unmetered_one() {
        let dir = tempfile::tempdir().unwrap();
        let path = car_topology::journal_path(dir.path());
        let mut journal = RecordJournal::open(&path).unwrap();

        assert!(record_run(
            &mut journal,
            "mini-lm",
            "t1",
            vec![0.5, 0.5],
            CoordinationShape::Debate,
            4,
            &[output("a", Some((400, 100)))],
            1.0,
        )
        .unwrap());

        assert!(!record_run(
            &mut journal,
            "mini-lm",
            "t2",
            vec![0.5, 0.5],
            CoordinationShape::Pipeline,
            4,
            &[output("a", None)],
            1.0,
        )
        .unwrap());

        drop(journal);
        let entries = RecordJournal::load(&path).unwrap();
        assert_eq!(entries.len(), 1, "only the metered run is recorded");
        assert_eq!(entries[0].record.tokens, 500);
        assert_eq!(entries[0].embedder, "mini-lm");
    }

    /// End to end through the public surface: record a batch of runs, fit a
    /// selector on them, and get a shape back that a `car-multi` caller can
    /// execute.
    #[test]
    fn recorded_runs_fit_a_selector_that_returns_an_executable_shape() {
        let team_size = 4;
        let mut records = Vec::new();
        for i in 0..6 {
            let drift = i as f32 * 0.01;
            // Debate is cheap on "math"-flavoured queries, dear on "code" ones.
            for (family, query, cheap, dear) in [
                (
                    "math",
                    vec![1.0, 0.0, drift],
                    CoordinationShape::Debate,
                    CoordinationShape::Pipeline,
                ),
                (
                    "code",
                    vec![0.0, 1.0, drift],
                    CoordinationShape::Pipeline,
                    CoordinationShape::Debate,
                ),
            ] {
                let task = format!("{family}{i}");
                records.push(
                    execution_record(
                        &task,
                        query.clone(),
                        cheap,
                        team_size,
                        &[output("a", Some((400, 200)))],
                        1.0,
                    )
                    .unwrap()
                    .unwrap(),
                );
                records.push(
                    execution_record(
                        &task,
                        query,
                        dear,
                        team_size,
                        &[output("a", Some((1600, 800)))],
                        1.0,
                    )
                    .unwrap()
                    .unwrap(),
                );
            }
        }

        let selector = TopologySelector::fit(
            &RecordSet::new(records).unwrap(),
            &SelectorConfig::default(),
        )
        .unwrap();

        let (math, _) =
            select_shape(&selector, &[1.0, 0.0, 0.0], CoordinationShape::Swarm).unwrap();
        let (code, _) =
            select_shape(&selector, &[0.0, 1.0, 0.0], CoordinationShape::Swarm).unwrap();
        assert_eq!(math, CoordinationShape::Debate);
        assert_eq!(code, CoordinationShape::Pipeline);
    }
}