somatize_runtime/

executor.rs

//! Plan executor — walks [`ExecutionPlan`] trees and runs filter nodes.
//!
//! Handles sequential, parallel (scoped threads), cached, remote, loop,
//! and branch execution. Uses [`GraphInfo`] for topology-aware input resolution.

use crate::event_bus::EventBus;
use crate::node_catalog::{NodeCatalog, NodeImpl};
use somatize_compiler::ExecutionPlan;
use somatize_core::cache::CacheStore;
use somatize_core::control::{
    LoopCondition, LoopSignal, is_default_arm, read_arm_selector, read_loop_signal,
};
use somatize_core::error::{Result, SomaError};
use somatize_core::event::Event;
use somatize_core::node::NodeOutcome;
use somatize_core::store::DataStore;
use somatize_core::value::Value;
use somatize_core::virtual_value::VirtualValue;
use std::collections::HashMap;
use std::sync::Arc;
use std::time::Instant;

/// Graph topology information for input resolution.
///
/// Maps each node to its predecessor node IDs so the executor knows
/// where to read inputs from in the context store.
#[derive(Debug, Clone, Default)]
pub struct GraphInfo {
    /// node_id → list of predecessor node IDs
    predecessors: HashMap<String, Vec<String>>,
}

impl GraphInfo {
    /// An empty topology; every node resolves to no predecessors until
    /// [`Self::set_predecessors`] says otherwise.
    pub fn new() -> Self {
        Self::default()
    }

    /// Register predecessors for a node.
    pub fn set_predecessors(&mut self, node_id: impl Into<String>, preds: Vec<String>) {
        self.predecessors.insert(node_id.into(), preds);
    }

    /// Build GraphInfo from a somatize_core::graph::Graph.
    pub fn from_graph(graph: &somatize_core::graph::Graph) -> Self {
        let mut info = Self::new();
        for node in &graph.nodes {
            let preds: Vec<String> = graph
                .predecessors(&node.id)
                .into_iter()
                .map(|s| s.to_string())
                .collect();
            info.set_predecessors(node.id.clone(), preds);
        }
        info
    }

    /// Build GraphInfo for a linear pipeline (each node depends on the previous).
    pub fn for_linear(node_ids: &[&str]) -> Self {
        let mut info = Self::new();
        for (i, &id) in node_ids.iter().enumerate() {
            let preds = if i > 0 {
                vec![node_ids[i - 1].to_string()]
            } else {
                vec![]
            };
            info.set_predecessors(id, preds);
        }
        info
    }

    /// Get predecessors for a node.
    pub fn predecessors(&self, node_id: &str) -> &[String] {
        self.predecessors
            .get(node_id)
            .map(|v| v.as_slice())
            .unwrap_or(&[])
    }
}

/// What a run does to the nodes it visits.
///
/// Fitting and forwarding differ in exactly one way — whether a trainable
/// node learns its state before it computes — and in nothing else. They
/// used to be two whole execution loops: `run`/`forward` walked the plan
/// through `run_node`, while `fit` had a second, filter-only walk that
/// flattened the plan and re-implemented input resolution, events, caching
/// and panic handling. Making the difference a *value* is what lets both
/// go through one site.
#[derive(Clone, Debug, Default)]
pub enum RunMode {
    /// Every node computes with the state it already has.
    #[default]
    Forward,
    /// A trainable node learns its state from its resolved input and these
    /// labels, then computes with it. Labels are shared by the whole run;
    /// `None` means unsupervised.
    Fit {
        /// The run's labels; `None` means unsupervised.
        y: Option<Value>,
    },
}

impl RunMode {
    /// The labels, if this is a fit.
    fn labels(&self) -> Option<&Value> {
        match self {
            Self::Forward => None,
            Self::Fit { y } => y.as_ref(),
        }
    }

    fn is_fit(&self) -> bool {
        matches!(self, Self::Fit { .. })
    }
}

/// Execution context passed to filters during runtime.
///
/// Node outputs are stored as [`VirtualValue`]s — they may be materialized
/// in memory, cached on disk, or deferred (not yet computed). The executor
/// resolves them on demand when a downstream node needs the data.
pub struct Context {
    /// Fit or forward. See [`RunMode`].
    pub mode: RunMode,
    /// Node outputs as virtual values (may be lazy).
    ///
    /// Private, together with `execution_order`: the two are a pair.
    /// `execute_parallel` works out what a branch contributed by diffing
    /// `execution_order`, so a write that reached one and not the other
    /// is silently dropped at the join. Going through [`Context::set`]
    /// and [`Context::set_virtual`] is what keeps them in step.
    store: HashMap<String, VirtualValue>,
    /// Event bus for emitting runtime events.
    pub event_bus: Arc<EventBus>,
    /// Current run ID.
    pub run_id: String,
    /// Track execution order. Private for the reason above.
    execution_order: Vec<String>,
    /// Graph topology for input resolution.
    pub graph_info: GraphInfo,
    /// Optional transport for distributed plans.
    pub transport: Option<Arc<dyn crate::runner::Transport>>,
    /// Optional data store for persisting intermediate results.
    pub data_store: Option<Arc<dyn DataStore>>,
    /// Minimum value size (bytes) to spill to DataStore instead of keeping in memory.
    /// Default: 0 (disabled — all values stay in memory).
    pub spill_threshold: usize,
    /// Memoized content hashes of node outputs, keyed by node id.
    /// Invalidated whenever a node's output is (re)stored, so Loop
    /// iterations that overwrite an output never reuse a stale hash.
    output_hashes: HashMap<String, somatize_core::cache::CacheKey>,
    /// Experiment seed for this run. Hashed into every cache key so
    /// each seed owns an independent cache line (a 5-seed study is 5
    /// resumable computations, not one).
    pub seed: Option<i64>,
    /// Performs and journals step effects. Only needed when the plan
    /// contains a step; a purely computational graph leaves it unset.
    ///
    /// The steps themselves are not here: they live in the same
    /// [`NodeCatalog`] as the filters,
    /// which the executor already receives. Keeping a second registry in
    /// the context is what let the branch arm decide a node's kind by
    /// asking whether it happened to be in it.
    pub driver: Option<crate::effects::EffectDriver>,
}

impl Context {
    /// A forward-mode context with empty topology and no optional
    /// components; the `with_*` builders add what the run needs.
    pub fn new(event_bus: Arc<EventBus>, run_id: impl Into<String>) -> Self {
        Self {
            mode: RunMode::Forward,
            store: HashMap::new(),
            event_bus,
            run_id: run_id.into(),
            execution_order: Vec::new(),
            graph_info: GraphInfo::new(),
            transport: None,
            data_store: None,
            spill_threshold: 0,
            output_hashes: HashMap::new(),
            seed: None,
            driver: None,
        }
    }

    /// Register the effect driver an effectful plan needs.
    ///
    /// The driver should already carry its catalog
    /// ([`crate::effects::EffectDriver::with_catalog`]) if a step may fan
    /// out dynamically — whoever builds the driver knows which catalog it
    /// serves; the context does not.
    pub fn with_driver(mut self, driver: crate::effects::EffectDriver) -> Self {
        self.driver = Some(driver);
        self
    }

    /// Set the topology used for input resolution.
    pub fn with_graph_info(mut self, info: GraphInfo) -> Self {
        self.graph_info = info;
        self
    }

    /// Make this a fit: trainable nodes learn from `y` before computing.
    pub fn fitting(mut self, y: Option<Value>) -> Self {
        self.mode = RunMode::Fit { y };
        self
    }

    /// Record a state a node just learned.
    ///
    /// Stored under the same `__state_{id}` key the worker and the session
    /// already read, and appended to `execution_order` like any other
    /// write: that list is how `execute_parallel` works out what a branch
    /// contributed, so a state written inside a branch that skipped it
    /// would be dropped at the join. Readers asking "which node ran last"
    /// filter reserved keys out — see [`somatize_core::keys::is_reserved`].
    pub fn record_state(&mut self, node_id: &str, state: Value) {
        self.set(somatize_core::keys::state_key(node_id), state);
    }

    /// Set the experiment seed (hashed into every cache key).
    pub fn with_seed(mut self, seed: Option<i64>) -> Self {
        self.seed = seed;
        self
    }

    /// Set the transport a plan with `Remote` nodes executes through.
    pub fn with_transport(mut self, transport: Arc<dyn crate::runner::Transport>) -> Self {
        self.transport = Some(transport);
        self
    }

    /// Set the data store used for spilling and remote data movement.
    pub fn with_data_store(mut self, store: Arc<dyn DataStore>) -> Self {
        self.data_store = Some(store);
        self
    }

    /// Set spill threshold: values larger than this (in bytes) are offloaded
    /// to the DataStore and replaced with a VirtualValue::Cached reference.
    /// Requires a DataStore to be set via `with_data_store()`.
    pub fn with_spill_threshold(mut self, bytes: usize) -> Self {
        self.spill_threshold = bytes;
        self
    }

    /// If a DataStore and spill threshold are configured, check if the value
    /// should be offloaded. Returns VirtualValue (materialized or cached ref).
    fn maybe_spill(&self, node_id: &str, value: Value) -> VirtualValue {
        if self.spill_threshold > 0
            && let Some(store) = &self.data_store
        {
            let size = value.size() * 8; // approximate bytes (f64 = 8 bytes)
            if size >= self.spill_threshold {
                let key = somatize_core::cache::CacheKey::from_parts(&[
                    self.run_id.as_bytes(),
                    node_id.as_bytes(),
                ]);
                let vv_for_schema = VirtualValue::materialized(value.clone());
                let schema = vv_for_schema.schema().clone();
                if let Ok(_data_ref) = store.put(&key, &value) {
                    tracing::debug!("spilled node `{node_id}` ({size} bytes) to DataStore");
                    return VirtualValue::cached(key, schema);
                }
            }
        }
        VirtualValue::materialized(value)
    }

    /// The nodes that ran, in the order they ran.
    ///
    /// Includes the run's reserved keys (see [`somatize_core::keys`]);
    /// filter them out with `keys::is_reserved` if you want node ids only.
    pub fn execution_order(&self) -> &[String] {
        &self.execution_order
    }

    /// Every materialized value this run produced, keyed by node id.
    ///
    /// Consumes the context, because the point of asking is that the run
    /// is over. Lazy values that were never resolved are skipped.
    pub fn into_outputs(self) -> HashMap<String, Value> {
        self.store
            .into_iter()
            .filter_map(|(k, vv)| vv.as_value().cloned().map(|v| (k, v)))
            .collect()
    }

    /// Get the materialized Value for a node, if present and materialized.
    pub fn get(&self, node_id: &str) -> Option<&Value> {
        self.store.get(node_id).and_then(|vv| vv.as_value())
    }

    /// Get the raw VirtualValue for a node.
    pub fn get_virtual(&self, node_id: &str) -> Option<&VirtualValue> {
        self.store.get(node_id)
    }

    /// Store a materialized value for a node.
    pub fn set(&mut self, node_id: impl Into<String>, value: Value) {
        let id = node_id.into();
        self.execution_order.push(id.clone());
        self.output_hashes.remove(&id);
        self.store.insert(id, VirtualValue::materialized(value));
    }

    /// Store a virtual value (which may be deferred or cached).
    pub fn set_virtual(&mut self, node_id: impl Into<String>, vv: VirtualValue) {
        let id = node_id.into();
        self.execution_order.push(id.clone());
        self.output_hashes.remove(&id);
        self.store.insert(id, vv);
    }

    /// Content hash of a node's resolved input, memoized through the
    /// single-predecessor fast path: the input of a 1-pred node IS that
    /// predecessor's output, so sibling consumers (diamonds) reuse the
    /// hash instead of re-serializing a potentially large value.
    fn input_hash(&mut self, node_id: &str, input: &Value) -> somatize_core::cache::CacheKey {
        let preds = self.graph_info.predecessors(node_id);
        let single_pred = match preds {
            [only] => Some(only.clone()),
            _ => None,
        };
        if let Some(pred) = single_pred {
            if let Some(h) = self.output_hashes.get(&pred) {
                return h.clone();
            }
            // Only trust the memo association when the pred's output is
            // actually what resolve_input handed us (it may have fallen
            // back to Value::Empty if the pred produced nothing).
            if self.store.contains_key(&pred) {
                let h = somatize_core::cache::CacheKey::for_value(input);
                self.output_hashes.insert(pred, h.clone());
                return h;
            }
        }
        somatize_core::cache::CacheKey::for_value(input)
    }

    fn snapshot(&self) -> Self {
        Self {
            mode: self.mode.clone(),
            store: self.store.clone(),
            event_bus: self.event_bus.clone(),
            run_id: self.run_id.clone(),
            execution_order: self.execution_order.clone(),
            graph_info: self.graph_info.clone(),
            transport: self.transport.clone(),
            data_store: self.data_store.clone(),
            spill_threshold: self.spill_threshold,
            output_hashes: self.output_hashes.clone(),
            seed: self.seed,
            driver: self.driver.clone(),
        }
    }
}

/// Execute a compiled plan.
///
/// Every arm delegates. The variants that need real work — a node, a loop,
/// a branch, a fan-out — each have a function, so this reads as the list of
/// things a plan can be rather than as their implementations.
pub fn execute(
    plan: &ExecutionPlan,
    ctx: &mut Context,
    catalog: &NodeCatalog,
    cache: &dyn CacheStore,
) -> Result<()> {
    match plan {
        ExecutionPlan::Empty => Ok(()),

        // One call for both: a filter simply declares no handoffs.
        ExecutionPlan::Execute { node_id } => execute_node(node_id, &[], ctx, catalog, cache),

        ExecutionPlan::Step { node_id, handoffs } => {
            execute_node(node_id, handoffs, ctx, catalog, cache)
        }

        ExecutionPlan::Sequence(steps) => {
            for step in steps {
                execute(step, ctx, catalog, cache)?;
            }
            Ok(())
        }

        ExecutionPlan::Parallel(branches) => execute_parallel(branches, ctx, catalog, cache),

        ExecutionPlan::Loop {
            node_id,
            body,
            max_iterations,
            until,
            carry_from,
        } => execute_loop(
            node_id,
            body,
            *max_iterations,
            until,
            carry_from.as_deref(),
            ctx,
            catalog,
            cache,
        ),

        ExecutionPlan::Branch { node_id, arms } => {
            execute_branch(node_id, arms, ctx, catalog, cache)
        }

        ExecutionPlan::Remote {
            node_id,
            target: _,
            plan,
        } => execute_remote(node_id, plan, ctx, catalog, cache),

        ExecutionPlan::Composite { node_ids } => {
            // A composite block exists so that a set of differentiable
            // filters can be fitted together in one process, with tensors
            // passed directly and autograd intact. That is a property of
            // the *block*, not of any node in it, which is why it is
            // handled here and not inside `run_node`.
            if ctx.mode.is_fit() && composite_fit(node_ids, ctx, catalog)? {
                return Ok(());
            }
            // Otherwise, and always when forwarding: each node in order.
            for nid in node_ids {
                execute_node(nid, &[], ctx, catalog, cache)?;
            }
            Ok(())
        }

        ExecutionPlan::Stream {
            node_ids,
            chunk_size,
        } => execute_stream(node_ids, *chunk_size, ctx, catalog, cache),

        // `ExecutionPlan` is `#[non_exhaustive]`, so this arm is reachable
        // from a plan built by a newer compiler — deserialized from a
        // worker, say. It used to `warn!` and return `Ok(())`: a plan the
        // runtime did not understand was reported as having run, naming
        // neither the variant nor the node.
        other => Err(SomaError::Execution {
            node_id: other
                .node_ids()
                .first()
                .map_or_else(|| "<plan>".to_string(), |id| (*id).to_string()),
            message: format!(
                "this runtime does not know how to execute `{other:?}`. It was \
                 probably compiled by a newer version"
            ),
        }),
    }
}

/// Iterate `body` until the condition says stop, or the count runs out.
#[allow(clippy::too_many_arguments)]
fn execute_loop(
    node_id: &str,
    body: &ExecutionPlan,
    max_iterations: Option<usize>,
    until: &LoopCondition,
    carry_from: Option<&str>,
    ctx: &mut Context,
    catalog: &NodeCatalog,
    cache: &dyn CacheStore,
) -> Result<()> {
    let max = max_iterations.unwrap_or(100);
    let mut ran = 0usize;

    // The loop node's value is its *carry*: what the body reads on each
    // pass. A body entry's only predecessor is the loop node itself (that
    // control edge is what makes it the body), so without seeding this the
    // first iteration would run on `Empty`. Seed it with the loop's own
    // input; after each iteration the condition node's output replaces it,
    // which is what makes a refine loop actually refine rather than
    // redraft the same thing N times.
    let seed = resolve_input(node_id, ctx);
    ctx.set(node_id.to_string(), seed);

    for i in 0..max {
        execute(body, ctx, catalog, cache)?;
        ran = i + 1;

        // Advance the carry before testing the condition: even a loop with
        // no stop signal has to move forward, or every pass repeats the
        // first one.
        if let Some(source) = carry_from
            && let Some(value) = ctx.get(source).cloned()
        {
            ctx.set(node_id.to_string(), value);
        }

        // Termination is read from the node the compiler resolved, never
        // from whichever node happened to run last — with a parallel body
        // "last" is a race.
        let LoopCondition::WhenSignaled(cond_node) = until else {
            continue; // Exhaust: always run the full count
        };

        let value = ctx.get(cond_node).ok_or_else(|| SomaError::Execution {
            node_id: node_id.to_string(),
            message: format!(
                "loop condition node `{cond_node}` produced no output on iteration {ran}"
            ),
        })?;

        let signal = read_loop_signal(value).ok_or_else(|| SomaError::Execution {
            node_id: node_id.to_string(),
            message: format!(
                "loop condition node `{cond_node}` produced `{}`, which carries no \
                 termination signal. Return a bool, \"done\"/\"stop\", or \
                 {{\"done\": bool}}",
                value.type_name()
            ),
        })?;

        if signal == LoopSignal::Stop {
            emit_control_completed(ctx, node_id, format!("Loop terminated at iteration {ran}"));
            return Ok(());
        }
    }

    emit_control_completed(ctx, node_id, format!("Loop exhausted {ran} iterations"));
    Ok(())
}

/// Run the condition node, then the one arm it names.
fn execute_branch(
    node_id: &str,
    arms: &[(String, ExecutionPlan)],
    ctx: &mut Context,
    catalog: &NodeCatalog,
    cache: &dyn CacheStore,
) -> Result<()> {
    // The condition node may be an ordinary filter or an effectful step:
    // an LLM deciding where a request goes is the routing case agentic
    // graphs are built for. Which it is no longer needs asking — one call
    // runs either, and the outcome says how the arm was chosen.
    let request = resolve_input(node_id, ctx);

    let selector = match run_node(node_id, ctx, catalog, cache)? {
        // A routing step names its arm directly. This used to be
        // unreachable: the branch called the step with no handoffs — its
        // control edges having been consumed as the branch's arms — so a
        // `Goto` always errored.
        NodeOutcome::HandOff { target, .. } => target,

        NodeOutcome::Produced(condition) => {
            read_arm_selector(&condition).ok_or_else(|| SomaError::Execution {
                node_id: node_id.to_string(),
                message: format!(
                    "branch condition produced `{}`, which names no arm. Return the \
                     arm's label as a string, a bool, or {{\"branch\": \"<label>\"}}",
                    condition.type_name()
                ),
            })?
        }

        NodeOutcome::Paused { turn, reason } => {
            return Err(SomaError::Suspended {
                run_id: ctx.run_id.clone(),
                node_id: node_id.to_string(),
                turn,
                reason: Box::new(reason),
            });
        }
    };

    let (label, plan) = arms
        .iter()
        .find(|(label, _)| label == &selector)
        .or_else(|| arms.iter().find(|(label, _)| is_default_arm(label)))
        .ok_or_else(|| SomaError::Execution {
            node_id: node_id.to_string(),
            message: format!(
                "branch selected `{selector}`, which matches no arm ({}) and there is \
                 no `default` arm",
                arms.iter()
                    .map(|(l, _)| l.as_str())
                    .collect::<Vec<_>>()
                    .join(", ")
            ),
        })?;

    emit_control_completed(ctx, node_id, format!("Branch selected: {label}"));

    // The selector is control, not data. An arm's only predecessor is the
    // branch node, so leaving the label there would hand the chosen agent
    // the string "billing" instead of the customer's question — the
    // handoff-context loss the multi-agent failure literature keeps
    // finding. The branch passes its input through instead; put a filter
    // *before* it if the request needs transforming.
    ctx.set(node_id.to_string(), request);
    execute(plan, ctx, catalog, cache)
}

/// Hand a node to a worker, or run it here if there is no transport.
fn execute_remote(
    node_id: &str,
    plan: &ExecutionPlan,
    ctx: &mut Context,
    catalog: &NodeCatalog,
    cache: &dyn CacheStore,
) -> Result<()> {
    let Some(transport) = ctx.transport.clone() else {
        return execute(plan, ctx, catalog, cache);
    };
    let input = ctx
        .graph_info
        .predecessors(node_id)
        .first()
        .and_then(|pred| ctx.get(pred));
    let result = transport.execute_node(node_id, input)?;
    ctx.set(node_id.to_string(), result);
    Ok(())
}

/// A control-flow construct finishing. It has no duration of its own —
/// the time is in the nodes it ran.
fn emit_control_completed(ctx: &Context, node_id: &str, summary: String) {
    ctx.event_bus.emit(Event::NodeCompleted {
        run_id: ctx.run_id.clone(),
        node_id: node_id.to_string(),
        duration: std::time::Duration::ZERO,
        output_summary: summary,
    });
}

/// Salt a cache key with the run's experiment seed, when set.
/// `None` leaves the key untouched (and distinct from any seeded key).
pub(crate) fn salt_with_seed(
    key: somatize_core::cache::CacheKey,
    seed: Option<i64>,
) -> somatize_core::cache::CacheKey {
    match seed {
        Some(s) => somatize_core::cache::CacheKey::from_parts(&[b"seed", &s.to_le_bytes(), &key.0]),
        None => key,
    }
}

/// What a panic payload says, when it says anything at all.
///
/// `panic!("...")` with arguments produces a `String`; a bare literal
/// produces a `&str`. Anything else — `panic_any` with a custom type —
/// carries no message we can read.
pub(crate) fn panic_message(payload: &(dyn std::any::Any + Send)) -> &str {
    payload
        .downcast_ref::<String>()
        .map(|s| s.as_str())
        .or_else(|| payload.downcast_ref::<&str>().copied())
        .unwrap_or("unknown panic")
}

// ── The primitives every execution path shares ──
//
// `run_node` composes these for the topological walk; the stream driver
// composes the same three per chunk. Anything that must be true of every
// node execution — the memoization guard, the one key derivation, panic
// containment, provenance on writes — lives here and nowhere else.

/// The node's output key, or `None` when it must not be memoized.
///
/// The single owner of the `cacheable && deterministic` guard and of the
/// derivation `hash(config + state + input)`, salted with the run seed.
/// Nondeterministic forwards are excluded because serving a recorded
/// result would silently freeze what the user expects to vary.
pub(crate) fn output_key(
    node: &NodeImpl,
    meta: &somatize_core::node::NodeMeta,
    state: &Value,
    input_key: &somatize_core::cache::CacheKey,
    seed: Option<i64>,
) -> Option<somatize_core::cache::CacheKey> {
    if !(meta.cacheable && meta.deterministic) {
        return None;
    }
    let key = somatize_core::cache::CacheKey::for_output(
        &node.config_hash(),
        &somatize_core::cache::CacheKey::for_value(state),
        input_key,
    );
    Some(salt_with_seed(key, seed))
}

/// Run the node's own computation, containing any panic.
///
/// The `catch_unwind` around [`run_node_inner`] — a panic in user code
/// (a Python filter or step alike) must not crash the process.
pub(crate) fn compute_node(
    node: &NodeImpl,
    node_id: &str,
    ctx: &Context,
    input: &Value,
    state: &Value,
) -> Result<NodeOutcome> {
    let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
        run_node_inner(node, node_id, ctx, input, state)
    }));
    match result {
        Ok(inner) => inner,
        Err(panic) => {
            let msg = panic_message(&*panic);
            tracing::error!(node_id, "node panicked: {msg}");
            Err(SomaError::Execution {
                node_id: node_id.to_string(),
                message: format!("node panicked: {msg}"),
            })
        }
    }
}

/// Store a computed output with its provenance. Best-effort: a full
/// cache disk must never fail the run.
pub(crate) fn store_output(
    cache: &dyn CacheStore,
    key: &somatize_core::cache::CacheKey,
    output: &Value,
    node_id: &str,
    run_id: &str,
    duration: std::time::Duration,
    deterministic: bool,
) {
    let origin = somatize_core::cache::Origin::Computed {
        node_id: node_id.to_string(),
        run_id: run_id.to_string(),
    };
    if let Err(e) = cache.put_computed(key, output, &origin, duration, deterministic) {
        tracing::warn!(node_id, error = %e, "failed to cache node output");
    }
}

/// Run one node and act on how it finished.
///
/// The single entry point for both kinds. Everything that does not depend
/// on which kind ran — resolving the input, the output cache, containing a
/// panic, the start/complete/fail events — happens once, in [`run_node`].
/// What is left here is the control flow only a step can ask for.
fn execute_node(
    node_id: &str,
    handoffs: &[(String, ExecutionPlan)],
    ctx: &mut Context,
    catalog: &NodeCatalog,
    cache: &dyn CacheStore,
) -> Result<()> {
    match run_node(node_id, ctx, catalog, cache)? {
        NodeOutcome::Produced(_) => Ok(()),

        // A handoff: the node is finished and names who continues.
        NodeOutcome::HandOff { target, .. } => {
            let plan = select_handoff(node_id, &target, handoffs)?;
            execute(plan, ctx, catalog, cache)
        }

        // Not a failure: the run paused. It travels as an error so the
        // rest of the plan does not execute, and carries everything the
        // caller needs to answer and resume.
        NodeOutcome::Paused { turn, reason } => Err(SomaError::Suspended {
            run_id: ctx.run_id.clone(),
            node_id: node_id.to_string(),
            turn,
            reason: Box::new(reason),
        }),
    }
}

/// Which sub-plan a handoff target names.
fn select_handoff<'p>(
    node_id: &str,
    target: &str,
    handoffs: &'p [(String, ExecutionPlan)],
) -> Result<&'p ExecutionPlan> {
    handoffs
        .iter()
        .find(|(t, _)| t == target)
        .map(|(_, p)| p)
        .ok_or_else(|| SomaError::Execution {
            node_id: node_id.to_string(),
            message: if handoffs.is_empty() {
                format!(
                    "step handed control to `{target}`, but it declares no \
                     handoffs. Add a control edge from `{node_id}` to `{target}`"
                )
            } else {
                format!(
                    "step handed control to `{target}`, which is not among its \
                     declared handoffs ({})",
                    handoffs
                        .iter()
                        .map(|(t, _)| t.as_str())
                        .collect::<Vec<_>>()
                        .join(", ")
                )
            },
        })
}

/// Everything running a node involves that is the same for both kinds.
///
/// Cacheable nodes are memoized: the output key is
/// `hash(config + state + input)` — if a previous run (possibly in another
/// process, via a persistent cache) already computed this exact forward,
/// the stored output is used and the node never runs.
///
/// A step never reaches that path, and not because of a check here: its
/// [`NodeMeta`](somatize_core::node::NodeMeta) declares
/// `cacheable: false`, so the guard below skips it the way it skips a
/// filter that declared the same. What makes a step re-runnable instead is
/// the effect journal, which the driver consults per effect.
///
/// The produced value — or a handoff's carry — is stored under `node_id`
/// before returning, so a successor resolves it as an ordinary predecessor
/// output.
fn run_node(
    node_id: &str,
    ctx: &mut Context,
    catalog: &NodeCatalog,
    cache: &dyn CacheStore,
) -> Result<NodeOutcome> {
    let start = Instant::now();

    let node = catalog
        .node(node_id)
        .ok_or_else(|| SomaError::NodeNotFound(node_id.to_string()))?
        .clone();
    let meta = node.meta();

    let _span = tracing::info_span!("run_node", %node_id).entered();

    let input = resolve_input(node_id, ctx);

    // In a fit, a trainable node learns before it computes. Everything
    // after this point is identical to a forward — which is the whole
    // reason fit no longer needs a walk of its own.
    let fitted = fit_state_if_needed(node_id, &node, &meta, &input, ctx, cache)?;

    // Borrow state via Arc — cloning the inner Value here would deep-copy
    // potentially huge tensors (encoder outputs, model weights) on every
    // forward call. Arc::clone is a cheap atomic increment.
    let state = catalog.get_state(node_id);
    let state_ref: &Value = fitted
        .as_ref()
        .or(state.as_deref())
        .unwrap_or(&Value::Empty);

    // Nondeterministic forwards are excluded by `output_key` — their fit
    // STATES still cache: any recorded training result is acceptable,
    // constructive-trace semantics.
    let out_key = output_key(
        &node,
        &meta,
        state_ref,
        &ctx.input_hash(node_id, &input),
        ctx.seed,
    );

    // `get_located`, not `get`: which tier served the value is the whole
    // content of this event, and hardcoding `Memory` made every per-tier
    // statistic report the same thing.
    if let Some(key) = &out_key
        && let Ok(Some((cached, tier))) = cache.get_located(key)
    {
        ctx.set(node_id.to_string(), cached.clone());
        ctx.event_bus.emit(Event::NodeCacheHit {
            run_id: ctx.run_id.clone(),
            node_id: node_id.to_string(),
            key: key.clone(),
            tier,
            load_time: start.elapsed(),
        });
        return Ok(NodeOutcome::Produced(cached));
    }

    if let Some(key) = &out_key {
        ctx.event_bus.emit(Event::NodeCacheMiss {
            run_id: ctx.run_id.clone(),
            node_id: node_id.to_string(),
            key: key.clone(),
        });
    }

    ctx.event_bus.emit(Event::NodeStarted {
        run_id: ctx.run_id.clone(),
        node_id: node_id.to_string(),
        kind: meta.kind,
        effectful: meta.effectful,
    });

    let outcome = match compute_node(&node, node_id, ctx, &input, state_ref) {
        Ok(outcome) => outcome,
        Err(e) => {
            tracing::error!(node_id, error = %e, "node execution failed");
            ctx.event_bus.emit(Event::NodeFailed {
                run_id: ctx.run_id.clone(),
                node_id: node_id.to_string(),
                error: e.to_string(),
            });
            return Err(e);
        }
    };

    let duration = start.elapsed();
    match &outcome {
        NodeOutcome::Produced(output) => {
            let summary = format!("{output}");
            if let Some(key) = &out_key {
                store_output(
                    cache,
                    key,
                    output,
                    node_id,
                    &ctx.run_id,
                    duration,
                    meta.deterministic,
                );
            }
            let vv = ctx.maybe_spill(node_id, output.clone());
            ctx.set_virtual(node_id, vv);
            ctx.event_bus.emit(Event::NodeCompleted {
                run_id: ctx.run_id.clone(),
                node_id: node_id.to_string(),
                duration,
                output_summary: summary,
            });
        }

        // The carried value is stored under *this* node, so the target
        // resolves it as an ordinary predecessor output — no special path.
        NodeOutcome::HandOff { target, carry } => {
            ctx.set(node_id, carry.clone());
            ctx.event_bus.emit(Event::NodeCompleted {
                run_id: ctx.run_id.clone(),
                node_id: node_id.to_string(),
                duration,
                output_summary: format!("handed off to {target}"),
            });
        }

        // Nothing to store: the node did not finish.
        NodeOutcome::Paused { .. } => {}
    }

    Ok(outcome)
}

/// Fit a whole composite block through the first filter's `composite_fit`.
///
/// `Ok(false)` means the block was not fitted as a block — a node is
/// missing, or the filter declined — and the caller runs the nodes one by
/// one instead. `Ok(true)` means the results are already stored.
fn composite_fit(node_ids: &[String], ctx: &mut Context, catalog: &NodeCatalog) -> Result<bool> {
    let Some(first) = node_ids.first() else {
        return Ok(false);
    };
    // A Composite block is built by the compiler from differentiable
    // filters. A step inside one is a broken plan, and falling back to
    // one-by-one execution would paper over it.
    if let Some(step_id) = node_ids.iter().find(|id| catalog.step(id).is_some()) {
        return Err(SomaError::Execution {
            node_id: step_id.to_string(),
            message: "a Composite block contains a step; composite fit is defined \
                      only over differentiable filters"
                .into(),
        });
    }
    let peers: Option<Vec<(String, Arc<dyn somatize_core::filter::Filter>)>> = node_ids
        .iter()
        .map(|id| catalog.get(id).map(|f| (id.clone(), f)))
        .collect();
    let (Some(peers), Some(filter)) = (peers, catalog.get(first)) else {
        return Ok(false);
    };

    let input = resolve_input(first, ctx);
    let y = ctx.mode.labels().cloned();
    let Some(result) = filter.composite_fit(&peers, &input, y.as_ref()) else {
        return Ok(false);
    };
    let (output, states) = result?;

    for (id, state) in states {
        ctx.record_state(&id, state);
    }
    if let Some(last) = node_ids.last() {
        ctx.set(last.clone(), output);
    }
    Ok(true)
}

/// Learn this node's state, if the run is a fit and the node is trainable.
///
/// Returns the fitted state, or `None` when there is nothing to fit — a
/// forward run, a stateless or library-state filter, or a step (a step's
/// re-run semantics belong to the journal, not to a state cache).
///
/// The state cache key includes the labels on purpose: the same features
/// trained against different labels must not collide, and it is salted with
/// the run seed so a 5-seed study is five independent computations rather
/// than one recorded five times.
fn fit_state_if_needed(
    node_id: &str,
    node: &NodeImpl,
    meta: &somatize_core::node::NodeMeta,
    input: &Value,
    ctx: &mut Context,
    cache: &dyn CacheStore,
) -> Result<Option<Value>> {
    if !ctx.mode.is_fit() || !meta.trainable() {
        return Ok(None);
    }
    // The metadata already said "trainable", which a step's meta never
    // does — this extraction is structural, not a second decision.
    let NodeImpl::Filter(filter) = node else {
        return Ok(None);
    };

    let y = ctx.mode.labels().cloned();
    let key = salt_with_seed(
        somatize_core::cache::CacheKey::for_state(
            &filter.config_hash(),
            &somatize_core::cache::CacheKey::for_value(input),
            y.as_ref()
                .map(somatize_core::cache::CacheKey::for_value)
                .as_ref(),
        ),
        ctx.seed,
    );

    let state = match cache.get(&key)? {
        Some(cached) => cached,
        None => {
            let start = Instant::now();
            let learned = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
                filter.fit(input, y.as_ref())
            }))
            .map_err(|panic| SomaError::Execution {
                node_id: node_id.to_string(),
                message: format!("fit panicked: {}", panic_message(&*panic)),
            })??;
            let origin = somatize_core::cache::Origin::Computed {
                node_id: node_id.to_string(),
                run_id: ctx.run_id.clone(),
            };
            if let Err(e) = cache.put_computed(&key, &learned, &origin, start.elapsed(), true) {
                tracing::warn!(node_id, error = %e, "failed to cache fitted state");
            }
            learned
        }
    };

    ctx.record_state(node_id, state.clone());
    Ok(Some(state))
}

/// The only place in the runtime that knows a filter from a step.
fn run_node_inner(
    node: &NodeImpl,
    node_id: &str,
    ctx: &Context,
    input: &Value,
    state: &Value,
) -> Result<NodeOutcome> {
    match node {
        NodeImpl::Filter(filter) => filter.forward(input, state).map(NodeOutcome::Produced),

        NodeImpl::Step(step) => {
            let driver = ctx.driver.as_ref().ok_or_else(|| SomaError::Execution {
                node_id: node_id.to_string(),
                message: "the plan contains a step but no effect driver was registered; \
                          build the context with `with_driver(...)`"
                    .into(),
            })?;
            driver.run(step.as_ref(), &ctx.run_id, node_id, input)
        }
    }
}

/// Execute parallel branches concurrently using std::thread::scope.
///
/// Each branch gets a snapshot of the context. After all branches complete,
/// their new outputs are merged back into the main context.
fn execute_parallel(
    branches: &[ExecutionPlan],
    ctx: &mut Context,
    catalog: &NodeCatalog,
    cache: &dyn CacheStore,
) -> Result<()> {
    // What a branch contributes is what it *wrote*, not what happens to be
    // absent from the parent. Those differ the second time a parallel block
    // runs: inside a `Loop`, every body node already has a value from the
    // previous iteration, so filtering by "not already present" discarded
    // every fresh result and left downstream fan-in reading iteration one
    // forever — the nodes re-ran and their new outputs went nowhere.
    //
    // `execution_order` is appended to by every `set`/`set_virtual`, and a
    // snapshot copies it, so whatever a branch appended past this mark is
    // exactly its write set.
    let order_mark = ctx.execution_order.len();

    // Use scoped threads for true parallelism without Send requirements
    let results: Vec<Result<Vec<(String, VirtualValue)>>> = std::thread::scope(|s| {
        let handles: Vec<_> = branches
            .iter()
            .map(|branch| {
                let mut branch_ctx = ctx.snapshot();
                s.spawn(move || {
                    execute(branch, &mut branch_ctx, catalog, cache)?;
                    let written: std::collections::HashSet<&String> =
                        branch_ctx.execution_order[order_mark..].iter().collect();
                    let new_entries: Vec<(String, VirtualValue)> = written
                        .into_iter()
                        .filter_map(|k| branch_ctx.store.get(k).map(|v| (k.clone(), v.clone())))
                        .collect();
                    Ok(new_entries)
                })
            })
            .collect();

        // A branch thread that panicked comes back as an Err from `join`.
        // Unwrapping it here re-panics on the *parent* thread, which
        // aborts the process and undoes the `catch_unwind` that
        // `execute_node` installs precisely so a user filter cannot.
        handles
            .into_iter()
            .map(|h| match h.join() {
                Ok(result) => result,
                Err(panic) => {
                    let msg = panic_message(&*panic);
                    tracing::error!("parallel branch panicked: {msg}");
                    Err(SomaError::Execution {
                        node_id: "<parallel branch>".to_string(),
                        message: format!("parallel branch panicked: {msg}"),
                    })
                }
            })
            .collect()
    });

    // Merge results and propagate first error
    for result in results {
        let entries = result?;
        for (key, vv) in entries {
            ctx.set_virtual(key, vv);
        }
    }

    Ok(())
}

/// Resolve a VirtualValue to a concrete Value, loading from DataStore if needed.
fn resolve_value(vv: &VirtualValue, data_store: &Option<Arc<dyn DataStore>>) -> Option<Value> {
    match vv {
        VirtualValue::Materialized { value, .. } => Some(value.clone()),
        VirtualValue::Cached { key, .. } => {
            // Try to load from DataStore
            if let Some(store) = data_store {
                let data_ref = somatize_core::store::DataRef::Cached {
                    cache_key: key.clone(),
                };
                store.get(&data_ref).ok()
            } else {
                None
            }
        }
        _ => None,
    }
}

/// Resolve the input for a node from the context store using graph topology.
/// If a predecessor was spilled to DataStore, loads it back.
pub(crate) fn resolve_input(node_id: &str, ctx: &Context) -> Value {
    let preds = ctx.graph_info.predecessors(node_id);

    let resolve_node = |id: &str| -> Option<Value> {
        ctx.store
            .get(id)
            .and_then(|vv| resolve_value(vv, &ctx.data_store))
    };

    match preds.len() {
        0 => ctx
            .execution_order
            .last()
            .and_then(|id| resolve_node(id))
            .unwrap_or(Value::Empty),
        1 => resolve_node(&preds[0]).unwrap_or(Value::Empty),
        _ => {
            let mut merged = serde_json::Map::new();
            for pred_id in preds {
                if let Some(val) = resolve_node(pred_id) {
                    let json_val = val.to_plain_json();
                    merged.insert(pred_id.clone(), json_val);
                }
            }
            Value::json(serde_json::Value::Object(merged))
        }
    }
}

/// Execute a stream plan: chunk the input and drive it through
/// [`StreamRun`](crate::executors::stream), which runs every chunk of
/// every node through the same primitives `run_node` composes. Events
/// are per node — `NodeStarted` at the first chunk, `NodeCompleted`
/// after the flush with an aggregated summary, a real `NodeFailed` on
/// error — so a stream run reads back like any other run.
fn execute_stream(
    node_ids: &[String],
    chunk_size: usize,
    ctx: &mut Context,
    catalog: &NodeCatalog,
    cache: &dyn CacheStore,
) -> Result<()> {
    use crate::executors::stream::StreamRun;

    // Streaming has no training semantics; leaving this undefined would
    // silently skip every fit. Nothing invokes it today — keep it that
    // way explicitly.
    if matches!(ctx.mode, RunMode::Fit { .. }) {
        return Err(SomaError::Execution {
            node_id: node_ids.first().cloned().unwrap_or_default(),
            message: "a stream plan cannot run in fit mode: fit the graph first, \
                      then stream the forward"
                .into(),
        });
    }

    // Resolve input from the first node's predecessors.
    let first_id = node_ids
        .first()
        .ok_or_else(|| SomaError::Other("stream plan has no nodes".into()))?;
    let input = resolve_input(first_id, ctx);

    // Chunk the input along the first tensor dimension.
    let chunks = chunk_value(&input, chunk_size);

    let last_id = node_ids.last().unwrap().clone();
    let mut run = StreamRun::new(node_ids, catalog)?;

    // Incremental concatenation — bounded memory, see `StreamOutput`.
    let mut output = crate::executors::StreamOutput::new();

    for (i, chunk) in chunks.into_iter().enumerate() {
        tracing::debug!(node_id = %last_id, chunk = i, "streaming chunk");
        if let Some(out) = run.process_chunk(chunk, ctx, cache)? {
            output.push(out);
        }
    }

    // Flush barrier filters.
    if let Some(flushed) = run.flush(ctx, cache)? {
        output.push(flushed);
    }

    tracing::debug!(node_id = %last_id, chunks = run.chunks_processed(), "stream done");
    run.finish(ctx);

    ctx.set(last_id, output.finish());
    Ok(())
}

/// Split a Value::Tensor along the first dimension into chunks.
fn chunk_value(x: &Value, chunk_size: usize) -> Vec<Value> {
    match x {
        Value::Tensor { values, shape } if !values.is_empty() && chunk_size > 0 => {
            let row_size = if shape.len() > 1 {
                shape[1..].iter().product()
            } else {
                1
            };
            let n_rows = shape[0];
            let mut chunks = Vec::new();
            for start in (0..n_rows).step_by(chunk_size) {
                let end = (start + chunk_size).min(n_rows);
                let flat_start = start * row_size;
                let flat_end = end * row_size;
                let chunk_vals = values[flat_start..flat_end].to_vec();
                let mut chunk_shape = shape.clone();
                chunk_shape[0] = end - start;
                chunks.push(Value::tensor(chunk_vals, chunk_shape));
            }
            chunks
        }
        _ => vec![x.clone()],
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::cache::MemoryCache;
    use somatize_core::cache::CacheKey;
    use somatize_core::filter::{Filter, FilterKind, FilterMeta, StreamMode};

    /// Panics in `meta()`, which — unlike `forward()` — runs outside the
    /// `catch_unwind` in `execute_node`, so the unwind reaches the thread
    /// boundary.
    struct PanicsInMeta;

    impl Filter for PanicsInMeta {
        fn config_hash(&self) -> CacheKey {
            CacheKey::from_parts(&[b"PanicsInMeta"])
        }
        fn fit(&self, _x: &Value, _y: Option<&Value>) -> Result<Value> {
            Ok(Value::Empty)
        }
        fn forward(&self, x: &Value, _state: &Value) -> Result<Value> {
            Ok(x.clone())
        }
        fn meta(&self) -> FilterMeta {
            panic!("meta blew up");
        }
    }

    /// `execute_parallel` used to `join().unwrap()`, which re-raises a
    /// branch's panic on the parent thread. Inside `std::thread::scope`
    /// that aborts the process — so a panic the runtime deliberately
    /// contains everywhere else took the whole host down when it happened
    /// in a parallel branch.
    #[test]
    fn a_panicking_parallel_branch_becomes_an_error() {
        let mut lib = NodeCatalog::new();
        lib.register("boom", Box::new(PanicsInMeta));
        lib.register("fine", Box::new(DoublerFilter));

        let cache = MemoryCache::default();
        let bus = Arc::new(EventBus::new(64));
        let mut ctx = Context::new(bus, "run-panic");
        ctx.set("input".to_string(), Value::tensor(vec![1.0], vec![1]));

        let plan = ExecutionPlan::Parallel(vec![
            ExecutionPlan::Execute {
                node_id: "boom".into(),
            },
            ExecutionPlan::Execute {
                node_id: "fine".into(),
            },
        ]);

        // Keep the default hook from printing the backtrace this test
        // provokes on purpose.
        let previous = std::panic::take_hook();
        std::panic::set_hook(Box::new(|_| {}));
        let result = execute(&plan, &mut ctx, &lib, &cache);
        std::panic::set_hook(previous);

        let err = result.expect_err("a panicking branch must not be a success");
        assert!(
            err.to_string().contains("meta blew up"),
            "the panic message should survive; got: {err}"
        );
    }

    struct DoublerFilter;

    impl Filter for DoublerFilter {
        fn config_hash(&self) -> CacheKey {
            CacheKey::from_parts(&[b"Doubler"])
        }
        fn fit(&self, _x: &Value, _y: Option<&Value>) -> Result<Value> {
            Ok(Value::Empty)
        }
        fn forward(&self, x: &Value, _state: &Value) -> Result<Value> {
            match x {
                Value::Tensor { values, shape } => {
                    let doubled: Vec<f64> = values.iter().map(|v| v * 2.0).collect();
                    Ok(Value::tensor(doubled, shape.clone()))
                }
                _ => Ok(x.clone()),
            }
        }
        fn meta(&self) -> FilterMeta {
            FilterMeta {
                name: "Doubler".into(),
                kind: FilterKind::Stateless,
                cacheable: true,
                differentiable: true,
                deterministic: true,
                stream_mode: StreamMode::FixedState,
                distribution: somatize_core::filter::Distribution::Local,
                input_schema: None,
                output_schema: None,
            }
        }
    }

    struct AdderFilter {
        amount: f64,
    }

    impl Filter for AdderFilter {
        fn config_hash(&self) -> CacheKey {
            CacheKey::from_parts(&[b"Adder", &self.amount.to_le_bytes()])
        }
        fn fit(&self, _x: &Value, _y: Option<&Value>) -> Result<Value> {
            Ok(Value::Empty)
        }
        fn forward(&self, x: &Value, _state: &Value) -> Result<Value> {
            match x {
                Value::Tensor { values, shape } => {
                    let added: Vec<f64> = values.iter().map(|v| v + self.amount).collect();
                    Ok(Value::tensor(added, shape.clone()))
                }
                _ => Ok(x.clone()),
            }
        }
        fn meta(&self) -> FilterMeta {
            FilterMeta {
                name: "Adder".into(),
                kind: FilterKind::Stateless,
                cacheable: true,
                differentiable: true,
                deterministic: true,
                stream_mode: StreamMode::FixedState,
                distribution: somatize_core::filter::Distribution::Local,
                input_schema: None,
                output_schema: None,
            }
        }
    }

    /// Slow filter that sleeps to verify parallelism.
    struct SlowFilter {
        id: String,
        delay_ms: u64,
    }

    impl Filter for SlowFilter {
        fn config_hash(&self) -> CacheKey {
            CacheKey::from_parts(&[b"Slow", self.id.as_bytes()])
        }
        fn fit(&self, _x: &Value, _y: Option<&Value>) -> Result<Value> {
            Ok(Value::Empty)
        }
        fn forward(&self, x: &Value, _state: &Value) -> Result<Value> {
            std::thread::sleep(std::time::Duration::from_millis(self.delay_ms));
            Ok(x.clone())
        }
        fn meta(&self) -> FilterMeta {
            FilterMeta {
                name: format!("Slow_{}", self.id),
                kind: FilterKind::Stateless,
                cacheable: false,
                differentiable: true,
                deterministic: true,
                stream_mode: StreamMode::FixedState,
                distribution: somatize_core::filter::Distribution::Local,
                input_schema: None,
                output_schema: None,
            }
        }
    }

    fn setup() -> (Arc<EventBus>, MemoryCache) {
        (Arc::new(EventBus::new(64)), MemoryCache::default())
    }

    #[test]
    fn execute_single_node() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_1");
        ctx.set("input", Value::tensor(vec![1.0, 2.0, 3.0], vec![3]));
        ctx.graph_info
            .set_predecessors("doubler", vec!["input".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("doubler", Box::new(DoublerFilter));

        let plan = ExecutionPlan::Execute {
            node_id: "doubler".into(),
        };

        execute(&plan, &mut ctx, &filters, &cache).unwrap();

        let result = ctx.get("doubler").unwrap();
        let (data, _) = result.as_tensor().unwrap();
        assert_eq!(data, &[2.0, 4.0, 6.0]);
    }

    #[test]
    fn execute_sequence_with_graph_info() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_1");
        ctx.set("input", Value::tensor(vec![1.0, 2.0], vec![2]));

        let graph_info = GraphInfo::for_linear(&["input", "add", "double"]);
        ctx.graph_info = graph_info;

        let mut filters = NodeCatalog::new();
        filters.register("add", Box::new(AdderFilter { amount: 10.0 }));
        filters.register("double", Box::new(DoublerFilter));

        let plan = ExecutionPlan::Sequence(vec![
            ExecutionPlan::Execute {
                node_id: "add".into(),
            },
            ExecutionPlan::Execute {
                node_id: "double".into(),
            },
        ]);

        execute(&plan, &mut ctx, &filters, &cache).unwrap();

        let result = ctx.get("double").unwrap();
        let (data, _) = result.as_tensor().unwrap();
        assert_eq!(data, &[22.0, 24.0]);
    }

    #[test]
    fn execute_emits_events() {
        let bus = Arc::new(EventBus::new(64));
        let cache = MemoryCache::default();
        let mut rx = bus.subscribe();

        let mut ctx = Context::new(bus, "run_1");
        ctx.set("input", Value::tensor(vec![1.0], vec![1]));
        ctx.graph_info
            .set_predecessors("double", vec!["input".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));

        execute(
            &ExecutionPlan::Execute {
                node_id: "double".into(),
            },
            &mut ctx,
            &filters,
            &cache,
        )
        .unwrap();

        // Cacheable node, cold cache: miss → started → completed.
        let e1 = rx.try_recv().unwrap();
        assert!(matches!(e1, Event::NodeCacheMiss { .. }), "got {e1:?}");
        let e2 = rx.try_recv().unwrap();
        assert!(matches!(e2, Event::NodeStarted { .. }), "got {e2:?}");
        let e3 = rx.try_recv().unwrap();
        assert!(matches!(e3, Event::NodeCompleted { .. }), "got {e3:?}");
    }

    #[test]
    fn execute_missing_filter_errors() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_1");
        let filters = NodeCatalog::new();

        let result = execute(
            &ExecutionPlan::Execute {
                node_id: "nonexistent".into(),
            },
            &mut ctx,
            &filters,
            &cache,
        );
        assert!(matches!(result, Err(SomaError::NodeNotFound(_))));
    }

    #[test]
    fn execute_empty_plan() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_1");
        let filters = NodeCatalog::new();
        execute(&ExecutionPlan::Empty, &mut ctx, &filters, &cache).unwrap();
    }

    #[test]
    fn parallel_merge_keeps_rerun_outputs() {
        // Running the same parallel block twice must leave the *second*
        // results in the context. Merging by "keys the parent lacks" passed
        // the first pass and silently dropped every later one, so anything
        // downstream of a parallel body inside a `Loop` read iteration one
        // forever while the nodes dutifully re-ran.
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_1");
        ctx.graph_info
            .set_predecessors("double", vec!["input".into()]);
        ctx.graph_info.set_predecessors("add", vec!["input".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));
        filters.register("add", Box::new(AdderFilter { amount: 100.0 }));

        let plan = ExecutionPlan::Parallel(vec![
            ExecutionPlan::Execute {
                node_id: "double".into(),
            },
            ExecutionPlan::Execute {
                node_id: "add".into(),
            },
        ]);

        ctx.set("input", Value::tensor(vec![5.0], vec![1]));
        execute(&plan, &mut ctx, &filters, &cache).unwrap();
        assert_eq!(ctx.get("double").unwrap().as_tensor().unwrap().0, &[10.0]);

        // Same plan, new input — as a second loop iteration would.
        ctx.set("input", Value::tensor(vec![7.0], vec![1]));
        execute(&plan, &mut ctx, &filters, &cache).unwrap();

        assert_eq!(
            ctx.get("double").unwrap().as_tensor().unwrap().0,
            &[14.0],
            "second pass output was discarded by the merge"
        );
        assert_eq!(ctx.get("add").unwrap().as_tensor().unwrap().0, &[107.0]);
    }

    #[test]
    fn execute_parallel_branches_merge_outputs() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_1");
        ctx.set("input", Value::tensor(vec![5.0], vec![1]));
        ctx.graph_info
            .set_predecessors("double", vec!["input".into()]);
        ctx.graph_info.set_predecessors("add", vec!["input".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));
        filters.register("add", Box::new(AdderFilter { amount: 100.0 }));

        let plan = ExecutionPlan::Parallel(vec![
            ExecutionPlan::Execute {
                node_id: "double".into(),
            },
            ExecutionPlan::Execute {
                node_id: "add".into(),
            },
        ]);

        execute(&plan, &mut ctx, &filters, &cache).unwrap();

        let double_out = ctx.get("double").unwrap().as_tensor().unwrap().0;
        assert_eq!(double_out, &[10.0]);
        let add_out = ctx.get("add").unwrap().as_tensor().unwrap().0;
        assert_eq!(add_out, &[105.0]);
    }

    #[test]
    fn parallel_branches_run_concurrently() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_1");
        ctx.set("input", Value::tensor(vec![1.0], vec![1]));
        ctx.graph_info
            .set_predecessors("slow_a", vec!["input".into()]);
        ctx.graph_info
            .set_predecessors("slow_b", vec!["input".into()]);

        let mut filters = NodeCatalog::new();
        filters.register(
            "slow_a",
            Box::new(SlowFilter {
                id: "a".into(),
                delay_ms: 200,
            }),
        );
        filters.register(
            "slow_b",
            Box::new(SlowFilter {
                id: "b".into(),
                delay_ms: 200,
            }),
        );

        let plan = ExecutionPlan::Parallel(vec![
            ExecutionPlan::Execute {
                node_id: "slow_a".into(),
            },
            ExecutionPlan::Execute {
                node_id: "slow_b".into(),
            },
        ]);

        let start = Instant::now();
        execute(&plan, &mut ctx, &filters, &cache).unwrap();
        let elapsed = start.elapsed();

        // If truly parallel: ~200ms. If sequential: ~400ms. The wide
        // margin keeps the discrimination robust on loaded CI runners.
        assert!(
            elapsed.as_millis() < 350,
            "parallel branches took {}ms, expected <350ms (sequential would be ~400ms)",
            elapsed.as_millis()
        );

        assert!(ctx.get("slow_a").is_some());
        assert!(ctx.get("slow_b").is_some());
    }

    #[test]
    fn resolve_input_single_predecessor() {
        let bus = Arc::new(EventBus::new(8));
        let mut ctx = Context::new(bus, "r");
        ctx.set("A", Value::tensor(vec![42.0], vec![1]));
        ctx.graph_info.set_predecessors("B", vec!["A".into()]);

        let input = resolve_input("B", &ctx);
        let (data, _) = input.as_tensor().unwrap();
        assert_eq!(data, &[42.0]);
    }

    #[test]
    fn resolve_input_multiple_predecessors() {
        let bus = Arc::new(EventBus::new(8));
        let mut ctx = Context::new(bus, "r");
        ctx.set("A", Value::tensor(vec![1.0], vec![1]));
        ctx.set("B", Value::tensor(vec![2.0], vec![1]));
        ctx.graph_info
            .set_predecessors("C", vec!["A".into(), "B".into()]);

        let input = resolve_input("C", &ctx);
        let json = input.as_json().unwrap();
        assert!(json.get("A").is_some());
        assert!(json.get("B").is_some());
    }

    #[test]
    fn resolve_input_no_predecessors_fallback() {
        let bus = Arc::new(EventBus::new(8));
        let mut ctx = Context::new(bus, "r");
        ctx.set("prev", Value::tensor(vec![7.0], vec![1]));

        let input = resolve_input("root", &ctx);
        let (data, _) = input.as_tensor().unwrap();
        assert_eq!(data, &[7.0]);
    }

    #[test]
    fn graph_info_from_linear() {
        let info = GraphInfo::for_linear(&["a", "b", "c"]);
        assert!(info.predecessors("a").is_empty());
        assert_eq!(info.predecessors("b"), &["a"]);
        assert_eq!(info.predecessors("c"), &["b"]);
    }

    #[test]
    fn execute_stream_chunks_input() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_stream");
        // 6-element input, chunk_size=2 → 3 chunks
        ctx.set(
            "__input__",
            Value::tensor(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], vec![6]),
        );
        ctx.graph_info
            .set_predecessors("double", vec!["__input__".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));

        let plan = ExecutionPlan::Stream {
            node_ids: vec!["double".into()],
            chunk_size: 2,
        };

        execute(&plan, &mut ctx, &filters, &cache).unwrap();

        let result = ctx.get("double").unwrap();
        let (data, shape) = result.as_tensor().unwrap();
        assert_eq!(data, &[2.0, 4.0, 6.0, 8.0, 10.0, 12.0]);
        assert_eq!(shape, &[6]);
    }

    #[test]
    fn execute_stream_chain() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_stream_chain");
        ctx.set(
            "__input__",
            Value::tensor(vec![1.0, 2.0, 3.0, 4.0], vec![4]),
        );
        ctx.graph_info
            .set_predecessors("double", vec!["__input__".into()]);
        ctx.graph_info
            .set_predecessors("add", vec!["double".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));
        filters.register("add", Box::new(AdderFilter { amount: 10.0 }));

        let plan = ExecutionPlan::Stream {
            node_ids: vec!["double".into(), "add".into()],
            chunk_size: 2,
        };

        execute(&plan, &mut ctx, &filters, &cache).unwrap();

        // double → add: [1,2,3,4] → [2,4,6,8] → [12,14,16,18]
        let result = ctx.get("add").unwrap();
        let (data, shape) = result.as_tensor().unwrap();
        assert_eq!(data, &[12.0, 14.0, 16.0, 18.0]);
        assert_eq!(shape, &[4]);
    }

    /// Counts forward() invocations — the probe for cache-hit tests.
    struct CountingFilter {
        forwards: Arc<std::sync::atomic::AtomicUsize>,
        cacheable: bool,
        config: f64,
    }

    impl Filter for CountingFilter {
        fn config_hash(&self) -> CacheKey {
            CacheKey::from_parts(&[b"Counting", &self.config.to_le_bytes()])
        }
        fn fit(&self, _x: &Value, _y: Option<&Value>) -> Result<Value> {
            Ok(Value::Empty)
        }
        fn forward(&self, x: &Value, _state: &Value) -> Result<Value> {
            self.forwards
                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
            match x {
                Value::Tensor { values, shape } => {
                    let out: Vec<f64> = values.iter().map(|v| v + self.config).collect();
                    Ok(Value::tensor(out, shape.clone()))
                }
                _ => Ok(x.clone()),
            }
        }
        fn meta(&self) -> FilterMeta {
            FilterMeta {
                name: "Counting".into(),
                kind: FilterKind::Stateless,
                cacheable: self.cacheable,
                differentiable: true,
                deterministic: true,
                stream_mode: StreamMode::FixedState,
                distribution: somatize_core::filter::Distribution::Local,
                input_schema: None,
                output_schema: None,
            }
        }
    }

    fn counting_setup(
        cacheable: bool,
    ) -> (NodeCatalog, Arc<std::sync::atomic::AtomicUsize>, GraphInfo) {
        let forwards = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let mut filters = NodeCatalog::new();
        filters.register(
            "a",
            Box::new(CountingFilter {
                forwards: forwards.clone(),
                cacheable,
                config: 1.0,
            }),
        );
        filters.register(
            "b",
            Box::new(CountingFilter {
                forwards: forwards.clone(),
                cacheable,
                config: 2.0,
            }),
        );
        let info = GraphInfo::for_linear(&["input", "a", "b"]);
        (filters, forwards, info)
    }

    fn run_chain(cache: &dyn CacheStore, filters: &NodeCatalog, info: &GraphInfo) -> Value {
        let bus = Arc::new(EventBus::new(64));
        let mut ctx = Context::new(bus, "run").with_graph_info(info.clone());
        ctx.set("input", Value::tensor(vec![1.0, 2.0], vec![2]));
        let plan = ExecutionPlan::Sequence(vec![
            ExecutionPlan::Execute {
                node_id: "a".into(),
            },
            ExecutionPlan::Execute {
                node_id: "b".into(),
            },
        ]);
        execute(&plan, &mut ctx, filters, cache).unwrap();
        ctx.get("b").unwrap().clone()
    }

    #[test]
    fn second_run_hits_cache_and_skips_execution() {
        let (filters, forwards, info) = counting_setup(true);
        let cache = MemoryCache::default();

        let first = run_chain(&cache, &filters, &info);
        assert_eq!(forwards.load(std::sync::atomic::Ordering::SeqCst), 2);

        let second = run_chain(&cache, &filters, &info);
        assert_eq!(
            forwards.load(std::sync::atomic::Ordering::SeqCst),
            2,
            "second run must not execute any filter"
        );
        assert_eq!(first, second);
    }

    #[test]
    fn uncacheable_filter_always_executes() {
        let (filters, forwards, info) = counting_setup(false);
        let cache = MemoryCache::default();

        run_chain(&cache, &filters, &info);
        run_chain(&cache, &filters, &info);
        assert_eq!(forwards.load(std::sync::atomic::Ordering::SeqCst), 4);
    }

    #[test]
    fn cache_survives_process_restart() {
        use crate::cache::LocalCache;
        let dir = std::env::temp_dir().join(format!(
            "soma_exec_restart_{}_{}",
            std::process::id(),
            std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap()
                .as_nanos()
        ));
        let (filters, forwards, info) = counting_setup(true);

        {
            let cache = LocalCache::new(&dir).unwrap();
            run_chain(&cache, &filters, &info);
        }
        assert_eq!(forwards.load(std::sync::atomic::Ordering::SeqCst), 2);

        // "Restart": a fresh cache instance over the same directory.
        {
            let cache = LocalCache::new(&dir).unwrap();
            run_chain(&cache, &filters, &info);
        }
        assert_eq!(
            forwards.load(std::sync::atomic::Ordering::SeqCst),
            2,
            "after restart the persisted cache must serve both nodes"
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    #[test]
    fn different_input_misses_cache() {
        let (filters, forwards, info) = counting_setup(true);
        let cache = MemoryCache::default();

        run_chain(&cache, &filters, &info);

        let bus = Arc::new(EventBus::new(64));
        let mut ctx = Context::new(bus, "run2").with_graph_info(info.clone());
        ctx.set("input", Value::tensor(vec![9.0, 9.0], vec![2]));
        let plan = ExecutionPlan::Sequence(vec![
            ExecutionPlan::Execute {
                node_id: "a".into(),
            },
            ExecutionPlan::Execute {
                node_id: "b".into(),
            },
        ]);
        execute(&plan, &mut ctx, &filters, &cache).unwrap();
        assert_eq!(
            forwards.load(std::sync::atomic::Ordering::SeqCst),
            4,
            "different input data must not hit the cache"
        );
    }

    #[test]
    fn cache_hit_emits_cache_hit_event() {
        let (filters, _forwards, info) = counting_setup(true);
        let cache = MemoryCache::default();
        run_chain(&cache, &filters, &info);

        let bus = Arc::new(EventBus::new(64));
        let mut rx = bus.subscribe();
        let mut ctx = Context::new(bus, "run2").with_graph_info(info.clone());
        ctx.set("input", Value::tensor(vec![1.0, 2.0], vec![2]));
        execute(
            &ExecutionPlan::Execute {
                node_id: "a".into(),
            },
            &mut ctx,
            &filters,
            &cache,
        )
        .unwrap();

        let event = rx.try_recv().unwrap();
        assert!(
            matches!(event, Event::NodeCacheHit { ref node_id, .. } if node_id == "a"),
            "expected NodeCacheHit for `a`, got: {event:?}"
        );
    }

    #[test]
    fn spill_roundtrip_through_datastore() {
        use somatize_core::store::LocalDataStore;
        let dir = std::env::temp_dir().join(format!(
            "soma_spill_test_{}_{}",
            std::process::id(),
            std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap()
                .as_nanos()
        ));
        let store: Arc<dyn DataStore> = Arc::new(LocalDataStore::new(&dir));

        let (filters, _forwards, info) = counting_setup(true);
        let bus = Arc::new(EventBus::new(64));
        let mut ctx = Context::new(bus, "run_spill")
            .with_graph_info(info.clone())
            .with_data_store(store)
            .with_spill_threshold(1); // spill everything
        ctx.set("input", Value::tensor(vec![1.0, 2.0], vec![2]));

        let plan = ExecutionPlan::Sequence(vec![
            ExecutionPlan::Execute {
                node_id: "a".into(),
            },
            ExecutionPlan::Execute {
                node_id: "b".into(),
            },
        ]);
        execute(&plan, &mut ctx, &filters, &cache_for_spill())
            .expect("spilled intermediate must be readable downstream");

        // `a`'s output was spilled; `b` must still have received it:
        // input + 1.0 + 2.0 = [4.0, 5.0]
        let out = resolve_value(ctx.get_virtual("b").unwrap(), &ctx.data_store).unwrap();
        let (data, _) = out.as_tensor().unwrap();
        assert_eq!(data, &[4.0, 5.0]);

        let _ = std::fs::remove_dir_all(&dir);
    }

    fn cache_for_spill() -> MemoryCache {
        MemoryCache::default()
    }

    /// Config carries a salt that does NOT affect the output — models a
    /// cosmetic/irrelevant config change upstream.
    struct SaltedFilter {
        salt: f64,
        forwards: Arc<std::sync::atomic::AtomicUsize>,
    }

    impl Filter for SaltedFilter {
        fn config_hash(&self) -> CacheKey {
            CacheKey::from_parts(&[b"Salted", &self.salt.to_le_bytes()])
        }
        fn fit(&self, _x: &Value, _y: Option<&Value>) -> Result<Value> {
            Ok(Value::Empty)
        }
        fn forward(&self, x: &Value, _state: &Value) -> Result<Value> {
            self.forwards
                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
            match x {
                Value::Tensor { values, shape } => Ok(Value::tensor(
                    values.iter().map(|v| v + 1.0).collect(),
                    shape.clone(),
                )),
                _ => Ok(x.clone()),
            }
        }
        fn meta(&self) -> FilterMeta {
            FilterMeta {
                name: "Salted".into(),
                kind: FilterKind::Stateless,
                cacheable: true,
                differentiable: true,
                deterministic: true,
                stream_mode: StreamMode::FixedState,
                distribution: somatize_core::filter::Distribution::Local,
                input_schema: None,
                output_schema: None,
            }
        }
    }

    #[test]
    fn early_cutoff_downstream_hits_when_upstream_output_unchanged() {
        // Downstream keys derive from input CONTENT hashes, not from
        // upstream provenance — so a config change in A that produces
        // identical bytes must not invalidate B (rustc/salsa-style
        // early cutoff; impossible under the old deep-provenance keys).
        let a_forwards = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let b_forwards = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let cache = MemoryCache::default();
        let info = GraphInfo::for_linear(&["input", "a", "b"]);
        let plan = ExecutionPlan::Sequence(vec![
            ExecutionPlan::Execute {
                node_id: "a".into(),
            },
            ExecutionPlan::Execute {
                node_id: "b".into(),
            },
        ]);

        let run = |salt: f64| {
            let mut filters = NodeCatalog::new();
            filters.register(
                "a",
                Box::new(SaltedFilter {
                    salt,
                    forwards: a_forwards.clone(),
                }),
            );
            filters.register(
                "b",
                Box::new(CountingFilter {
                    forwards: b_forwards.clone(),
                    cacheable: true,
                    config: 2.0,
                }),
            );
            let bus = Arc::new(EventBus::new(64));
            let mut ctx = Context::new(bus, "run").with_graph_info(info.clone());
            ctx.set("input", Value::tensor(vec![1.0, 2.0], vec![2]));
            execute(&plan, &mut ctx, &filters, &cache).unwrap();
        };

        run(1.0);
        assert_eq!(a_forwards.load(std::sync::atomic::Ordering::SeqCst), 1);
        assert_eq!(b_forwards.load(std::sync::atomic::Ordering::SeqCst), 1);

        // A's config changed (new salt) → A re-executes. Its output is
        // byte-identical, so B must hit.
        run(2.0);
        assert_eq!(
            a_forwards.load(std::sync::atomic::Ordering::SeqCst),
            2,
            "A's config changed, it must re-execute"
        );
        assert_eq!(
            b_forwards.load(std::sync::atomic::Ordering::SeqCst),
            1,
            "B's input content is unchanged — early cutoff must serve it from cache"
        );
    }

    #[test]
    fn nondeterministic_filter_is_never_cached() {
        struct RandomishFilter {
            forwards: Arc<std::sync::atomic::AtomicUsize>,
        }
        impl Filter for RandomishFilter {
            fn config_hash(&self) -> CacheKey {
                CacheKey::from_parts(&[b"Randomish"])
            }
            fn fit(&self, _x: &Value, _y: Option<&Value>) -> Result<Value> {
                Ok(Value::Empty)
            }
            fn forward(&self, x: &Value, _state: &Value) -> Result<Value> {
                self.forwards
                    .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
                Ok(x.clone())
            }
            fn meta(&self) -> FilterMeta {
                FilterMeta {
                    name: "Randomish".into(),
                    kind: FilterKind::Stateless,
                    cacheable: true,
                    differentiable: false,
                    deterministic: false, // declared nondeterministic
                    stream_mode: StreamMode::FixedState,
                    distribution: somatize_core::filter::Distribution::Local,
                    input_schema: None,
                    output_schema: None,
                }
            }
        }

        let forwards = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let mut filters = NodeCatalog::new();
        filters.register(
            "rng",
            Box::new(RandomishFilter {
                forwards: forwards.clone(),
            }),
        );
        let cache = MemoryCache::default();
        let info = GraphInfo::for_linear(&["input", "rng"]);
        for _ in 0..2 {
            let bus = Arc::new(EventBus::new(64));
            let mut ctx = Context::new(bus, "run").with_graph_info(info.clone());
            ctx.set("input", Value::tensor(vec![1.0], vec![1]));
            execute(
                &ExecutionPlan::Execute {
                    node_id: "rng".into(),
                },
                &mut ctx,
                &filters,
                &cache,
            )
            .unwrap();
        }
        assert_eq!(
            forwards.load(std::sync::atomic::Ordering::SeqCst),
            2,
            "a filter declared nondeterministic must run every time"
        );
    }

    #[test]
    fn execute_stream_single_chunk() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_stream_single");
        ctx.set("__input__", Value::tensor(vec![5.0, 10.0], vec![2]));
        ctx.graph_info
            .set_predecessors("double", vec!["__input__".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));

        // chunk_size larger than input → single chunk
        let plan = ExecutionPlan::Stream {
            node_ids: vec!["double".into()],
            chunk_size: 1000,
        };

        execute(&plan, &mut ctx, &filters, &cache).unwrap();

        let result = ctx.get("double").unwrap();
        let (data, _) = result.as_tensor().unwrap();
        assert_eq!(data, &[10.0, 20.0]);
    }

    /// Fails on any chunk containing a value >= its threshold.
    struct Tripwire {
        at: f64,
    }
    impl Filter for Tripwire {
        fn config_hash(&self) -> CacheKey {
            CacheKey::from_parts(&[b"Tripwire", &self.at.to_le_bytes()])
        }
        fn fit(&self, _x: &Value, _y: Option<&Value>) -> Result<Value> {
            Ok(Value::Empty)
        }
        fn forward(&self, x: &Value, _state: &Value) -> Result<Value> {
            if let Value::Tensor { values, .. } = x
                && values.iter().any(|v| *v >= self.at)
            {
                return Err(SomaError::Other(format!("tripped at {}", self.at)));
            }
            Ok(x.clone())
        }
        fn meta(&self) -> FilterMeta {
            DoublerFilter.meta()
        }
    }

    fn stream_events(
        rx: &mut tokio::sync::broadcast::Receiver<Event>,
    ) -> Vec<(String, &'static str)> {
        let mut seen = Vec::new();
        while let Ok(event) = rx.try_recv() {
            match event {
                Event::NodeStarted { node_id, .. } => seen.push((node_id, "started")),
                Event::NodeCompleted { node_id, .. } => seen.push((node_id, "completed")),
                Event::NodeFailed { node_id, .. } => seen.push((node_id, "failed")),
                _ => {}
            }
        }
        seen
    }

    /// T1: one bracket per NODE, not one per plan and not one per chunk.
    /// Three chunks through two nodes is exactly two started/completed
    /// pairs, both under real node ids.
    #[test]
    fn stream_emits_one_bracket_per_node() {
        let (bus, cache) = setup();
        let mut rx = bus.subscribe();
        let mut ctx = Context::new(bus, "run_stream_events");
        ctx.set(
            "__input__",
            Value::tensor(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], vec![6]),
        );
        ctx.graph_info
            .set_predecessors("double", vec!["__input__".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));
        filters.register("add", Box::new(AdderFilter { amount: 1.0 }));

        let plan = ExecutionPlan::Stream {
            node_ids: vec!["double".into(), "add".into()],
            chunk_size: 2,
        };
        execute(&plan, &mut ctx, &filters, &cache).unwrap();

        let seen = stream_events(&mut rx);
        for node in ["double", "add"] {
            assert_eq!(
                seen.iter()
                    .filter(|(id, kind)| id == node && *kind == "started")
                    .count(),
                1,
                "{node}: exactly one NodeStarted, got {seen:?}"
            );
            assert_eq!(
                seen.iter()
                    .filter(|(id, kind)| id == node && *kind == "completed")
                    .count(),
                1,
                "{node}: exactly one NodeCompleted, got {seen:?}"
            );
        }
        assert!(
            seen.iter().all(|(id, _)| id == "double" || id == "add"),
            "no made-up node ids: {seen:?}"
        );
    }

    /// T2: the failing node emits a real NodeFailed naming the chunk;
    /// the upstream node's span stays open — it died mid-node, which is
    /// literally true.
    #[test]
    fn stream_node_failed_names_the_chunk() {
        let (bus, cache) = setup();
        let mut rx = bus.subscribe();
        let mut ctx = Context::new(bus, "run_stream_fail");
        ctx.set("__input__", Value::tensor(vec![1.0, 3.0], vec![2]));
        ctx.graph_info
            .set_predecessors("double", vec!["__input__".into()]);

        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));
        // Doubling 3.0 -> 6.0 trips the wire on the second chunk.
        filters.register("trip", Box::new(Tripwire { at: 5.0 }));

        let plan = ExecutionPlan::Stream {
            node_ids: vec!["double".into(), "trip".into()],
            chunk_size: 1,
        };
        let err = execute(&plan, &mut ctx, &filters, &cache).unwrap_err();
        assert!(err.to_string().contains("tripped"), "{err}");

        let mut failed = None;
        let mut double_completed = false;
        while let Ok(event) = rx.try_recv() {
            match event {
                Event::NodeFailed { node_id, error, .. } => failed = Some((node_id, error)),
                Event::NodeCompleted { node_id, .. } if node_id == "double" => {
                    double_completed = true;
                }
                _ => {}
            }
        }
        let (node_id, error) = failed.expect("no NodeFailed was emitted");
        assert_eq!(node_id, "trip");
        assert!(error.contains("chunk 1"), "should name the chunk: {error}");
        assert!(
            !double_completed,
            "the upstream span must stay open: the run died mid-node"
        );
    }

    /// T6: the derivation is shared, so a single-chunk stream and the
    /// standard path land on ONE cache line — the second is a hit.
    #[test]
    fn stream_and_standard_share_one_cache_line() {
        let (bus, cache) = setup();
        let input = Value::tensor(vec![1.0, 2.0], vec![2]);

        let mut ctx = Context::new(bus.clone(), "run_standard");
        ctx.set("__input__", input.clone());
        ctx.graph_info
            .set_predecessors("double", vec!["__input__".into()]);
        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));
        let standard = ExecutionPlan::Execute {
            node_id: "double".into(),
        };
        execute(&standard, &mut ctx, &filters, &cache).unwrap();
        assert_eq!(cache.len(), 1);

        let mut rx = bus.subscribe();
        let mut ctx2 = Context::new(bus, "run_streamed");
        ctx2.set("__input__", input);
        ctx2.graph_info
            .set_predecessors("double", vec!["__input__".into()]);
        let streamed = ExecutionPlan::Stream {
            node_ids: vec!["double".into()],
            chunk_size: 1000, // single chunk == the standard input
        };
        execute(&streamed, &mut ctx2, &filters, &cache).unwrap();

        assert_eq!(
            cache.len(),
            1,
            "the stream must read the standard path's line, not mint its own"
        );
        let mut completed_summary = String::new();
        while let Ok(event) = rx.try_recv() {
            if let Event::NodeCompleted {
                node_id,
                output_summary,
                ..
            } = event
                && node_id == "double"
            {
                completed_summary = output_summary;
            }
        }
        assert!(
            completed_summary.contains("1 hits"),
            "the chunk should have been a cache hit: {completed_summary}"
        );
    }

    /// T11: for a plain FixedState chain, the stream path's event set is
    /// the standard path's, modulo the per-chunk hit/miss events the
    /// stream deliberately aggregates.
    #[test]
    fn stream_events_match_standard_for_fixed_chains() {
        let run = |streamed: bool| -> Vec<(String, &'static str)> {
            let (bus, cache) = setup();
            let mut rx = bus.subscribe();
            let mut ctx = Context::new(bus, "run_compare");
            ctx.set("__input__", Value::tensor(vec![1.0, 2.0], vec![2]));
            ctx.graph_info
                .set_predecessors("double", vec!["__input__".into()]);
            ctx.graph_info
                .set_predecessors("add", vec!["double".into()]);
            let mut filters = NodeCatalog::new();
            filters.register("double", Box::new(DoublerFilter));
            filters.register("add", Box::new(AdderFilter { amount: 1.0 }));
            let plan = if streamed {
                ExecutionPlan::Stream {
                    node_ids: vec!["double".into(), "add".into()],
                    chunk_size: 1,
                }
            } else {
                ExecutionPlan::Sequence(vec![
                    ExecutionPlan::Execute {
                        node_id: "double".into(),
                    },
                    ExecutionPlan::Execute {
                        node_id: "add".into(),
                    },
                ])
            };
            execute(&plan, &mut ctx, &filters, &cache).unwrap();
            let mut seen = stream_events(&mut rx);
            seen.sort();
            seen
        };

        assert_eq!(
            run(false),
            run(true),
            "same nodes, same brackets, whichever path executed them"
        );
    }

    /// D10: a stream plan in fit mode is an explicit error, not an
    /// undefined skip of every fit.
    #[test]
    fn stream_refuses_fit_mode() {
        let (bus, cache) = setup();
        let mut ctx = Context::new(bus, "run_stream_fit");
        ctx.mode = RunMode::Fit { y: None };
        ctx.set("__input__", Value::tensor(vec![1.0], vec![1]));
        ctx.graph_info
            .set_predecessors("double", vec!["__input__".into()]);
        let mut filters = NodeCatalog::new();
        filters.register("double", Box::new(DoublerFilter));

        let plan = ExecutionPlan::Stream {
            node_ids: vec!["double".into()],
            chunk_size: 2,
        };
        let err = execute(&plan, &mut ctx, &filters, &cache).unwrap_err();
        assert!(err.to_string().contains("fit"), "{err}");
    }
}