Skip to main content

harn_vm/vm/
state.rs

1use std::collections::{BTreeMap, HashMap, HashSet};
2use std::path::PathBuf;
3use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
4use std::sync::Arc;
5use std::time::Instant;
6
7use crate::chunk::{Chunk, ChunkRef, Constant};
8use crate::runtime_limits::RuntimeLimits;
9use crate::value::{
10    ModuleFunctionRegistry, VmAsyncBuiltinFn, VmBuiltinFn, VmClosure, VmEnv, VmError, VmMutex,
11    VmTaskHandle, VmValue,
12};
13use crate::BuiltinId;
14
15use super::debug::DebugHook;
16use super::modules::LoadedModule;
17use super::VmBuiltinMetadata;
18
19/// A lazy callable's resolved export set together with the module graph that
20/// was loaded to produce it.
21///
22/// The exported closures — and every function they can transitively reach
23/// through imports — hold only `Weak`s into their home module's function
24/// registry and module state. The child VM that first loaded the graph is
25/// normally their sole strong owner (via its `module_cache`), and it dies once
26/// the hook fire completes. A later child VM that hits this cache never
27/// re-imports the graph, so a transitively imported callee's sibling `pub fn`
28/// would fall through name resolution to host-bridge dispatch (`Undefined
29/// builtin`). Retaining the complete loaded graph here for the cache entry's
30/// lifetime keeps every transitively reachable registry and module state
31/// upgradeable at call time, for every fire — not just the first.
32pub(crate) struct ResolvedLazyCallable {
33    pub(crate) exports: BTreeMap<String, Arc<VmClosure>>,
34    /// Intentionally unread: retained solely to keep the loaded module graph's
35    /// function registries and module states alive for this cache entry (the
36    /// same liveness role [`crate::value::RetainedModuleScope`] plays for a
37    /// single retained closure, generalized across the whole import graph).
38    #[allow(dead_code)]
39    pub(crate) retained_module_graph: Arc<BTreeMap<PathBuf, LoadedModule>>,
40}
41
42pub(crate) type LazyCallableResolution = Arc<ResolvedLazyCallable>;
43pub(crate) type LazyCallableModuleCache =
44    Arc<VmMutex<BTreeMap<PathBuf, Arc<tokio::sync::OnceCell<LazyCallableResolution>>>>>;
45
46/// RAII guard that starts a tracing span on creation and ends it on drop.
47pub(crate) struct ScopeSpan(u64);
48
49impl ScopeSpan {
50    pub(crate) fn new(kind: crate::tracing::SpanKind, name: String) -> Self {
51        Self(crate::tracing::span_start(kind, name))
52    }
53}
54
55/// Cancellation-safe host deadline state. The guard owns the shared state so
56/// dropping an in-flight `execute_with_timeout` future restores the previous
57/// deadline and poisons the abandoned VM even though the future still holds
58/// `&mut Vm`.
59pub(crate) struct ExecutionDeadlineState {
60    origin: Instant,
61    /// Nanoseconds from `origin`, plus one so zero remains the inactive value.
62    deadline_offset: AtomicU64,
63    /// Set when the host drops a polled execution future. Arbitrary async
64    /// cancellation cannot unwind interpreter state, so subsequent execution
65    /// entries fail loudly instead of resuming a partial frame.
66    abandoned: AtomicBool,
67}
68
69impl ExecutionDeadlineState {
70    pub(crate) fn new(origin: Instant, deadline: Option<Instant>) -> Arc<Self> {
71        Arc::new(Self {
72            origin,
73            deadline_offset: AtomicU64::new(Self::encode(origin, deadline)),
74            abandoned: AtomicBool::new(false),
75        })
76    }
77
78    #[inline]
79    pub(crate) fn is_active(&self) -> bool {
80        self.deadline_offset.load(Ordering::Acquire) != 0
81    }
82
83    #[inline]
84    pub(crate) fn is_abandoned(&self) -> bool {
85        self.abandoned.load(Ordering::Acquire)
86    }
87
88    pub(crate) fn fork(&self) -> Arc<Self> {
89        let state = Self::new(self.origin, self.current());
90        state
91            .abandoned
92            .store(self.is_abandoned(), Ordering::Release);
93        state
94    }
95
96    pub(crate) fn current(&self) -> Option<Instant> {
97        let encoded = self.deadline_offset.load(Ordering::Acquire);
98        (encoded != 0)
99            .then(|| self.origin + std::time::Duration::from_nanos(encoded.saturating_sub(1)))
100    }
101
102    pub(crate) fn install(self: &Arc<Self>, deadline: Instant) -> ExecutionDeadlineGuard {
103        let previous = self.deadline_offset.load(Ordering::Acquire);
104        let requested = Self::encode(self.origin, Some(deadline));
105        let active = if previous == 0 {
106            requested
107        } else {
108            previous.min(requested)
109        };
110        self.deadline_offset.store(active, Ordering::Release);
111        ExecutionDeadlineGuard {
112            state: Arc::clone(self),
113            previous,
114            completed: false,
115        }
116    }
117
118    fn encode(origin: Instant, deadline: Option<Instant>) -> u64 {
119        deadline.map_or(0, |deadline| {
120            let nanos = deadline.saturating_duration_since(origin).as_nanos();
121            u64::try_from(nanos)
122                .unwrap_or(u64::MAX - 1)
123                .saturating_add(1)
124        })
125    }
126}
127
128pub(crate) struct ExecutionDeadlineGuard {
129    state: Arc<ExecutionDeadlineState>,
130    previous: u64,
131    completed: bool,
132}
133
134impl ExecutionDeadlineGuard {
135    /// Mark an awaited execution as terminal before restoring its prior host
136    /// deadline. Dropping without this acknowledgement poisons the VM.
137    pub(crate) fn complete(mut self) {
138        self.completed = true;
139    }
140}
141
142impl Drop for ExecutionDeadlineGuard {
143    fn drop(&mut self) {
144        self.state
145            .deadline_offset
146            .store(self.previous, Ordering::Release);
147        if !self.completed {
148            self.state.abandoned.store(true, Ordering::Release);
149            crate::orchestration::clear_pipeline_on_finish();
150        }
151    }
152}
153
154impl Drop for ScopeSpan {
155    fn drop(&mut self) {
156        crate::tracing::span_end(self.0);
157    }
158}
159
160#[derive(Clone)]
161pub(crate) struct LocalSlot {
162    pub(crate) value: VmValue,
163    pub(crate) initialized: bool,
164    pub(crate) synced: bool,
165}
166
167impl Drop for LocalSlot {
168    fn drop(&mut self) {
169        // Slot locals hold script values directly (e.g. a `var` bound to a
170        // deeply nested list). When a frame is torn down, the default
171        // recursive drop of such a value would overflow the native stack and
172        // abort the process. For the overwhelmingly common scalar slot this is
173        // a single `matches!` check and then the normal trivial drop; only a
174        // nested container is moved out and torn down iteratively, so hot
175        // frame teardown is unaffected.
176        if crate::value::recursion::is_recursive_container(&self.value) {
177            crate::value::recursion::dismantle(std::mem::replace(&mut self.value, VmValue::Nil));
178        }
179    }
180}
181
182#[derive(Clone)]
183pub(crate) struct InterruptHandler {
184    pub(crate) handle: i64,
185    pub(crate) signals: Vec<String>,
186    pub(crate) once: bool,
187    pub(crate) graceful_timeout_ms: Option<u64>,
188    pub(crate) handler: VmValue,
189}
190
191/// Call frame for function execution.
192pub(crate) struct CallFrame {
193    pub(crate) chunk: ChunkRef,
194    /// VM-local inline-cache set for this frame's chunk. Computed once at
195    /// frame entry so hot opcode dispatch can index cache feedback directly
196    /// instead of hashing the chunk id on every cached opcode.
197    pub(crate) inline_cache_set: usize,
198    pub(crate) ip: usize,
199    pub(crate) stack_base: usize,
200    pub(crate) saved_env: VmEnv,
201    /// Env snapshot captured at call-time, *after* argument binding. Used
202    /// by the debugger's `restartFrame` to rewind this frame to its
203    /// entry state (re-binding args from the original values) without
204    /// re-entering the call site. Cheap to clone because `VmEnv` is
205    /// already cloned into `saved_env` on every call. `None` for
206    /// scratch frames (evaluate, import init) where restart isn't
207    /// meaningful.
208    pub(crate) initial_env: Option<VmEnv>,
209    pub(crate) initial_local_slots: Option<Vec<LocalSlot>>,
210    /// Iterator stack depth to restore when this frame unwinds.
211    pub(crate) saved_iterator_depth: usize,
212    /// Function name for stack traces (empty for top-level pipeline).
213    pub(crate) fn_name: String,
214    /// Number of arguments actually passed by the caller (for default arg support).
215    pub(crate) argc: usize,
216    /// Saved VM_SOURCE_DIR to restore when this frame is popped.
217    /// Set when entering a closure that originated from an imported module.
218    pub(crate) saved_source_dir: Option<std::path::PathBuf>,
219    /// Module-local named functions available to symbolic calls within this frame.
220    pub(crate) module_functions: Option<ModuleFunctionRegistry>,
221    /// Shared module-level env for top-level `var` / `let` bindings of
222    /// this frame's originating module. Looked up after `self.env` and
223    /// before `self.globals` by `GetVar` / `SetVar`, giving each module
224    /// its own live static state that persists across calls. See the
225    /// `module_state` field on `VmClosure` for the full rationale.
226    pub(crate) module_state: Option<crate::value::ModuleState>,
227    /// Slot-indexed locals for compiler-resolved names in this frame.
228    pub(crate) local_slots: Vec<LocalSlot>,
229    /// Env scope index that corresponds to compiler local scope depth 0.
230    pub(crate) local_scope_base: usize,
231    /// Current compiler local scope depth, updated by PushScope/PopScope.
232    pub(crate) local_scope_depth: usize,
233}
234
235pub(crate) struct InlineCacheSite {
236    pub(crate) cache_set: usize,
237    pub(crate) slot_count: usize,
238    pub(crate) slot: Option<usize>,
239}
240
241impl CallFrame {
242    #[inline]
243    pub(crate) fn inline_cache_site_for_previous_op(&self) -> InlineCacheSite {
244        let op_offset = self.ip.saturating_sub(1);
245        InlineCacheSite {
246            cache_set: self.inline_cache_set,
247            slot_count: self.chunk.inline_cache_slot_count(),
248            slot: self.chunk.inline_cache_slot(op_offset),
249        }
250    }
251}
252
253/// Exception handler for try/catch.
254pub(crate) struct ExceptionHandler {
255    pub(crate) catch_ip: usize,
256    pub(crate) stack_depth: usize,
257    pub(crate) frame_depth: usize,
258    pub(crate) env_scope_depth: usize,
259    /// When present, this catch only handles errors whose enum_name matches.
260    pub(crate) error_type: Option<crate::value::HarnStr>,
261}
262
263/// A structured-concurrency nursery (`scope { }`). Tasks spawned while this
264/// scope is innermost record their id here; `TaskScopeExit` joins them.
265pub(crate) struct TaskScope {
266    /// Ids of tasks spawned in this scope that have not been explicitly
267    /// `await`ed away. Joined (normal exit) or cancelled (unwind) on close.
268    pub(crate) task_ids: Vec<String>,
269    /// Frame depth at which the scope was opened, for unwind pruning.
270    pub(crate) frame_depth: usize,
271    /// Env scope depth at open, for unwind pruning.
272    pub(crate) env_scope_depth: usize,
273}
274
275/// Iterator state for for-in loops.
276pub(crate) enum IterState {
277    Vec {
278        items: Arc<Vec<VmValue>>,
279        idx: usize,
280    },
281    Dict {
282        entries: Arc<crate::value::DictMap>,
283        keys: Vec<String>,
284        idx: usize,
285    },
286    Channel {
287        receiver: std::sync::Arc<tokio::sync::Mutex<tokio::sync::mpsc::Receiver<VmValue>>>,
288        close: std::sync::Arc<crate::value::VmChannelCloseState>,
289    },
290    Generator {
291        gen: Arc<crate::value::VmGenerator>,
292    },
293    Stream {
294        stream: Arc<crate::value::VmStream>,
295    },
296    /// Step through a lazy range without materializing a Vec.
297    /// Inclusive ranges keep `end` as an actual value so `i64::MAX to i64::MAX`
298    /// still yields one item instead of overflowing a one-past-end sentinel.
299    Range {
300        next: i64,
301        end: i64,
302        inclusive: bool,
303        done: bool,
304    },
305    VmIter {
306        handle: crate::vm::iter::VmIterHandle,
307    },
308}
309
310#[derive(Clone)]
311pub(crate) enum VmBuiltinDispatch {
312    Sync(VmBuiltinFn),
313    Async(VmAsyncBuiltinFn),
314}
315
316#[derive(Clone)]
317pub(crate) struct VmBuiltinEntry {
318    pub(crate) name: Arc<str>,
319    pub(crate) dispatch: VmBuiltinDispatch,
320}
321
322/// The Harn bytecode virtual machine.
323pub struct Vm {
324    pub(crate) stack: Vec<VmValue>,
325    pub(crate) env: VmEnv,
326    pub(crate) output: String,
327    pub(crate) builtins: Arc<BTreeMap<String, VmBuiltinFn>>,
328    pub(crate) async_builtins: Arc<BTreeMap<String, VmAsyncBuiltinFn>>,
329    pub(crate) builtin_metadata: Arc<BTreeMap<String, VmBuiltinMetadata>>,
330    /// Numeric side index for builtins. Name-keyed maps remain authoritative;
331    /// this index is the hot path for direct builtin bytecode and callback refs.
332    pub(crate) builtins_by_id: Arc<HashMap<BuiltinId, VmBuiltinEntry>>,
333    /// IDs with detected name collisions. Collided names safely fall back to
334    /// the authoritative name-keyed lookup path.
335    pub(crate) builtin_id_collisions: Arc<HashSet<BuiltinId>>,
336    /// Iterator state for for-in loops.
337    pub(crate) iterators: Vec<IterState>,
338    /// Call frame stack.
339    pub(crate) frames: Vec<CallFrame>,
340    /// Exception handler stack.
341    pub(crate) exception_handlers: Vec<ExceptionHandler>,
342    /// Spawned async task handles.
343    pub(crate) spawned_tasks: BTreeMap<String, VmTaskHandle>,
344    /// Shared process-local synchronization primitives inherited by child VMs.
345    pub(crate) sync_runtime: Arc<crate::synchronization::VmSyncRuntime>,
346    /// Shared process-local cells, maps, and mailboxes inherited by child VMs.
347    pub(crate) shared_state_runtime: Arc<crate::shared_state::VmSharedStateRuntime>,
348    /// Per-isolate inline cache entries. `inline_cache_set_by_chunk` maps a
349    /// compiled chunk identity to an index in this vector at frame entry; the
350    /// dispatch loop uses the frame-local index for per-op reads/writes.
351    pub(crate) inline_cache_sets: Vec<Vec<crate::chunk::InlineCacheEntry>>,
352    pub(crate) inline_cache_set_by_chunk: HashMap<u64, usize>,
353    /// VM-scoped pool registry inherited by child VMs and scoped into Tokio tasks.
354    pub(crate) pool_registry: Arc<crate::stdlib::pool::PoolRegistry>,
355    /// Shared task/channel wait graph for this VM execution tree.
356    pub(crate) wait_for_graph: Arc<crate::wait_for_graph::VmWaitForGraph>,
357    /// Permits acquired by lexical synchronization blocks in this VM.
358    pub(crate) held_sync_guards: Vec<crate::synchronization::VmSyncHeldGuard>,
359    /// Locks held by an ancestor VM that is *suspended on this VM's execution*:
360    /// an inline async-builtin child runs while its parent is parked
361    /// mid-instruction still holding these permits. Re-acquiring more permits
362    /// than the primitive can grant is a provably-unresolvable self-deadlock, so
363    /// HARN-ORC-011 fires across the child boundary. Empty for new concurrent
364    /// tasks (`spawn`/`parallel`/triggers), where the parent keeps running and
365    /// blocking can be legitimately resolvable.
366    pub(crate) inherited_held_keys: Arc<Vec<crate::synchronization::VmSyncHeldKey>>,
367    /// Structured-concurrency nursery stack. Each `scope { }` block pushes a
368    /// `TaskScope`; tasks spawned while it is innermost register their id here.
369    /// On normal exit (`TaskScopeExit`) the scope's tasks are joined and the
370    /// first error propagates; on unwind they are cancelled. Modeled on
371    /// `held_sync_guards` (push on enter, prune/cancel on frame/handler exit).
372    pub(crate) task_scopes: Vec<TaskScope>,
373    /// Counter for generating unique task IDs.
374    pub(crate) task_counter: u64,
375    /// Counter for logical runtime-context task groups.
376    pub(crate) runtime_context_counter: u64,
377    /// Logical runtime task context visible through `runtime_context()`.
378    pub(crate) runtime_context: crate::runtime_context::RuntimeContext,
379    /// Active deadline stack: (deadline_instant, frame_depth).
380    pub(crate) deadlines: Vec<(Instant, usize)>,
381    /// Uncatchable wall-clock deadline imposed by the embedding host.
382    pub(crate) execution_deadline: Arc<ExecutionDeadlineState>,
383    /// Breakpoints, keyed by source-file path so a breakpoint at line N
384    /// in `auto.harn` doesn't also fire when execution hits line N in an
385    /// imported lib. The empty-string key is a wildcard used by callers
386    /// that don't track source paths (legacy `set_breakpoints` API).
387    pub(crate) breakpoints: BTreeMap<String, std::collections::BTreeSet<usize>>,
388    /// Function-name breakpoints. Any closure call whose
389    /// `CompiledFunction.name` matches an entry here raises a stop on
390    /// entry, regardless of the call site's file or line. Lets the IDE
391    /// break on `llm_call` / `host_run_pipeline` / any user pipeline
392    /// function without pinning down a source location first.
393    pub(crate) function_breakpoints: std::collections::BTreeSet<String>,
394    /// Latched on `push_closure_frame` when the callee's name matches
395    /// `function_breakpoints`; consumed by the next step so the stop is
396    /// reported with reason="function breakpoint" and the breakpoint
397    /// name available for the DAP `stopped` event.
398    pub(crate) pending_function_bp: Option<String>,
399    /// Whether the VM is in step mode.
400    pub(crate) step_mode: bool,
401    /// The frame depth at which stepping started (for step-over).
402    pub(crate) step_frame_depth: usize,
403    /// Whether the VM is currently stopped at a debug point.
404    pub(crate) stopped: bool,
405    /// Last source line executed (to detect line changes).
406    pub(crate) last_line: usize,
407    /// Source directory for resolving imports.
408    pub(crate) source_dir: Option<std::path::PathBuf>,
409    /// Modules currently being imported (cycle prevention).
410    pub(crate) imported_paths: Vec<std::path::PathBuf>,
411    /// Imports that hit an in-progress module (an import cycle) and so could
412    /// not bind inline. Drained by `flush_deferred_cyclic_imports` once the
413    /// involved modules finish loading.
414    pub(crate) deferred_cyclic_imports: Vec<super::modules::DeferredCyclicImport>,
415    /// Loaded module cache keyed by canonical or synthetic module path.
416    pub(crate) module_cache: Arc<BTreeMap<std::path::PathBuf, LoadedModule>>,
417    /// Lazy manifest modules initialized by any child in this execution tree.
418    /// The complete export set is retained so handlers and predicates from the
419    /// same module share one module state across repeated child invocations.
420    pub(crate) lazy_callable_modules: LazyCallableModuleCache,
421    /// Source text keyed by canonical or synthetic module path for debugger retrieval.
422    pub(crate) source_cache: Arc<BTreeMap<std::path::PathBuf, String>>,
423    /// Source file path for error reporting.
424    pub(crate) source_file: Option<String>,
425    /// Source text for error reporting.
426    pub(crate) source_text: Option<String>,
427    /// Line-coverage accumulator. `Some` only while a coverage session is
428    /// active (see [`crate::coverage`]); folded into the global report on drop.
429    pub(crate) coverage: Option<crate::coverage::Coverage>,
430    /// Optional bridge for delegating unknown builtins in bridge mode.
431    pub(crate) bridge: Option<Arc<crate::bridge::HostBridge>>,
432    /// Builtins denied by sandbox mode (`--deny` / `--allow` flags).
433    pub(crate) denied_builtins: Arc<HashSet<String>>,
434    /// Cancellation token for cooperative graceful shutdown (set by parent).
435    pub(crate) cancel_token: Option<std::sync::Arc<std::sync::atomic::AtomicBool>>,
436    pub(crate) interrupt_signal_token: Option<std::sync::Arc<std::sync::Mutex<Option<String>>>>,
437    /// Remaining instruction-boundary checks before a requested host
438    /// cancellation is forcefully raised. This gives `is_cancelled()` loops a
439    /// deterministic chance to return cleanly without letting non-cooperative
440    /// CPU-bound code run forever.
441    pub(crate) cancel_grace_instructions_remaining: Option<usize>,
442    /// User-visible interrupt handlers registered through `std/signal`.
443    pub(crate) interrupt_handlers: Vec<InterruptHandler>,
444    pub(crate) next_interrupt_handle: i64,
445    pub(crate) pending_interrupt_signal: Option<String>,
446    pub(crate) interrupted: bool,
447    pub(crate) dispatching_interrupt: bool,
448    pub(crate) interrupt_handler_deadline: Option<Instant>,
449    /// Captured stack trace from the most recent error (fn_name, line, col).
450    pub(crate) error_stack_trace: Vec<(String, usize, usize, Option<String>)>,
451    /// Yield channel sender for generator execution. When set, `Op::Yield`
452    /// sends values through this channel instead of being a no-op.
453    pub(crate) yield_sender: Option<tokio::sync::mpsc::Sender<Result<VmValue, VmError>>>,
454    /// Project root directory (detected via harn.toml).
455    /// Used as base directory for metadata, store, and checkpoint operations.
456    pub(crate) project_root: Option<std::path::PathBuf>,
457    /// Global constants (e.g. `pi`, `e`). Checked as a fallback in `GetVar`
458    /// after the environment, so user-defined variables can shadow them.
459    pub(crate) globals: Arc<crate::value::DictMap>,
460    /// Optional debugger hook invoked when execution advances to a new source line.
461    pub(crate) debug_hook: Option<parking_lot::Mutex<Box<DebugHook>>>,
462    /// Effective runtime ceilings for this VM execution.
463    pub(crate) runtime_limits: RuntimeLimits,
464}
465
466/// Reusable VM baseline for hosts that need many clean executions with the
467/// same stable builtin/source setup.
468///
469/// The baseline intentionally does not snapshot execution state. Each
470/// instantiation gets fresh stacks, frames, tasks, cancellation fields, sync
471/// primitives, shared cells/maps/mailboxes, and debug state. Builtin tables are
472/// shared through `Arc` until a per-execution rebind needs copy-on-write.
473#[derive(Clone)]
474pub struct VmBaseline {
475    builtins: Arc<BTreeMap<String, VmBuiltinFn>>,
476    async_builtins: Arc<BTreeMap<String, VmAsyncBuiltinFn>>,
477    builtin_metadata: Arc<BTreeMap<String, VmBuiltinMetadata>>,
478    builtins_by_id: Arc<HashMap<BuiltinId, VmBuiltinEntry>>,
479    builtin_id_collisions: Arc<HashSet<BuiltinId>>,
480    source_dir: Option<std::path::PathBuf>,
481    source_file: Option<String>,
482    source_text: Option<String>,
483    project_root: Option<std::path::PathBuf>,
484    globals: Arc<crate::value::DictMap>,
485    denied_builtins: Arc<HashSet<String>>,
486    runtime_limits: RuntimeLimits,
487}
488
489impl VmBaseline {
490    pub fn from_vm(vm: &Vm) -> Self {
491        Self {
492            builtins: Arc::clone(&vm.builtins),
493            async_builtins: Arc::clone(&vm.async_builtins),
494            builtin_metadata: Arc::clone(&vm.builtin_metadata),
495            builtins_by_id: Arc::clone(&vm.builtins_by_id),
496            builtin_id_collisions: Arc::clone(&vm.builtin_id_collisions),
497            source_dir: vm.source_dir.clone(),
498            source_file: vm.source_file.clone(),
499            source_text: vm.source_text.clone(),
500            project_root: vm.project_root.clone(),
501            globals: Arc::clone(&vm.globals),
502            denied_builtins: Arc::clone(&vm.denied_builtins),
503            runtime_limits: vm.runtime_limits,
504        }
505    }
506
507    pub fn instantiate(&self) -> Vm {
508        let mut source_cache = BTreeMap::new();
509        if let (Some(file), Some(text)) = (&self.source_file, &self.source_text) {
510            source_cache.insert(std::path::PathBuf::from(file), text.clone());
511        }
512        if let Some(dir) = &self.source_dir {
513            crate::stdlib::set_thread_source_dir(dir);
514        }
515
516        let mut vm = Vm {
517            stack: Vec::with_capacity(256),
518            env: VmEnv::new(),
519            output: String::new(),
520            builtins: Arc::clone(&self.builtins),
521            async_builtins: Arc::clone(&self.async_builtins),
522            builtin_metadata: Arc::clone(&self.builtin_metadata),
523            builtins_by_id: Arc::clone(&self.builtins_by_id),
524            builtin_id_collisions: Arc::clone(&self.builtin_id_collisions),
525            iterators: Vec::new(),
526            frames: Vec::new(),
527            exception_handlers: Vec::new(),
528            spawned_tasks: BTreeMap::new(),
529            sync_runtime: Arc::new(crate::synchronization::VmSyncRuntime::new()),
530            shared_state_runtime: Arc::new(crate::shared_state::VmSharedStateRuntime::new()),
531            inline_cache_sets: Vec::new(),
532            inline_cache_set_by_chunk: HashMap::new(),
533            pool_registry: crate::stdlib::pool::new_pool_registry(),
534            wait_for_graph: Arc::new(crate::wait_for_graph::VmWaitForGraph::new()),
535            held_sync_guards: Vec::new(),
536            inherited_held_keys: Arc::new(Vec::new()),
537            task_scopes: Vec::new(),
538            task_counter: 0,
539            runtime_context_counter: 0,
540            runtime_context: crate::runtime_context::RuntimeContext::root(),
541            deadlines: Vec::new(),
542            execution_deadline: super::execution::new_execution_deadline_state(None),
543            breakpoints: BTreeMap::new(),
544            function_breakpoints: std::collections::BTreeSet::new(),
545            pending_function_bp: None,
546            step_mode: false,
547            step_frame_depth: 0,
548            stopped: false,
549            last_line: 0,
550            source_dir: self.source_dir.clone(),
551            imported_paths: Vec::new(),
552            deferred_cyclic_imports: Vec::new(),
553            module_cache: Arc::new(BTreeMap::new()),
554            lazy_callable_modules: Arc::new(crate::value::VmMutex::new(BTreeMap::new())),
555            source_cache: Arc::new(source_cache),
556            source_file: self.source_file.clone(),
557            source_text: self.source_text.clone(),
558            coverage: crate::coverage::for_primary(self.source_file.as_deref()),
559            bridge: None,
560            denied_builtins: Arc::clone(&self.denied_builtins),
561            cancel_token: None,
562            interrupt_signal_token: None,
563            cancel_grace_instructions_remaining: None,
564            interrupt_handlers: Vec::new(),
565            next_interrupt_handle: 1,
566            pending_interrupt_signal: None,
567            interrupted: false,
568            dispatching_interrupt: false,
569            interrupt_handler_deadline: None,
570            error_stack_trace: Vec::new(),
571            yield_sender: None,
572            project_root: self.project_root.clone(),
573            globals: Arc::clone(&self.globals),
574            debug_hook: None,
575            runtime_limits: self.runtime_limits,
576        };
577
578        crate::stdlib::rebind_execution_state_builtins(&mut vm);
579        vm
580    }
581}
582
583impl Vm {
584    pub(crate) fn ensure_execution_available(&self) -> Result<(), VmError> {
585        if self.execution_deadline.is_abandoned() {
586            return Err(VmError::AbandonedExecution);
587        }
588        Ok(())
589    }
590
591    pub(crate) fn fresh_local_slots(chunk: &Chunk) -> Vec<LocalSlot> {
592        chunk
593            .local_slots
594            .iter()
595            .map(|_| LocalSlot {
596                value: VmValue::Nil,
597                initialized: false,
598                synced: false,
599            })
600            .collect()
601    }
602
603    pub(crate) fn bind_param_slots(
604        slots: &mut [LocalSlot],
605        func: &crate::chunk::CompiledFunction,
606        args: &[VmValue],
607        synced: bool,
608    ) {
609        Self::bind_param_slots_args(slots, func, &super::CallArgs::Slice(args), synced);
610    }
611
612    pub(crate) fn bind_param_slots_args(
613        slots: &mut [LocalSlot],
614        func: &crate::chunk::CompiledFunction,
615        args: &super::CallArgs<'_>,
616        synced: bool,
617    ) {
618        let param_count = func.params.len();
619        for (i, _param) in func.params.iter().enumerate() {
620            if i >= slots.len() {
621                break;
622            }
623            if func.has_rest_param && i == param_count - 1 {
624                let rest_args = args.to_vec_from(i);
625                slots[i].value = VmValue::List(std::sync::Arc::new(rest_args));
626                slots[i].initialized = true;
627                slots[i].synced = synced;
628            } else if let Some(arg) = args.get(i) {
629                slots[i].value = arg.clone();
630                slots[i].initialized = true;
631                slots[i].synced = synced;
632            }
633        }
634    }
635
636    pub(crate) fn visible_variables(&self) -> crate::value::DictMap {
637        let mut vars = self.env.all_variables();
638        let Some(frame) = self.frames.last() else {
639            return vars;
640        };
641        for (slot, info) in frame.local_slots.iter().zip(frame.chunk.local_slots.iter()) {
642            if slot.initialized && info.scope_depth <= frame.local_scope_depth {
643                vars.insert(crate::value::intern_key(&info.name), slot.value.clone());
644            }
645        }
646        vars
647    }
648
649    pub(crate) fn sync_current_frame_locals_to_env(&mut self) {
650        let frames = &mut self.frames;
651        let env = &mut self.env;
652        let Some(frame) = frames.last_mut() else {
653            return;
654        };
655        let local_scope_base = frame.local_scope_base;
656        let local_scope_depth = frame.local_scope_depth;
657        for (slot, info) in frame
658            .local_slots
659            .iter_mut()
660            .zip(frame.chunk.local_slots.iter())
661        {
662            if slot.initialized && !slot.synced && info.scope_depth <= local_scope_depth {
663                slot.synced = true;
664                let scope_idx = local_scope_base + info.scope_depth;
665                while env.scopes.len() <= scope_idx {
666                    env.push_scope();
667                }
668                // Slot-backed locals are never boxed captures (those resolve
669                // to `None` in the compiler and are defined straight into the
670                // env as `Cell`s), so syncing a slot always yields a `Value`.
671                Arc::make_mut(&mut env.scopes[scope_idx].vars).insert(
672                    info.name.clone(),
673                    crate::value::Binding::Value {
674                        value: slot.value.clone(),
675                        mutable: info.mutable,
676                    },
677                );
678            }
679        }
680    }
681
682    pub(crate) fn closure_call_env_for_current_frame(
683        &self,
684        closure: &crate::value::VmClosure,
685    ) -> VmEnv {
686        if closure.module_state().is_some() {
687            return closure.env.cloned_for_call();
688        }
689        let call_env = Self::closure_call_env(&self.env, closure);
690        // Same compile-time short-circuit as the env walk in
691        // `closure_call_env`: when the callee body never resolves an
692        // outer name through the env, injecting closure-typed *slot*
693        // locals from the caller's frame is wasted work too.
694        if !closure.func.chunk.references_outer_names {
695            return call_env;
696        }
697        let mut call_env = call_env;
698        let Some(frame) = self.frames.last() else {
699            return call_env;
700        };
701        for (slot, info) in frame
702            .local_slots
703            .iter()
704            .zip(frame.chunk.local_slots.iter())
705            .filter(|(slot, info)| slot.initialized && info.scope_depth <= frame.local_scope_depth)
706        {
707            if matches!(slot.value, VmValue::Closure(_)) && !call_env.contains(&info.name) {
708                let _ = call_env.define(&info.name, slot.value.clone(), info.mutable);
709            }
710        }
711        call_env
712    }
713
714    pub(crate) fn active_local_slot_value(&self, name: &str) -> Option<VmValue> {
715        let frame = self.frames.last()?;
716        let idx = self.active_local_slot_index(name)?;
717        frame.local_slots.get(idx).map(|slot| slot.value.clone())
718    }
719
720    /// Returns the slot index of an initialized active local with the given
721    /// name, walking from innermost to outermost scope. Used by legacy by-name
722    /// hot paths that still want to mutate the slot value in place without
723    /// paying a defensive `VmValue::clone` first.
724    pub(crate) fn active_local_slot_index(&self, name: &str) -> Option<usize> {
725        let frame = self.frames.last()?;
726        for (idx, info) in frame.chunk.local_slots.iter().enumerate().rev() {
727            if info.name == name && info.scope_depth <= frame.local_scope_depth {
728                if let Some(slot) = frame.local_slots.get(idx) {
729                    if slot.initialized {
730                        return Some(idx);
731                    }
732                }
733            }
734        }
735        None
736    }
737
738    pub(crate) fn assign_active_local_slot(
739        &mut self,
740        name: &str,
741        value: VmValue,
742        debug: bool,
743    ) -> Result<bool, VmError> {
744        let Some(frame) = self.frames.last_mut() else {
745            return Ok(false);
746        };
747        for (idx, info) in frame.chunk.local_slots.iter().enumerate().rev() {
748            if info.name == name && info.scope_depth <= frame.local_scope_depth {
749                if !debug && !info.mutable {
750                    return Err(VmError::ImmutableAssignment(name.to_string()));
751                }
752                if let Some(slot) = frame.local_slots.get_mut(idx) {
753                    crate::value::recursion::dismantle(std::mem::replace(&mut slot.value, value));
754                    slot.initialized = true;
755                    slot.synced = false;
756                    return Ok(true);
757                }
758            }
759        }
760        Ok(false)
761    }
762
763    pub fn new() -> Self {
764        Self {
765            stack: Vec::with_capacity(256),
766            env: VmEnv::new(),
767            output: String::new(),
768            builtins: Arc::new(BTreeMap::new()),
769            async_builtins: Arc::new(BTreeMap::new()),
770            builtin_metadata: Arc::new(BTreeMap::new()),
771            builtins_by_id: Arc::new(HashMap::new()),
772            builtin_id_collisions: Arc::new(HashSet::new()),
773            iterators: Vec::new(),
774            frames: Vec::new(),
775            exception_handlers: Vec::new(),
776            spawned_tasks: BTreeMap::new(),
777            sync_runtime: Arc::new(crate::synchronization::VmSyncRuntime::new()),
778            shared_state_runtime: Arc::new(crate::shared_state::VmSharedStateRuntime::new()),
779            inline_cache_sets: Vec::new(),
780            inline_cache_set_by_chunk: HashMap::new(),
781            pool_registry: crate::stdlib::pool::new_pool_registry(),
782            wait_for_graph: Arc::new(crate::wait_for_graph::VmWaitForGraph::new()),
783            held_sync_guards: Vec::new(),
784            inherited_held_keys: Arc::new(Vec::new()),
785            task_scopes: Vec::new(),
786            task_counter: 0,
787            runtime_context_counter: 0,
788            runtime_context: crate::runtime_context::RuntimeContext::root(),
789            deadlines: Vec::new(),
790            execution_deadline: super::execution::new_execution_deadline_state(None),
791            breakpoints: BTreeMap::new(),
792            function_breakpoints: std::collections::BTreeSet::new(),
793            pending_function_bp: None,
794            step_mode: false,
795            step_frame_depth: 0,
796            stopped: false,
797            last_line: 0,
798            source_dir: None,
799            imported_paths: Vec::new(),
800            deferred_cyclic_imports: Vec::new(),
801            module_cache: Arc::new(BTreeMap::new()),
802            lazy_callable_modules: Arc::new(crate::value::VmMutex::new(BTreeMap::new())),
803            source_cache: Arc::new(BTreeMap::new()),
804            source_file: None,
805            source_text: None,
806            coverage: crate::coverage::for_primary(None),
807            bridge: None,
808            denied_builtins: Arc::new(HashSet::new()),
809            cancel_token: None,
810            interrupt_signal_token: None,
811            cancel_grace_instructions_remaining: None,
812            interrupt_handlers: Vec::new(),
813            next_interrupt_handle: 1,
814            pending_interrupt_signal: None,
815            interrupted: false,
816            dispatching_interrupt: false,
817            interrupt_handler_deadline: None,
818            error_stack_trace: Vec::new(),
819            yield_sender: None,
820            project_root: None,
821            globals: Arc::new(crate::value::DictMap::new()),
822            debug_hook: None,
823            runtime_limits: RuntimeLimits::default(),
824        }
825    }
826
827    pub fn baseline(&self) -> VmBaseline {
828        VmBaseline::from_vm(self)
829    }
830
831    /// Return the effective runtime limit profile for this VM.
832    pub fn runtime_limits(&self) -> RuntimeLimits {
833        self.runtime_limits
834    }
835
836    /// Return a host/debug report describing the VM's effective runtime limits.
837    pub fn runtime_limit_report(&self) -> crate::RuntimeLimitsReport {
838        self.runtime_limits.report()
839    }
840
841    /// Returns true if any debugging affordance is active — DAP hook,
842    /// line breakpoints, or function breakpoints. Call-site code uses
843    /// this to decide whether to capture per-frame restart snapshots
844    /// (`initial_env`, `initial_local_slots`); without a debugger those
845    /// snapshots are dead weight, so skipping them removes two
846    /// allocations from every function call hot path.
847    ///
848    /// All three signals are stable across a function call's lifetime
849    /// (they're set before pipeline execution starts), so the gate is
850    /// consistent between frame creation and any later `restart_frame`
851    /// invocation. The three `is_empty` checks compile to a handful of
852    /// branch-predicted memory probes — cheaper than a single
853    /// `BTreeMap` clone, which is what we're avoiding.
854    #[inline]
855    pub(crate) fn debugger_attached(&self) -> bool {
856        self.debug_hook.is_some()
857            || !self.breakpoints.is_empty()
858            || !self.function_breakpoints.is_empty()
859    }
860
861    /// Set the bridge for delegating unknown builtins in bridge mode.
862    pub fn set_bridge(&mut self, bridge: Arc<crate::bridge::HostBridge>) {
863        self.bridge = Some(bridge);
864    }
865
866    /// Set builtins that are denied in sandbox mode.
867    /// When called, the given builtin names will produce a permission error.
868    pub fn set_denied_builtins(&mut self, denied: HashSet<String>) {
869        self.denied_builtins = Arc::new(denied);
870    }
871
872    /// Set source info for error reporting (file path and source text).
873    pub fn set_source_info(&mut self, file: &str, text: &str) {
874        self.source_file = Some(file.to_string());
875        self.source_text = Some(text.to_string());
876        if let Some(cov) = self.coverage.as_mut() {
877            cov.set_primary_file(file);
878        }
879        Arc::make_mut(&mut self.source_cache)
880            .insert(std::path::PathBuf::from(file), text.to_string());
881    }
882
883    /// Initialize execution (push the initial frame).
884    pub fn start(&mut self, chunk: &Chunk) -> Result<(), VmError> {
885        self.ensure_execution_available()?;
886        // The top-level pipeline frame captures env at start so
887        // restartFrame on the outermost frame rewinds to the
888        // pre-pipeline state — basically "restart session" in
889        // debugger terms. Skipped when no debugger is attached:
890        // the snapshot is dead weight in that case and dominates
891        // call-overhead bench numbers (~5-10%).
892        let debugger = self.debugger_attached();
893        let initial_env = if debugger {
894            Some(self.env.clone())
895        } else {
896            None
897        };
898        let initial_local_slots = if debugger {
899            Some(Self::fresh_local_slots(chunk))
900        } else {
901            None
902        };
903        let chunk = Arc::new(chunk.clone());
904        let local_slots = Self::fresh_local_slots(&chunk);
905        let inline_cache_set = self.inline_cache_set_index_for_chunk(&chunk);
906        self.frames.push(CallFrame {
907            chunk,
908            inline_cache_set,
909            ip: 0,
910            stack_base: self.stack.len(),
911            saved_env: self.env.clone(),
912            initial_env,
913            initial_local_slots,
914            saved_iterator_depth: self.iterators.len(),
915            fn_name: String::new(),
916            argc: 0,
917            saved_source_dir: None,
918            module_functions: None,
919            module_state: None,
920            local_slots,
921            local_scope_base: self.env.scope_depth().saturating_sub(1),
922            local_scope_depth: 0,
923        });
924        Ok(())
925    }
926
927    /// Create a child VM that shares builtins and env but has fresh execution state.
928    /// Used for parallel/spawn to fork the VM for concurrent tasks.
929    pub(crate) fn child_vm(&self) -> Vm {
930        Vm {
931            stack: Vec::with_capacity(64),
932            env: self.env.clone(),
933            output: String::new(),
934            builtins: Arc::clone(&self.builtins),
935            async_builtins: Arc::clone(&self.async_builtins),
936            builtin_metadata: Arc::clone(&self.builtin_metadata),
937            builtins_by_id: Arc::clone(&self.builtins_by_id),
938            builtin_id_collisions: Arc::clone(&self.builtin_id_collisions),
939            iterators: Vec::new(),
940            frames: Vec::new(),
941            exception_handlers: Vec::new(),
942            spawned_tasks: BTreeMap::new(),
943            sync_runtime: self.sync_runtime.clone(),
944            shared_state_runtime: self.shared_state_runtime.clone(),
945            inline_cache_sets: Vec::new(),
946            inline_cache_set_by_chunk: HashMap::new(),
947            pool_registry: self.pool_registry.clone(),
948            wait_for_graph: self.wait_for_graph.clone(),
949            held_sync_guards: Vec::new(),
950            inherited_held_keys: Arc::new(Vec::new()),
951            task_scopes: Vec::new(),
952            task_counter: 0,
953            runtime_context_counter: self.runtime_context_counter,
954            runtime_context: self.runtime_context.clone(),
955            deadlines: self.deadlines.clone(),
956            execution_deadline: self.execution_deadline.fork(),
957            breakpoints: BTreeMap::new(),
958            function_breakpoints: std::collections::BTreeSet::new(),
959            pending_function_bp: None,
960            step_mode: false,
961            step_frame_depth: 0,
962            stopped: false,
963            last_line: 0,
964            source_dir: self.source_dir.clone(),
965            imported_paths: Vec::new(),
966            deferred_cyclic_imports: Vec::new(),
967            module_cache: Arc::clone(&self.module_cache),
968            lazy_callable_modules: Arc::clone(&self.lazy_callable_modules),
969            source_cache: Arc::clone(&self.source_cache),
970            source_file: self.source_file.clone(),
971            source_text: self.source_text.clone(),
972            coverage: crate::coverage::for_primary(self.source_file.as_deref()),
973            bridge: self.bridge.clone(),
974            denied_builtins: Arc::clone(&self.denied_builtins),
975            cancel_token: self.cancel_token.clone(),
976            interrupt_signal_token: self.interrupt_signal_token.clone(),
977            cancel_grace_instructions_remaining: None,
978            interrupt_handlers: Vec::new(),
979            next_interrupt_handle: 1,
980            pending_interrupt_signal: None,
981            interrupted: self.interrupted,
982            dispatching_interrupt: false,
983            interrupt_handler_deadline: None,
984            error_stack_trace: Vec::new(),
985            yield_sender: None,
986            project_root: self.project_root.clone(),
987            globals: Arc::clone(&self.globals),
988            debug_hook: None,
989            runtime_limits: self.runtime_limits,
990        }
991    }
992
993    /// Create a child VM for external adapters that need to invoke Harn
994    /// closures while sharing the parent's builtins, globals, and module state.
995    pub(crate) fn child_vm_for_host(&self) -> Vm {
996        self.child_vm()
997    }
998
999    /// Request cancellation for every outstanding child task owned by this VM
1000    /// and then abort the join handles. This prevents un-awaited spawned tasks
1001    /// from outliving their parent execution scope.
1002    pub(crate) fn cancel_spawned_tasks(&mut self) {
1003        for (_, task) in std::mem::take(&mut self.spawned_tasks) {
1004            task.cancel_token
1005                .store(true, std::sync::atomic::Ordering::SeqCst);
1006            task.handle.abort();
1007        }
1008    }
1009
1010    /// Set the source directory for import resolution and introspection.
1011    /// Also auto-detects the project root if not already set.
1012    pub fn set_source_dir(&mut self, dir: &std::path::Path) {
1013        let dir = crate::stdlib::process::normalize_context_path(dir);
1014        self.source_dir = Some(dir.clone());
1015        crate::stdlib::set_thread_source_dir(&dir);
1016        // Auto-detect project root if not explicitly set.
1017        if self.project_root.is_none() {
1018            self.project_root = crate::stdlib::process::find_project_root(&dir);
1019        }
1020    }
1021
1022    /// Explicitly set the project root directory.
1023    /// Used by ACP/CLI to override auto-detection.
1024    pub fn set_project_root(&mut self, root: &std::path::Path) {
1025        self.project_root = Some(root.to_path_buf());
1026    }
1027
1028    /// Get only the explicit or auto-detected project root, without falling
1029    /// back to `source_dir`.
1030    pub(crate) fn explicit_project_root(&self) -> Option<&std::path::Path> {
1031        self.project_root.as_deref()
1032    }
1033
1034    /// Get the project root directory, falling back to source_dir.
1035    pub fn project_root(&self) -> Option<&std::path::Path> {
1036        self.project_root.as_deref().or(self.source_dir.as_deref())
1037    }
1038
1039    /// Return all registered builtin names (sync + async).
1040    pub fn builtin_names(&self) -> Vec<String> {
1041        let mut names: Vec<String> = self.builtins.keys().cloned().collect();
1042        names.extend(self.async_builtins.keys().cloned());
1043        names
1044    }
1045
1046    /// Return discoverable metadata for registered builtins.
1047    pub fn builtin_metadata(&self) -> Vec<VmBuiltinMetadata> {
1048        self.builtin_metadata.values().cloned().collect()
1049    }
1050
1051    /// Return discoverable metadata for a registered builtin name.
1052    pub fn builtin_metadata_for(&self, name: &str) -> Option<&VmBuiltinMetadata> {
1053        self.builtin_metadata.get(name)
1054    }
1055
1056    /// Set a global constant (e.g. `pi`, `e`).
1057    /// Stored separately from the environment so user-defined variables can shadow them.
1058    pub fn set_global(&mut self, name: &str, value: VmValue) {
1059        Arc::make_mut(&mut self.globals).insert(crate::value::intern_key(name), value);
1060    }
1061
1062    /// Read a previously-installed global (the value `set_global` /
1063    /// `set_harness` recorded). Returns `None` for unknown names.
1064    /// Hosts use this to look up runtime-installed capability handles
1065    /// (e.g. the `harness` slot) without having to track them
1066    /// separately.
1067    pub fn global(&self, name: &str) -> Option<&VmValue> {
1068        self.globals.get(name)
1069    }
1070
1071    /// Install the script's `Harness` capability handle as the `harness`
1072    /// global so the auto-call emitted by `Compiler::compile()` (for
1073    /// `fn main(harness: Harness)` entrypoints) can read it. Hosts that
1074    /// drive the VM directly (CLI, MCP server, composition runtime) call
1075    /// this once before `execute()`.
1076    pub fn set_harness(&mut self, harness: crate::harness::Harness) {
1077        self.set_global("harness", harness.into_vm_value());
1078    }
1079
1080    /// Get the captured output.
1081    pub fn output(&self) -> &str {
1082        &self.output
1083    }
1084
1085    /// Drain and return the captured output, leaving the buffer empty.
1086    /// Used by the async-builtin dispatch path to forward closure output
1087    /// from a child VM back to its parent.
1088    pub fn take_output(&mut self) -> String {
1089        std::mem::take(&mut self.output)
1090    }
1091
1092    /// Append text to this VM's captured output. Used to forward output
1093    /// from child VMs (e.g. closures invoked via `call_closure_pub`)
1094    /// back into the parent stream.
1095    pub fn append_output(&mut self, text: &str) {
1096        self.output.push_str(text);
1097    }
1098
1099    pub(crate) fn pop(&mut self) -> Result<VmValue, VmError> {
1100        self.stack.pop().ok_or(VmError::StackUnderflow)
1101    }
1102
1103    pub(crate) fn peek(&self) -> Result<&VmValue, VmError> {
1104        self.stack.last().ok_or(VmError::StackUnderflow)
1105    }
1106
1107    pub(crate) fn const_str(c: &Constant) -> Result<&str, VmError> {
1108        match c {
1109            Constant::String(s) => Ok(s.as_str()),
1110            _ => Err(VmError::TypeError("expected string constant".into())),
1111        }
1112    }
1113
1114    pub(crate) fn release_sync_guards_for_current_scope(&mut self) {
1115        let depth = self.env.scope_depth();
1116        self.held_sync_guards
1117            .retain(|guard| guard.env_scope_depth < depth);
1118        // A `scope { }` torn down without a normal `TaskScopeExit` (break /
1119        // continue out of it) leaves a dangling nursery — cancel its tasks.
1120        self.cancel_task_scopes_where(|s| s.env_scope_depth >= depth);
1121    }
1122
1123    pub(crate) fn release_sync_guards_after_unwind(
1124        &mut self,
1125        frame_depth: usize,
1126        env_scope_depth: usize,
1127    ) {
1128        self.held_sync_guards.retain(|guard| {
1129            guard.frame_depth <= frame_depth && guard.env_scope_depth <= env_scope_depth
1130        });
1131        // Cancel nurseries opened above the catch handler (a `throw` unwound
1132        // past their `TaskScopeExit`).
1133        self.cancel_task_scopes_where(|s| {
1134            !(s.frame_depth <= frame_depth && s.env_scope_depth <= env_scope_depth)
1135        });
1136    }
1137
1138    pub(crate) fn release_sync_guards_for_frame(&mut self, frame_depth: usize) {
1139        self.held_sync_guards
1140            .retain(|guard| guard.frame_depth != frame_depth);
1141        // Cancel any nursery whose `scope {}` block belonged to the frame being
1142        // torn down (e.g. a `return` jumped past its `TaskScopeExit`).
1143        self.cancel_task_scopes_where(|s| s.frame_depth == frame_depth);
1144    }
1145
1146    pub(crate) fn adopt_sync_permit_for_current_scope(
1147        &mut self,
1148        permit: crate::value::VmSyncPermitHandle,
1149    ) {
1150        if permit.is_released()
1151            || self
1152                .held_sync_guards
1153                .iter()
1154                .any(|guard| guard._permit.same_lease(&permit))
1155        {
1156            return;
1157        }
1158        self.held_sync_guards
1159            .push(crate::synchronization::VmSyncHeldGuard {
1160                _permit: permit,
1161                frame_depth: self.frames.len(),
1162                env_scope_depth: self.env.scope_depth(),
1163            });
1164    }
1165
1166    /// Deregister a task id from every open nursery (it was explicitly
1167    /// `await`ed, so it must not be double-joined or cancelled at scope exit).
1168    pub(crate) fn deregister_task_from_scopes(&mut self, id: &str) {
1169        for scope in &mut self.task_scopes {
1170            scope.task_ids.retain(|t| t != id);
1171        }
1172    }
1173
1174    /// Cancel and remove every task scope matching `doomed`, aborting its bound
1175    /// tasks (used when a `scope {}` is torn down without a normal join).
1176    fn cancel_task_scopes_where<F: Fn(&TaskScope) -> bool>(&mut self, doomed: F) {
1177        let mut i = 0;
1178        while i < self.task_scopes.len() {
1179            if doomed(&self.task_scopes[i]) {
1180                let scope = self.task_scopes.remove(i);
1181                for id in &scope.task_ids {
1182                    if let Some(task) = self.spawned_tasks.remove(id) {
1183                        task.cancel_token
1184                            .store(true, std::sync::atomic::Ordering::SeqCst);
1185                        task.handle.abort();
1186                    }
1187                }
1188            } else {
1189                i += 1;
1190            }
1191        }
1192    }
1193
1194    /// Total live permits this VM already holds for `kind:key`. The held-set is
1195    /// tiny (bounded by lexical nesting and explicit sync acquisitions), so this
1196    /// scan is cheap and only runs on the rare blocking-acquire path.
1197    pub(crate) fn held_permits_for(&self, kind: &str, key: &str) -> u32 {
1198        let own: u32 = self
1199            .held_sync_guards
1200            .iter()
1201            .filter(|guard| {
1202                !guard._permit.is_released()
1203                    && guard._permit.kind() == kind
1204                    && guard._permit.key() == key
1205            })
1206            .map(|guard| guard._permit.permits())
1207            .sum();
1208        let inherited: u32 = self
1209            .inherited_held_keys
1210            .iter()
1211            .filter(|held| held.kind == kind && held.key == key)
1212            .map(|held| held.permits)
1213            .sum();
1214        own + inherited
1215    }
1216
1217    /// Every live sync permit held by this VM *and* its suspended ancestors: the
1218    /// transitive held-set seen by an inline child.
1219    pub(crate) fn combined_held_keys(&self) -> Vec<crate::synchronization::VmSyncHeldKey> {
1220        let mut keys: Vec<crate::synchronization::VmSyncHeldKey> = self
1221            .held_sync_guards
1222            .iter()
1223            .filter_map(|guard| crate::synchronization::VmSyncHeldKey::from_permit(&guard._permit))
1224            .collect();
1225        keys.extend(self.inherited_held_keys.iter().cloned());
1226        keys
1227    }
1228
1229    /// Clone a child VM for an **inline, same-task** execution (an async builtin
1230    /// awaited while this VM is parked, or a user closure that builtin runs and
1231    /// awaits). The child inherits this VM's transitive held-lock keys so a
1232    /// re-acquire of a parent-held lock is caught as a self-deadlock
1233    /// (HARN-ORC-011). Use plain `child_vm()` for new concurrent tasks.
1234    pub(crate) fn child_vm_inline(&self) -> Vm {
1235        let mut child = self.child_vm();
1236        child.inherited_held_keys = Arc::new(self.combined_held_keys());
1237        child
1238    }
1239}
1240
1241impl Drop for Vm {
1242    fn drop(&mut self) {
1243        if let Some(coverage) = self.coverage.take() {
1244            crate::coverage::merge_into_global(coverage);
1245        }
1246        self.cancel_spawned_tasks();
1247    }
1248}
1249
1250impl Default for Vm {
1251    fn default() -> Self {
1252        Self::new()
1253    }
1254}
1255
1256#[cfg(test)]
1257mod tests {
1258
1259    use super::*;
1260
1261    fn baseline_with_stdlib(source: &str) -> VmBaseline {
1262        let mut vm = Vm::new();
1263        crate::register_vm_stdlib(&mut vm);
1264        vm.set_source_info("baseline_test.harn", source);
1265        vm.set_global(
1266            "stable_global",
1267            VmValue::String(arcstr::ArcStr::from("baseline")),
1268        );
1269        vm.baseline()
1270    }
1271
1272    #[test]
1273    fn vm_baseline_instantiates_clean_mutable_execution_state() {
1274        let baseline = baseline_with_stdlib("pipeline main() { __io_println(stable_global) }");
1275
1276        let mut dirty = baseline.instantiate();
1277        dirty.stack.push(VmValue::Int(42));
1278        dirty.output.push_str("dirty");
1279        dirty.task_counter = 9;
1280        dirty.runtime_context_counter = 7;
1281        dirty
1282            .error_stack_trace
1283            .push(("main".to_string(), 1, 1, None));
1284
1285        let clean = baseline.instantiate();
1286        assert!(clean.stack.is_empty());
1287        assert!(clean.output.is_empty());
1288        assert!(clean.frames.is_empty());
1289        assert!(clean.exception_handlers.is_empty());
1290        assert!(clean.spawned_tasks.is_empty());
1291        assert!(clean.held_sync_guards.is_empty());
1292        assert_eq!(clean.task_counter, 0);
1293        assert_eq!(clean.runtime_context_counter, 0);
1294        assert!(clean.deadlines.is_empty());
1295        assert!(clean.cancel_token.is_none());
1296        assert!(clean.interrupt_handlers.is_empty());
1297        assert!(clean.error_stack_trace.is_empty());
1298        assert!(clean.bridge.is_none());
1299        assert!(clean
1300            .globals
1301            .get("stable_global")
1302            .is_some_and(|value| value.display() == "baseline"));
1303    }
1304
1305    #[tokio::test]
1306    async fn inline_child_inherits_held_lock_keys_but_concurrent_child_does_not() {
1307        let mut parent = Vm::new();
1308        let permit = parent
1309            .sync_runtime
1310            .acquire("mutex", "v:test", 1, 1, None, None)
1311            .await
1312            .unwrap()
1313            .unwrap();
1314        parent
1315            .held_sync_guards
1316            .push(crate::synchronization::VmSyncHeldGuard {
1317                _permit: permit,
1318                frame_depth: 0,
1319                env_scope_depth: 0,
1320            });
1321        assert_eq!(parent.held_permits_for("mutex", "v:test"), 1);
1322
1323        // An inline child (async builtin awaited while the parent is parked, or
1324        // a closure the builtin runs inline) inherits the held key, so a
1325        // re-acquire is caught as a cross-context self-deadlock (HARN-ORC-011)
1326        // — even transitively through a further inline child.
1327        let inline = parent.child_vm_inline();
1328        assert_eq!(inline.held_permits_for("mutex", "v:test"), 1);
1329        assert_eq!(
1330            inline.child_vm_inline().held_permits_for("mutex", "v:test"),
1331            1
1332        );
1333
1334        // A new concurrent task (spawn / parallel / trigger) does NOT inherit:
1335        // blocking on a parent-held lock there is legitimately resolvable, so
1336        // flagging it would be a false positive.
1337        let concurrent = parent.child_vm();
1338        assert_eq!(concurrent.held_permits_for("mutex", "v:test"), 0);
1339    }
1340
1341    #[test]
1342    fn vm_reports_effective_runtime_limits() {
1343        let vm = Vm::new();
1344
1345        assert_eq!(vm.runtime_limits(), RuntimeLimits::default());
1346        assert_eq!(
1347            vm.runtime_limit_report().entries.len(),
1348            crate::RUNTIME_LIMIT_DESCRIPTIONS.len()
1349        );
1350        assert_eq!(vm.child_vm().runtime_limits(), vm.runtime_limits());
1351        assert_eq!(
1352            vm.baseline().instantiate().runtime_limits(),
1353            vm.runtime_limits()
1354        );
1355    }
1356
1357    #[tokio::test(flavor = "current_thread")]
1358    async fn vm_baseline_rebinds_shared_state_builtins_per_instance() {
1359        let local = tokio::task::LocalSet::new();
1360        local
1361            .run_until(async {
1362                let source = r#"
1363pipeline main() {
1364  const cell = shared_cell({scope: "task_group", key: "turn", initial: 0})
1365  __io_println(shared_get(cell))
1366  shared_set(cell, shared_get(cell) + 1)
1367}"#;
1368                let chunk = crate::compile_source(source).expect("compile");
1369                let baseline = baseline_with_stdlib(source);
1370
1371                let mut first = baseline.instantiate();
1372                first.execute(&chunk).await.expect("first execute");
1373                assert_eq!(first.output(), "0\n");
1374
1375                let mut second = baseline.instantiate();
1376                second.execute(&chunk).await.expect("second execute");
1377                assert_eq!(
1378                    second.output(),
1379                    "0\n",
1380                    "shared state created by the first VM must not leak into the next baseline instance"
1381                );
1382            })
1383            .await;
1384    }
1385}