ternlang-core 1.2.9

pub mod bet;

use crate::trit::Trit;
use crate::vm::bet::{unpack_trits, BetFault};

use std::fmt;
use std::sync::Arc;

// ─── Remote transport trait ───────────────────────────────────────────────────

pub trait RemoteTransport: Send + Sync {
    fn remote_send(&self, node_addr: &str, agent_id: usize, trit: i8) -> std::io::Result<()>;
    fn remote_await(&self, node_addr: &str, agent_id: usize) -> std::io::Result<i8>;
}

/// Maximum call depth before the VM returns a `CallStackOverflow` error.
/// Prevents OOM/freeze when programs contain unbounded recursion or
/// mutual recursion across imported modules.
const MAX_CALL_DEPTH: usize = 4096;

#[derive(Debug, PartialEq, Eq)]
pub enum VmError {
    StackUnderflow,
    BetFault(BetFault),
    Halt,
    InvalidOpcode(u8),
    InvalidRegister(u8),
    PcOutOfBounds(usize),
    TypeMismatch { expected: String, found: String },
    // ── Tensor errors ────────────────────────────────────────────────────────
    TensorIndexOutOfBounds { tensor_id: usize, index: usize, size: usize },
    TensorNotAllocated(usize),
    // ── Agent errors ─────────────────────────────────────────────────────────
    AgentTypeNotRegistered(u16),
    AgentIdInvalid(usize),
    RuntimeError(String),
    CallStackOverflow,
    // ── File I/O errors ──────────────────────────────────────────────────────
    FileOpenError(String),
    FileReadError(String),
    FileWriteError(String),
    FileNotOpen(usize),
    AssertionFailed,
}

impl fmt::Display for VmError {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            VmError::StackUnderflow =>
                write!(f, "[BET-001] Stack underflow — you tried to pop a truth that wasn't there.\n          → details: stdlib/errors/BET-001.tern  |  ternlang errors BET-001"),
            VmError::BetFault(fault) =>
                write!(f, "[BET-002] BET encoding fault: {fault:?}. The 0b00 state is forbidden — only 01/10/11 are valid trit bits.\n          → details: stdlib/errors/BET-002.tern  |  ternlang errors BET-002"),
            VmError::Halt =>
                write!(f, "[BET-003] VM halted cleanly. Execution reached the end. This is not an error — this is peace.\n          → details: stdlib/errors/BET-003.tern  |  ternlang errors BET-003"),
            VmError::InvalidOpcode(op) =>
                write!(f, "[BET-004] Unknown opcode 0x{op:02x} — the machine has never seen this instruction. Delete cached .ternbc files and recompile.\n          → details: stdlib/errors/BET-004.tern  |  ternlang errors BET-004"),
            VmError::InvalidRegister(reg) =>
                write!(f, "[BET-005] Register {reg} is out of range. The BET has exactly 27 registers (0–26). That's 3³. No more.\n          → details: stdlib/errors/BET-005.tern  |  ternlang errors BET-005"),
            VmError::PcOutOfBounds(pc) =>
                write!(f, "[BET-006] PC {pc} is out of bounds — you jumped outside the known universe. Recompile from source.\n          → details: stdlib/errors/BET-006.tern  |  ternlang errors BET-006"),
            VmError::TypeMismatch { expected, found } =>
                write!(f, "[BET-007] Runtime type mismatch — expected {expected} but found {found}. Square peg, round hole.\n          → details: stdlib/errors/BET-007.tern  |  ternlang errors BET-007"),
            VmError::TensorIndexOutOfBounds { tensor_id, index, size } =>
                write!(f, "[BET-008] Tensor[{tensor_id}]: index {index} is out of bounds — tensor only has {size} element(s). Trittensors don't grow on access.\n          → details: stdlib/errors/BET-008.tern  |  ternlang errors BET-008"),
            VmError::TensorNotAllocated(idx) =>
                write!(f, "[BET-009] TensorRef({idx}) doesn't exist — you never allocated it. TALLOC first, then TIDX.\n          → details: stdlib/errors/BET-009.tern  |  ternlang errors BET-009"),
            VmError::AgentTypeNotRegistered(type_id) =>
                write!(f, "[BET-010] Agent type_id 0x{type_id:04x} was never registered. You can't spawn what was never declared.\n          → details: stdlib/errors/BET-010.tern  |  ternlang errors BET-010"),
            VmError::AgentIdInvalid(id) =>
                write!(f, "[BET-011] Agent #{id} doesn't exist — no agent was spawned at this ID. TSEND and TAWAIT require a live agent.\n          → details: stdlib/errors/BET-011.tern  |  ternlang errors BET-011"),
            VmError::RuntimeError(msg) =>
                write!(f, "[BET-012] Runtime error: {msg}"),
            VmError::CallStackOverflow =>
                write!(f, "[BET-013] Call stack overflow — max depth ({MAX_CALL_DEPTH}) exceeded. Infinite recursion or unbounded cross-module mutual calls detected.\n          → details: stdlib/errors/BET-013.tern  |  ternlang errors BET-013"),
            VmError::FileOpenError(e) =>
                write!(f, "[IO-001] File open error: {e}"),
            VmError::FileReadError(e) =>
                write!(f, "[IO-002] File read error: {e}"),
            VmError::FileWriteError(e) =>
                write!(f, "[IO-003] File write error: {e}"),
            VmError::FileNotOpen(id) =>
                write!(f, "[IO-004] File handle {id} is not open or was closed."),
            VmError::AssertionFailed =>
                write!(f, "[ASSERT-001] Assertion failed: an assert() condition evaluated to reject or tend."),
        }
    }
}

#[derive(Debug, Clone, PartialEq)]
pub enum Value {
    Trit(Trit),
    Int(i64),
    Float(f64),
    String(String),
    TensorRef(usize),
    TensorView {
        tensor_id: usize,
        offset: usize,
        length: usize,
        stride: usize,
    },
    AgentRef(usize, Option<String>),
    Struct(std::collections::HashMap<String, Value>),
}

impl Default for Value {
    fn default() -> Self {
        Value::Trit(Trit::Tend)
    }
}

enum TensorData {
    Trit(Vec<Trit>),
    PackedTrit(Vec<u8>, usize),
    Float(Vec<f64>),
    Int(Vec<i64>),
}

impl TensorData {
    fn len(&self) -> usize {
        match self {
            TensorData::Trit(v) => v.len(),
            TensorData::PackedTrit(_, len) => *len,
            TensorData::Float(v) => v.len(),
            TensorData::Int(v) => v.len(),
        }
    }
}

struct TensorInstance {
    data: TensorData,
    rows: usize,
    cols: usize,
}

struct AgentInstance {
    handler_addr: usize,
    mailbox: std::collections::VecDeque<Value>,
}

pub struct BetVm {
    /// Dynamic register file — grows on demand so programs with > 27 locals work correctly
    /// instead of silently dropping stores and returning zero on reads.
    registers: Vec<Value>,
    register_stack: Vec<Vec<Value>>,
    carry_reg: Trit,
    stack: Vec<Value>,
    call_stack: Vec<usize>,
    tensors: Vec<TensorInstance>,
    agents: Vec<AgentInstance>,
    agent_types: std::collections::HashMap<u16, usize>,
    pc: usize,
    code: Vec<u8>,
    node_id: String,
    pub sparse_dropped: bool,
    remote: Option<Arc<dyn RemoteTransport>>,
    open_files: Vec<Option<std::fs::File>>,
    bindings: std::collections::HashMap<usize, Value>,
    _instructions_count: u64,
    pub print_log: Vec<String>,
}

impl BetVm {
    pub fn new(code: Vec<u8>) -> Self {
        Self {
            registers: vec![Value::default(); 27],
            register_stack: Vec::new(),
            carry_reg: Trit::Tend,
            stack: Vec::new(),
            call_stack: Vec::new(),
            tensors: Vec::new(),
            agents: Vec::new(),
            agent_types: std::collections::HashMap::new(),
            pc: 0,
            code,
            node_id: "127.0.0.1".into(),
            sparse_dropped: false,
            remote: None,
            open_files: Vec::new(),
            bindings: std::collections::HashMap::new(),
            _instructions_count: 0,
            print_log: Vec::new(),
        }
    }


    /// Drain all lines printed by `print()`/`println()` during execution.
    pub fn take_output(&mut self) -> Vec<String> {
        std::mem::take(&mut self.print_log)
    }

    pub fn set_node_id(&mut self, node_id: String) {
        self.node_id = node_id;
    }

    pub fn set_remote(&mut self, transport: Arc<dyn RemoteTransport>) {
        self.remote = Some(transport);
    }

    pub fn register_agent_type(&mut self, type_id: u16, handler_addr: usize) {
        self.agent_types.insert(type_id, handler_addr);
    }

    pub fn peek_stack(&self) -> Option<Value> {
        self.stack.last().cloned()
    }

    pub fn get_registers(&self) -> Vec<Value> {
        self.registers.clone()
    }

    pub fn get_register(&self, reg: u8) -> Value {
        self.registers.get(reg as usize).cloned().unwrap_or_default()
    }

    pub fn node_id(&self) -> &str {
        &self.node_id
    }

    pub fn run(&mut self) -> Result<(), VmError> {
        loop {
            if self.pc >= self.code.len() { break; }
            let opcode = self.code[self.pc];
            self.pc += 1;

            match opcode {
                0x01 => { // Tpush
                    let packed = self.read_u8()?;
                    let trits = unpack_trits(&[packed], 1).map_err(VmError::BetFault)?;
                    self.stack.push(Value::Trit(trits[0]));
                }
                0x02 => { // Tadd
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match (a.clone(), b.clone()) {
                        (Value::Trit(av), Value::Trit(bv)) => {
                            let (sum, carry) = av + bv;
                            self.stack.push(Value::Trit(sum));
                            self.carry_reg = carry;
                        }
                        (Value::Int(av), Value::Int(bv)) => self.stack.push(Value::Int(av + bv)),
                        (Value::Float(av), Value::Float(bv)) => self.stack.push(Value::Float(av + bv)),
                        (Value::Int(av), Value::Trit(bv)) => self.stack.push(Value::Int(av + bv as i64)),
                        (Value::Trit(av), Value::Int(bv)) => self.stack.push(Value::Int(av as i64 + bv)),
                        (Value::Float(av), Value::Trit(bv)) => self.stack.push(Value::Float(av + (bv as i8 as f64))),
                        (Value::Trit(av), Value::Float(bv)) => self.stack.push(Value::Float((av as i8 as f64) + bv)),
                        (Value::Float(av), Value::Int(bv)) => self.stack.push(Value::Float(av + (bv as f64))),
                        (Value::Int(av), Value::Float(bv)) => self.stack.push(Value::Float((av as f64) + bv)),
                        // PARSER-STR-001: string concatenation via + operator
                        (Value::String(av), Value::String(bv)) => self.stack.push(Value::String(av + &bv)),
                        _ => return Err(VmError::TypeMismatch { expected: "Numeric".into(), found: format!("{:?}", (a, b)) }),
                    }
                }
                0x03 => { // Tmul
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match (a.clone(), b.clone()) {
                        (Value::Trit(av), Value::Trit(bv)) => self.stack.push(Value::Trit(av * bv)),
                        (Value::Int(av), Value::Int(bv)) => self.stack.push(Value::Int(av * bv)),
                        (Value::Float(av), Value::Float(bv)) => self.stack.push(Value::Float(av * bv)),
                        (Value::Int(av), Value::Trit(bv)) => self.stack.push(Value::Int(av * bv as i64)),
                        (Value::Trit(av), Value::Int(bv)) => self.stack.push(Value::Int(av as i64 * bv)),
                        (Value::Float(av), Value::Trit(bv)) => self.stack.push(Value::Float(av * (bv as i8 as f64))),
                        (Value::Trit(av), Value::Float(bv)) => self.stack.push(Value::Float((av as i8 as f64) * bv)),
                        (Value::Float(av), Value::Int(bv)) => self.stack.push(Value::Float(av * (bv as f64))),
                        (Value::Int(av), Value::Float(bv)) => self.stack.push(Value::Float((av as f64) * bv)),
                        _ => return Err(VmError::TypeMismatch { expected: "Numeric".into(), found: format!("{:?}", (a, b)) }),
                    }
                }
                0x04 => { // Tneg
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match a.clone() {
                        Value::Trit(av) => self.stack.push(Value::Trit(-av)),
                        Value::Int(av) => self.stack.push(Value::Int(-av)),
                        Value::Float(av) => self.stack.push(Value::Float(-av)),
                        _ => return Err(VmError::TypeMismatch { expected: "Numeric".into(), found: format!("{:?}", a) }),
                    }
                }
                0x05 => { // TjmpPos — jumps if top is positive (Affirm / any int > 0)
                    let addr = self.read_u16()?;
                    let val = self.stack.last().ok_or(VmError::StackUnderflow)?;
                    let is_pos = match val {
                        Value::Trit(Trit::Affirm) => true,
                        Value::Int(v) => *v > 0,
                        Value::Float(f) => *f > 0.0,
                        _ => false,
                    };
                    if is_pos { self.pc = addr as usize; }
                }
                0x06 => { // TjmpZero — jumps if top is zero (Tend / 0)
                    let addr = self.read_u16()?;
                    let val = self.stack.last().ok_or(VmError::StackUnderflow)?;
                    let is_zero = match val {
                        Value::Trit(Trit::Tend) => true,
                        Value::Int(v) => *v == 0,
                        Value::Float(f) => *f == 0.0,
                        _ => false,
                    };
                    if is_zero { self.pc = addr as usize; }
                }
                0x07 => { // TjmpNeg — jumps if top is negative (Reject / any int < 0)
                    let addr = self.read_u16()?;
                    let val = self.stack.last().ok_or(VmError::StackUnderflow)?;
                    let is_neg = match val {
                        Value::Trit(Trit::Reject) => true,
                        Value::Int(v) => *v < 0,
                        Value::Float(f) => *f < 0.0,
                        _ => false,
                    };
                    if is_neg { self.pc = addr as usize; }
                }
                0x08 => { // Tstore
                    let reg = self.read_u8()? as usize;
                    let val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    self.bindings.remove(&reg);
                    if reg >= self.registers.len() { self.registers.resize(reg + 1, Value::default()); }
                    self.registers[reg] = val;
                }
                0x09 => { // Tload
                    let reg = self.read_u8()? as usize;
                    let val = self.bindings.get(&reg).cloned().unwrap_or_else(|| {
                        if reg >= self.registers.len() { self.registers.resize(reg + 1, Value::default()); }
                        self.registers[reg].clone()
                    });
                    self.stack.push(val);
                }
                0x0a => { // Tdup
                    let val = self.stack.last().ok_or(VmError::StackUnderflow)?;
                    self.stack.push(val.clone());
                }
                0x0b => { // Tjmp
                    let addr = self.read_u16()?;
                    self.pc = addr as usize;
                }
                0x0c => { // Tpop
                    self.stack.pop().ok_or(VmError::StackUnderflow)?;
                }
                0x0e => { // Tcons
                    let b_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    
                    let a = match a_val {
                        Value::Trit(t) => t,
                        Value::Int(v) if v == 1 => Trit::Affirm,
                        Value::Int(v) if v == 0 => Trit::Tend,
                        Value::Int(v) if v == -1 => Trit::Reject,
                        _ => return Err(VmError::TypeMismatch { expected: "Trit or Int(-1..1)".into(), found: format!("{:?}", a_val) }),
                    };
                    let b = match b_val {
                        Value::Trit(t) => t,
                        Value::Int(v) if v == 1 => Trit::Affirm,
                        Value::Int(v) if v == 0 => Trit::Tend,
                        Value::Int(v) if v == -1 => Trit::Reject,
                        _ => return Err(VmError::TypeMismatch { expected: "Trit or Int(-1..1)".into(), found: format!("{:?}", b_val) }),
                    };

                    let result = match (a, b) {
                        (Trit::Affirm, Trit::Affirm) => Trit::Affirm,
                        (Trit::Reject, Trit::Reject) => Trit::Reject,
                        (Trit::Tend, x) => x,
                        (x, Trit::Tend) => x,
                        _ => Trit::Tend,
                    };
                    self.stack.push(Value::Trit(result));
                }
                0x0f => { // Talloc (trit tensor)
                    let rows = self.read_u32()? as usize;
                    let cols = self.read_u32()? as usize;
                    let size = rows * cols;
                    let idx = self.tensors.len();
                    self.tensors.push(TensorInstance {
                        data: TensorData::Trit(vec![Trit::Tend; size]),
                        rows,
                        cols,
                    });
                    self.stack.push(Value::TensorRef(idx));
                }
                0x3c => { // Talloc_Int (int tensor)
                    let rows = self.read_u32()? as usize;
                    let cols = self.read_u32()? as usize;
                    let size = rows * cols;
                    let idx = self.tensors.len();
                    self.tensors.push(TensorInstance {
                        data: TensorData::Int(vec![0i64; size]),
                        rows,
                        cols,
                    });
                    self.stack.push(Value::TensorRef(idx));
                }
                0x3d => { // Talloc_Float (float tensor)
                    let rows = self.read_u32()? as usize;
                    let cols = self.read_u32()? as usize;
                    let size = rows * cols;
                    let idx = self.tensors.len();
                    self.tensors.push(TensorInstance {
                        data: TensorData::Float(vec![0.0f64; size]),
                        rows,
                        cols,
                    });
                    self.stack.push(Value::TensorRef(idx));
                }
                0x10 => { // Tcall
                    if self.call_stack.len() >= MAX_CALL_DEPTH {
                        return Err(VmError::CallStackOverflow);
                    }
                    let addr = self.read_u16()? as usize;
                    self.register_stack.push(self.registers.clone());
                    self.call_stack.push(self.pc);
                    self.pc = addr;
                }
                0x11 => { // Tret
                    if let Some(prev) = self.register_stack.pop() {
                        self.registers = prev;
                    }
                    match self.call_stack.pop() {
                        Some(ret) => self.pc = ret,
                        None => return Ok(()),
                    }
                }
                0x14 => { // Tless
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match (a.clone(), b.clone()) {
                        (Value::Int(x), Value::Int(y)) => {
                            let r = if x < y { Trit::Affirm } else if x == y { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Float(x), Value::Float(y)) => {
                            let r = if x < y { Trit::Affirm } else if (x - y).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Int(x), Value::Trit(y)) => {
                            let bv = y as i64;
                            let r = if x < bv { Trit::Affirm } else if x == bv { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Trit(x), Value::Int(y)) => {
                            let av = x as i64;
                            let r = if av < y { Trit::Affirm } else if av == y { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Int(av), Value::Float(bv)) => {
                            let a_val = av as f64;
                            let r = if a_val < bv { Trit::Affirm } else if (a_val - bv).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Float(av), Value::Int(bv)) => {
                            let b_val = bv as f64;
                            let r = if av < b_val { Trit::Affirm } else if (av - b_val).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Trit(av), Value::Float(bv)) => {
                            let a_val = av as i8 as f64;
                            let r = if a_val < bv { Trit::Affirm } else if (a_val - bv).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Float(av), Value::Trit(bv)) => {
                            let b_val = bv as i8 as f64;
                            let r = if av < b_val { Trit::Affirm } else if (av - b_val).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Trit(x), Value::Trit(y)) => {
                            let av = x as i64;
                            let bv = y as i64;
                            let r = if av < bv { Trit::Affirm } else if av == bv { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        _ => return Err(VmError::TypeMismatch { expected: "Numeric".into(), found: format!("{:?}", (a, b)) }),
                    }
                }
                0x15 => { // Tgreater
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match (a.clone(), b.clone()) {
                        (Value::Int(x), Value::Int(y)) => {
                            let r = if x > y { Trit::Affirm } else if x == y { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Float(x), Value::Float(y)) => {
                            let r = if x > y { Trit::Affirm } else if (x - y).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Int(x), Value::Trit(y)) => {
                            let bv = y as i64;
                            let r = if x > bv { Trit::Affirm } else if x == bv { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Trit(x), Value::Int(y)) => {
                            let av = x as i64;
                            let r = if av > y { Trit::Affirm } else if av == y { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Int(av), Value::Float(bv)) => {
                            let a_val = av as f64;
                            let r = if a_val > bv { Trit::Affirm } else if (a_val - bv).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Float(av), Value::Int(bv)) => {
                            let b_val = bv as f64;
                            let r = if av > b_val { Trit::Affirm } else if (av - b_val).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Trit(av), Value::Float(bv)) => {
                            let a_val = av as i8 as f64;
                            let r = if a_val > bv { Trit::Affirm } else if (a_val - bv).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Float(av), Value::Trit(bv)) => {
                            let b_val = bv as i8 as f64;
                            let r = if av > b_val { Trit::Affirm } else if (av - b_val).abs() < f64::EPSILON { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        (Value::Trit(x), Value::Trit(y)) => {
                            let av = x as i64;
                            let bv = y as i64;
                            let r = if av > bv { Trit::Affirm } else if av == bv { Trit::Tend } else { Trit::Reject };
                            self.stack.push(Value::Trit(r));
                        }
                        _ => return Err(VmError::TypeMismatch { expected: "Numeric".into(), found: format!("{:?}", (a, b)) }),
                    }
                }
                0x16 => { // Teq
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let is_eq = match (a.clone(), b.clone()) {
                        (Value::Int(av), Value::Trit(bv)) => av == bv as i64,
                        (Value::Trit(av), Value::Int(bv)) => av as i64 == bv,
                        (Value::Float(av), Value::Float(bv)) => (av - bv).abs() < f64::EPSILON,
                        (Value::Float(av), Value::Trit(bv)) => (av - (bv as i8 as f64)).abs() < f64::EPSILON,
                        (Value::Trit(av), Value::Float(bv)) => ((av as i8 as f64) - bv).abs() < f64::EPSILON,
                        (Value::Float(av), Value::Int(bv)) => (av - (bv as f64)).abs() < f64::EPSILON,
                        (Value::Int(av), Value::Float(bv)) => ((av as f64) - bv).abs() < f64::EPSILON,
                        _ => a == b,
                    };
                    let r = if is_eq { Trit::Affirm } else { Trit::Reject };
                    self.stack.push(Value::Trit(r));
                }
                0x17 => { // TpushInt
                    let mut b = [0u8; 8];
                    for i in 0..8 { b[i] = self.read_u8()?; }
                    self.stack.push(Value::Int(i64::from_le_bytes(b)));
                }
                0x18 => { // TaddInt
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match (a.clone(), b.clone()) {
                        (Value::Int(x), Value::Int(y)) => self.stack.push(Value::Int(x + y)),
                        _ => return Err(VmError::TypeMismatch { expected: "Int".into(), found: format!("{:?}", (a, b)) }),
                    }
                }
                0x19 => { // TpushFloat
                    let mut b = [0u8; 8];
                    for i in 0..8 { b[i] = self.read_u8()?; }
                    self.stack.push(Value::Float(f64::from_le_bytes(b)));
                }
                0x1e => { // Tdiv
                    let b_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match (a_val.clone(), b_val.clone()) {
                        (Value::Int(av), Value::Int(bv)) => {
                            if bv == 0 { return Err(VmError::RuntimeError("Division by zero".into())); }
                            self.stack.push(Value::Int(av / bv));
                        }
                        (Value::Float(av), Value::Float(bv)) => {
                            if bv == 0.0 { return Err(VmError::RuntimeError("Division by zero".into())); }
                            self.stack.push(Value::Float(av / bv));
                        }
                        (Value::Int(av), Value::Trit(bv)) => {
                            let b = bv as i64;
                            if b == 0 { return Err(VmError::RuntimeError("Division by zero".into())); }
                            self.stack.push(Value::Int(av / b));
                        }
                        (Value::Trit(av), Value::Int(bv)) => {
                            if bv == 0 { return Err(VmError::RuntimeError("Division by zero".into())); }
                            self.stack.push(Value::Int(av as i64 / bv));
                        }
                        (Value::Float(av), Value::Trit(bv)) => {
                            let b = bv as i8 as f64;
                            if b == 0.0 { return Err(VmError::RuntimeError("Division by zero".into())); }
                            self.stack.push(Value::Float(av / b));
                        }
                        (Value::Trit(av), Value::Float(bv)) => {
                            if bv == 0.0 { return Err(VmError::RuntimeError("Division by zero".into())); }
                            self.stack.push(Value::Float(av as i8 as f64 / bv));
                        }
                        (Value::Float(av), Value::Int(bv)) => {
                            let b = bv as f64;
                            if b == 0.0 { return Err(VmError::RuntimeError("Division by zero".into())); }
                            self.stack.push(Value::Float(av / b));
                        }
                        (Value::Int(av), Value::Float(bv)) => {
                            if bv == 0.0 { return Err(VmError::RuntimeError("Division by zero".into())); }
                            self.stack.push(Value::Float(av as f64 / bv));
                        }
                        _ => return Err(VmError::TypeMismatch { expected: "Numeric".into(), found: format!("{:?}", (a_val, b_val)) }),
                    }
                }
                0x1f => { // Tmod
                    let b_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match (a_val.clone(), b_val.clone()) {
                        (Value::Int(av), Value::Int(bv)) => {
                            if bv == 0 { return Err(VmError::RuntimeError("Modulo by zero".into())); }
                            self.stack.push(Value::Int(av % bv));
                        }
                        (Value::Int(av), Value::Trit(bv)) => {
                            let b = bv as i64;
                            if b == 0 { return Err(VmError::RuntimeError("Modulo by zero".into())); }
                            self.stack.push(Value::Int(av % b));
                        }
                        (Value::Trit(av), Value::Int(bv)) => {
                            if bv == 0 { return Err(VmError::RuntimeError("Modulo by zero".into())); }
                            self.stack.push(Value::Int(av as i64 % bv));
                        }
                        (Value::Float(av), Value::Float(bv)) => {
                             if bv == 0.0 { return Err(VmError::RuntimeError("Modulo by zero".into())); }
                             self.stack.push(Value::Float(av % bv));
                        }
                        (Value::Float(av), Value::Trit(bv)) => {
                             let b = bv as i8 as f64;
                             if b == 0.0 { return Err(VmError::RuntimeError("Modulo by zero".into())); }
                             self.stack.push(Value::Float(av % b));
                        }
                        (Value::Trit(av), Value::Float(bv)) => {
                             if bv == 0.0 { return Err(VmError::RuntimeError("Modulo by zero".into())); }
                             self.stack.push(Value::Float(av as i8 as f64 % bv));
                        }
                        (Value::Float(av), Value::Int(bv)) => {
                             let b = bv as f64;
                             if b == 0.0 { return Err(VmError::RuntimeError("Modulo by zero".into())); }
                             self.stack.push(Value::Float(av % b));
                        }
                        (Value::Int(av), Value::Float(bv)) => {
                             if bv == 0.0 { return Err(VmError::RuntimeError("Modulo by zero".into())); }
                             self.stack.push(Value::Float(av as f64 % bv));
                        }
                        _ => return Err(VmError::TypeMismatch { expected: "Int or Trit".into(), found: format!("{:?}", (a_val, b_val)) }),
                    }
                }
                0x20 => { // Tprint
                    let val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let line = match &val {
                        Value::Trit(t) => format!("{:?}", t),
                        Value::Int(i) => format!("{}", i),
                        Value::Float(f) => format!("{}", f),
                        Value::String(s) => s.clone(),
                        Value::TensorRef(idx) => format!("TensorRef({})", idx),
                        Value::TensorView { tensor_id, offset, length, .. } => format!("TensorView({}[{}..{}])", tensor_id, offset, offset + length),
                        Value::AgentRef(idx, addr) => format!("AgentRef({}, {:?})", idx, addr),
                        Value::Struct(fields) => format!("Struct({:?})", fields),
                    };
                    println!("{}", line);
                    self.print_log.push(line);
                }
                0x21 => { // TpushString
                    let len = self.read_u16()? as usize;
                    let mut bytes = vec![0u8; len];
                    for i in 0..len { bytes[i] = self.read_u8()?; }
                    let s = String::from_utf8(bytes).map_err(|_| VmError::RuntimeError("Invalid UTF-8 string".into()))?;
                    self.stack.push(Value::String(s));
                }
                0x22 => { // Tidx
                    let col = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let row = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let rf = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let r = match row { Value::Int(v) => v, Value::Trit(t) => t as i64, _ => return Err(VmError::TypeMismatch { expected: "Int or Trit".into(), found: format!("{:?}", row) }) };
                    let c = match col { Value::Int(v) => v, Value::Trit(t) => t as i64, _ => return Err(VmError::TypeMismatch { expected: "Int or Trit".into(), found: format!("{:?}", col) }) };
                    
                    let (idx, pos) = self.get_pos(&rf, r, c)?;
                    let tensor = &self.tensors[idx];
                    let data_len = tensor.data.len();
                    if pos >= data_len {
                        return Err(VmError::TensorIndexOutOfBounds { tensor_id: idx, index: pos, size: data_len });
                    }
                    let pushed = match &tensor.data {
                        TensorData::Trit(v) => Value::Trit(v[pos]),
                        TensorData::PackedTrit(v, _) => {
                            let byte_idx = pos / 5;
                            let trit_idx = pos % 5;
                            let trits = crate::trit::unpack_5_trits(v[byte_idx]);
                            Value::Trit(trits[trit_idx])
                        }
                        TensorData::Float(v) => Value::Float(v[pos]),
                        TensorData::Int(v) => Value::Int(v[pos]),
                    };
                    self.stack.push(pushed);
                }
                0x23 => { // Tset
                    let val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let col = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let row = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let rf = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let r = match row { Value::Int(v) => v, Value::Trit(t) => t as i64, _ => return Err(VmError::TypeMismatch { expected: "Int or Trit".into(), found: format!("{:?}", row) }) };
                    let c = match col { Value::Int(v) => v, Value::Trit(t) => t as i64, _ => return Err(VmError::TypeMismatch { expected: "Int or Trit".into(), found: format!("{:?}", col) }) };
                    
                    let (idx, pos) = self.get_pos(&rf, r, c)?;
                    let tensor = &mut self.tensors[idx];
                    let data_len = tensor.data.len();
                    if pos >= data_len { return Err(VmError::TensorIndexOutOfBounds { tensor_id: idx, index: pos, size: data_len }); }
                    match (&mut tensor.data, val.clone()) {
                        (TensorData::Trit(v), Value::Trit(t)) => v[pos] = t,
                        (TensorData::Trit(v), Value::Int(i)) => v[pos] = if i > 0 { Trit::Affirm } else if i < 0 { Trit::Reject } else { Trit::Tend },
                        (TensorData::PackedTrit(v, _), val_v) => {
                            let byte_idx = pos / 5;
                            let trit_idx = pos % 5;
                            let mut trits = crate::trit::unpack_5_trits(v[byte_idx]);
                            trits[trit_idx] = match val_v {
                                Value::Trit(t) => t,
                                Value::Int(i) => if i > 0 { Trit::Affirm } else if i < 0 { Trit::Reject } else { Trit::Tend },
                                _ => return Err(VmError::TypeMismatch { expected: "Trit or Int".into(), found: format!("{:?}", val_v) }),
                            };
                            v[byte_idx] = crate::trit::pack_5_trits(trits);
                        }
                        (TensorData::Float(v), Value::Float(f)) => v[pos] = f,
                        (TensorData::Float(v), Value::Int(i)) => v[pos] = i as f64,
                        (TensorData::Int(v), Value::Int(i)) => v[pos] = i,
                        (TensorData::Int(v), Value::Float(f)) => v[pos] = f as i64,
                        (TensorData::Int(v), Value::Trit(t)) => v[pos] = t as i64,
                        _ => return Err(VmError::TypeMismatch { expected: "compatible value for tensor type".into(), found: format!("{:?}", val) }),
                    }
                }
                0x24 => { // Tshape — TensorRef/View → (rows, cols); String → (len, 1)
                    let rf = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    match rf {
                        Value::TensorRef(idx) => {
                            if idx >= self.tensors.len() { return Err(VmError::TensorNotAllocated(idx)); }
                            let tensor = &self.tensors[idx];
                            self.stack.push(Value::Int(tensor.rows as i64));
                            self.stack.push(Value::Int(tensor.cols as i64));
                        }
                        Value::TensorView { length, .. } => {
                            self.stack.push(Value::Int(length as i64));
                            self.stack.push(Value::Int(1));
                        }
                        Value::String(s) => {
                            let n = s.chars().count() as i64;
                            self.stack.push(Value::Int(n));
                            self.stack.push(Value::Int(1));
                        }
                        _ => return Err(VmError::TypeMismatch { expected: "TensorRef, TensorView, or String".into(), found: format!("{:?}", rf) }),
                    }
                }
                0x30 => { // Tspawn — (type_id) → AgentRef
                    let type_id = self.read_u16()?;
                    if let Some(&handler_addr) = self.agent_types.get(&type_id) {
                        let id = self.agents.len();
                        self.agents.push(AgentInstance { handler_addr, mailbox: Default::default() });
                        self.stack.push(Value::AgentRef(id, None));
                    } else {
                        return Err(VmError::AgentTypeNotRegistered(type_id));
                    }
                }
                0x31 => { // Tsend — msg, target → void
                    let msg = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let target = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let Value::AgentRef(id, None) = target {
                        if id < self.agents.len() {
                            self.agents[id].mailbox.push_back(msg);
                        } else {
                            return Err(VmError::AgentIdInvalid(id));
                        }
                    } else {
                        return Err(VmError::TypeMismatch { expected: "Local AgentRef".into(), found: format!("{:?}", target) });
                    }
                }
                0x32 => { // Tawait — target → result
                    let target = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let Value::AgentRef(id, None) = target {
                        if id < self.agents.len() {
                            if self.call_stack.len() >= MAX_CALL_DEPTH {
                                return Err(VmError::CallStackOverflow);
                            }
                            let handler_addr = self.agents[id].handler_addr;
                            let msg = self.agents[id].mailbox.pop_front().unwrap_or(Value::default());
                            // Synchronous handler dispatch — identical to TCALL
                            self.register_stack.push(self.registers.clone());
                            self.call_stack.push(self.pc);
                            self.pc = handler_addr;
                            self.stack.push(msg);
                        } else {
                            return Err(VmError::AgentIdInvalid(id));
                        }
                    } else {
                        return Err(VmError::TypeMismatch { expected: "Local AgentRef".into(), found: format!("{:?}", target) });
                    }
                }
                0x25 => { // TjmpEqInt — imm_int, imm_addr → peek, jumps if eq
                    let mut b = [0u8; 8];
                    for i in 0..8 { b[i] = self.read_u8()?; }
                    let target_val = i64::from_le_bytes(b);
                    let addr = self.read_u16()?;
                    let val = self.stack.last().ok_or(VmError::StackUnderflow)?;
                    let is_eq = match val {
                        Value::Int(v) => *v == target_val,
                        Value::Trit(t) => (*t as i8) as i64 == target_val,
                        _ => false,
                    };
                    if is_eq { self.pc = addr as usize; }
                }
                0x26 => { // TlessEqual
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let is_le = match (a.clone(), b.clone()) {
                        (Value::Int(x), Value::Int(y)) => x <= y,
                        (Value::Float(x), Value::Float(y)) => x <= y || (x - y).abs() < f64::EPSILON,
                        (Value::Int(x), Value::Trit(y)) => x <= y as i64,
                        (Value::Trit(x), Value::Int(y)) => (x as i64) <= y,
                        (Value::Trit(x), Value::Trit(y)) => (x as i64) <= (y as i64),
                        _ => false,
                    };
                    self.stack.push(Value::Trit(if is_le { Trit::Affirm } else { Trit::Reject }));
                }
                0x27 => { // TgreaterEqual
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let is_ge = match (a.clone(), b.clone()) {
                        (Value::Int(x), Value::Int(y)) => x >= y,
                        (Value::Float(x), Value::Float(y)) => x >= y || (x - y).abs() < f64::EPSILON,
                        (Value::Int(x), Value::Trit(y)) => x >= y as i64,
                        (Value::Trit(x), Value::Int(y)) => (x as i64) >= y,
                        (Value::Trit(x), Value::Trit(y)) => (x as i64) >= (y as i64),
                        _ => false,
                    };
                    self.stack.push(Value::Trit(if is_ge { Trit::Affirm } else { Trit::Reject }));
                }
                0x28 => { // Tand — min(a, b) in balanced ternary (logical AND)
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let to_trit = |v: Value| -> Result<Trit, VmError> {
                        match v {
                            Value::Trit(t) => Ok(t),
                            Value::Int(n) if n > 0 => Ok(Trit::Affirm),
                            Value::Int(0) => Ok(Trit::Tend),
                            Value::Int(_) => Ok(Trit::Reject),
                            other => Err(VmError::TypeMismatch { expected: "Trit or Int".into(), found: format!("{:?}", other) }),
                        }
                    };
                    let ta = to_trit(a)?;
                    let tb = to_trit(b)?;
                    let result = if (ta as i8) <= (tb as i8) { ta } else { tb };
                    self.stack.push(Value::Trit(result));
                }
                0x29 => { // Tor — max(a, b) in balanced ternary (logical OR)
                    let b = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let to_trit = |v: Value| -> Result<Trit, VmError> {
                        match v {
                            Value::Trit(t) => Ok(t),
                            Value::Int(n) if n > 0 => Ok(Trit::Affirm),
                            Value::Int(0) => Ok(Trit::Tend),
                            Value::Int(_) => Ok(Trit::Reject),
                            other => Err(VmError::TypeMismatch { expected: "Trit or Int".into(), found: format!("{:?}", other) }),
                        }
                    };
                    let ta = to_trit(a)?;
                    let tb = to_trit(b)?;
                    let result = if (ta as i8) >= (tb as i8) { ta } else { tb };
                    self.stack.push(Value::Trit(result));
                }
                0x2a => { // TjmpEqFloat — peek stack, jump if top Float equals embedded f64 literal
                    // Emitted by betbc.rs for float match-arm patterns (PARSER-002 fix).
                    // Layout: [opcode: u8] [target_f64: 8 bytes LE] [jump_addr: u16 LE]
                    // Peeks the top of stack (does NOT consume it), jumps if value matches
                    // within machine epsilon.
                    let mut fb = [0u8; 8];
                    for i in 0..8 { fb[i] = self.read_u8()?; }
                    let target_f = f64::from_le_bytes(fb);
                    let addr = self.read_u16()?;
                    let val = self.stack.last().ok_or(VmError::StackUnderflow)?;
                    if let Value::Float(f) = val {
                        if (f - target_f).abs() < 1e-9 {
                            self.pc = addr as usize;
                        }
                    }
                }
                0x33 => { // Topent — path_str, mode_int → handle_int
                    let mode = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let path = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let (Value::String(p), Value::Int(m)) = (path, mode) {
                        use std::fs::OpenOptions;
                        let mut options = OpenOptions::new();
                        match m {
                            0 => { options.read(true); } // Read
                            1 => { options.write(true).create(true).truncate(true); } // Write
                            2 => { options.append(true).create(true); } // Append
                            _ => return Err(VmError::RuntimeError(format!("Invalid file mode: {m}"))),
                        }
                        let file = options.open(&p).map_err(|e| VmError::FileOpenError(e.to_string()))?;
                        let handle = self.open_files.len();
                        self.open_files.push(Some(file));
                        self.stack.push(Value::Int(handle as i64));
                    } else {
                        return Err(VmError::TypeMismatch { expected: "String, Int".into(), found: "Unknown".into() });
                    }
                }
                0x34 => { // Treadt — handle_int → trit
                    let handle_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let Value::Int(h) = handle_val {
                        let h = h as usize;
                        if h >= self.open_files.len() || self.open_files[h].is_none() {
                            return Err(VmError::FileNotOpen(h));
                        }
                        let file = self.open_files[h].as_mut().unwrap();
                        let mut buf = [0u8; 1];
                        use std::io::Read;
                        match file.read_exact(&mut buf) {
                            Ok(_) => {
                                let t = match buf[0] {
                                    b'+' | b'1' => Trit::Affirm,
                                    b'-' => Trit::Reject,
                                    _ => Trit::Tend,
                                };
                                self.stack.push(Value::Trit(t));
                            }
                            Err(e) if e.kind() == std::io::ErrorKind::UnexpectedEof => {
                                self.stack.push(Value::Trit(Trit::Tend)); // EOF as Tend
                            }
                            Err(e) => return Err(VmError::FileReadError(e.to_string())),
                        }
                    } else {
                        return Err(VmError::TypeMismatch { expected: "Int".into(), found: format!("{:?}", handle_val) });
                    }
                }
                0x35 => { // Twritet — handle_int, trit → void
                    let t_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let h_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let (Value::Int(h), Value::Trit(t)) = (h_val, t_val) {
                        let h = h as usize;
                        if h >= self.open_files.len() || self.open_files[h].is_none() {
                            return Err(VmError::FileNotOpen(h));
                        }
                        let file = self.open_files[h].as_mut().unwrap();
                        let out = match t {
                            Trit::Affirm => b'+',
                            Trit::Reject => b'-',
                            Trit::Tend   => b'0',
                        };
                        use std::io::Write;
                        file.write_all(&[out]).map_err(|e| VmError::FileWriteError(e.to_string()))?;
                    } else {
                        return Err(VmError::TypeMismatch { expected: "Int, Trit".into(), found: "Unknown".into() });
                    }
                }
                0x36 => { // Tnodeid — push this node's runtime address as a String
                    // Defers the binding to runtime so --node-addr is respected.
                    // Previously, Expr::NodeId emitted a hardcoded "127.0.0.1:7373"
                    // string at compile time, ignoring vm.set_node_id().
                    self.stack.push(Value::String(self.node_id.clone()));
                }
                0x37 => { // Tassert
                    let val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let is_affirm = match val {
                        Value::Trit(Trit::Affirm) => true,
                        Value::Int(1) => true,
                        _ => false,
                    };
                    if !is_affirm {
                        return Err(VmError::AssertionFailed);
                    }
                }
                0x38 => { // TSparseMatmul (@sparseskip)
                    // Layout: [opcode] [a_rows: u8] [a_cols: u8] [b_cols: u8]
                    // Pops: A_tensor, B_tensor
                    // Pushes: Result_tensor
                    let a_rows = self.read_u8()? as usize;
                    let a_cols = self.read_u8()? as usize;
                    let b_cols = self.read_u8()? as usize;
                    let b_ref = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a_ref = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    
                    if let (Value::TensorRef(a_idx), Value::TensorRef(b_idx)) = (a_ref, b_ref) {
                        let (a_data, b_data) = {
                            let a = self.tensors.get(a_idx).ok_or(VmError::TensorNotAllocated(a_idx))?;
                            let b = self.tensors.get(b_idx).ok_or(VmError::TensorNotAllocated(b_idx))?;
                            
                            let a_data = match &a.data {
                                TensorData::Trit(v) => v,
                                _ => return Err(VmError::TypeMismatch { expected: "TritTensor".into(), found: "Other".into() }),
                            };
                            let b_data = match &b.data {
                                TensorData::Trit(v) => v,
                                _ => return Err(VmError::TypeMismatch { expected: "TritTensor".into(), found: "Other".into() }),
                            };
                            (a_data.clone(), b_data.clone())
                        };

                        let mut result = vec![Trit::Tend; a_rows * b_cols];
                        let mut skipped = false;

                        for i in 0..a_rows {
                            for k in 0..a_cols {
                                let a_val = a_data[i * a_cols + k];
                                if a_val == Trit::Tend {
                                    skipped = true;
                                    continue;
                                }
                                for j in 0..b_cols {
                                    let b_val = b_data[k * b_cols + j];
                                    if b_val == Trit::Tend { continue; }
                                    let prod = a_val * b_val;
                                    let (sum, _) = result[i * b_cols + j] + prod;
                                    result[i * b_cols + j] = sum;
                                }
                            }
                        }

                        if skipped { self.sparse_dropped = true; }
                        let res_idx = self.tensors.len();
                        self.tensors.push(TensorInstance {
                            data: TensorData::Trit(result),
                            rows: a_rows,
                            cols: b_cols,
                        });
                        self.stack.push(Value::TensorRef(res_idx));
                    } else {
                        return Err(VmError::TypeMismatch { expected: "TensorRef, TensorRef".into(), found: "Unknown".into() });
                    }
                }
                0x40 => { // Tstruct
                    let num_fields = self.read_u8()? as usize;
                    let mut fields = std::collections::HashMap::new();
                    for _ in 0..num_fields {
                        let name_len = self.read_u8()? as usize;
                        let mut name_bytes = vec![0u8; name_len];
                        for i in 0..name_len { name_bytes[i] = self.read_u8()?; }
                        let name = String::from_utf8(name_bytes).unwrap();
                        let val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                        fields.insert(name, val);
                    }
                    self.stack.push(Value::Struct(fields));
                }
                0x41 => { // Tfield
                    let name_len = self.read_u8()? as usize;
                    let mut name_bytes = vec![0u8; name_len];
                    for i in 0..name_len { name_bytes[i] = self.read_u8()?; }
                    let name = String::from_utf8(name_bytes).unwrap();
                    let obj = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let Value::Struct(fields) = obj {
                        let val = fields.get(&name).cloned().unwrap_or_default();
                        self.stack.push(val);
                    } else {
                        return Err(VmError::TypeMismatch { expected: "Struct".into(), found: format!("{:?}", obj) });
                    }
                }
                0x42 => { // TBIND reg, view_reg
                    let reg = self.read_u8()? as usize;
                    let view_val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let Value::TensorView { .. } = view_val {
                        self.bindings.insert(reg, view_val);
                    } else {
                        return Err(VmError::TypeMismatch { expected: "TensorView".into(), found: format!("{:?}", view_val) });
                    }
                }
                0x55 => { // TVIEW tensor_ref/view, offset, length, stride → TensorView
                    let stride = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let length = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let offset = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let rf = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    
                    if let (Value::Int(o), Value::Int(l), Value::Int(s)) = (&offset, &length, &stride) {
                        match rf {
                            Value::TensorRef(id) => {
                                self.stack.push(Value::TensorView {
                                    tensor_id: id,
                                    offset: *o as usize,
                                    length: *l as usize,
                                    stride: *s as usize,
                                });
                            }
                            Value::TensorView { tensor_id, offset: v_off, stride: v_stride, .. } => {
                                // Slicing a view: absolute offset = view_offset + (slice_offset * view_stride)
                                // New stride = view_stride * slice_stride
                                self.stack.push(Value::TensorView {
                                    tensor_id,
                                    offset: v_off + (*o as usize * v_stride),
                                    length: *l as usize,
                                    stride: v_stride * (*s as usize),
                                });
                            }
                            _ => return Err(VmError::TypeMismatch { expected: "TensorRef or TensorView".into(), found: format!("{:?}", rf) }),
                        }
                    } else {
                        return Err(VmError::TypeMismatch { expected: "Int, Int, Int".into(), found: format!("{:?}, {:?}, {:?}", offset, length, stride) });
                    }
                }
                0x50 => { // TPACK: pops 5 trits, pushes 1 packed byte (as Int)
                    let mut trits = [Trit::Tend; 5];
                    for i in (0..5).rev() {
                        let t = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                        trits[i] = match t {
                            Value::Trit(tv) => tv,
                            _ => return Err(VmError::TypeMismatch { expected: "Trit".into(), found: format!("{:?}", t) }),
                        };
                    }
                    let packed = crate::trit::pack_5_trits(trits);
                    self.stack.push(Value::Int(packed as i64));
                }
                0x51 => { // TUNPACK: pops 1 packed byte, pushes 5 trits
                    let val = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let Value::Int(packed) = val {
                        let trits = crate::trit::unpack_5_trits(packed as u8);
                        for t in trits {
                            self.stack.push(Value::Trit(t));
                        }
                    } else {
                        return Err(VmError::TypeMismatch { expected: "Int (packed byte)".into(), found: format!("{:?}", val) });
                    }
                }
                0x56 => { // TALLOC_PACKED (packed trit tensor)
                    let rows = self.read_u32()? as usize;
                    let cols = self.read_u32()? as usize;
                    let size = rows * cols;
                    let num_bytes = (size + 4) / 5;
                    let idx = self.tensors.len();
                    self.tensors.push(TensorInstance {
                        data: TensorData::PackedTrit(vec![0x00; num_bytes], size), // 0x00 is Reject Reject Reject Reject Reject
                        rows,
                        cols,
                    });
                    self.stack.push(Value::TensorRef(idx));
                }
                0x52 => { // TV_ADD: Vectorized addition of two packed tensors
                    let b_ref = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a_ref = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let (Value::TensorRef(a_idx), Value::TensorRef(b_idx)) = (a_ref, b_ref) {
                        let res_idx = self.tensors.len();
                        let (rows, cols, data) = {
                            let a = self.tensors.get(a_idx).ok_or(VmError::TensorNotAllocated(a_idx))?;
                            let b = self.tensors.get(b_idx).ok_or(VmError::TensorNotAllocated(b_idx))?;
                            if a.rows != b.rows || a.cols != b.cols {
                                return Err(VmError::RuntimeError("Tensor dimension mismatch in TV_ADD".into()));
                            }
                            match (&a.data, &b.data) {
                                (TensorData::PackedTrit(av, alen), TensorData::PackedTrit(bv, _)) => {
                                    let mut res_v = vec![0u8; av.len()];
                                    for i in 0..av.len() {
                                        res_v[i] = crate::trit::packed_add(av[i], bv[i]);
                                    }
                                    (a.rows, a.cols, TensorData::PackedTrit(res_v, *alen))
                                }
                                _ => return Err(VmError::TypeMismatch { expected: "PackedTrit tensors".into(), found: "Other".into() }),
                            }
                        };
                        self.tensors.push(TensorInstance { data, rows, cols });
                        self.stack.push(Value::TensorRef(res_idx));
                    } else {
                        return Err(VmError::TypeMismatch { expected: "TensorRef, TensorRef".into(), found: "Unknown".into() });
                    }
                }
                0x53 => { // TV_NEG: Vectorized negation of a packed tensor
                    let a_ref = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let Value::TensorRef(idx) = a_ref {
                        let res_idx = self.tensors.len();
                        let (rows, cols, data) = {
                            let a = self.tensors.get(idx).ok_or(VmError::TensorNotAllocated(idx))?;
                            match &a.data {
                                TensorData::PackedTrit(v, len) => {
                                    let mut res_v = vec![0u8; v.len()];
                                    for i in 0..v.len() {
                                        res_v[i] = crate::trit::packed_neg(v[i]);
                                    }
                                    (a.rows, a.cols, TensorData::PackedTrit(res_v, *len))
                                }
                                _ => return Err(VmError::TypeMismatch { expected: "PackedTrit tensor".into(), found: "Other".into() }),
                            }
                        };
                        self.tensors.push(TensorInstance { data, rows, cols });
                        self.stack.push(Value::TensorRef(res_idx));
                    } else {
                        return Err(VmError::TypeMismatch { expected: "TensorRef".into(), found: format!("{:?}", a_ref) });
                    }
                }
                0x54 => { // TV_CON: Vectorized consensus of two packed tensors
                    let b_ref = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    let a_ref = self.stack.pop().ok_or(VmError::StackUnderflow)?;
                    if let (Value::TensorRef(a_idx), Value::TensorRef(b_idx)) = (a_ref, b_ref) {
                        let res_idx = self.tensors.len();
                        let (rows, cols, data) = {
                            let a = self.tensors.get(a_idx).ok_or(VmError::TensorNotAllocated(a_idx))?;
                            let b = self.tensors.get(b_idx).ok_or(VmError::TensorNotAllocated(b_idx))?;
                            if a.rows != b.rows || a.cols != b.cols {
                                return Err(VmError::RuntimeError("Tensor dimension mismatch in TV_CON".into()));
                            }
                            match (&a.data, &b.data) {
                                (TensorData::PackedTrit(av, alen), TensorData::PackedTrit(bv, _)) => {
                                    let mut res_v = vec![0u8; av.len()];
                                    for i in 0..av.len() {
                                        res_v[i] = crate::trit::packed_consensus(av[i], bv[i]);
                                    }
                                    (a.rows, a.cols, TensorData::PackedTrit(res_v, *alen))
                                }
                                _ => return Err(VmError::TypeMismatch { expected: "PackedTrit tensors".into(), found: "Other".into() }),
                            }
                        };
                        self.tensors.push(TensorInstance { data, rows, cols });
                        self.stack.push(Value::TensorRef(res_idx));
                    } else {
                        return Err(VmError::TypeMismatch { expected: "TensorRef, TensorRef".into(), found: "Unknown".into() });
                    }
                }
                0x00 => return Ok(()),
                _ => return Err(VmError::InvalidOpcode(opcode)),
            }
        }
        Ok(())
    }

    fn read_u8(&mut self) -> Result<u8, VmError> {
        if self.pc >= self.code.len() { return Err(VmError::PcOutOfBounds(self.pc)); }
        let val = self.code[self.pc];
        self.pc += 1;
        Ok(val)
    }

    fn read_u16(&mut self) -> Result<u16, VmError> {
        if self.pc + 1 >= self.code.len() { return Err(VmError::PcOutOfBounds(self.pc)); }
        let val = u16::from_le_bytes([self.code[self.pc], self.code[self.pc + 1]]);
        self.pc += 2;
        Ok(val)
    }

    fn read_u32(&mut self) -> Result<u32, VmError> {
        if self.pc + 3 >= self.code.len() { return Err(VmError::PcOutOfBounds(self.pc)); }
        let val = u32::from_le_bytes([
            self.code[self.pc], self.code[self.pc + 1],
            self.code[self.pc + 2], self.code[self.pc + 3]
        ]);
        self.pc += 4;
        Ok(val)
    }

    fn get_pos(&self, rf: &Value, row: i64, col: i64) -> Result<(usize, usize), VmError> {
        match rf {
            Value::TensorRef(idx) => {
                let tensor = self.tensors.get(*idx).ok_or(VmError::TensorNotAllocated(*idx))?;
                let pos = if tensor.cols > 1 && col >= 0 {
                    row as usize * tensor.cols + col as usize
                } else {
                    row as usize
                };
                Ok((*idx, pos))
            }
            Value::TensorView { tensor_id, offset, stride, .. } => {
                let pos = *offset + (row as usize * *stride);
                Ok((*tensor_id, pos))
            }
            _ => Err(VmError::TypeMismatch { expected: "TensorRef or TensorView".into(), found: format!("{:?}", rf) }),
        }
    }
}