logicaffeine-compile 0.10.1

//! UI bridge for web interface integration.
//!
//! This module provides high-level, UI-friendly wrappers around the compilation
//! pipeline, returning structured results suitable for display in the browser.
//!
//! # Key Functions
//!
//! | Function | Purpose |
//! |----------|---------|
//! | [`compile_for_ui`] | Compile LOGOS to FOL (UI format) |
//! | [`compile_for_proof`] | Compile and search for proofs |
//! | [`compile_theorem_for_ui`] | Compile theorems with derivation trees |
//! | [`verify_theorem`] | Verify a theorem is provable |
//! | [`interpret_for_ui`] | Run imperative code and return output |
//! | [`generate_rust_code`] | Generate Rust source (requires `codegen` feature) |
//!
//! # Result Types
//!
//! All functions return serializable result types:
//! - [`CompileResult`] - Tokens, FOL, AST, errors
//! - [`ProofCompileResult`] - FOL with proof search results
//! - [`TheoremCompileResult`] - Theorem verification with derivation tree
//! - [`InterpreterResult`] - Program output lines and errors
//!
//! # Token Categories
//!
//! The [`TokenCategory`] enum provides syntactic highlighting categories:
//! Quantifier, Noun, Verb, Adjective, Connective, Determiner, etc.

use serde::{Deserialize, Serialize};
use std::collections::HashSet;

use logicaffeine_base::{Arena, Interner};
use logicaffeine_language::{
    ast::{self, LogicExpr, Term},
    analysis::DiscoveryPass,
    arena_ctx::AstContext,
    compile::{compile_forest, compile_forest_with_options},
    drs,
    error::socratic_explanation,
    lexer::Lexer,
    mwe,
    parser::Parser,
    pragmatics,
    registry::SymbolRegistry,
    semantics,
    token::TokenType,
    CompileOptions, OutputFormat, ParseError,
};
use logicaffeine_proof::{DerivationTree, ProofExpr, ProofTerm};

// Re-export interpreter result from our interpreter module
pub use crate::interpreter::InterpreterResult;

// ═══════════════════════════════════════════════════════════════════
// Token Visualization
// ═══════════════════════════════════════════════════════════════════

/// Syntactic category of a token for UI highlighting.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub enum TokenCategory {
    /// Universal/existential quantifiers: every, some, no, most
    Quantifier,
    /// Common nouns: dog, cat, person
    Noun,
    /// Main verbs: runs, loves, gives
    Verb,
    /// Adjective modifiers: tall, happy, red
    Adjective,
    /// Logical connectives: and, or, not, if
    Connective,
    /// Articles and determiners: a, an, the
    Determiner,
    /// Prepositional words: in, on, to, with
    Preposition,
    /// Personal and relative pronouns: he, she, who
    Pronoun,
    /// Modal auxiliaries: can, must, might
    Modal,
    /// Sentence-ending punctuation
    Punctuation,
    /// Proper names (capitalized)
    Proper,
    /// Uncategorized tokens
    Other,
}

/// Token information for UI display with position and category.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct TokenInfo {
    /// Byte offset of token start in source string.
    pub start: usize,
    /// Byte offset of token end (exclusive) in source string.
    pub end: usize,
    /// The actual text of the token.
    pub text: String,
    /// Syntactic category for highlighting.
    pub category: TokenCategory,
}

fn categorize_token(kind: &TokenType, _interner: &Interner) -> TokenCategory {
    match kind {
        TokenType::All | TokenType::Some | TokenType::No | TokenType::Any
        | TokenType::Most | TokenType::Few | TokenType::Many
        | TokenType::Cardinal(_) | TokenType::AtLeast(_) | TokenType::AtMost(_) => TokenCategory::Quantifier,
        TokenType::Noun(_) => TokenCategory::Noun,
        TokenType::Verb { .. } => TokenCategory::Verb,
        TokenType::Adjective(_) | TokenType::NonIntersectiveAdjective(_) => TokenCategory::Adjective,
        TokenType::And | TokenType::Or | TokenType::Not | TokenType::If | TokenType::Then
        | TokenType::Iff | TokenType::Because => TokenCategory::Connective,
        TokenType::Article(_) => TokenCategory::Determiner,
        TokenType::Preposition(_) => TokenCategory::Preposition,
        TokenType::Pronoun { .. } => TokenCategory::Pronoun,
        TokenType::Must | TokenType::Can | TokenType::Should | TokenType::Shall
        | TokenType::Would | TokenType::Could | TokenType::May | TokenType::Cannot
        | TokenType::Might => TokenCategory::Modal,
        TokenType::Period | TokenType::Comma => TokenCategory::Punctuation,
        TokenType::ProperName(_) => TokenCategory::Proper,
        _ => TokenCategory::Other,
    }
}

/// Tokenizes input text and returns token information for UI display.
///
/// Each token includes its byte position, text content, and syntactic
/// category for syntax highlighting in the browser interface.
pub fn tokenize_for_ui(input: &str) -> Vec<TokenInfo> {
    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();

    tokens.iter().map(|t| TokenInfo {
        start: t.span.start,
        end: t.span.end,
        text: input[t.span.start..t.span.end].to_string(),
        category: categorize_token(&t.kind, &interner),
    }).collect()
}

// ═══════════════════════════════════════════════════════════════════
// AST Visualization
// ═══════════════════════════════════════════════════════════════════

/// AST node for tree visualization in the UI.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub struct AstNode {
    /// Display label for this node (e.g., "∀x", "Run(x)").
    pub label: String,
    /// Node type for styling (e.g., "quantifier", "predicate", "connective").
    pub node_type: String,
    /// Child nodes in the AST.
    pub children: Vec<AstNode>,
}

impl AstNode {
    pub fn leaf(label: &str, node_type: &str) -> Self {
        AstNode {
            label: label.to_string(),
            node_type: node_type.to_string(),
            children: Vec::new(),
        }
    }

    pub fn with_children(label: &str, node_type: &str, children: Vec<AstNode>) -> Self {
        AstNode {
            label: label.to_string(),
            node_type: node_type.to_string(),
            children,
        }
    }
}

/// Converts a logic expression to an AST node for tree visualization.
///
/// Recursively builds a tree structure with labeled nodes suitable for
/// rendering in the UI. Each node includes a display label, node type
/// for styling, and child nodes.
pub fn expr_to_ast_node(expr: &LogicExpr, interner: &Interner) -> AstNode {
    match expr {
        LogicExpr::Predicate { name, args, .. } => {
            let name_str = interner.resolve(*name);
            let arg_nodes: Vec<AstNode> = args.iter()
                .map(|t| term_to_ast_node(t, interner))
                .collect();
            AstNode::with_children(
                &format!("{}({})", name_str, args.len()),
                "predicate",
                arg_nodes,
            )
        }
        LogicExpr::Quantifier { kind, variable, body, .. } => {
            let var_str = interner.resolve(*variable);
            let symbol = match kind {
                ast::QuantifierKind::Universal => "∀",
                ast::QuantifierKind::Existential => "∃",
                ast::QuantifierKind::Most => "MOST",
                ast::QuantifierKind::Few => "FEW",
                ast::QuantifierKind::Many => "MANY",
                ast::QuantifierKind::Cardinal(n) => return AstNode::with_children(
                    &format!("∃={}{}", n, var_str),
                    "quantifier",
                    vec![expr_to_ast_node(body, interner)],
                ),
                ast::QuantifierKind::AtLeast(n) => return AstNode::with_children(
                    &format!("∃≥{}{}", n, var_str),
                    "quantifier",
                    vec![expr_to_ast_node(body, interner)],
                ),
                ast::QuantifierKind::AtMost(n) => return AstNode::with_children(
                    &format!("∃≤{}{}", n, var_str),
                    "quantifier",
                    vec![expr_to_ast_node(body, interner)],
                ),
                ast::QuantifierKind::Generic => "GEN",
            };
            AstNode::with_children(
                &format!("{}{}", symbol, var_str),
                "quantifier",
                vec![expr_to_ast_node(body, interner)],
            )
        }
        LogicExpr::BinaryOp { left, op, right } => {
            let op_str = match op {
                TokenType::And => "∧",
                TokenType::Or => "∨",
                TokenType::If | TokenType::Then => "→",
                TokenType::Iff => "↔",
                _ => "?",
            };
            AstNode::with_children(
                op_str,
                "binary_op",
                vec![
                    expr_to_ast_node(left, interner),
                    expr_to_ast_node(right, interner),
                ],
            )
        }
        LogicExpr::UnaryOp { op, operand } => {
            let op_str = match op {
                TokenType::Not => "¬",
                _ => "?",
            };
            AstNode::with_children(
                op_str,
                "unary_op",
                vec![expr_to_ast_node(operand, interner)],
            )
        }
        LogicExpr::Identity { left, right } => {
            AstNode::with_children(
                "=",
                "identity",
                vec![
                    term_to_ast_node(left, interner),
                    term_to_ast_node(right, interner),
                ],
            )
        }
        LogicExpr::Modal { vector, operand } => {
            AstNode::with_children(
                &format!("□{:?}", vector.domain),
                "modal",
                vec![expr_to_ast_node(operand, interner)],
            )
        }
        LogicExpr::Lambda { variable, body } => {
            let var_str = interner.resolve(*variable);
            AstNode::with_children(
                &format!("λ{}", var_str),
                "lambda",
                vec![expr_to_ast_node(body, interner)],
            )
        }
        LogicExpr::SpeechAct { performer, act_type, content } => {
            AstNode::with_children(
                &format!(
                    "{}!{}",
                    interner.resolve(*act_type),
                    interner.resolve(*performer)
                ),
                "speech_act",
                vec![expr_to_ast_node(content, interner)],
            )
        }
        _ => AstNode::leaf(&format!("{:?}", expr), "other"),
    }
}

fn term_to_ast_node(term: &Term, interner: &Interner) -> AstNode {
    match term {
        Term::Constant(sym) => AstNode::leaf(interner.resolve(*sym), "constant"),
        Term::Variable(sym) => AstNode::leaf(interner.resolve(*sym), "variable"),
        Term::Function(name, args) => {
            let name_str = interner.resolve(*name);
            let arg_nodes: Vec<AstNode> = args.iter()
                .map(|t| term_to_ast_node(t, interner))
                .collect();
            AstNode::with_children(&format!("{}()", name_str), "function", arg_nodes)
        }
        Term::Group(terms) => {
            let term_nodes: Vec<AstNode> = terms.iter()
                .map(|t| term_to_ast_node(t, interner))
                .collect();
            AstNode::with_children("⊕", "group", term_nodes)
        }
        _ => AstNode::leaf(&format!("{:?}", term), "term"),
    }
}

// ═══════════════════════════════════════════════════════════════════
// Compilation Results
// ═══════════════════════════════════════════════════════════════════

#[derive(Debug, Clone, Serialize, Deserialize)]
/// Result of compiling English input to FOL with UI metadata.
pub struct CompileResult {
    /// Primary FOL output (Unicode format), if compilation succeeded.
    pub logic: Option<String>,
    /// Simplified FOL with modals stripped (for verification).
    pub simple_logic: Option<String>,
    /// Kripke semantics output with explicit world quantification.
    pub kripke_logic: Option<String>,
    /// AST tree representation for visualization.
    pub ast: Option<AstNode>,
    /// All scope readings in Unicode format.
    pub readings: Vec<String>,
    /// All scope readings in simplified format.
    pub simple_readings: Vec<String>,
    /// All scope readings in Kripke format.
    pub kripke_readings: Vec<String>,
    /// Tokenization with categories for syntax highlighting.
    pub tokens: Vec<TokenInfo>,
    /// Parse/compile error message, if any.
    pub error: Option<String>,
}

/// Compile English input to FOL with full UI metadata.
pub fn compile_for_ui(input: &str) -> CompileResult {
    let tokens = tokenize_for_ui(input);
    let readings = compile_forest(input);

    // Generate Simple readings (modals stripped) - deduplicated
    let simple_readings: Vec<String> = {
        let raw = compile_forest_with_options(input, CompileOptions { format: OutputFormat::SimpleFOL, pragmatic: false });
        let mut seen = HashSet::new();
        raw.into_iter().filter(|r| seen.insert(r.clone())).collect()
    };

    // Generate Kripke readings with explicit world quantification
    let kripke_readings = compile_forest_with_options(input, CompileOptions { format: OutputFormat::Kripke, pragmatic: false });

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let lex_tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let lex_tokens = mwe::apply_mwe_pipeline(lex_tokens, &mwe_trie, &mut interner);

    // Pass 1: Discovery
    let type_registry = {
        let mut discovery = DiscoveryPass::new(&lex_tokens, &mut interner);
        discovery.run()
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();

    let ctx = AstContext::new(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
    );

    // Pass 2: Parse
    let mut world_state = drs::WorldState::new();
    let mut parser = Parser::new(lex_tokens, &mut world_state, &mut interner, ctx, type_registry);

    match parser.parse() {
        Ok(ast) => {
            let ast = semantics::apply_axioms(ast, ctx.exprs, ctx.terms, &mut interner);
            let ast = pragmatics::apply_pragmatics(ast, ctx.exprs, &interner);
            let ast_node = expr_to_ast_node(ast, &interner);
            let mut registry = SymbolRegistry::new();
            let logic = ast.transpile_discourse(&mut registry, &interner, OutputFormat::Unicode);
            let simple_logic = ast.transpile_discourse(&mut registry, &interner, OutputFormat::SimpleFOL);

            let kripke_ast = semantics::apply_kripke_lowering(ast, ctx.exprs, ctx.terms, &mut interner);
            let kripke_logic = kripke_ast.transpile_discourse(&mut registry, &interner, OutputFormat::Kripke);

            CompileResult {
                logic: Some(logic),
                simple_logic: Some(simple_logic),
                kripke_logic: Some(kripke_logic),
                ast: Some(ast_node),
                readings,
                simple_readings,
                kripke_readings,
                tokens,
                error: None,
            }
        }
        Err(e) => {
            let advice = socratic_explanation(&e, &interner);
            CompileResult {
                logic: None,
                simple_logic: None,
                kripke_logic: None,
                ast: None,
                readings: Vec::new(),
                simple_readings: Vec::new(),
                kripke_readings: Vec::new(),
                tokens,
                error: Some(advice),
            }
        }
    }
}

// ═══════════════════════════════════════════════════════════════════
// Proof Integration
// ═══════════════════════════════════════════════════════════════════

/// Result of compiling English to a proof expression.
///
/// Used by the proof engine to search for derivations. Contains the
/// converted proof expression, the simplified FOL string representation,
/// and any compilation error.
#[derive(Debug, Clone)]
pub struct ProofCompileResult {
    /// The compiled proof expression, or `None` on error.
    pub proof_expr: Option<ProofExpr>,
    /// Simplified FOL string representation for display.
    pub logic_string: Option<String>,
    /// Error message if compilation failed.
    pub error: Option<String>,
}

/// Compile English input to ProofExpr for the proof engine.
pub fn compile_for_proof(input: &str) -> ProofCompileResult {
    use logicaffeine_language::proof_convert::logic_expr_to_proof_expr;

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let lex_tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let lex_tokens = mwe::apply_mwe_pipeline(lex_tokens, &mwe_trie, &mut interner);

    let type_registry = {
        let mut discovery = DiscoveryPass::new(&lex_tokens, &mut interner);
        discovery.run()
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();

    let ctx = AstContext::new(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
    );

    let mut world_state = drs::WorldState::new();
    let mut parser = Parser::new(lex_tokens, &mut world_state, &mut interner, ctx, type_registry);

    match parser.parse() {
        Ok(ast) => {
            let ast = semantics::apply_axioms(ast, ctx.exprs, ctx.terms, &mut interner);
            let ast = pragmatics::apply_pragmatics(ast, ctx.exprs, &interner);

            let mut registry = SymbolRegistry::new();
            let logic_string = ast.transpile(&mut registry, &interner, OutputFormat::SimpleFOL);
            let proof_expr = logic_expr_to_proof_expr(ast, &interner);

            ProofCompileResult {
                proof_expr: Some(proof_expr),
                logic_string: Some(logic_string),
                error: None,
            }
        }
        Err(e) => {
            let advice = socratic_explanation(&e, &interner);
            ProofCompileResult {
                proof_expr: None,
                logic_string: None,
                error: Some(advice),
            }
        }
    }
}

/// Result of compiling and verifying a theorem block.
///
/// Contains the parsed theorem structure (name, premises, goal) along
/// with the proof derivation tree if automatic proof search succeeded.
#[derive(Debug, Clone)]
pub struct TheoremCompileResult {
    /// The theorem's declared name.
    pub name: String,
    /// Compiled premise expressions (axioms).
    pub premises: Vec<ProofExpr>,
    /// The goal expression to prove, or `None` on parse error.
    pub goal: Option<ProofExpr>,
    /// Simplified FOL string of the goal for display.
    pub goal_string: Option<String>,
    /// Derivation tree from backward chaining, if proof found.
    pub derivation: Option<DerivationTree>,
    /// True iff the derivation was certified AND kernel type-checked.
    /// A derivation alone (`derivation.is_some()`) never implies this.
    pub verified: bool,
    /// Where verification broke (certification or type-check), if it did.
    pub verification_error: Option<String>,
    /// For a wh-question goal ("Who is in Florida?"), the certified witness(es); `None`
    /// when the goal is a closed proposition.
    pub answer: Option<Vec<String>>,
    /// For a recognized finite-domain grid, the cells the certified prover could fill —
    /// attached whenever the premises take the grid form, no flag required.
    pub grid: Option<SolvedGrid>,
    /// Error message if compilation or proof failed.
    pub error: Option<String>,
}

/// A solved (or partially solved) logic grid: the row entities plus one column per
/// declared category closure. Every filled cell is an entailment the no-Z3 certified
/// prover (CDCL+RUP / kernel) closed, so the whole table is independently checkable.
#[derive(Debug, Clone, PartialEq)]
pub struct SolvedGrid {
    /// Display header for the identity column — the row sort (e.g. "Trip").
    pub row_label: String,
    /// The row entities in declaration order (e.g. Alpha, Beta, Gamma, Delta).
    pub rows: Vec<String>,
    /// One column per category closure (year, state, friend, activity, …).
    pub columns: Vec<GridColumn>,
}

/// One category column of a [`SolvedGrid`].
#[derive(Debug, Clone, PartialEq)]
pub struct GridColumn {
    /// Category header (the value sort, e.g. "Year"), best-effort from the declarations.
    pub label: String,
    /// The category's domain values in closure order.
    pub values: Vec<String>,
    /// The determined value for each row (parallel to [`SolvedGrid::rows`]), or `None`
    /// where the prover could not force a unique value.
    pub cells: Vec<Option<String>>,
}

/// A parsed theorem block: premises and goal as owned (arena-free) `ProofExpr`s, the goal's
/// display string, and whether the strategy is `Auto`. Factored out of
/// [`compile_theorem_for_ui`] so the grounded grid problem can feed every trust tier and the
/// benchmark, not just the kernel path.
struct ParsedTheorem {
    name: String,
    premises: Vec<ProofExpr>,
    goal: ProofExpr,
    goal_string: String,
    is_auto: bool,
    /// An explicit `Proof:` tactic script (English-esque vernacular), if the block
    /// gave one instead of `Auto`. Run by the tactic framework, kernel-certified.
    script: Option<String>,
    /// Optional names for the premises (`Given (h): …`), parallel to `premises`, so a
    /// script can refer to a premise as `h` rather than `hp0`.
    premise_names: Vec<Option<String>>,
}

fn parse_theorem(input: &str) -> Result<ParsedTheorem, String> {
    use logicaffeine_language::proof_convert::logic_expr_to_proof_expr;

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);

    let type_registry = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        discovery.run()
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();

    let ctx = AstContext::new(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
    );

    let mut world_state = drs::WorldState::new();
    let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);

    let statements = parser
        .parse_program()
        .map_err(|e| format!("Parse error: {:?}", e))?;

    let theorem = statements
        .iter()
        .find_map(|stmt| if let ast::Stmt::Theorem(t) = stmt { Some(t) } else { None })
        .ok_or_else(|| "No theorem block found".to_string())?;

    let premises: Vec<ProofExpr> = theorem
        .premises
        .iter()
        .map(|p| logic_expr_to_proof_expr(p, &interner))
        .collect();
    let goal = logic_expr_to_proof_expr(theorem.goal, &interner);

    let mut registry = SymbolRegistry::new();
    let goal_string = theorem.goal.transpile(&mut registry, &interner, OutputFormat::SimpleFOL);

    let script = match &theorem.strategy {
        ast::theorem::ProofStrategy::Script(s) => Some(s.clone()),
        _ => None,
    };

    Ok(ParsedTheorem {
        name: theorem.name.clone(),
        premises,
        goal,
        goal_string,
        is_auto: matches!(theorem.strategy, ast::theorem::ProofStrategy::Auto),
        script,
        premise_names: theorem.premise_names.clone(),
    })
}

/// A single theorem's verdict within a compiled formal development.
#[derive(Debug, Clone)]
pub struct TheoryTheoremResult {
    pub name: String,
    pub verified: bool,
    pub error: Option<String>,
}

/// The result of compiling a program's formal development — its `## Axiom` declarations
/// (a shared premise base) plus the theorems of its `## Theory` block, each discharged in
/// citation order and kernel-certified by the multi-theorem driver.
#[derive(Debug, Clone)]
pub struct TheoryCompileResult {
    pub theory_name: Option<String>,
    pub axiom_count: usize,
    pub theorems: Vec<TheoryTheoremResult>,
    pub parse_error: Option<String>,
}

impl TheoryCompileResult {
    /// Whether every theorem in the development was kernel-certified (and it parsed).
    pub fn all_verified(&self) -> bool {
        self.parse_error.is_none()
            && !self.theorems.is_empty()
            && self.theorems.iter().all(|t| t.verified)
    }
}

/// Compile a program's formal development end to end: collect every `## Axiom` (a shared
/// premise base) and every `## Theory` block's `Axiom`/`Theorem` declarations, then prove
/// the theorems in citation order, each kernel-certified by the multi-theorem driver. This
/// is the seam that turns a Tarski-style source file into a verified development.
pub fn compile_theory_for_ui(input: &str) -> TheoryCompileResult {
    use logicaffeine_proof::development::parse_development;
    use logicaffeine_proof::formula::parse_formula;
    use logicaffeine_proof::verify::{prove_library_with_axioms, LibraryTheorem};

    // Parse the program with the same wiring as `parse_theorem`.
    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();
    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);
    let type_registry = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        discovery.run()
    };
    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();
    let ctx = AstContext::new(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
    );
    let mut world_state = drs::WorldState::new();
    let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);
    let statements = match parser.parse_program() {
        Ok(s) => s,
        Err(e) => {
            return TheoryCompileResult {
                theory_name: None,
                axiom_count: 0,
                theorems: Vec::new(),
                parse_error: Some(format!("Parse error: {:?}", e)),
            }
        }
    };

    // Collect the development: standalone `## Axiom` blocks + `## Theory` bodies.
    let mut axioms: Vec<ProofExpr> = Vec::new();
    let mut theorems: Vec<LibraryTheorem> = Vec::new();
    let mut theory_name: Option<String> = None;

    for stmt in &statements {
        match stmt {
            ast::Stmt::Axiom(a) => match parse_formula(&a.formula) {
                Ok(expr) => axioms.push(expr),
                Err(e) => {
                    return TheoryCompileResult {
                        theory_name,
                        axiom_count: axioms.len(),
                        theorems: Vec::new(),
                        parse_error: Some(format!("Axiom '{}': {}", a.name, e)),
                    }
                }
            },
            ast::Stmt::Theory(t) => {
                if theory_name.is_none() {
                    theory_name = Some(t.name.clone());
                }
                if !t.body.trim().is_empty() {
                    match parse_development(&t.body) {
                        Ok(dev) => {
                            axioms.extend(dev.axiom_exprs());
                            theorems.extend(dev.theorems);
                        }
                        Err(e) => {
                            return TheoryCompileResult {
                                theory_name,
                                axiom_count: axioms.len(),
                                theorems: Vec::new(),
                                parse_error: Some(format!("Theory '{}': {}", t.name, e)),
                            }
                        }
                    }
                }
            }
            _ => {}
        }
    }

    let results = prove_library_with_axioms(&axioms, &theorems);
    let theorem_results = theorems
        .iter()
        .zip(results)
        .map(|(t, r)| TheoryTheoremResult {
            name: t.name.clone(),
            verified: r.verified,
            error: r.verification_error,
        })
        .collect();

    TheoryCompileResult {
        theory_name,
        axiom_count: axioms.len(),
        theorems: theorem_results,
        parse_error: None,
    }
}

/// The GROUNDED, quantifier-free grid problem `(solver_input, goal)` for an `Auto` grid
/// theorem — the EXACT problem [`compile_theorem_for_ui`] solves, exposed so the trust tiers
/// (Untrusted CDCL / RUP / Kernel) and the benchmark all run the same input. `None` if the
/// theorem is not an `Auto` finite-domain grid.
pub fn grounded_grid_problem(input: &str) -> Option<(Vec<ProofExpr>, ProofExpr)> {
    let parsed = parse_theorem(input).ok()?;
    if !parsed.is_auto || !looks_like_grid(&parsed.premises) {
        return None;
    }
    let solver_input = prepare_premises_opts(&parsed.premises, true);
    let g = erase_tense(&parsed.goal);
    Some((solver_input, g))
}

/// Compile a theorem block for UI display.
pub fn compile_theorem_for_ui(input: &str) -> TheoremCompileResult {
    let parsed = match parse_theorem(input) {
        Ok(p) => p,
        Err(e) => {
            return TheoremCompileResult {
                name: String::new(),
                premises: Vec::new(),
                goal: None,
                goal_string: None,
                derivation: None,
                verified: false,
                verification_error: None,
                answer: None,
                grid: None,
                error: Some(e),
            };
        }
    };
    let ParsedTheorem { name, premises, goal, goal_string, is_auto, script, premise_names } =
        parsed;

    // The easter egg: when the premises take the grid form (declared bijection +
    // disjunctive closures), fill every cell the certified prover can force — no flag,
    // the structure alone triggers it. Attached to every grid result so the studio can
    // render the solved table alongside the headline goal.
    let grid = if is_auto && looks_like_grid(&premises) {
        solve_grid_from_premises(&premises, input)
    } else {
        None
    };

    let (derivation, verified, verification_error, answer) =
        if is_auto {
            // A wh-question goal ("Who is in Florida?") is an open ∃-form; the closed-goal
            // prover cannot discharge it and grinds a deep, futile search. Recognize the
            // form and ANSWER it by enumerating witnesses through the same certified no-Z3
            // path the puzzle uses — each candidate proved (or refuted) in ~1ms.
            if let ProofExpr::Exists { variable, body } = &goal {
                let witnesses = answer_wh(&premises, variable, body);
                let verified = !witnesses.is_empty();
                let verr = (!verified)
                    .then(|| "no individual in the domain satisfies the question".to_string());
                (None, verified, verr, Some(witnesses))
            } else {
            // A finite-domain GRID (a declared closure / "exactly one") is proved over
            // its GROUNDED, quantifier-free form, so the certified kernel can close it
            // in the browser (no Z3). An OPEN syllogism (Socrates: "every man is
            // mortal") is left untouched, keeping its `UniversalInst`/`ModusPonens`
            // trace. Both remain kernel-certified.
            let outcome = if looks_like_grid(&premises) {
                let g = erase_tense(&goal);
                // The incremental certified grid solver (trail propagation + DPLL) emits
                // a DerivationTree the kernel re-checks; it is faster and avoids the
                // general prover's re-saturation, and is the only path given the
                // FUNCTIONALITY exclusions. It is strictly additive: only a VERIFIED
                // solver tree is used, otherwise we fall back to the bounded backward
                // chainer (without functionality), so the solver can never regress.
                let solver_input = prepare_premises_opts(&premises, true);
                let solve = || {
                    let trace = std::env::var("LOGOS_TRACE").is_ok();
                    let t_solve = trace.then(std::time::Instant::now);
                    let tree = logicaffeine_proof::grid_solver::grid_prove(&solver_input, &g);
                    if let Some(t_solve) = t_solve {
                        let n = tree.as_ref().map(count_tree_nodes).unwrap_or(0);
                        eprintln!("[grid] solve+emit {:.2?} ({} tree nodes)", t_solve.elapsed(), n);
                    }
                    let t_cert = trace.then(std::time::Instant::now);
                    let solved = tree
                        .map(|tree| logicaffeine_proof::verify::check_derivation(&solver_input, &g, tree))
                        .filter(|vp| vp.verified);
                    if let Some(t_cert) = t_cert {
                        eprintln!("[grid] kernel-certify {:.2?} (verified={})", t_cert.elapsed(), solved.is_some());
                    }
                    match solved {
                        Some(vp) => vp,
                        None => {
                            // The grid solver produced no certified tree. Before the
                            // expensive bounded chainer, ask the fast certified RUP tier
                            // whether the goal is even entailed — a non-entailed goal is
                            // decided in ~1ms instead of grinding a deep, futile search.
                            // (RUP runs the SAME `solver_input`, so a NotEntailed verdict is
                            // sound: the goal is not entailed by the premises.)
                            if matches!(
                                logicaffeine_proof::rup::entails_certified(&solver_input, &g),
                                Some(logicaffeine_proof::rup::Verdict::NotEntailed)
                            ) {
                                return logicaffeine_proof::verify::VerifiedProof {
                                    derivation: None,
                                    proof_term: None,
                                    kernel_ctx: Default::default(),
                                    verified: false,
                                    verification_error: Some(
                                        "goal is not entailed (RUP-certified)".to_string(),
                                    ),
                                };
                            }
                            let grounded = prepare_premises(&premises);
                            logicaffeine_proof::verify::prove_certify_check_bounded(&grounded, &g, 60)
                        }
                    }
                };
                // A full multi-category grid's certified derivation recurses deeply
                // (case-analysis over every interdependent clue), so run the solve on a
                // generous stack — well past the default 8 MiB — on native targets.
                #[cfg(not(target_arch = "wasm32"))]
                {
                    std::thread::scope(|s| {
                        std::thread::Builder::new()
                            .stack_size(512 * 1024 * 1024)
                            .spawn_scoped(s, solve)
                            .expect("spawn grid-solve thread")
                            .join()
                            .expect("grid-solve thread panicked")
                    })
                }
                #[cfg(target_arch = "wasm32")]
                {
                    solve()
                }
            } else {
                logicaffeine_proof::verify::prove_certify_check(&premises, &goal)
            };
            (outcome.derivation, outcome.verified, outcome.verification_error, None)
            }
        } else if let Some(src) = &script {
            // An explicit `Proof:` tactic script (English-esque vernacular): run it
            // through the tactic framework on a generous stack, kernel-certified.
            use logicaffeine_proof::tactic::ProofState;
            let run = || {
                let mut st =
                    ProofState::start_with_names(premises.clone(), &premise_names, goal.clone());
                match st.run_script(src) {
                    Ok(_) => match st.qed() {
                        Ok(vp) => (vp.derivation, vp.verified, vp.verification_error),
                        Err(e) => (None, false, Some(format!("{e:?}"))),
                    },
                    Err(e) => (None, false, Some(e.to_string())),
                }
            };
            #[cfg(not(target_arch = "wasm32"))]
            let (derivation, verified, verr) = std::thread::scope(|s| {
                std::thread::Builder::new()
                    .stack_size(256 * 1024 * 1024)
                    .spawn_scoped(s, run)
                    .expect("spawn proof-script thread")
                    .join()
                    .expect("proof-script thread panicked")
            });
            #[cfg(target_arch = "wasm32")]
            let (derivation, verified, verr) = run();
            (derivation, verified, verr, None)
        } else {
            (None, false, None, None)
        };

    TheoremCompileResult {
        name,
        premises,
        goal: Some(goal),
        goal_string: Some(goal_string),
        derivation,
        verified,
        verification_error,
        answer,
        grid,
        error: None,
    }
}

fn count_tree_nodes(t: &logicaffeine_proof::DerivationTree) -> usize {
    1 + t.premises.iter().map(count_tree_nodes).sum::<usize>()
}

// ═══════════════════════════════════════════════════════════════════
// Code Generation
// ═══════════════════════════════════════════════════════════════════

/// Generate Rust code from LOGOS source — the Studio Code path. Mixed-document
/// aware: a source interleaving imperative code with Coq-style math
/// (`Definition`/`## Theorem:`/…) extracts the math into a bundled `mod proven` that
/// the imperative half calls into. Pure imperative source is a no-op partition.
#[cfg(feature = "codegen")]
pub fn generate_rust_code(source: &str) -> Result<String, ParseError> {
    let (imperative_src, math_src) = partition_mixed(source);
    let proven = math_src.as_deref().and_then(mixed_proven_module);
    generate_rust_code_with_proven(&imperative_src, proven.as_deref())
}

/// [`generate_rust_code`] with an already-extracted proven module bundled in.
#[cfg(feature = "codegen")]
pub fn generate_rust_code_with_proven(source: &str, proven: Option<&str>) -> Result<String, ParseError> {
    use logicaffeine_language::ast::stmt::{Stmt, Expr, TypeExpr};

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(source, &mut interner);
    let tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);

    let (type_registry, policy_registry) = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        let result = discovery.run_full();
        (result.types, result.policies)
    };
    let codegen_registry = type_registry.clone();
    let codegen_policies = policy_registry.clone();

    let mut world_state = drs::WorldState::new();
    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();
    let stmt_arena: Arena<Stmt> = Arena::new();
    let imperative_expr_arena: Arena<Expr> = Arena::new();
    let type_expr_arena: Arena<TypeExpr> = Arena::new();

    let ast_ctx = AstContext::with_types(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
        &stmt_arena,
        &imperative_expr_arena,
        &type_expr_arena,
    );

    let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ast_ctx, type_registry);
    let stmts = parser.parse_program()?;

    let type_env = crate::analysis::types::TypeEnv::infer_program(&stmts, &interner, &codegen_registry);
    let rust_code = crate::codegen::codegen_program_with_proven(&stmts, &codegen_registry, &codegen_policies, &interner, &type_env, &crate::optimization::OptimizationConfig::from_env(), "proven", proven);
    Ok(rust_code)
}

// ═══════════════════════════════════════════════════════════════════
// Interpreter (async)
// ═══════════════════════════════════════════════════════════════════

/// Interpret LOGOS imperative code and return output lines.
///
/// Same engine dispatch as [`interpret_for_ui_sync`]: the bytecode VM runs
/// synchronous programs (with the tree-walker as the debug shadow oracle);
/// programs needing async (file I/O, sleep, mount) run on the tree-walker's
/// async executor.
/// The first Send/escape-analysis violation in `stmts`, packaged as a ready
/// rejection — or `None` if the program respects the message-passing + CRDT
/// memory model. This is the static concurrency gate (Phase 4): a program whose
/// concurrent branches share non-CRDT mutable state is refused *before* any tier
/// runs, so neither the VM nor the tree-walker executes a data race.
fn send_escape_rejection(stmts: &[logicaffeine_language::ast::stmt::Stmt]) -> Option<InterpreterResult> {
    crate::concurrency::check_send_escape(stmts)
        .first()
        .map(|d| InterpreterResult { lines: Vec::new(), error: Some(d.message.clone()) })
}

/// Reject a dimension-incoherent program (`2 meters + 1 gram`) BEFORE it runs, so the interpreter
/// tier matches the AOT compiler: a dimension mismatch is an analysis error on every tier, surfaced
/// before any output is produced — not a mid-execution failure. Mirrors `send_escape_rejection`.
fn send_dimension_rejection(
    stmts: &[logicaffeine_language::ast::stmt::Stmt],
    interner: &Interner,
) -> Option<InterpreterResult> {
    crate::analysis::dimension_check::DimensionChecker::new(interner)
        .check_program(stmts)
        .err()
        .map(|e| InterpreterResult { lines: Vec::new(), error: Some(e.message) })
}

pub async fn interpret_for_ui(input: &str) -> InterpreterResult {
    interpret_for_ui_with_args(input, &[]).await
}

/// Like [`interpret_for_ui`], but supplies the program's argument vector to the
/// `args()` system native. `program_args` is the full argv (index 0 is the
/// program name), matching the compiled binary's `env::args()`.
pub async fn interpret_for_ui_with_args(
    input: &str,
    program_args: &[String],
) -> InterpreterResult {
    // The synchronous dispatcher covers every non-async program; for async
    // ones it uses block_on, which would nest inside this future — so handle
    // the async case here with a real await.
    let needs_async = with_parsed_program(input, |parsed, _| match parsed {
        Ok((stmts, _, _)) => crate::interpreter::needs_async(stmts),
        Err(_) => false,
    });
    if !needs_async {
        return interpret_for_ui_sync_with_args(input, program_args);
    }

    use logicaffeine_language::ast::stmt::{Stmt, Expr, TypeExpr};

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);

    let (type_registry, policy_registry) = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        let result = discovery.run_full();
        (result.types, result.policies)
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();
    let stmt_arena: Arena<Stmt> = Arena::new();
    let imperative_expr_arena: Arena<Expr> = Arena::new();
    let type_expr_arena: Arena<TypeExpr> = Arena::new();

    let ctx = AstContext::with_types(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
        &stmt_arena,
        &imperative_expr_arena,
        &type_expr_arena,
    );

    let mut world_state = drs::WorldState::new();
    let type_registry_for_interp = type_registry.clone();
    let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);

    match parser.parse_program() {
        Ok(stmts) => {
            if let Some(rejection) = send_escape_rejection(&stmts) {
                return rejection;
            }
            // A program that uses channels/tasks AND needs async (e.g. networking) must run
            // on the cooperative scheduler — `interp.run` alone installs none, so a channel op
            // would panic "outside a scheduler context". The async drive loop services both
            // the channel ops and the network awaits (the latter over the host reactor). A
            // pure-channel program (no async) never reaches here; it takes the sync path.
            if crate::concurrency::uses_scheduler(&stmts) {
                return run_program_concurrent_streaming(
                    &stmts,
                    &type_registry_for_interp,
                    policy_registry,
                    &interner,
                    program_args,
                    None,
                    None,
                    None,
                    0,
                )
                .await;
            }
            let mut interp = crate::interpreter::Interpreter::new(&interner)
                .with_type_registry(&type_registry_for_interp)
                .with_policies(policy_registry)
                .with_program_args(program_args.to_vec());
            match interp.run(&stmts).await {
                Ok(()) => InterpreterResult {
                    lines: interp.output,
                    error: None,
                },
                Err(e) => InterpreterResult {
                    lines: interp.output,
                    error: Some(e),
                },
            }
        }
        Err(e) => {
            let advice = socratic_explanation(&e, &interner);
            InterpreterResult {
                lines: vec![],
                error: Some(advice),
            }
        }
    }
}

/// The execution front-end shared by every engine: lex → MWE → discovery →
/// parse. The tree-walker and the bytecode VM MUST both come through here so
/// they see the identical token stream, type registry, and policies — a
/// differential test that parses differently per engine is comparing two
/// different programs.
///
/// Parsed statements borrow stack-local arenas, so they are handed to the
/// closure rather than returned. A parse failure is delivered as
/// `Err(socratic advice text)` — also identical for both engines.
pub fn with_parsed_program<R>(
    input: &str,
    f: impl for<'a> FnOnce(
        Result<
            (
                &'a [logicaffeine_language::ast::stmt::Stmt<'a>],
                &'a logicaffeine_language::analysis::TypeRegistry,
                logicaffeine_language::analysis::PolicyRegistry,
            ),
            String,
        >,
        &'a Interner,
    ) -> R,
) -> R {
    use logicaffeine_language::ast::stmt::{Expr, Stmt, TypeExpr};

    // A bare script (no `##` headers, imperative opening) runs as `## Main`.
    let implicit = implicit_main(input);
    let input = implicit.as_deref().unwrap_or(input);

    // Phase 10: auto-prepend the stdlib modules the program references (no-op when
    // it uses no stdlib vocabulary, so such programs stay byte-identical).
    let prelude_src = crate::loader::apply_prelude(input);
    let input = prelude_src.as_ref();

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);

    let (type_registry, policy_registry) = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        let result = discovery.run_full();
        (result.types, result.policies)
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();
    let stmt_arena: Arena<Stmt> = Arena::new();
    let imperative_expr_arena: Arena<Expr> = Arena::new();
    let type_expr_arena: Arena<TypeExpr> = Arena::new();

    let ctx = AstContext::with_types(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
        &stmt_arena,
        &imperative_expr_arena,
        &type_expr_arena,
    );

    let mut world_state = drs::WorldState::new();
    let type_registry_for_engines = type_registry.clone();
    let (parsed, opt_flags) = {
        let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);
        let stmts = parser.parse_program();
        let flags = parser.program_opt_flags();
        (stmts, flags)
    };

    match parsed {
        Ok(stmts) => {
            // Strength-reduce accumulator recursion to a constant-stack `while`
            // loop so the VM and tree-walker match the AOT (and never hit the
            // call-depth limit on `Return n + f(n-1)`-shaped recursion).
            // Type-directed division: rewrite `Divide → ExactDivide` in Rational
            // contexts (default stays floor), then accumulator-TCO the resolved AST.
            // The constant Rational fold honors the `Opt::Comptime` toggle (env +
            // in-source `## No comptime`), consistent with the AOT and tiered paths.
            let mut run_cfg =
                crate::optimization::OptimizationConfig::from_env().merged(&opt_flags);
            run_cfg.normalize();
            let resolved = crate::resolve_division::resolve_divisions(
                &stmts,
                &stmt_arena,
                &imperative_expr_arena,
                &interner,
                run_cfg.is_on(crate::optimization::Opt::Comptime),
            );
            let pre = resolved.unwrap_or(stmts.as_slice());
            match crate::tail_call::rewrite_accumulators(
                pre,
                &stmt_arena,
                &imperative_expr_arena,
                &mut interner,
            ) {
                Some(rw) => f(Ok((rw, &type_registry_for_engines, policy_registry)), &interner),
                None => f(Ok((pre, &type_registry_for_engines, policy_registry)), &interner),
            }
        }
        Err(e) => {
            let advice = socratic_explanation(&e, &interner);
            f(Err(advice), &interner)
        }
    }
}

/// [`with_parsed_program`], but the statements pass through the RUN-PATH
/// optimizer (EXODIA D1: the Futamura residual — PE, GVN, LICM, closed-form,
/// deforestation, interval analysis, DCE) before reaching the closure. The
/// live `largo run` engines consume this; the differential seams stay on the
/// raw variant so optimizer correctness is itself differentially gated
/// (optimized-VM vs raw-tree-walker).
pub fn with_optimized_program<R>(
    input: &str,
    f: impl for<'a> FnOnce(
        Result<
            (
                &'a [logicaffeine_language::ast::stmt::Stmt<'a>],
                &'a logicaffeine_language::analysis::TypeRegistry,
                logicaffeine_language::analysis::PolicyRegistry,
            ),
            String,
        >,
        &'a Interner,
    ) -> R,
) -> R {
    // The run-path upfront tier is the env-configured mode (default Eager → T3, i.e.
    // today's behavior, bit-for-bit). `LOGOS_TIER_PROFILE`/`LOGOS_FORCE_TIER` select
    // Tiered/Baseline for the benchmark A/B and the T_optimize measurement (§12.1).
    let tier = crate::optimization::HotswapConfig::from_env().run_tier();
    with_optimized_program_tiered(input, tier, f)
}

/// [`with_optimized_program`] gated by an explicit hotness `tier` (HOTSWAP §4): the
/// statements pass through [`crate::optimize::optimize_for_run_tiered`] at `tier`.
/// `Tier::T3` reproduces `with_optimized_program` exactly; `Tier::T0` skips the
/// optimizer (the baseline tier). The accumulator→loop TCO rewrite is a LANGUAGE
/// SEMANTIC, not an optimization, so it runs at every tier — deep recursion stays
/// constant-stack regardless of the optimization tier.
pub fn with_optimized_program_tiered<R>(
    input: &str,
    tier: crate::optimization::Tier,
    f: impl for<'a> FnOnce(
        Result<
            (
                &'a [logicaffeine_language::ast::stmt::Stmt<'a>],
                &'a logicaffeine_language::analysis::TypeRegistry,
                logicaffeine_language::analysis::PolicyRegistry,
            ),
            String,
        >,
        &'a Interner,
    ) -> R,
) -> R {
    use logicaffeine_language::ast::stmt::{Expr, Stmt, TypeExpr};

    // A bare script (no `##` headers, imperative opening) runs as `## Main`.
    let implicit = implicit_main(input);
    let input = implicit.as_deref().unwrap_or(input);

    // Phase 10: auto-prepend the stdlib modules the program references (no-op when
    // it uses no stdlib vocabulary, so such programs stay byte-identical).
    let prelude_src = crate::loader::apply_prelude(input);
    let input = prelude_src.as_ref();

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);

    let (type_registry, policy_registry) = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        let result = discovery.run_full();
        (result.types, result.policies)
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();
    let stmt_arena: Arena<Stmt> = Arena::new();
    let imperative_expr_arena: Arena<Expr> = Arena::new();
    let type_expr_arena: Arena<TypeExpr> = Arena::new();

    let ctx = AstContext::with_types(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
        &stmt_arena,
        &imperative_expr_arena,
        &type_expr_arena,
    );

    let mut world_state = drs::WorldState::new();
    let type_registry_for_engines = type_registry.clone();
    let (parsed, opt_flags, tier_pins) = {
        let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);
        let stmts = parser.parse_program();
        let flags = parser.program_opt_flags();
        let pins = parser.program_tier_pins();
        (stmts, flags, pins)
    };

    match parsed {
        Ok(stmts) => {
            // Respect file-level `## No <opt>` decorators on the run path too, so
            // optimization toggles behave consistently with the AOT compile.
            let mut run_cfg =
                crate::optimization::OptimizationConfig::from_env().merged(&opt_flags);
            run_cfg.normalize();
            // Tiered-optimizer pins: ambient env (`LOGOS_TIER_PIN`) overlaid by the
            // program's in-source `## Tier` decorators (the decorator wins).
            let mut hotswap = crate::optimization::HotswapConfig::from_env();
            hotswap.pins.overlay(&tier_pins);
            // Type-directed division (Divide → ExactDivide in Rational contexts) runs
            // BEFORE optimization so the optimizer sees ExactDivide (opaque, never
            // floor-folded), not a bare `7 / 2` it would fold to `3`.
            let resolved = crate::resolve_division::resolve_divisions(
                &stmts,
                &stmt_arena,
                &imperative_expr_arena,
                &interner,
                run_cfg.is_on(crate::optimization::Opt::Comptime),
            );
            let pre: Vec<_> = match resolved {
                Some(rw) => rw.to_vec(),
                None => stmts,
            };
            let optimized = crate::optimize::optimize_for_run_tiered(
                pre,
                &imperative_expr_arena,
                &stmt_arena,
                &mut interner,
                &run_cfg,
                &hotswap,
                tier,
            );
            // Accumulator recursion → constant-stack loop (matches the AOT and
            // the raw-parse engines).
            match crate::tail_call::rewrite_accumulators(
                &optimized,
                &stmt_arena,
                &imperative_expr_arena,
                &mut interner,
            ) {
                Some(rw) => f(Ok((rw, &type_registry_for_engines, policy_registry)), &interner),
                None => f(Ok((&optimized, &type_registry_for_engines, policy_registry)), &interner),
            }
        }
        Err(e) => {
            let advice = socratic_explanation(&e, &interner);
            f(Err(advice), &interner)
        }
    }
}

/// Like [`with_optimized_program`], but the program runs through the
/// ARCHITECT pipeline (`optimize_program_v2`): equality saturation with
/// kernel-certified rewrites in place of GVN + LICM + closed-form. The
/// differential gates in `phase_exodia_architect.rs` hold this path to the
/// raw tree-walker's outcomes.
pub fn with_v2_optimized_program<R>(
    input: &str,
    f: impl for<'a> FnOnce(
        Result<
            (
                &'a [logicaffeine_language::ast::stmt::Stmt<'a>],
                &'a logicaffeine_language::analysis::TypeRegistry,
                logicaffeine_language::analysis::PolicyRegistry,
            ),
            String,
        >,
        &'a Interner,
    ) -> R,
) -> R {
    use logicaffeine_language::ast::stmt::{Expr, Stmt, TypeExpr};

    // A bare script (no `##` headers, imperative opening) runs as `## Main`.
    let implicit = implicit_main(input);
    let input = implicit.as_deref().unwrap_or(input);

    // Phase 10: auto-prepend the stdlib modules the program references (no-op when
    // it uses no stdlib vocabulary, so such programs stay byte-identical).
    let prelude_src = crate::loader::apply_prelude(input);
    let input = prelude_src.as_ref();

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);

    let (type_registry, policy_registry) = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        let result = discovery.run_full();
        (result.types, result.policies)
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();
    let stmt_arena: Arena<Stmt> = Arena::new();
    let imperative_expr_arena: Arena<Expr> = Arena::new();
    let type_expr_arena: Arena<TypeExpr> = Arena::new();

    let ctx = AstContext::with_types(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
        &stmt_arena,
        &imperative_expr_arena,
        &type_expr_arena,
    );

    let mut world_state = drs::WorldState::new();
    let type_registry_for_engines = type_registry.clone();
    let parsed = {
        let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);
        parser.parse_program()
    };

    match parsed {
        Ok(stmts) => {
            let optimized = crate::optimize::optimize_program(
                stmts,
                &imperative_expr_arena,
                &stmt_arena,
                &mut interner,
                &crate::optimization::OptimizationConfig::from_env(),
            );
            f(Ok((&optimized, &type_registry_for_engines, policy_registry)), &interner)
        }
        Err(e) => {
            let advice = socratic_explanation(&e, &interner);
            f(Err(advice), &interner)
        }
    }
}

/// Interpret LOGOS imperative code synchronously (no async runtime needed).
///
/// The bytecode VM is the LIVE engine for the synchronous path (which is also
/// the browser/WASM path). The tree-walker remains fully wired as: the async
/// path (file I/O, sleep, mount), the shadow oracle in debug builds (every
/// program the corpus runs is differentially checked against it), and the
/// fallback for anything the VM compiler rejects.
pub fn interpret_for_ui_sync(input: &str) -> InterpreterResult {
    interpret_for_ui_sync_with_args(input, &[])
}

#[cfg(not(target_arch = "wasm32"))]
thread_local! {
    /// Compiled-native functions to install into the next VM run on this thread
    /// (HOTSWAP §Axis-3). `largo run` loads a `.logos-native` bundle into this; the VM
    /// seam below drains it and installs each via `install_aot_native`, so a `native`-
    /// annotated function dispatches to `rustc -O3` machine code from its first call.
    /// Empty for every other caller — no AOT, behavior unchanged.
    static PENDING_AOT: std::cell::RefCell<Vec<(String, Box<dyn crate::vm::NativeFn>)>> =
        const { std::cell::RefCell::new(Vec::new()) };
}

/// Queue compiled-native functions for the next VM run on this thread (HOTSWAP §Axis-3).
/// Consumed by the next `interpret_for_ui_*` VM run.
#[cfg(not(target_arch = "wasm32"))]
pub fn set_pending_aot_natives(natives: Vec<(String, Box<dyn crate::vm::NativeFn>)>) {
    PENDING_AOT.with(|p| *p.borrow_mut() = natives);
}

/// Drain the pending compiled-native functions and install them into `vm`, resolving
/// each by name to its function index. No-op when none are queued.
#[cfg(not(target_arch = "wasm32"))]
fn install_pending_aot_natives(
    vm: &mut crate::vm::Vm,
    program: &crate::vm::CompiledProgram,
    interner: &Interner,
) {
    let pending = PENDING_AOT.with(|p| std::mem::take(&mut *p.borrow_mut()));
    for (name, nf) in pending {
        if let Some(fi) = program
            .fn_index
            .iter()
            .find(|(s, _)| interner.resolve(**s) == name)
            .map(|(_, i)| *i as usize)
        {
            vm.install_aot_native(fi, nf);
            if std::env::var_os("LOGOS_ENGINE_TRACE").is_some() {
                eprintln!("logos-engine: aot-native installed for '{name}'");
            }
        }
    }
}

/// Like [`interpret_for_ui_sync`], but supplies the program's argument vector
/// to the `args()` system native. `program_args` is the full argv (index 0 is
/// the program name), matching the compiled binary's `env::args()` — this is
/// what `largo run --interpret N` forwards so an args-driven `main.lg` runs the
/// same on the VM/JIT as on the native binary.
pub fn interpret_for_ui_sync_with_args(input: &str, program_args: &[String]) -> InterpreterResult {
    // Diagnostic for benchmarking: with LOGOS_ENGINE_TRACE set, report which
    // engine actually ran a program to stderr, so a silent tree-walker fallback
    // is never mistaken for the VM+JIT.
    let trace = |engine: &str| {
        if std::env::var_os("LOGOS_ENGINE_TRACE").is_some() {
            eprintln!("logos-engine: {engine}");
        }
    };
    // The LIVE path runs the Futamura residual: both the VM and the debug
    // shadow oracle receive the SAME optimized program (two engines, one
    // program — the optimizer is differentially gated elsewhere against the
    // raw tree-walker).
    with_optimized_program(input, |parsed, interner| match parsed {
        Ok((stmts, type_registry, policies)) => {
            if let Some(rejection) = send_escape_rejection(stmts) {
                return rejection;
            }
            if let Some(rejection) = send_dimension_rejection(stmts, interner) {
                return rejection;
            }
            if crate::interpreter::needs_async(stmts) || crate::concurrency::uses_scheduler(stmts) {
                trace("treewalker (async)");
                return run_treewalker(stmts, type_registry, policies, interner, true, program_args);
            }
            // Oracle range analysis (M9) over the EXACT snapshot being
            // compiled → bounds-check elimination in the JIT.
            let oracle = crate::optimize::oracle_analyze_with(stmts, interner);
            match crate::vm::Compiler::compile_with_oracle(
                stmts,
                interner,
                Some(type_registry),
                Some(oracle),
            ) {
                Ok(program) => {
                    trace("vm+jit");
                    let mut vm = crate::vm::Vm::new(&program)
                        .with_policy_ctx(&policies, interner)
                        .with_program_args(program_args.to_vec());
                    if let Some(tier) = crate::vm::installed_native_tier() {
                        vm = vm.with_native_tier(tier);
                    }
                    // Compiled-native tier (HOTSWAP §Axis-3): install any AOT functions
                    // the run was given, so they dispatch to rustc -O3 machine code.
                    #[cfg(not(target_arch = "wasm32"))]
                    install_pending_aot_natives(&mut vm, &program, interner);
                    let error = vm.run().err();
                    let result = InterpreterResult { lines: vm.into_lines(), error };

                    // Debug-build shadow oracle: the SAME program runs on the
                    // tree-walker and the full outcome must match — this turns
                    // the entire existing test corpus into a differential
                    // suite. (Skipped on wasm to keep dev builds light.)
                    #[cfg(all(debug_assertions, not(target_arch = "wasm32")))]
                    {
                        let shadow = run_treewalker(
                            stmts,
                            type_registry,
                            policies.clone(),
                            interner,
                            false,
                            program_args,
                        );
                        assert_eq!(
                            (&result.lines, &result.error),
                            (&shadow.lines, &shadow.error),
                            "VM diverged from the tree-walker oracle for:\n{input}"
                        );
                    }
                    result
                }
                // The VM compiler rejects only constructs outside the parser's
                // reach; run them on the tree-walker rather than failing.
                Err(_) => {
                    trace("treewalker (vm-reject)");
                    run_treewalker(stmts, type_registry, policies, interner, false, program_args)
                }
            }
        }
        Err(advice) => InterpreterResult { lines: vec![], error: Some(advice) },
    })
}

/// Baseline-tier interpret for the interactive UI (the Studio) — the cold-start
/// engine. Runs on the bytecode VM with NO run-path optimizer
/// (`with_parsed_program`, not `with_optimized_program`) and NO oracle
/// (`compile_with_types`, not `compile_with_oracle`), so there is no
/// optimize/analysis latency before execution. `largo run` and the benchmarks
/// keep [`interpret_for_ui`], which optimizes ahead of execution for peak
/// throughput. Output is identical — the VM is differentially gated against the
/// tree-walker on the SAME parsed statements (see the debug shadow-oracle assert
/// in [`interpret_for_ui_baseline_sync_with_args`]); only the startup latency
/// differs. This is the baseline half of the EXODIA tier split; promoting hot
/// code to the optimized path is a later step.
pub async fn interpret_for_ui_baseline(input: &str) -> InterpreterResult {
    interpret_for_ui_baseline_with_args(input, &[]).await
}

/// Like [`interpret_for_ui_baseline`], but supplies the program's argument
/// vector to the `args()` system native.
pub async fn interpret_for_ui_baseline_with_args(
    input: &str,
    program_args: &[String],
) -> InterpreterResult {
    // Async programs (file I/O, sleep, mount) run on the tree-walker's async
    // executor regardless of tier — the VM is sync-only — so that case is
    // identical to the optimized entry (and must `await` rather than nest a
    // `block_on`); delegate it.
    let needs_async = with_parsed_program(input, |parsed, _| match parsed {
        Ok((stmts, _, _)) => crate::interpreter::needs_async(stmts),
        Err(_) => false,
    });
    if needs_async {
        return interpret_for_ui_with_args(input, program_args).await;
    }
    interpret_for_ui_baseline_sync_with_args(input, program_args)
}

/// The synchronous baseline core: parse (UNoptimized) → bytecode VM (no oracle)
/// → run. Mirrors [`interpret_for_ui_sync_with_args`] exactly, minus
/// `optimize_for_run` and the oracle range analysis.
pub fn interpret_for_ui_baseline_sync_with_args(
    input: &str,
    program_args: &[String],
) -> InterpreterResult {
    let trace = |engine: &str| {
        if std::env::var_os("LOGOS_ENGINE_TRACE").is_some() {
            eprintln!("logos-engine: {engine}");
        }
    };
    with_parsed_program(input, |parsed, interner| match parsed {
        Ok((stmts, type_registry, policies)) => {
            if let Some(rejection) = send_escape_rejection(stmts) {
                return rejection;
            }
            if let Some(rejection) = send_dimension_rejection(stmts, interner) {
                return rejection;
            }
            if crate::interpreter::needs_async(stmts) || crate::concurrency::uses_scheduler(stmts) {
                trace("treewalker (async)");
                return run_treewalker(stmts, type_registry, policies, interner, true, program_args);
            }
            match crate::vm::Compiler::compile_with_types(stmts, interner, Some(type_registry)) {
                Ok(program) => {
                    trace("vm (baseline)");
                    let mut vm = crate::vm::Vm::new(&program)
                        .with_policy_ctx(&policies, interner)
                        .with_program_args(program_args.to_vec());
                    if let Some(tier) = crate::vm::installed_native_tier() {
                        vm = vm.with_native_tier(tier);
                    }
                    // Compiled-native tier (HOTSWAP §Axis-3): install any AOT functions
                    // the run was given, so they dispatch to rustc -O3 machine code.
                    #[cfg(not(target_arch = "wasm32"))]
                    install_pending_aot_natives(&mut vm, &program, interner);
                    let error = vm.run().err();
                    let result = InterpreterResult { lines: vm.into_lines(), error };

                    // The same debug differential net as the optimized path: the
                    // baseline VM and the tree-walker must agree on the SAME
                    // (unoptimized) statements. (Skipped on wasm to keep dev
                    // builds light.)
                    #[cfg(all(debug_assertions, not(target_arch = "wasm32")))]
                    {
                        let shadow = run_treewalker(
                            stmts,
                            type_registry,
                            policies.clone(),
                            interner,
                            false,
                            program_args,
                        );
                        assert_eq!(
                            (&result.lines, &result.error),
                            (&shadow.lines, &shadow.error),
                            "baseline VM diverged from the tree-walker oracle for:\n{input}"
                        );
                    }
                    result
                }
                // The VM compiler rejects only constructs outside the parser's
                // reach; run them on the tree-walker rather than failing.
                Err(_) => {
                    trace("treewalker (vm-reject)");
                    run_treewalker(stmts, type_registry, policies, interner, false, program_args)
                }
            }
        }
        Err(advice) => InterpreterResult { lines: vec![], error: Some(advice) },
    })
}

/// Run a parsed program on the TREE-WALKER (the oracle engine). `force_async`
/// selects the async executor; otherwise the sync path is used.
/// Run a concurrent program on the deterministic scheduler: spawn the main block
/// as a task (which may spawn more), drive the scheduler to quiescence, and
/// collect the output merged through a shared callback.
fn run_program_concurrent<'a>(
    stmts: &'a [logicaffeine_language::ast::stmt::Stmt<'a>],
    type_registry: &logicaffeine_language::analysis::TypeRegistry,
    policies: logicaffeine_language::analysis::PolicyRegistry,
    interner: &'a Interner,
    program_args: &[String],
    vfs: Option<std::sync::Arc<dyn logicaffeine_system::fs::Vfs>>,
    stream: Option<crate::interpreter::OutputCallback>,
    seed: u64,
) -> InterpreterResult {
    use crate::concurrency::bridge::YieldState;
    use crate::concurrency::driver::InterpreterTask;
    use logicaffeine_runtime::{run_with_seed, RunOutcome, SchedSeed, SchedulerConfig};
    use std::cell::RefCell;
    use std::rc::Rc;

    let output_sink: Rc<RefCell<Vec<String>>> = Rc::new(RefCell::new(Vec::new()));
    let sink = output_sink.clone();
    // Collect every task's output for the final result, and — when streaming —
    // forward each line to the live callback as it is produced (the Studio's
    // real-time display).
    let callback: crate::interpreter::OutputCallback =
        Rc::new(RefCell::new(move |line: String| {
            if let Some(s) = &stream {
                (s.borrow_mut())(line.clone());
            }
            sink.borrow_mut().push(line);
        }));
    let err_sink: crate::concurrency::driver::ErrSink = Rc::new(RefCell::new(None));

    let mut main = crate::interpreter::Interpreter::new(interner)
        .with_type_registry(type_registry)
        .with_policies(policies)
        .with_program_args(program_args.to_vec())
        .with_output_callback(callback);
    if let Some(v) = vfs {
        main = main.with_vfs(v);
    }
    let main_ys = Rc::new(RefCell::new(YieldState::new()));
    main.install_yield_state(main_ys.clone());

    let main_fut = Box::pin(async move { main.run(stmts).await });
    let main_task = InterpreterTask::new(main_fut, main_ys, Some(err_sink.clone()));

    let (outcome, _trace) =
        run_with_seed(SchedulerConfig::default(), SchedSeed(seed), move |sched| {
            sched.spawn_main(Box::new(main_task));
        });

    let mut error = err_sink.borrow().clone();
    if error.is_none() {
        match outcome {
            RunOutcome::Deadlock => {
                error = Some("deadlock: every task is blocked waiting".to_string());
            }
            // The synchronous scheduler has no reactor; a program that needs network I/O
            // must run on the async drive loop (it routes there via `needs_async`). Reaching
            // here means it was driven on the wrong tier — surface it instead of half-running.
            RunOutcome::WaitingForIo => {
                error = Some(
                    "networking requires the async runtime; this program was run on the \
                     synchronous scheduler"
                        .to_string(),
                );
            }
            RunOutcome::Done(_) => {}
        }
    }
    let lines = output_sink.borrow().clone();
    InterpreterResult { lines, error }
}

/// A snapshot sink for the Studio's Tasks/Channels strip — invoked between scheduler
/// slices with the live task/channel state.
pub type ObserverCallback = std::rc::Rc<std::cell::RefCell<dyn FnMut(logicaffeine_runtime::SchedSnapshot)>>;

/// Yield a macrotask so the host event loop (Dioxus) can repaint between scheduler
/// slices. On wasm this is a real `setTimeout(0)`; on native there is no event loop to
/// yield to, so it returns immediately and the drive loop continues — the output is
/// identical, only the interleaving with repaints differs.
async fn yield_macrotask() {
    #[cfg(target_arch = "wasm32")]
    {
        gloo_timers::future::TimeoutFuture::new(0).await;
    }
}

/// Yield control back to the host async runtime once, so its reactor can advance a pending
/// network future before the scheduler re-polls the task that awaits it. Runtime-agnostic
/// (no tokio dependency): a future that wakes itself and returns `Pending` exactly once — the
/// same shape as `tokio::task::yield_now`, which lets the current-thread runtime poll its I/O
/// driver between turns. On wasm a `setTimeout(0)` lets the event loop fire the socket
/// callbacks. This is how a [`logicaffeine_runtime::RunOutcome::WaitingForIo`] is serviced.
async fn yield_to_reactor() {
    #[cfg(target_arch = "wasm32")]
    {
        gloo_timers::future::TimeoutFuture::new(0).await;
    }
    #[cfg(not(target_arch = "wasm32"))]
    {
        use std::future::Future;
        use std::pin::Pin;
        use std::task::{Context, Poll};
        struct YieldOnce(bool);
        impl Future for YieldOnce {
            type Output = ();
            fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<()> {
                if self.0 {
                    Poll::Ready(())
                } else {
                    self.0 = true;
                    cx.waker().wake_by_ref();
                    Poll::Pending
                }
            }
        }
        YieldOnce(false).await;
    }
}

/// The browser-facing async driver for concurrent programs. Identical semantics to
/// [`run_program_concurrent`] (same deterministic scheduler, same seed, same output), but
/// it advances the scheduler in bounded slices and yields a macrotask between them so the
/// Studio repaints instead of freezing — and emits a [`logicaffeine_runtime::SchedSnapshot`]
/// to `observer` after each slice to drive the Tasks/Channels strip.
#[allow(clippy::too_many_arguments)]
async fn run_program_concurrent_streaming<'a>(
    stmts: &'a [logicaffeine_language::ast::stmt::Stmt<'a>],
    type_registry: &logicaffeine_language::analysis::TypeRegistry,
    policies: logicaffeine_language::analysis::PolicyRegistry,
    interner: &'a Interner,
    program_args: &[String],
    vfs: Option<std::sync::Arc<dyn logicaffeine_system::fs::Vfs>>,
    stream: Option<crate::interpreter::OutputCallback>,
    observer: Option<ObserverCallback>,
    seed: u64,
) -> InterpreterResult {
    use crate::concurrency::bridge::YieldState;
    use crate::concurrency::driver::InterpreterTask;
    use logicaffeine_runtime::{Chooser, RunOutcome, SchedSeed, Scheduler, SchedulerConfig};
    use std::cell::RefCell;
    use std::rc::Rc;

    // The slice budget: how many scheduler steps to run before yielding a macrotask. Large
    // enough that a quiescent program finishes in one slice; small enough that a busy one
    // still repaints. It does not affect output (the scheduler is deterministic).
    const SLICE_STEPS: usize = 256;

    let output_sink: Rc<RefCell<Vec<String>>> = Rc::new(RefCell::new(Vec::new()));
    let sink = output_sink.clone();
    let callback: crate::interpreter::OutputCallback = Rc::new(RefCell::new(move |line: String| {
        if let Some(s) = &stream {
            (s.borrow_mut())(line.clone());
        }
        sink.borrow_mut().push(line);
    }));
    let err_sink: crate::concurrency::driver::ErrSink = Rc::new(RefCell::new(None));

    let mut main = crate::interpreter::Interpreter::new(interner)
        .with_type_registry(type_registry)
        .with_policies(policies)
        .with_program_args(program_args.to_vec())
        .with_output_callback(callback);
    if let Some(v) = vfs {
        main = main.with_vfs(v);
    }
    let main_ys = Rc::new(RefCell::new(YieldState::new()));
    main.install_yield_state(main_ys.clone());

    let main_fut = Box::pin(async move { main.run(stmts).await });
    let main_task = InterpreterTask::new(main_fut, main_ys, Some(err_sink.clone()));

    let mut sched = Scheduler::new(SchedulerConfig::default(), Chooser::record(SchedSeed(seed)));
    sched.spawn_main(Box::new(main_task));

    let outcome = loop {
        match sched.run_slice(SLICE_STEPS) {
            Some(RunOutcome::WaitingForIo) => {
                // A task is awaiting external I/O (a network op). Let the host reactor make
                // progress, then re-poll the parked tasks — they re-drive their network
                // futures, which now observe completion.
                if let Some(ob) = &observer {
                    (ob.borrow_mut())(sched.snapshot());
                }
                yield_to_reactor().await;
                sched.wake_io();
            }
            Some(o) => break o,
            None => {
                if let Some(ob) = &observer {
                    (ob.borrow_mut())(sched.snapshot());
                }
                yield_macrotask().await;
            }
        }
    };

    let mut error = err_sink.borrow().clone();
    if error.is_none() {
        if let RunOutcome::Deadlock = outcome {
            error = Some("deadlock: every task is blocked waiting".to_string());
        }
    }
    let lines = output_sink.borrow().clone();
    InterpreterResult { lines, error }
}

/// Parse `input` and run its concurrent program on the **tree-walker** scheduler
/// under an explicit `seed` — the seeded sibling of the default
/// `run_program_concurrent` entry, used by the cross-tier seeded differential.
pub fn run_treewalker_concurrent_seeded(input: &str, seed: u64) -> InterpreterResult {
    with_parsed_program(input, |parsed, interner| match parsed {
        Ok((stmts, type_registry, policies)) => {
            run_program_concurrent(stmts, type_registry, policies, interner, &[], None, None, seed)
        }
        Err(advice) => InterpreterResult { lines: vec![], error: Some(advice) },
    })
}

/// Run a concurrent program on the **bytecode VM** under the deterministic
/// scheduler (T11): compile to opcodes, then drive the main `Vm` (and any
/// spawned per-task VMs) through `run_until_block` via a `VmTask`. This is the
/// VM analog of [`run_program_concurrent`]; the default routing still sends
/// concurrent programs to the tree-walker, so this is the explicit VM entry used
/// to exercise and (Phase 5c) differentially compare the VM concurrency tier.
pub fn run_vm_concurrent(input: &str) -> InterpreterResult {
    run_vm_concurrent_seeded(input, 0)
}

/// [`run_vm_concurrent`] under an explicit scheduler `seed` — the VM sibling of
/// [`run_treewalker_concurrent_seeded`] for the cross-tier seeded differential.
pub fn run_vm_concurrent_seeded(input: &str, seed: u64) -> InterpreterResult {
    use crate::concurrency::vm_driver::VmTask;
    use logicaffeine_runtime::{run_with_seed, RunOutcome, SchedSeed, SchedulerConfig};
    use std::cell::RefCell;
    use std::rc::Rc;

    with_parsed_program(input, |parsed, interner| match parsed {
        Ok((stmts, type_registry, policies)) => {
            let program = match crate::vm::Compiler::compile_with_types(
                stmts,
                interner,
                Some(type_registry),
            ) {
                Ok(p) => p,
                Err(e) => return InterpreterResult { lines: vec![], error: Some(e) },
            };
            let output: Rc<RefCell<Vec<String>>> = Rc::new(RefCell::new(Vec::new()));
            let err_sink: crate::concurrency::driver::ErrSink = Rc::new(RefCell::new(None));
            let mut vm = crate::vm::Vm::new(&program).with_policy_ctx(&policies, interner);
            // Task bodies tier exactly like the main program: a hot integer loop
            // inside a task JIT-compiles. Concurrency ops are JIT-ineligible (not
            // integer ops ⇒ never region-selected), so a tiered region is
            // yield-free and a task only ever suspends on the bytecode path.
            // Spawned children inherit this tier via `spawn_task_vm`.
            #[cfg(not(target_arch = "wasm32"))]
            if let Some(tier) = crate::vm::installed_native_tier() {
                vm = vm.with_native_tier(tier);
            }
            let main_task = VmTask::new(vm, output.clone(), Some(err_sink.clone()));
            let (outcome, _trace) =
                run_with_seed(SchedulerConfig::default(), SchedSeed(seed), move |sched| {
                    sched.spawn_main(Box::new(main_task));
                });
            let mut error = err_sink.borrow().clone();
            if error.is_none() {
                if let RunOutcome::Deadlock = outcome {
                    error = Some("deadlock: every task is blocked waiting".to_string());
                }
            }
            let lines = output.borrow().clone();
            InterpreterResult { lines, error }
        }
        Err(advice) => InterpreterResult { lines: vec![], error: Some(advice) },
    })
}

/// One cooperative poll interval for the VM net runner — a short async sleep so the relay delivers
/// (and, in the browser, the event loop runs) between `Await` polls. Mirrors the tree-walker's.
async fn vm_net_poll_tick() {
    #[cfg(not(target_arch = "wasm32"))]
    logicaffeine_system::tokio::time::sleep(std::time::Duration::from_millis(2)).await;
    #[cfg(target_arch = "wasm32")]
    gloo_timers::future::TimeoutFuture::new(2).await;
}

/// A peer value → its relay topic (mirrors the tree-walker's `Interpreter::peer_topic_of`).
fn vm_peer_topic_of(value: &crate::interpreter::RuntimeValue) -> Result<String, String> {
    use crate::interpreter::RuntimeValue;
    match value {
        RuntimeValue::Peer(topic) => Ok((**topic).clone()),
        RuntimeValue::Text(s) => Ok(logicaffeine_system::addr::canonical_topic(s)),
        other => Err(format!(
            "Send/Await expects a PeerAgent or address string, got {}",
            other.type_name()
        )),
    }
}

/// Service one suspended VM networking request through the shared [`NetInbox`] — the SAME inbox the
/// tree-walker uses, so the wire encode/decode is identical and a `Send`/`Await` produces the same
/// value on both tiers. `Connect`/`Listen` do the async relay I/O; `Send`/`Stream` encode + publish;
/// `Await` drains-and-takes (yielding between polls) until the peer's message arrives.
async fn service_vm_net_block(
    vm: &mut crate::vm::Vm<'_>,
    netbox: &mut crate::concurrency::net_inbox::NetInbox,
    req: crate::vm::VmBlock,
) -> Result<(), String> {
    use crate::concurrency::marshal::{self};
    use crate::interpreter::RuntimeValue;
    use crate::vm::{Value, VmBlock};
    match req {
        VmBlock::NetConnect(url_payload) => {
            let raw = marshal::rebuild(url_payload).to_display_string();
            let url = logicaffeine_system::addr::multiaddr_to_ws_url(&raw).map_err(|e| {
                format!("Connect address '{raw}' is not a ws:// URL or supported multiaddr: {e}")
            })?;
            // OFFLINE mode: `Connect` is a single-node local no-op (mirrors the tree-walker's offline
            // `Connect`), so the VM net oracle agrees with it. A relay-connected driver dials for real.
            if !crate::concurrency::net_inbox::net_is_offline() {
                let net = logicaffeine_system::net::Net::connect(&url)
                    .await
                    .map_err(|e| format!("Connect to relay '{url}' failed: {e}"))?;
                netbox.net = Some(net);
            }
            vm.deliver_resume(Value::nothing());
            Ok(())
        }
        VmBlock::NetListen(topic_payload) => {
            let raw = marshal::rebuild(topic_payload).to_display_string();
            let topic = logicaffeine_system::addr::canonical_topic(&raw);
            let hs_topic = crate::concurrency::net_inbox::handshake_topic_for(&topic);
            // LOCAL/OFFLINE mode (no relay): declare the inbox identity locally, skip the relay
            // subscribe — mirrors the tree-walker's offline `Listen` so both tiers agree.
            if let Some(net) = netbox.net.as_mut() {
                net.subscribe(&topic).await?;
                // Also receive peers' capability handshakes on our dedicated handshake topic (kept off
                // the data topic so they never interleave with messages).
                net.subscribe(&hs_topic).await?;
            }
            netbox.inbox = Some(std::rc::Rc::new(topic));
            vm.deliver_resume(Value::nothing());
            Ok(())
        }
        VmBlock::NetSend(to_payload, msg_payload) => {
            let topic = vm_peer_topic_of(&marshal::rebuild(to_payload))?;
            // Advertise our surface to this peer on first contact — on its HANDSHAKE topic, never the
            // data topic — so it can negotiate back. Absorbed by the peer's drain, invisible to data.
            if let Some(hs) = netbox.first_contact_handshake(&topic) {
                netbox.publish(&crate::concurrency::net_inbox::handshake_topic_for(&topic), hs)?;
            }
            let value = marshal::rebuild(msg_payload);
            let from = netbox.inbox.as_ref().map(|t| t.to_string()).unwrap_or_default();
            // The negotiated maximal crush — byte-identical to the tree-walker's plain send (both call
            // `encode_negotiated` with the same program registry). Type-id fires only on a matching
            // peer epoch, so a raw / different-program peer still gets the plain encoding.
            let bytes = netbox.encode_negotiated(&from, &value, &topic, netbox.my_registry.clone())?;
            netbox.publish(&topic, bytes)?;
            vm.deliver_resume(Value::nothing());
            Ok(())
        }
        VmBlock::NetStream(to_payload, values_payload) => {
            let topic = vm_peer_topic_of(&marshal::rebuild(to_payload))?;
            if let Some(hs) = netbox.first_contact_handshake(&topic) {
                netbox.publish(&crate::concurrency::net_inbox::handshake_topic_for(&topic), hs)?;
            }
            let list = marshal::rebuild(values_payload);
            let items = match &list {
                RuntimeValue::List(rc) => rc.borrow().to_values(),
                other => vec![other.clone()],
            };
            let from = netbox.inbox.as_ref().map(|t| t.to_string()).unwrap_or_default();
            let blob = marshal::frame_stream_message(&from, &items)?;
            netbox.publish(&topic, blob)?;
            vm.deliver_resume(Value::nothing());
            Ok(())
        }
        VmBlock::NetAwait(from_payload, stream) => {
            let want = vm_peer_topic_of(&marshal::rebuild(from_payload))?;
            if netbox.inbox.is_none() {
                return Err("Await requires a prior Listen to establish an inbox".to_string());
            }
            // OFFLINE (no relay): drain our own loopback outbox (the `Send … to <self>` delivered
            // locally) rather than requiring a relay — the VM net oracle mirrors the tree-walker.
            loop {
                let got = if stream {
                    netbox.try_take_stream(&want)
                } else {
                    netbox.try_take_message(&want, false)
                };
                if let Some(v) = got {
                    vm.deliver_resume(Value::from_runtime(v));
                    return Ok(());
                }
                // No type-id registry: the VM net runner sends self-describing (no `shared` dial),
                // so an empty registry decodes everything correctly.
                netbox.drain(netbox.my_registry.clone());
                let got = if stream {
                    netbox.try_take_stream(&want)
                } else {
                    netbox.try_take_message(&want, false)
                };
                if let Some(v) = got {
                    vm.deliver_resume(Value::from_runtime(v));
                    return Ok(());
                }
                vm_net_poll_tick().await;
            }
        }
        VmBlock::NetMakePeer(addr_payload) => {
            let raw = marshal::rebuild(addr_payload).to_display_string();
            let topic = logicaffeine_system::addr::canonical_topic(&raw);
            vm.deliver_resume(Value::from_runtime(RuntimeValue::Peer(std::rc::Rc::new(topic))));
            Ok(())
        }
        VmBlock::NetSync(topic_payload, current_payload) => {
            let topic = marshal::rebuild(topic_payload).to_display_string();
            let current = marshal::rebuild(current_payload);
            // Encode the counter (Int) or counter-struct as the relay wire form; `None` ⇒ nothing
            // to publish yet. Then merge whatever has arrived since the last sync point.
            let publish_bytes = crate::semantics::arith::crdt_to_wire(&current);
            // LOCAL/OFFLINE mode: single-node sync is a no-op — the local CRDT value stands (mirrors
            // the tree-walker's offline `Sync`). A relay-connected node publishes + merges.
            let merged = if let Some(net) = netbox.net.as_mut() {
                net.subscribe(&topic).await?;
                if let Some(bytes) = publish_bytes {
                    net.publish(&topic, bytes)?;
                }
                let incoming = net.drain();
                let mut merged = current;
                for (_t, data) in incoming {
                    merged = crate::semantics::arith::crdt_merge_wire(merged, &data);
                }
                merged
            } else {
                current
            };
            vm.deliver_resume(Value::from_runtime(merged));
            Ok(())
        }
        _ => Err("run_vm_net_async services only peer-networking blocks; this program mixes \
                  channels/tasks with networking, which the VM net runner does not yet drive"
            .to_string()),
    }
}

/// Run a peer-networking program on the BYTECODE VM (single task, no channels): drive the resumable
/// VM and service each `VmBlock::Net*` through the shared `NetInbox` — the SAME inbox the
/// tree-walker uses — so `Connect`/`Listen`/`Send`/`Stream`/`Await` run byte-identically on both
/// tiers. The VM analog of the tree-walker's networking path; the cross-tier lock proves they agree.
pub async fn run_vm_net_async(input: &str) -> InterpreterResult {
    use logicaffeine_language::ast::stmt::{Expr, Stmt, TypeExpr};

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();
    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);
    let (type_registry, policy_registry) = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        let result = discovery.run_full();
        (result.types, result.policies)
    };
    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();
    let stmt_arena: Arena<Stmt> = Arena::new();
    let imperative_expr_arena: Arena<Expr> = Arena::new();
    let type_expr_arena: Arena<TypeExpr> = Arena::new();
    let ctx = AstContext::with_types(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
        &stmt_arena,
        &imperative_expr_arena,
        &type_expr_arena,
    );
    let mut world_state = drs::WorldState::new();
    let type_registry_for_vm = type_registry.clone();
    let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);
    let stmts = match parser.parse_program() {
        Ok(s) => s,
        Err(e) => return InterpreterResult { lines: vec![], error: Some(format!("{e:?}")) },
    };
    let program =
        match crate::vm::Compiler::compile_with_types(&stmts, &interner, Some(&type_registry_for_vm)) {
            Ok(p) => p,
            Err(e) => return InterpreterResult { lines: vec![], error: Some(e) },
        };
    let mut vm = crate::vm::Vm::new(&program).with_policy_ctx(&policy_registry, &interner);
    let mut netbox = crate::concurrency::net_inbox::NetInbox::new();
    // Build the wire type registry from the analysis types — IDENTICALLY to the tree-walker's
    // `build_wire_type_registry` (both derive from the same analysis table) — so two same-program peers
    // advertise matching registry epochs and may elide type NAMES from the wire (type-id).
    {
        let mut structs: Vec<(String, Vec<String>)> = Vec::new();
        let mut enums: Vec<(String, Vec<String>)> = Vec::new();
        for (name_sym, type_def) in type_registry_for_vm.iter_types() {
            let type_name = interner.resolve(*name_sym).to_string();
            match type_def {
                crate::analysis::registry::TypeDef::Struct { fields, .. } => {
                    structs.push((type_name, fields.iter().map(|f| interner.resolve(f.name).to_string()).collect()));
                }
                crate::analysis::registry::TypeDef::Enum { variants, .. } => {
                    enums.push((type_name, variants.iter().map(|v| interner.resolve(v.name).to_string()).collect()));
                }
                _ => {}
            }
        }
        netbox.set_registry(
            crate::concurrency::marshal::WireTypeRegistry::new(structs).with_enums(enums),
        );
    }
    let mut lines: Vec<String> = Vec::new();
    let mut error = None;
    loop {
        let step = vm.run_until_block();
        lines.extend(vm.drain_lines());
        match step {
            Ok(crate::vm::VmStep::Done(_)) => break,
            Ok(crate::vm::VmStep::Blocked) => {
                let req = match vm.take_pending() {
                    Some(r) => r,
                    None => break,
                };
                if let Err(e) = service_vm_net_block(&mut vm, &mut netbox, req).await {
                    error = Some(e);
                    break;
                }
            }
            Ok(crate::vm::VmStep::Paused) => break,
            Err(e) => {
                error = Some(e);
                break;
            }
        }
    }
    InterpreterResult { lines, error }
}

/// [`run_vm_concurrent_seeded`] under the **work-stealing M:N driver**: `workers`
/// OS-thread workers poll task bodies in parallel while one coordinator owns the
/// scheduler and applies channel ops + flushes output in deterministic pick
/// order. The observable result is byte-identical to the cooperative driver at the
/// same seed (`diff_cooperative_eq_workstealing_seeded`) — the difference is that
/// task bodies genuinely run on multiple cores.
///
/// The executor uses scoped threads, so each worker *borrows* the one shared,
/// immutable program (+ policies + interner) — no clone, no leak. Only a `Send`
/// `SpawnDesc` crosses a worker boundary; the worker rebuilds the `!Send` task
/// body locally from it.
#[cfg(not(target_arch = "wasm32"))]
pub fn run_vm_workstealing_seeded(input: &str, seed: u64, workers: usize) -> InterpreterResult {
    use crate::concurrency::vm_driver::VmTask;
    use logicaffeine_runtime::{
        run_workstealing_seeded, RunOutcome, SchedSeed, SchedulerConfig, SpawnDesc, Task,
    };

    with_parsed_program(input, |parsed, interner| match parsed {
        Ok((stmts, type_registry, policies)) => {
            let program = match crate::vm::Compiler::compile_with_types(
                stmts,
                interner,
                Some(type_registry),
            ) {
                Ok(p) => p,
                Err(e) => return InterpreterResult { lines: vec![], error: Some(e) },
            };
            // Build a worker-local task body from a `Send` descriptor. Workers run
            // the bytecode path: the native JIT tier is per-thread state (its code
            // cache is `!Sync`), so no shared tier is installed across workers.
            // Tiering is output-preserving, so this stays byte-identical to the
            // cooperative (tiered) driver; per-worker JIT is a future optimization.
            // The parallelism is the task bodies running on separate cores, which
            // the bytecode interpreter already delivers. The main task runs the
            // top-level program (`Vm::new` positions there); a spawned task is
            // positioned at its function via `setup_task`.
            let build = |desc: SpawnDesc| -> Box<dyn Task<'_> + '_> {
                let mut vm = crate::vm::Vm::new(&program).with_policy_ctx(&policies, interner);
                if !desc.is_main {
                    vm.setup_task(desc.func, &desc.args);
                }
                Box::new(VmTask::work_stealing(vm, None))
            };
            let main = SpawnDesc { func: 0, args: vec![], priority: 0, is_main: true };
            let config = SchedulerConfig::default().with_workers(workers.max(1));
            let result = run_workstealing_seeded(config, SchedSeed(seed), main, build);
            let error = match result.outcome {
                RunOutcome::Deadlock => {
                    Some("deadlock: every task is blocked waiting".to_string())
                }
                _ => None,
            };
            InterpreterResult { lines: result.output, error }
        }
        Err(advice) => InterpreterResult { lines: vec![], error: Some(advice) },
    })
}

/// Run a program on the synchronous tree-walker and return its global
/// bindings as sorted `(name, type, value)` rows — the REPL's `:vars`
/// inspection. Returns `None` for programs the sync walker can't host
/// (async/concurrent) or that fail to parse or run.
pub fn repl_global_bindings(input: &str, program_args: &[String]) -> Option<Vec<(String, String, String)>> {
    with_parsed_program(input, |parsed, interner| match parsed {
        Ok((stmts, type_registry, policies)) => {
            if crate::interpreter::needs_async(stmts) || crate::concurrency::uses_scheduler(stmts) {
                return None;
            }
            let mut interp = crate::interpreter::Interpreter::new(interner)
                .with_type_registry(type_registry)
                .with_policies(policies)
                .with_program_args(program_args.to_vec());
            interp.run_sync(stmts).ok()?;
            Some(interp.global_bindings())
        }
        Err(_) => None,
    })
}

pub(crate) fn run_treewalker<'a>(
    stmts: &'a [logicaffeine_language::ast::stmt::Stmt<'a>],
    type_registry: &logicaffeine_language::analysis::TypeRegistry,
    policies: logicaffeine_language::analysis::PolicyRegistry,
    interner: &'a Interner,
    force_async: bool,
    program_args: &[String],
) -> InterpreterResult {
    if crate::concurrency::uses_scheduler(stmts) {
        return run_program_concurrent(
            stmts, type_registry, policies, interner, program_args, None, None, 0,
        );
    }
    let mut interp = crate::interpreter::Interpreter::new(interner)
        .with_type_registry(type_registry)
        .with_policies(policies)
        .with_program_args(program_args.to_vec());
    let run_result = if force_async {
        futures::executor::block_on(interp.run(stmts))
    } else {
        interp.run_sync(stmts)
    };
    match run_result {
        Ok(()) => InterpreterResult { lines: interp.output, error: None },
        Err(e) => InterpreterResult { lines: interp.output, error: Some(e) },
    }
}

/// Interpret LOGOS imperative code with streaming output.
///
/// The `on_output` callback is called each time `Show` executes, allowing
/// real-time output display like a REPL. The callback receives the output line.
///
/// # Example
/// ```no_run
/// use std::rc::Rc;
/// use std::cell::RefCell;
/// # use logicaffeine_compile::interpret_streaming;
///
/// # fn main() {}
/// # async fn example() {
/// # let source = "## Main\nShow \"hello\".";
/// let lines = Rc::new(RefCell::new(Vec::new()));
/// let lines_clone = lines.clone();
///
/// interpret_streaming(source, Rc::new(RefCell::new(move |line: String| {
///     lines_clone.borrow_mut().push(line);
/// }))).await;
/// # }
/// ```
pub async fn interpret_streaming<F>(input: &str, on_output: std::rc::Rc<std::cell::RefCell<F>>) -> InterpreterResult
where
    F: FnMut(String) + 'static,
{
    interpret_streaming_with_vfs(input, on_output, None).await
}

/// Like [`interpret_streaming`], but routes the interpreter's file I/O
/// (`Write`/`Read`/`Mount`) to `vfs`. The browser Studio passes its `WebVfs`
/// here so the standard-library I/O vocabulary works against OPFS/IndexedDB; a
/// `None` vfs leaves file I/O reporting "VFS not initialized" as before. The
/// clone-per-task model means concurrent tasks inherit the same VFS handle.
pub async fn interpret_streaming_with_vfs<F>(
    input: &str,
    on_output: std::rc::Rc<std::cell::RefCell<F>>,
    vfs: Option<std::sync::Arc<dyn logicaffeine_system::fs::Vfs>>,
) -> InterpreterResult
where
    F: FnMut(String) + 'static,
{
    interpret_streaming_impl(input, on_output, vfs, None).await
}

/// Like [`interpret_streaming_with_vfs`], but also emits a [`logicaffeine_runtime::SchedSnapshot`]
/// to `observer` after each scheduler slice — the Studio's Tasks/Channels strip subscribes
/// here to show a concurrent program's live task and channel state as it runs.
pub async fn interpret_streaming_with_vfs_observer<F>(
    input: &str,
    on_output: std::rc::Rc<std::cell::RefCell<F>>,
    vfs: Option<std::sync::Arc<dyn logicaffeine_system::fs::Vfs>>,
    observer: ObserverCallback,
) -> InterpreterResult
where
    F: FnMut(String) + 'static,
{
    interpret_streaming_impl(input, on_output, vfs, Some(observer)).await
}

async fn interpret_streaming_impl<F>(
    input: &str,
    on_output: std::rc::Rc<std::cell::RefCell<F>>,
    vfs: Option<std::sync::Arc<dyn logicaffeine_system::fs::Vfs>>,
    observer: Option<ObserverCallback>,
) -> InterpreterResult
where
    F: FnMut(String) + 'static,
{
    use logicaffeine_language::ast::stmt::{Stmt, Expr, TypeExpr};
    use crate::interpreter::OutputCallback;

    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);

    let (type_registry, policy_registry) = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        let result = discovery.run_full();
        (result.types, result.policies)
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();
    let stmt_arena: Arena<Stmt> = Arena::new();
    let imperative_expr_arena: Arena<Expr> = Arena::new();
    let type_expr_arena: Arena<TypeExpr> = Arena::new();

    let ctx = AstContext::with_types(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
        &stmt_arena,
        &imperative_expr_arena,
        &type_expr_arena,
    );

    let mut world_state = drs::WorldState::new();
    let type_registry_for_interp = type_registry.clone();
    let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);

    match parser.parse_program() {
        Ok(stmts) => {
            if let Some(rejection) = send_escape_rejection(&stmts) {
                return rejection;
            }
            // Create the callback wrapper that calls the user's callback
            let callback: OutputCallback = std::rc::Rc::new(std::cell::RefCell::new(move |line: String| {
                (on_output.borrow_mut())(line);
            }));

            // Concurrent programs run on the deterministic scheduler — the browser
            // concurrency path — driven in slices that yield a macrotask between them so
            // the UI repaints, streaming each line as it is produced and emitting a
            // snapshot to the observer (if any). (Without this, a concurrency op would
            // `yield_request` with no scheduler installed and panic.)
            if crate::concurrency::uses_scheduler(&stmts) {
                return run_program_concurrent_streaming(
                    &stmts,
                    &type_registry_for_interp,
                    policy_registry,
                    &interner,
                    &[],
                    vfs,
                    Some(callback),
                    observer,
                    0,
                )
                .await;
            }

            let mut interp = crate::interpreter::Interpreter::new(&interner)
                .with_type_registry(&type_registry_for_interp)
                .with_policies(policy_registry)
                .with_output_callback(callback);
            if let Some(v) = vfs {
                interp = interp.with_vfs(v);
            }

            match interp.run(&stmts).await {
                Ok(()) => InterpreterResult {
                    lines: interp.output,
                    error: None,
                },
                Err(e) => InterpreterResult {
                    lines: interp.output,
                    error: Some(e),
                },
            }
        }
        Err(e) => {
            let advice = socratic_explanation(&e, &interner);
            InterpreterResult {
                lines: vec![],
                error: Some(advice),
            }
        }
    }
}

// ═══════════════════════════════════════════════════════════════════
// Theorem Verification (Kernel-certified)
// ═══════════════════════════════════════════════════════════════════

use logicaffeine_language::ast::Stmt;
use logicaffeine_language::proof_convert::logic_expr_to_proof_expr;
use crate::kernel;

/// Phase 78: Verify a theorem with full kernel certification.
///
/// Pipeline:
/// 1. Parse theorem block
/// 2. Extract symbols and build kernel context
/// 3. Run proof engine
/// 4. Certify derivation tree to kernel term
/// 5. Type-check the term
/// 6. Return (proof_term, context)
pub fn verify_theorem(input: &str) -> Result<(kernel::Term, kernel::Context), ParseError> {
    let (proof_exprs, goal_expr, definitions) = theorem_proof_exprs_with_defs(input)?;

    // === STEPS 3-6: Prove → certify → type-check (the one canonical pipeline) ===
    let outcome = logicaffeine_proof::verify::prove_certify_check_with_defs(
        &proof_exprs,
        &goal_expr,
        &definitions,
    );
    if outcome.verified {
        Ok((
            outcome
                .proof_term
                .expect("a verified outcome always carries a proof term"),
            outcome.kernel_ctx,
        ))
    } else {
        Err(ParseError {
            kind: logicaffeine_language::error::ParseErrorKind::Custom(
                outcome
                    .verification_error
                    .unwrap_or_else(|| "Theorem verification failed".to_string()),
            ),
            span: logicaffeine_language::token::Span::default(),
        })
    }
}

/// An English theorem proved, with its FOL and a human-readable PROOF TRACE — the
/// "English in → proof out, see the trace" path. Parsing yields the premises and
/// goal as FOL; the kernel-certified backward chainer yields the derivation, which
/// renders as an indented step-by-step proof.
#[derive(Debug, Clone)]
pub struct TheoremTrace {
    /// Whether the goal was proved (kernel-certified).
    pub verified: bool,
    /// The premises, rendered as FOL — what the English compiled to.
    pub premises: Vec<String>,
    /// The goal, rendered as FOL.
    pub goal: String,
    /// The derivation rendered as an indented proof trace (`└─ [Rule] conclusion`
    /// per step). `None` when no derivation was found.
    pub trace: Option<String>,
    /// The verification error, when proving failed.
    pub error: Option<String>,
}

/// Prove an English `## Theorem` block and return its FOL plus a rendered PROOF
/// TRACE. Unlike [`verify_theorem`] (which returns the raw kernel term), this
/// surfaces the derivation tree so the proof steps are visible.
pub fn prove_theorem_trace(input: &str) -> Result<TheoremTrace, ParseError> {
    let (proof_exprs, goal_expr, definitions) = theorem_proof_exprs_with_defs(input)?;
    let outcome = logicaffeine_proof::verify::prove_certify_check_with_defs(
        &proof_exprs,
        &goal_expr,
        &definitions,
    );
    Ok(TheoremTrace {
        verified: outcome.verified,
        premises: proof_exprs.iter().map(|p| p.to_string()).collect(),
        goal: goal_expr.to_string(),
        trace: outcome.derivation.as_ref().map(|d| d.display_tree()),
        error: outcome.verification_error,
    })
}

// ═══════════════════════════════════════════════════════════════════
// Proof / Math → Rust extraction (the Curry-Howard "Forge", UI-facing)
// ═══════════════════════════════════════════════════════════════════

/// Names the user defined — present in `ctx` but absent from a fresh kernel
/// context (i.e. excluding the StandardLibrary baseline). Inductives are listed
/// before definitions so the extracted module reads top-down; StandardLibrary
/// items are pulled in by [`extract_math_rust`] only when transitively needed.
fn user_defined_entries(ctx: &kernel::Context) -> Vec<String> {
    let baseline = kernel::interface::Repl::new();
    let base = baseline.context();
    let mut base_names: HashSet<String> = HashSet::new();
    for (name, _) in base.iter_inductives() {
        base_names.insert(name.to_string());
    }
    for (name, _, _) in base.iter_definitions() {
        base_names.insert(name.to_string());
    }

    let mut entries = Vec::new();
    for (name, _) in ctx.iter_inductives() {
        // A user inductive extracts as an enum only if every constructor field is
        // a concrete Rust type. Dependent/indexed families (e.g. `Eq<Nat,n,…>`,
        // value-indexed proof types) and opaque/primitive types are skipped.
        if !base_names.contains(name) && inductive_is_emittable(ctx, name) {
            entries.push(name.to_string());
        }
    }
    for (name, ty, _) in ctx.iter_definitions() {
        // Only data-typed definitions whose body references nothing inextractable
        // (no Prop proofs, no axioms/tactics like `syn_diag`/`try_auto`) extract to
        // runnable Rust — so we never emit Rust that references undefined symbols.
        if !base_names.contains(name)
            && type_is_emittable(ctx, ty, &[])
            && crate::extraction::is_extractable(ctx, name)
        {
            entries.push(name.to_string());
        }
    }
    // Deterministic order (iter_* walk HashMaps with per-instance random seeds).
    entries.sort();
    entries.dedup();
    entries
}

/// Whether a kernel type extracts to a concrete Rust type: a mapped primitive, a
/// user inductive with emittable constructors (→ enum), a declared type parameter
/// (`generics`), or a function/application built from those. `Sort` (Prop/Type),
/// logical/opaque types, and undeclared type variables are not.
fn type_is_emittable(ctx: &kernel::Context, ty: &kernel::Term, generics: &[String]) -> bool {
    use kernel::Term;
    match ty {
        Term::Global(name) => {
            crate::extraction::primitive_rust_type(name).is_some()
                || (ctx.is_inductive(name)
                    && !ctx.get_constructors(name).is_empty()
                    && !crate::extraction::is_logical_type(name))
        }
        // A type variable is only a real Rust type if it's a declared generic of
        // the enclosing inductive — never in a bare definition type.
        Term::Var(v) => generics.iter().any(|g| g == v),
        Term::Pi { param_type, body_type, .. } => {
            type_is_emittable(ctx, param_type, generics) && type_is_emittable(ctx, body_type, generics)
        }
        Term::App(f, a) => {
            type_is_emittable(ctx, f, generics) && type_is_emittable(ctx, a, generics)
        }
        _ => false,
    }
}

/// The leading `: Type` parameters of an inductive (its Rust generics).
fn inductive_generics(ctx: &kernel::Context, ind: &str) -> Vec<String> {
    use kernel::Term;
    let mut names = Vec::new();
    let mut cur = match ctx.get_global(ind) {
        Some(t) => t,
        None => return names,
    };
    while let Term::Pi { param, param_type, body_type } = cur {
        if matches!(param_type.as_ref(), Term::Sort(_)) {
            names.push(param.clone());
            cur = body_type;
        } else {
            break;
        }
    }
    names
}

/// A user inductive is emittable iff it has constructors and every constructor
/// *field* (after the leading type parameters) is itself an emittable type.
fn inductive_is_emittable(ctx: &kernel::Context, ind: &str) -> bool {
    use kernel::Term;
    let ctors = ctx.get_constructors(ind);
    if ctors.is_empty() || crate::extraction::is_logical_type(ind) {
        return false;
    }
    let generics = inductive_generics(ctx, ind);
    for (_, ty) in &ctors {
        let mut cur = *ty;
        for _ in 0..generics.len() {
            if let Term::Pi { body_type, .. } = cur {
                cur = body_type;
            } else {
                break;
            }
        }
        while let Term::Pi { param_type, body_type, .. } = cur {
            if !type_is_emittable(ctx, param_type, &generics) {
                return false;
            }
            cur = body_type;
        }
    }
    true
}

/// Extract every user-defined inductive and definition in `ctx` into one Rust
/// module — the "compile my math to Rust" path for the Math studio. Shared and
/// StandardLibrary dependencies are pulled in only when a user definition
/// transitively needs them, and are emitted exactly once.
pub fn extract_math_rust(ctx: &kernel::Context) -> Result<String, String> {
    let entries = user_defined_entries(ctx);
    if entries.is_empty() {
        return Ok("// nothing defined yet — add a Definition or Inductive".to_string());
    }
    let module = extract_math_module(ctx)?;
    let checks: Vec<(String, Vec<kernel::Term>)> =
        property_checks(ctx).into_iter().map(|(n, _, p)| (n, p)).collect();
    // The "compiled mathematical object" runs and proves itself.
    Ok(format!("{module}{}", math_demo_main(ctx, &entries, &checks)))
}

/// Extract every user-defined inductive/definition in `ctx` into one Rust module —
/// types, functions, and `check_*` property fns from proven theorems — WITHOUT a
/// demo `main`. This is the linkable artifact bundled into an imperative program's
/// `mod proven`; [`extract_math_rust`] wraps it with a self-verifying `main` for the
/// standalone Math compile.
pub fn extract_math_module(ctx: &kernel::Context) -> Result<String, String> {
    let entries = user_defined_entries(ctx);
    if entries.is_empty() {
        return Ok("// nothing defined yet — add a Definition or Inductive".to_string());
    }
    let refs: Vec<&str> = entries.iter().map(|s| s.as_str()).collect();
    let mut module = crate::extraction::extract_programs(ctx, &refs).map_err(|e| e.to_string())?;
    // Proven theorems → runnable property checks over the extracted functions
    // (e.g. `∀n. add Zero n = n` → `fn check_…(n) -> bool { add(Zero, n) == n }`).
    // Appended in sorted order so the module is byte-identical across recompiles.
    let mut check_names: std::collections::HashSet<String> = std::collections::HashSet::new();
    for (name, check_fn, _) in property_checks(ctx) {
        module.push_str(&check_fn);
        check_names.insert(name);
    }
    // Honest notes for proof-irrelevant theorems (Gödel, consistency, `True`, …): a
    // definition whose TYPE is a proposition is a *proof*, and by Curry-Howard a proof
    // of a Prop has no computational content — so it has no runnable form. The
    // constructive dependencies it uses are still extracted above; the note explains
    // why the theorem itself isn't a function. Sorted for deterministic output.
    let base = baseline_names();
    let mut notes: Vec<String> = Vec::new();
    for (name, ty, _) in ctx.iter_definitions() {
        if base.contains(name) || check_names.contains(name) {
            continue;
        }
        if def_type_is_proposition(ty) {
            notes.push(format!(
                "// note: `{name}` is a proof of a proposition — proof-irrelevant (no \
                 computational content), so it has no runnable form; any constructive \
                 definitions it relies on are extracted above.\n"
            ));
        }
    }
    notes.sort();
    for n in notes {
        module.push_str(&n);
    }
    Ok(module)
}

/// Whether a definition's type is a proposition — i.e. the definition is a *proof*
/// (proof-irrelevant), not computational data. Peels any `∀` (Pi) binders, then
/// checks whether the head of the result is a logical/proof type (`Eq`/`And`/`True`/
/// `Syntax`/`Derivation`/…). Used to emit honest notes for theorems with no runnable form.
fn def_type_is_proposition(ty: &kernel::Term) -> bool {
    use kernel::Term;
    let mut cur = ty;
    while let Term::Pi { body_type, .. } = cur {
        cur = body_type;
    }
    let mut head = cur;
    while let Term::App(f, _) = head {
        head = f;
    }
    matches!(head, Term::Global(n) if crate::extraction::is_logical_type(n))
}

/// Proven theorems → runnable property checks: `(name, check_fn_source, param_types)`,
/// sorted by name for deterministic emission. A theorem yields a check iff
/// [`crate::extraction::emit_property_check`] produces a runnable predicate over the
/// extracted functions (it quantifies only over data types).
fn property_checks(ctx: &kernel::Context) -> Vec<(String, String, Vec<kernel::Term>)> {
    let base = baseline_names();
    let mut checks: Vec<(String, String, Vec<kernel::Term>)> = Vec::new();
    for (name, ty, _) in ctx.iter_definitions() {
        if base.contains(name) {
            continue;
        }
        if let Some(check_fn) = crate::extraction::emit_property_check(ctx, name, ty) {
            checks.push((name.to_string(), check_fn, pi_param_types(ty)));
        }
    }
    checks.sort_by(|a, b| a.0.cmp(&b.0));
    checks
}

/// Names present in a fresh `Repl::new()` (the StandardLibrary baseline).
fn baseline_names() -> std::collections::HashSet<String> {
    let baseline = kernel::interface::Repl::new();
    let base = baseline.context();
    let mut names = std::collections::HashSet::new();
    for (n, _) in base.iter_inductives() {
        names.insert(n.to_string());
    }
    for (n, _, _) in base.iter_definitions() {
        names.insert(n.to_string());
    }
    names
}

/// Whether a term applies the `div`/`mod` arithmetic builtins (whose extracted Rust
/// `/`/`%` panics on a zero divisor) — used to skip them in the self-verifying demo.
fn term_uses_div_or_mod(term: &kernel::Term) -> bool {
    use kernel::Term;
    match term {
        Term::Global(n) => n == "div" || n == "mod",
        Term::App(f, a) => term_uses_div_or_mod(f) || term_uses_div_or_mod(a),
        Term::Lambda { param_type, body, .. } => {
            term_uses_div_or_mod(param_type) || term_uses_div_or_mod(body)
        }
        Term::Pi { param_type, body_type, .. } => {
            term_uses_div_or_mod(param_type) || term_uses_div_or_mod(body_type)
        }
        Term::Fix { body, .. } => term_uses_div_or_mod(body),
        Term::Match { discriminant, motive, cases } => {
            term_uses_div_or_mod(discriminant)
                || term_uses_div_or_mod(motive)
                || cases.iter().any(term_uses_div_or_mod)
        }
        _ => false,
    }
}

/// A self-verifying demo `main`: each value/function result is `assert_eq!`d
/// against `kernel::normalize` (the kernel's evaluator) and printed; each proven
/// theorem's property check is run on a sample and asserted.
fn math_demo_main(
    ctx: &kernel::Context,
    entries: &[String],
    checks: &[(String, Vec<kernel::Term>)],
) -> String {
    use kernel::Term;
    let mut lines = Vec::new();
    for name in entries {
        let Some(body) = ctx.get_definition_body(name) else {
            continue; // an inductive, not a definition
        };
        // Skip exercising a def that uses `div`/`mod`: the demo samples Int as 0, and
        // `n / n` / `n % n` at n=0 is a compile-time divide-by-zero panic (and the
        // kernel leaves div-by-0 stuck, diverging from Rust's panic). The function is
        // still EXTRACTED — just not run in the self-verifying demo.
        if term_uses_div_or_mod(body) {
            continue;
        }
        if matches!(body, Term::Lambda { .. } | Term::Fix { .. }) {
            // A function: apply it to kernel-built sample arguments.
            let Some(ty) = ctx.get_definition_type(name) else { continue };
            let params = pi_param_types(ty);
            let samples: Option<Vec<Term>> =
                params.iter().map(|p| sample_value(ctx, p)).collect();
            let Some(samples) = samples else { continue }; // can't sample → skip
            if samples.is_empty() {
                continue;
            }
            let args_rust: Vec<String> =
                samples.iter().map(|s| crate::extraction::emit_value(ctx, s)).collect();
            let mut app = Term::Global(name.clone());
            for s in &samples {
                app = Term::App(Box::new(app), Box::new(s.clone()));
            }
            let expected = crate::extraction::emit_value(ctx, &kernel::normalize(ctx, &app));
            let call = format!("{}({})", name, args_rust.join(", "));
            lines.push(format!("    assert_eq!({call}, {expected});"));
            lines.push(format!("    println!(\"{name}(..) = {{:?}}\", {call});"));
        } else {
            // A value: evaluate and self-verify.
            let expected = crate::extraction::emit_value(ctx, &kernel::normalize(ctx, body));
            lines.push(format!("    assert_eq!({name}(), {expected});"));
            lines.push(format!("    println!(\"{name} = {{:?}}\", {name}());"));
        }
    }
    // Run each proven theorem's property check on a sample and assert it holds.
    for (name, param_types) in checks {
        let samples: Option<Vec<kernel::Term>> =
            param_types.iter().map(|p| sample_value(ctx, p)).collect();
        let Some(samples) = samples else { continue };
        let args: Vec<String> =
            samples.iter().map(|s| crate::extraction::emit_value(ctx, s)).collect();
        let call = format!("check_{}({})", name, args.join(", "));
        lines.push(format!("    assert!({call}, \"theorem {name} failed on sample\");"));
        lines.push(format!("    println!(\"\\u{{2713}} {name} holds (checked on a sample)\");"));
    }
    if lines.is_empty() {
        return "\nfn main() {}\n".to_string();
    }
    format!("\nfn main() {{\n{}\n}}\n", lines.join("\n"))
}

/// The parameter types of a (possibly curried) function type `A -> B -> C`.
fn pi_param_types(ty: &kernel::Term) -> Vec<kernel::Term> {
    use kernel::Term;
    let mut params = Vec::new();
    let mut cur = ty;
    while let Term::Pi { param_type, body_type, .. } = cur {
        params.push((**param_type).clone());
        cur = body_type;
    }
    params
}

/// A small sample value of a type, for the demo (a primitive zero, or the first
/// nullary constructor of an inductive). `None` if we can't build one cheaply.
fn sample_value(ctx: &kernel::Context, ty: &kernel::Term) -> Option<kernel::Term> {
    use kernel::{Literal, Term};
    match ty {
        Term::Global(name) => match name.as_str() {
            "Int" => Some(Term::Lit(Literal::Int(0))),
            "Float" => Some(Term::Lit(Literal::Float(0.0))),
            "Text" => Some(Term::Lit(Literal::Text(String::new()))),
            // First nullary constructor of a user inductive (Zero, Yes, MNil, …).
            _ if ctx.is_inductive(name) => ctx
                .get_constructors(name)
                .into_iter()
                .find(|(_, cty)| !matches!(cty, Term::Pi { .. }))
                .map(|(cname, _)| Term::Global(cname.to_string())),
            _ => None,
        },
        _ => None,
    }
}

/// Compile a Math-mode SOURCE program (the editor text) to a Rust module: split
/// it into vernacular statements, run them into a fresh kernel, then extract every
/// user-defined inductive/definition. This is the exact pipeline the Studio's
/// Math "🦀 Compile" button drives, exposed as one function so it is testable.
pub fn extract_math_rust_from_source(input: &str) -> String {
    let mut repl = kernel::interface::Repl::new();
    // Batch execution groups mutually-recursive inductives declared as separate
    // statements into one mutual block (see `Repl::execute_batch`); errors on individual
    // statements (e.g. `Check`/`Eval`) are ignored — only the resulting
    // definitions/inductives matter for extraction.
    let stmts = parse_math_statements(input);
    let _ = repl.execute_batch(&stmts);
    match extract_math_rust(repl.context()) {
        Ok(rust) => rust,
        Err(e) => format!("// extraction error: {e}"),
    }
}

/// Like [`extract_math_rust_from_source`], but produces the main-less module (the
/// linkable artifact bundled into an imperative program's `mod proven`).
pub fn extract_math_module_from_source(input: &str) -> String {
    let mut repl = kernel::interface::Repl::new();
    // Batch execution groups mutually-recursive inductives (`Tree`↔`Forest`) declared as
    // separate statements into one mutual block, so a forward reference between them
    // registers instead of failing its universe check. A source with no forward
    // references behaves exactly as the per-statement loop did.
    let stmts = parse_math_statements(input);
    let _ = repl.execute_batch(&stmts);
    match extract_math_module(repl.context()) {
        Ok(rust) => rust,
        Err(e) => format!("// extraction error: {e}"),
    }
}

/// Partition a possibly-MIXED document into its imperative and math streams.
///
/// A mixed document interleaves imperative LOGOS (`## To`/`## Main`/statements) with
/// Coq-style math (`Definition`/`Inductive`/`Axiom`/`Theorem`/`Lemma`/`Fixpoint`) and
/// literate `## Theorem:`/`## Lemma:` blocks. The math blocks feed the Forge (extracted
/// into `mod proven`); the imperative blocks compile normally and call into it.
///
/// Returns `(imperative_src, Some(math_src))` when math is present, or
/// `(source.to_string(), None)` for a pure imperative program — so that path is a
/// guaranteed no-op (byte-identical compile). In `imperative_src` the math lines are
/// BLANKED (kept as empty lines), not deleted, so imperative line numbers — and thus
/// error spans — stay aligned with the original source.
///
/// `## To`/`## A X is one of` stay imperative (only the Coq keywords and `## Theorem:`/
/// `## Lemma:` route to math), so this never steals an imperative function or enum.
/// IMPLICIT MAIN: a source with no `##` headers whose first non-blank line
/// reads as an imperative statement runs as its own `## Main` body — a bare
/// script no longer parses as logic prose and runs to empty output.
/// Natural-language documents (no statement-keyword opening) are untouched.
/// Returns the wrapped source, or `None` when the input is not a bare script.
pub fn implicit_main(source: &str) -> Option<String> {
    if source.lines().any(|l| l.trim_start().starts_with("##")) {
        return None;
    }
    let first = source.lines().find(|l| !l.trim().is_empty())?;
    const STATEMENT_HEADS: &[&str] = &[
        "Let ", "Show ", "Set ", "If ", "While ", "Repeat ", "Push ",
        "Pop ", "Add ", "Remove ", "Call ", "Return ", "Increase ",
        "Decrease ", "Break", "Inspect ", "Assert ",
    ];
    let t = first.trim_start();
    if STATEMENT_HEADS.iter().any(|h| t.starts_with(h)) {
        Some(format!("## Main\n{}", source))
    } else {
        None
    }
}

pub fn partition_mixed(source: &str) -> (String, Option<String>) {
    if let Some(wrapped) = implicit_main(source) {
        return (wrapped, None);
    }
    let lines: Vec<&str> = source.lines().collect();
    let mut is_math = vec![false; lines.len()];
    let mut i = 0;
    let mut any = false;
    while i < lines.len() {
        let t = lines[i].trim();
        if !is_math_block_start(t) {
            i += 1;
            continue;
        }
        any = true;
        if t.starts_with("## Theorem:") || t.starts_with("## Lemma:") {
            // Literate theorem: header + indented / `Statement:` / `Proof:` lines.
            is_math[i] = true;
            i += 1;
            while i < lines.len() {
                let nt = lines[i].trim();
                let indented = lines[i].starts_with(' ') || lines[i].starts_with('\t');
                if nt.is_empty() {
                    is_math[i] = true;
                    i += 1;
                    continue;
                }
                if indented || nt.starts_with("Statement:") || nt.starts_with("Proof:") {
                    is_math[i] = true;
                    i += 1;
                    if nt.starts_with("Proof:") && nt.ends_with('.') {
                        break;
                    }
                } else {
                    break;
                }
            }
        } else {
            // Coq-style statement: accumulate until a line ending with `.`. Defensive:
            // a `## ` block header (e.g. `## Main`, `## To`) ends the math block even
            // if the terminating `.` is missing/unterminated, so a malformed Definition
            // never swallows the imperative code that follows it.
            let mut ended = t.ends_with('.');
            is_math[i] = true;
            i += 1;
            while !ended && i < lines.len() {
                if lines[i].trim_start().starts_with("## ") {
                    break;
                }
                is_math[i] = true;
                ended = lines[i].trim().ends_with('.');
                i += 1;
            }
        }
    }
    if !any {
        return (source.to_string(), None);
    }
    let imp: Vec<&str> = lines
        .iter()
        .enumerate()
        .map(|(j, l)| if is_math[j] { "" } else { *l })
        .collect();
    let math: Vec<&str> = lines
        .iter()
        .enumerate()
        .filter(|(j, _)| is_math[*j])
        .map(|(_, l)| *l)
        .collect();
    (imp.join("\n"), Some(math.join("\n")))
}

fn is_math_block_start(trimmed: &str) -> bool {
    const COQ: [&str; 6] = [
        "Definition ", "Inductive ", "Axiom ", "Theorem ", "Lemma ", "Fixpoint ",
    ];
    COQ.iter().any(|k| trimmed.starts_with(k))
        || trimmed.starts_with("## Theorem:")
        || trimmed.starts_with("## Lemma:")
}

/// Extract a mixed document's math stream into a bundleable `mod proven` body, or
/// `None` if it has no public items to call into (e.g. only vacuous proofs / parse
/// errors) — so we never emit a `use proven::*;` over an empty module.
///
/// Unlike standalone math extraction, this ALSO emits a `Showable` impl for each
/// non-generic proven enum, so imperative `Show <proven value>` works. `Showable`
/// lives in `logicaffeine_system` (in scope inside `mod proven` via `use super::*;`),
/// which is ONLY linked when bundled into an imperative program — hence it is added
/// here, not in `extract_math_module` (whose output must also compile standalone, dep-free).
pub(crate) fn mixed_proven_module(math_src: &str) -> Option<String> {
    let mut repl = kernel::interface::Repl::new();
    // Batch execution groups mutually-recursive inductives (see `Repl::execute_batch`).
    let stmts = parse_math_statements(math_src);
    let _ = repl.execute_batch(&stmts);
    let ctx = repl.context();
    let mut rust = match extract_math_module(ctx) {
        Ok(r) => r,
        Err(_) => return None,
    };
    if !(rust.contains("pub fn") || rust.contains("pub enum") || rust.contains("pub struct")) {
        return None;
    }
    // Show-verb integration: bridge each non-generic proven enum to the runtime
    // `Showable` trait (via Debug). Sorted for deterministic output.
    let base = baseline_names();
    let mut inds: Vec<String> = Vec::new();
    for (name, _) in ctx.iter_inductives() {
        if !base.contains(name)
            && inductive_is_emittable(ctx, name)
            && inductive_generics(ctx, name).is_empty()
        {
            inds.push(name.to_string());
        }
    }
    inds.sort();
    for ind in inds {
        rust.push_str(&format!(
            "impl Showable for {ind} {{ fn format_show(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {{ write!(f, \"{{:?}}\", self) }} }}\n"
        ));
    }
    Some(rust)
}

/// Split a Math-mode program into complete vernacular statements.
///
/// Handles both Coq-style (period-terminated) and the Literate forms
/// (`## To …` functions, `## Theorem:` blocks, `A X is either …` inductives).
pub fn parse_math_statements(code: &str) -> Vec<String> {
    let mut statements = Vec::new();
    let lines: Vec<&str> = code.lines().collect();
    let mut i = 0;

    while i < lines.len() {
        let line = lines[i];
        let trimmed = line.trim();

        // Skip empty lines and comments
        if trimmed.is_empty() || trimmed.starts_with("--") {
            i += 1;
            continue;
        }

        // Literate function definition: "## To ..."
        if trimmed.starts_with("## To ") {
            let mut block = String::new();
            block.push_str(trimmed);
            i += 1;

            while i < lines.len() {
                let next_line = lines[i];
                let next_trimmed = next_line.trim();

                if next_trimmed.is_empty() {
                    i += 1;
                    continue;
                }
                if next_trimmed.starts_with("--") {
                    i += 1;
                    continue;
                }

                let is_indented = next_line.starts_with(' ') || next_line.starts_with('\t');
                let is_continuation = next_trimmed.starts_with("Consider ")
                    || next_trimmed.starts_with("When ")
                    || next_trimmed.starts_with("Yield ");

                if is_indented || is_continuation {
                    block.push(' ');
                    block.push_str(next_trimmed);
                    i += 1;
                } else {
                    break;
                }
            }

            statements.push(block);
            continue;
        }

        // Literate theorem: "## Theorem: ..." (header + Statement: + Proof:)
        if trimmed.starts_with("## Theorem:") {
            let mut block = String::new();
            block.push_str(trimmed);
            i += 1;

            while i < lines.len() {
                let next_line = lines[i];
                let next_trimmed = next_line.trim();

                if next_trimmed.is_empty() {
                    i += 1;
                    continue;
                }
                if next_trimmed.starts_with("--") {
                    i += 1;
                    continue;
                }

                let is_indented = next_line.starts_with(' ') || next_line.starts_with('\t');
                let is_theorem_part = next_trimmed.starts_with("Statement:")
                    || next_trimmed.starts_with("Proof:");

                if is_indented || is_theorem_part {
                    block.push('\n');
                    block.push_str(next_line);
                    i += 1;
                    if next_trimmed.starts_with("Proof:") && next_trimmed.ends_with('.') {
                        break;
                    }
                } else {
                    break;
                }
            }

            statements.push(block);
            continue;
        }

        // Literate inductive: "A X is either..." / "An X is either..."
        if (trimmed.starts_with("A ") || trimmed.starts_with("An ")) && trimmed.contains(" is either")
        {
            if trimmed.ends_with('.') && !trimmed.trim_end_matches('.').ends_with(':') {
                statements.push(trimmed.to_string());
                i += 1;
                continue;
            }

            let mut block = String::new();
            block.push_str(trimmed);
            i += 1;

            while i < lines.len() {
                let next_line = lines[i];
                let next_trimmed = next_line.trim();

                if next_trimmed.is_empty() {
                    i += 1;
                    continue;
                }
                if next_trimmed.starts_with("--") {
                    i += 1;
                    continue;
                }

                let is_indented = next_line.starts_with(' ') || next_line.starts_with('\t');
                let looks_like_variant = next_trimmed.starts_with("a ")
                    || next_trimmed
                        .chars()
                        .next()
                        .map(|c| c.is_uppercase())
                        .unwrap_or(false);

                if is_indented
                    || (looks_like_variant
                        && !next_trimmed.starts_with("A ")
                        && !next_trimmed.starts_with("An "))
                {
                    if !block.ends_with(':') {
                        block.push_str(" or ");
                    } else {
                        block.push(' ');
                    }
                    block.push_str(next_trimmed.trim_end_matches('.'));
                    i += 1;
                } else {
                    break;
                }
            }

            if !block.ends_with('.') {
                block.push('.');
            }
            statements.push(block);
            continue;
        }

        // Traditional Coq-style: accumulate until period
        let mut current_stmt = String::new();
        while i < lines.len() {
            let line = lines[i];
            let trimmed = line.trim();

            if trimmed.is_empty() || trimmed.starts_with("--") {
                i += 1;
                continue;
            }

            if !current_stmt.is_empty() {
                current_stmt.push(' ');
            }
            current_stmt.push_str(trimmed);
            i += 1;

            if trimmed.ends_with('.') {
                break;
            }
        }

        if !current_stmt.is_empty() {
            statements.push(current_stmt);
        }
    }

    statements
}

/// Compile a Logic-mode input to runnable Rust.
///
/// A `## Theorem:` block becomes a **model-checker** for `premises ⊨ goal`; a
/// plain sentence becomes a model-checker for that formula — a self-contained
/// `Model` + `holds(&Model) -> bool` + demo `main` you can compile and run
/// ([`crate::extraction::fol_model`]). The finite-domain *puzzle* (Simon grid) is
/// the exception: it would run the full solver synchronously and freeze the page,
/// so it bails with a note (use Execute to solve it).
pub fn extract_logic_rust(input: &str) -> Result<String, String> {
    extract_logic_impl(input, true)
}

/// Like [`extract_logic_rust`], but emits NO demo `main` — the linkable form bundled
/// into an imperative program's `mod proven` (which has its own `main`). Produces the
/// `World` + `holds` (+ `Monitor`) library items only.
pub fn extract_logic_module(input: &str) -> Result<String, String> {
    extract_logic_impl(input, false)
}

fn extract_logic_impl(input: &str, emit_main: bool) -> Result<String, String> {
    use crate::extraction::fol_model::{fol_to_model_checker, fol_to_model_checker_module};
    let emit = |premises: &[logicaffeine_proof::ProofExpr],
                goal: &logicaffeine_proof::ProofExpr,
                english: &str,
                fol: &str| {
        if emit_main {
            fol_to_model_checker(premises, goal, english, fol)
        } else {
            fol_to_model_checker_module(premises, goal, english, fol)
        }
    };

    // A `## Theorem:` block: model-check `premises ⊨ goal` — unless it is a grid
    // puzzle (bail before the solver runs, see above).
    if let Ok((premises, goal)) = theorem_proof_exprs(input) {
        if looks_like_grid(&premises) {
            return Ok("// this is a finite-domain puzzle — run it with Execute. \
                Compiling it to Rust would run the full solver synchronously."
                .to_string());
        }
        let fol = if premises.is_empty() {
            goal.to_string()
        } else {
            format!(
                "{} ⊢ {}",
                premises.iter().map(|p| p.to_string()).collect::<Vec<_>>().join(", "),
                goal
            )
        };
        return Ok(emit(&premises, &goal, input, &fol));
    }

    // A plain sentence: model-check the single formula.
    let proof = compile_for_proof(input);
    if let Some(expr) = proof.proof_expr {
        let fol = proof.logic_string.clone().unwrap_or_else(|| expr.to_string());
        return Ok(emit(&[], &expr, input, &fol));
    }

    Ok("// could not parse this input into a logical formula to compile".to_string())
}

/// Ask the Z3 oracle whether a theorem's premises semantically entail its goal.
///
/// Same `## Theorem:` block format as [`verify_theorem`], but the answer is an
/// [`SmtVerdict`](logicaffeine_proof::oracle::SmtVerdict) over the standard
/// translation (modal frame axioms, similarity relations, lattice axioms) —
/// **not** a kernel-certified proof. Use this for entailment questions the
/// monotonic kernel cannot express (modality, counterfactuals, mereology).
/// Assemble the lexicon-derived [`SmtTheory`](logicaffeine_proof::oracle::SmtTheory)
/// for a theorem: every mentioned predicate the lexicon tags as a MASS noun
/// becomes a cumulative (Link-lattice) predicate.
#[cfg(feature = "verification")]
fn smt_theory_for(
    premises: &[ProofExpr],
    goal: Option<&ProofExpr>,
) -> logicaffeine_proof::oracle::SmtTheory {
    let mut exprs: Vec<ProofExpr> = premises.to_vec();
    if let Some(g) = goal {
        exprs.push(g.clone());
    }
    let cumulative_predicates = logicaffeine_proof::oracle::predicate_names(&exprs)
        .into_iter()
        .filter(|name| logicaffeine_language::lexicon::is_mass_noun(name))
        .collect();
    logicaffeine_proof::oracle::SmtTheory {
        cumulative_predicates,
    }
}

#[cfg(feature = "verification")]
pub fn check_theorem_smt(
    input: &str,
) -> Result<logicaffeine_proof::oracle::SmtVerdict, ParseError> {
    let (proof_exprs, goal_expr) = theorem_proof_exprs(input)?;
    let theory = smt_theory_for(&proof_exprs, Some(&goal_expr));
    Ok(logicaffeine_proof::oracle::oracle_entails_with_theory(
        &proof_exprs,
        &goal_expr,
        &theory,
    ))
}

/// Ask the Z3 oracle whether a theorem's premises are jointly satisfiable
/// (the goal is parsed but ignored).
///
/// Every non-entailment claim in the test suite pairs with this check so an
/// inconsistent premise set cannot fake a `NotEntailed` via vacuity.
#[cfg(feature = "verification")]
pub fn check_theorem_premises_consistent(
    input: &str,
) -> Result<logicaffeine_proof::oracle::SmtConsistency, ParseError> {
    let (proof_exprs, _goal) = theorem_proof_exprs(input)?;
    let theory = smt_theory_for(&proof_exprs, None);
    Ok(logicaffeine_proof::oracle::oracle_consistent_with_theory(
        &proof_exprs,
        &theory,
    ))
}

/// Ask the defeasible (non-monotonic) layer whether a theorem's premises
/// defeasibly entail its goal: generics and implicatures license cancellable
/// inferences via circumscription over per-rule abnormality predicates.
///
/// The verdict is **not kernel-certified** and is strictly weaker than
/// classical entailment — a defeated default reads as `NotEntailed` while the
/// premise set stays [`SmtConsistency::Consistent`].
#[cfg(feature = "verification")]
pub fn check_theorem_defeasible(
    input: &str,
) -> Result<logicaffeine_proof::oracle::SmtVerdict, ParseError> {
    let (proof_exprs, goal, defaults, _definitions) = theorem_problem(input, true)?;
    let theory = smt_theory_for(&proof_exprs, Some(&goal));
    Ok(crate::defeasible::defeasible_entails(
        &proof_exprs,
        &goal,
        &defaults,
        &theory,
    ))
}

/// Consistency under the defeasible layer: defaults defeated by exceptions
/// must NOT read as contradictions.
#[cfg(feature = "verification")]
pub fn check_theorem_defeasible_consistent(
    input: &str,
) -> Result<logicaffeine_proof::oracle::SmtConsistency, ParseError> {
    let (proof_exprs, _goal, defaults, _definitions) = theorem_problem(input, true)?;
    let theory = smt_theory_for(&proof_exprs, None);
    Ok(crate::defeasible::defeasible_consistent(
        &proof_exprs,
        &defaults,
        &theory,
    ))
}

/// Parse a `## Theorem:` block and convert its premises and goal to
/// [`ProofExpr`]s — the shared front half of [`verify_theorem`] and
/// `check_theorem_smt`.
/// Parse a `## Theorem` document and convert its premises and goal to
/// [`ProofExpr`]. Public so a puzzle solver can obtain the parsed-FOL premises
/// (the Given clues/declarations) and feed them to the entailment oracle.
pub fn theorem_proof_exprs(input: &str) -> Result<(Vec<ProofExpr>, ProofExpr), ParseError> {
    let (premises, goal, _defaults, _definitions) = theorem_problem(input, false)?;
    Ok((premises, goal))
}

/// Like [`theorem_proof_exprs`] but also returns the document's `## Define`
/// blocks (Rung 0a) lowered to proof-layer definitions, so the prover can
/// δ-unfold them. This is the entry the kernel-certified theorem path uses.
pub fn theorem_proof_exprs_with_defs(
    input: &str,
) -> Result<
    (
        Vec<ProofExpr>,
        ProofExpr,
        Vec<logicaffeine_proof::verify::Definition>,
    ),
    ParseError,
> {
    let (premises, goal, _defaults, definitions) = theorem_problem(input, false)?;
    Ok((premises, goal, definitions))
}

/// The `uses` dependency graph (Rung 0b) for an English document: its `## Define`
/// blocks plus the theorem's premises and goal, lowered and analyzed. Each node
/// is a definition or the theorem; each edge is a `uses`. This is the structure
/// a `mathscrapes` node/edge compiles into.
pub fn theorem_dependency_graph(
    input: &str,
) -> Result<logicaffeine_proof::verify::DependencyGraph, ParseError> {
    let (premises, goal, definitions) = theorem_proof_exprs_with_defs(input)?;
    Ok(logicaffeine_proof::verify::dependency_graph(
        &definitions,
        &premises,
        &goal,
    ))
}

/// ANSWER a wh-question theorem ("Given: … Prove: Who is a lawyer?"). A wh-goal
/// converts to ∃x.φ(x); the ANSWER is the witness. We find it the same way the
/// puzzle is meant to be solved — by the GENERAL kernel-certified prover, applied
/// per candidate: enumerate the domain individuals named in the premises and prove
/// each candidate goal φ(c); the c's that prove are the answer (usually one, for a
/// puzzle with a unique solution). No question-specific reasoning — every step is
/// the same `prove_certify_check` that proves Socrates.
pub fn answer_question(input: &str) -> Result<Vec<String>, ParseError> {
    let (premises, goal) = theorem_proof_exprs(input)?;
    match &goal {
        ProofExpr::Exists { variable, body } => Ok(answer_wh(&premises, variable, body)),
        _ => Err(ParseError {
            kind: logicaffeine_language::error::ParseErrorKind::Custom(
                "Prove goal is not a question (expected a wh-question ∃-form)".to_string(),
            ),
            span: logicaffeine_language::token::Span::default(),
        }),
    }
}

/// The witnesses of a wh-question `∃var. body(var)`: enumerate the domain individuals
/// named in the premises and keep those `c` for which `body(c)` is entailed by the
/// certified no-Z3 path. Shared by [`answer_question`] and the studio entry
/// [`compile_theorem_for_ui`], so a wh-goal never reaches the closed-goal search.
fn answer_wh(premises: &[ProofExpr], var: &str, body: &ProofExpr) -> Vec<String> {
    let mut candidates: Vec<String> = Vec::new();
    for p in premises {
        collect_constants(p, &mut candidates);
    }
    candidates.sort();
    candidates.dedup();
    // Prepare the premise set ONCE (tense-erase + finite-domain grounding) — it is
    // identical for every candidate cell, so doing it per candidate would re-ground the
    // whole grid N times. This is the optimization that keeps bigger puzzles tractable.
    let trace = std::env::var("LOGOS_TRACE").is_ok();
    let t0 = trace.then(std::time::Instant::now);
    let prepared = prepare_premises(premises);
    if let Some(t0) = t0 {
        eprintln!(
            "[answer] {} premises → prepared ({} clauses) in {:.2?}; {} candidates",
            premises.len(),
            prepared.len(),
            t0.elapsed(),
            candidates.len()
        );
    }
    let mut answers = Vec::new();
    for c in &candidates {
        let tc = trace.then(std::time::Instant::now);
        let candidate_goal =
            logicaffeine_language::proof_convert::instantiate_var_with_constant(body, var, c);
        let ok = candidate_entailed_prepared(&prepared, &candidate_goal);
        if let Some(tc) = tc {
            eprintln!("[answer]   {:<14} {} ({:.2?})", c, ok, tc.elapsed());
        }
        if ok {
            answers.push(c.clone());
        }
    }
    if let Some(t0) = t0 {
        eprintln!("[answer] total {:.2?} → {:?}", t0.elapsed(), answers);
    }
    answers
}

/// SOLVE the whole grid for the studio (the form-recognized easter egg): re-parse the
/// theorem and, if it is a finite-domain grid, fill every cell the certified prover can
/// force. `None` for a non-grid theorem.
pub fn solve_grid(input: &str) -> Option<SolvedGrid> {
    let parsed = parse_theorem(input).ok()?;
    if !looks_like_grid(&parsed.premises) {
        return None;
    }
    solve_grid_from_premises(&parsed.premises, input)
}

/// One category closure `∀x(RowSort(x) → d₁ ∨ … ∨ dₙ)`: the bound variable, its row
/// sort, and each disjunct paired with the value it contributes.
struct GridClosure {
    var: String,
    row_sort: String,
    disjuncts: Vec<(String, ProofExpr)>,
}

/// Split a (possibly nested) disjunction into its leaf disjuncts.
fn flatten_or<'a>(e: &'a ProofExpr, out: &mut Vec<&'a ProofExpr>) {
    match e {
        ProofExpr::Or(l, r) => {
            flatten_or(l, out);
            flatten_or(r, out);
        }
        other => out.push(other),
    }
}

/// The unary SORT predicate guarding `var` (`Trip(x)` → "trip"), if the antecedent is
/// one — possibly under conjunction.
fn antecedent_sort(e: &ProofExpr, var: &str) -> Option<String> {
    let is_var = |t: &ProofTerm| matches!(t, ProofTerm::Variable(v) | ProofTerm::BoundVarRef(v) if v == var);
    match e {
        ProofExpr::Predicate { name, args, .. } if args.len() == 1 && is_var(&args[0]) => {
            Some(name.clone())
        }
        ProofExpr::And(l, r) => antecedent_sort(l, var).or_else(|| antecedent_sort(r, var)),
        _ => None,
    }
}

/// The value a disjunct contributes for `var`: the constant argument of a binary relation
/// (`in(x, Florida)` → "Florida") or the predicate name of a unary one (`cycling(x)` →
/// "cycling"). `None` if the disjunct is not a simple predicate over `var`.
fn disjunct_value(d: &ProofExpr, var: &str) -> Option<String> {
    let is_var = |t: &ProofTerm| matches!(t, ProofTerm::Variable(v) | ProofTerm::BoundVarRef(v) if v == var);
    match d {
        ProofExpr::Predicate { name, args, .. } => match args.as_slice() {
            [a] if is_var(a) => Some(name.clone()),
            [a, ProofTerm::Constant(c)] if is_var(a) => Some(c.clone()),
            [ProofTerm::Constant(c), a] if is_var(a) => Some(c.clone()),
            _ => None,
        },
        _ => None,
    }
}

/// Extract one [`GridClosure`] per disjunctive-closure premise (`∀x(Trip(x) → A ∨ B ∨
/// …)`). These closures ARE the grid's columns — the antecedent names the rows, the
/// disjuncts name the cells.
fn extract_grid_closures(premises: &[ProofExpr]) -> Vec<GridClosure> {
    fn from_forall(e: &ProofExpr, out: &mut Vec<GridClosure>) {
        if let ProofExpr::ForAll { variable, body } = e {
            match body.as_ref() {
                ProofExpr::Implies(ante, cons) => {
                    let mut leaves = Vec::new();
                    flatten_or(cons, &mut leaves);
                    let disjuncts: Vec<(String, ProofExpr)> = leaves
                        .iter()
                        .filter_map(|d| disjunct_value(d, variable).map(|v| (v, (*d).clone())))
                        .collect();
                    if let Some(row_sort) = antecedent_sort(ante, variable) {
                        // Only a closure whose every disjunct is a clean predicate over the
                        // row var defines a column (guards against a stray non-grid ∀).
                        if !disjuncts.is_empty() && disjuncts.len() == leaves.len() {
                            out.push(GridClosure { var: variable.clone(), row_sort, disjuncts });
                        }
                    }
                }
                ProofExpr::ForAll { .. } => from_forall(body, out),
                _ => {}
            }
        }
    }
    let mut out = Vec::new();
    for p in premises {
        from_forall(&erase_tense(p), &mut out);
    }
    out
}

/// Title-case a lowercased sort name for a column/row header ("trip" → "Trip").
fn title_case(s: &str) -> String {
    let mut chars = s.chars();
    match chars.next() {
        Some(first) => first.to_uppercase().collect::<String>() + chars.as_str(),
        None => String::new(),
    }
}

/// A human header for a category column: the declared sort whose domain covers the
/// column's values (case-insensitive), else a positional fallback.
fn grid_column_label(
    values: &[String],
    sorts: &std::collections::HashMap<String, Vec<ProofTerm>>,
    idx: usize,
) -> String {
    let want: Vec<String> = values.iter().map(|v| v.to_lowercase()).collect();
    let mut keys: Vec<&String> = sorts.keys().collect();
    keys.sort();
    for k in keys {
        let dom: std::collections::HashSet<String> = sorts[k]
            .iter()
            .filter_map(|t| match t {
                ProofTerm::Constant(c) => Some(c.to_lowercase()),
                _ => None,
            })
            .collect();
        if !dom.is_empty() && want.iter().all(|v| dom.contains(v)) {
            return title_case(k);
        }
    }
    format!("Category {}", idx + 1)
}

/// A coarse stem that bridges a gerund surface form and the base form it normalizes to
/// (`"Cycling"` and `"Cycle"` → `"cycl"`), so a grid value can be matched back to the word
/// the user actually wrote. Strips a trailing `-ing` then a trailing `-e`; leaves
/// already-base words (constants, years, names) effectively unchanged.
fn category_stem(s: &str) -> String {
    let mut w = s.to_lowercase();
    if w.len() > 5 && w.ends_with("ing") {
        w.truncate(w.len() - 3);
    }
    if w.len() > 3 && w.ends_with('e') {
        w.truncate(w.len() - 1);
    }
    w
}

/// Map each grid value back to the SURFACE word the user wrote. Category values
/// normalize to a base form in the FOL (a relation's constant like `Florida` survives, but
/// a gerund activity becomes `Cycle`); the original word ("cycling") lives only in the
/// input text. Key the input words by stem and recover them. First occurrence wins, so the
/// declaration's spelling is preferred.
fn surface_form_map(input: &str) -> std::collections::HashMap<String, String> {
    let mut map = std::collections::HashMap::new();
    for word in input.split(|c: char| !c.is_alphanumeric()) {
        if !word.is_empty() {
            map.entry(category_stem(word)).or_insert_with(|| word.to_string());
        }
    }
    map
}

/// Fill the grid from already-parsed premises: rows are the row sort's domain, columns
/// are the disjunctive closures, and each cell is the disjunct value the certified no-Z3
/// prover entails for that row (or `None` if undetermined). Same engine `answer_question`
/// uses per cell, swept across the whole table.
fn solve_grid_from_premises(premises: &[ProofExpr], input: &str) -> Option<SolvedGrid> {
    let closures = extract_grid_closures(premises);
    if closures.is_empty() {
        return None;
    }
    let untensed: Vec<ProofExpr> = premises.iter().map(erase_tense).collect();
    let sorts = logicaffeine_proof::grounding::sort_domains(&untensed);
    let row_sort = closures[0].row_sort.clone();
    let rows: Vec<String> = sorts
        .get(&row_sort)
        .map(|dom| {
            dom.iter()
                .filter_map(|t| match t {
                    ProofTerm::Constant(c) => Some(c.clone()),
                    _ => None,
                })
                .collect()
        })
        .unwrap_or_default();
    if rows.is_empty() {
        return None;
    }
    // Ground ONCE, with functionality, exactly as the fast certified grid path does — the
    // exclusion lemmas let RUP decide each non-entailed candidate in ~1ms.
    let prepared = prepare_premises_opts(premises, true);
    let surface = surface_form_map(input);
    let display = |v: &str| surface.get(&category_stem(v)).cloned().unwrap_or_else(|| v.to_string());
    let mut columns = Vec::new();
    for clo in &closures {
        if clo.row_sort != row_sort {
            continue;
        }
        let values: Vec<String> = clo.disjuncts.iter().map(|(v, _)| display(v)).collect();
        let mut cells = Vec::with_capacity(rows.len());
        for r in &rows {
            let mut found = None;
            for (label, dj) in &clo.disjuncts {
                let atom = erase_tense(
                    &logicaffeine_language::proof_convert::instantiate_var_with_constant(
                        dj, &clo.var, r,
                    ),
                );
                if candidate_entailed_prepared(&prepared, &atom) {
                    found = Some(display(label));
                    break;
                }
            }
            cells.push(found);
        }
        let label = grid_column_label(&values, &sorts, columns.len());
        columns.push(GridColumn { label, values, cells });
    }
    Some(SolvedGrid { row_label: title_case(&row_sort), rows, columns })
}

/// Prepare the premise set for the per-candidate solve, ONCE.
///
/// A logic grid is FINITE and STATIC, so we (1) ERASE TENSE — a grid is one static
/// scenario; the past-tense clue wrappers ("one WAS in Connecticut") carry no
/// information — and (2) GROUND the finite domain SORT-AWARE: each guarded quantifier
/// expands over its sort's declared domain (∀x(Trip(x)→…) over the four trips, not the
/// whole universe). The result is quantifier-free (decidable — neither our kernel nor
/// Z3 can instantiate ∀/∃ forever) and identical for every cell, so the per-cell
/// entailment check below sees a finite, decidable problem.
///
/// This is GENERAL over any declared finite-domain grid: the domains come from the
/// English declarations the parser produced (`sort_domains`), never from puzzle
/// knowledge. Grounding the KERNEL path (not just the Z3 path) is what lets the
/// certified prover close a full multi-category grid in the browser.
/// Does this premise set describe a finite-domain GRID — a declared bijection with a
/// closure (`∀x(Sort(x) → A ∨ B ∨ …)`) or an "exactly one" — as opposed to an open
/// syllogism? Only grids are grounded before proving (so Socrates keeps its
/// `UniversalInst` trace). Structural, never keyed to a particular puzzle.
fn looks_like_grid(premises: &[ProofExpr]) -> bool {
    fn has_disjunctive_closure(e: &ProofExpr) -> bool {
        match e {
            ProofExpr::ForAll { body, .. } => match body.as_ref() {
                ProofExpr::Implies(_, c) => matches!(c.as_ref(), ProofExpr::Or(..)),
                ProofExpr::ForAll { .. } => has_disjunctive_closure(body),
                _ => false,
            },
            ProofExpr::Temporal { body, .. } => has_disjunctive_closure(body),
            _ => false,
        }
    }
    !logicaffeine_proof::grounding::at_most_one_lemmas(premises).is_empty()
        || premises.iter().any(has_disjunctive_closure)
}

fn prepare_premises(premises: &[ProofExpr]) -> Vec<ProofExpr> {
    prepare_premises_opts(premises, false)
}

/// `prepare_premises`, optionally adding the bijection's FUNCTIONALITY half
/// (`∀x(Li → ¬Lj)`: a row takes at most one value per category). Functionality is
/// given ONLY to the incremental grid solver — it propagates the exclusions cheaply —
/// never to the re-saturating backward chainer, which it would slow to a crawl.
fn prepare_premises_opts(premises: &[ProofExpr], with_functionality: bool) -> Vec<ProofExpr> {
    let mut untensed: Vec<ProofExpr> = premises.iter().map(erase_tense).collect();
    // "Exactly one φ" ⟹ pairwise "at most one φ": the existence form grounds to a
    // disjunction the certified kernel cannot use directly; its entailed pairwise
    // uniqueness grounds to the conjunctive exclusion rules unit propagation runs on.
    let lemmas = logicaffeine_proof::grounding::at_most_one_lemmas(&untensed);
    untensed.extend(lemmas);
    // Definite-description clues ("the Florida trip is the hunting trip", "the trip with
    // Yvonne wasn't in Kentucky") name a unique row by a singleton value; rewrite them
    // to row-indexed implications the solver propagates, instead of an existential whose
    // grounding explodes to a row-disjunction.
    let defns = logicaffeine_proof::grounding::definite_property_implications(&untensed);
    untensed.extend(defns);
    if with_functionality {
        let func = logicaffeine_proof::grounding::functionality_lemmas(&untensed);
        untensed.extend(func);
        // HIDDEN SINGLE: the column closures (`⋁_r In(r, v)`) — the existence half of each
        // "exactly one" that `compile` otherwise drops. With functionality they let the
        // solver force a value's last remaining row, deepening the ROOT fixpoint so more
        // cells take the linear prove_var exit instead of search.
        let cols = logicaffeine_proof::grounding::column_closure_lemmas(&untensed);
        untensed.extend(cols);
    }
    let fallback = logicaffeine_proof::grounding::domain_constants(&untensed);
    let mut sorts = logicaffeine_proof::grounding::sort_domains(&untensed);
    bind_occasion_synonyms_to_row_domain(&untensed, &mut sorts);
    let grounded: Vec<ProofExpr> = untensed
        .iter()
        .map(|p| logicaffeine_proof::grounding::ground_sorted(p, &sorts, &fallback))
        .collect();
    // Discharge the now-true sort facts (`Trip(Alpha)`, …) so propagation runs on pure
    // value-literals: a closure becomes a bare disjunction, an at-most-one a one-step
    // exclusion rule. This is what keeps the full-grid search tractable.
    let discharged = logicaffeine_proof::grounding::discharge_unary_facts(&grounded);
    // Fold trivial reflexive identities: a grounded `∃x∃y` of-pair's diagonal
    // (`x = y = c`) carries `¬(c = c) = False`, so that disjunct drops.
    logicaffeine_proof::grounding::simplify_trivial_identities(&discharged)
}

/// Does the premise set entail the candidate answer goal?
///
/// Without `verification`: the KERNEL-CERTIFIED prover over the GROUNDED, tense-erased
/// grid. Grounding made the problem quantifier-free, so the kernel's case-split +
/// reductio strategies close each cell with a real proof term — no Z3, runs in WASM.
/// The bounded depth makes an unentailed cell fail fast rather than search to the hilt.
#[cfg(not(feature = "verification"))]
fn candidate_entailed_prepared(prepared: &[ProofExpr], goal: &ProofExpr) -> bool {
    let g = erase_tense(goal);
    // FAST certified tier first: the CDCL+RUP engine decides the grounded grid in ~1ms with
    // an independently-checked proof (and answers a NON-entailed candidate just as fast,
    // where the kernel path degenerates to a deep futile search). Fall back to the bounded
    // kernel chainer only when the problem isn't purely propositional (RUP returns `None`).
    match logicaffeine_proof::rup::entails_certified(prepared, &g) {
        Some(logicaffeine_proof::rup::Verdict::Entailed) => return true,
        Some(logicaffeine_proof::rup::Verdict::NotEntailed) => return false,
        None => {}
    }
    logicaffeine_proof::verify::prove_certify_check_bounded(prepared, &g, 100).verified
}

/// OCCASION soft-typing for the finite-domain solve.
///
/// In a logic grid the row entity is described by several OCCASION-sort head nouns
/// — "the hunting VACATION", "the Florida TRIP", "the 2004 HOLIDAY" — where the
/// modifier does the referring and the head is a soft type ([`Sort::is_occasion`],
/// the same principle [`drs::DiscourseModel::resolve_definite_by_modifier`] uses for
/// coreference). Only ONE of these head nouns is declared with a domain
/// (`Trip(Alpha)…`); the synonyms carry no domain, so an occasion guard like
/// `Vacation(x)` would ground over the WHOLE universe and a phantom non-row constant
/// (`Florida`, `2003`, …) could satisfy a synonym-headed clue. Point every
/// occasion-sorted guard at the shared row domain (the union of the declared
/// occasion-sort domains).
///
/// This is NOT global synonymy — no `Vacation ↔ Trip` axiom is asserted and "a work
/// trip" never collapses; the binding fires only for occasion-SORTED head predicates
/// inside the grounded solve, leaving ordinary parsing untouched.
fn bind_occasion_synonyms_to_row_domain(
    premises: &[ProofExpr],
    sorts: &mut std::collections::HashMap<String, Vec<logicaffeine_proof::ProofTerm>>,
) {
    use logicaffeine_language::lexicon::lookup_sort;
    use logicaffeine_proof::ProofTerm;
    let is_occasion = |n: &str| lookup_sort(n).map_or(false, |s| s.is_occasion());

    // The shared row domain: the union of every declared occasion-sort domain. In a
    // grid there is one row sort with a domain (`Trip`); its synonyms have none.
    let mut row_domain: Vec<ProofTerm> = Vec::new();
    for (name, dom) in sorts.iter() {
        if is_occasion(name) {
            for c in dom {
                if !row_domain.contains(c) {
                    row_domain.push(c.clone());
                }
            }
        }
    }
    if row_domain.is_empty() {
        return;
    }

    // Every occasion-sorted guard that appears in the premises but lacks its own
    // declared domain inherits the row domain.
    let mut names = HashSet::new();
    for p in premises {
        collect_unary_predicate_names(p, &mut names);
    }
    for name in names {
        if is_occasion(&name) {
            sorts.entry(name).or_insert_with(|| row_domain.clone());
        }
    }
}

/// Collect every UNARY predicate name appearing anywhere in `e` — the candidate sort
/// guards the occasion-binding inspects.
fn collect_unary_predicate_names(e: &ProofExpr, out: &mut HashSet<String>) {
    match e {
        ProofExpr::Predicate { name, args, .. } if args.len() == 1 => {
            out.insert(name.clone());
        }
        ProofExpr::And(l, r)
        | ProofExpr::Or(l, r)
        | ProofExpr::Implies(l, r)
        | ProofExpr::Iff(l, r) => {
            collect_unary_predicate_names(l, out);
            collect_unary_predicate_names(r, out);
        }
        ProofExpr::Counterfactual { antecedent, consequent } => {
            collect_unary_predicate_names(antecedent, out);
            collect_unary_predicate_names(consequent, out);
        }
        ProofExpr::Not(x)
        | ProofExpr::ForAll { body: x, .. }
        | ProofExpr::Exists { body: x, .. }
        | ProofExpr::Temporal { body: x, .. }
        | ProofExpr::Modal { body: x, .. } => collect_unary_predicate_names(x, out),
        _ => {}
    }
}

#[cfg(feature = "verification")]
fn candidate_entailed_prepared(prepared: &[ProofExpr], goal: &ProofExpr) -> bool {
    // The candidate goal is a ground predicate (no quantifiers) — just erase tense.
    let g = erase_tense(goal);
    // FAST certified tier first: the CDCL+RUP engine decides the grounded grid in ~1ms with
    // an independently-checked proof. Only if it can't encode the problem do we fall back to
    // the bounded kernel chainer, then Z3.
    match logicaffeine_proof::rup::entails_certified(prepared, &g) {
        Some(logicaffeine_proof::rup::Verdict::Entailed) => return true,
        Some(logicaffeine_proof::rup::Verdict::NotEntailed) => return false,
        None => {}
    }
    if logicaffeine_proof::verify::prove_certify_check_bounded(prepared, &g, 40).verified {
        return true;
    }
    matches!(
        logicaffeine_proof::oracle::oracle_entails(prepared, &g),
        logicaffeine_proof::oracle::SmtVerdict::Entailed
    )
}

/// Strip `Temporal` wrappers throughout a [`ProofExpr`] — a logic grid is one
/// static scenario, so tense carries no information and would otherwise block the
/// oracle from forcing values across past-tense clues.
fn erase_tense(e: &ProofExpr) -> ProofExpr {
    match e {
        ProofExpr::Temporal { body, .. } => erase_tense(body),
        ProofExpr::And(l, r) => {
            ProofExpr::And(Box::new(erase_tense(l)), Box::new(erase_tense(r)))
        }
        ProofExpr::Or(l, r) => ProofExpr::Or(Box::new(erase_tense(l)), Box::new(erase_tense(r))),
        ProofExpr::Implies(l, r) => {
            ProofExpr::Implies(Box::new(erase_tense(l)), Box::new(erase_tense(r)))
        }
        ProofExpr::Iff(l, r) => ProofExpr::Iff(Box::new(erase_tense(l)), Box::new(erase_tense(r))),
        ProofExpr::Not(x) => ProofExpr::Not(Box::new(erase_tense(x))),
        ProofExpr::ForAll { variable, body } => ProofExpr::ForAll {
            variable: variable.clone(),
            body: Box::new(erase_tense(body)),
        },
        ProofExpr::Exists { variable, body } => ProofExpr::Exists {
            variable: variable.clone(),
            body: Box::new(erase_tense(body)),
        },
        other => other.clone(),
    }
}

fn collect_constants(e: &ProofExpr, out: &mut Vec<String>) {
    use logicaffeine_proof::ProofTerm;
    fn term(t: &ProofTerm, out: &mut Vec<String>) {
        match t {
            ProofTerm::Constant(s) => out.push(s.clone()),
            ProofTerm::Function(_, args) | ProofTerm::Group(args) => {
                args.iter().for_each(|a| term(a, out))
            }
            _ => {}
        }
    }
    match e {
        ProofExpr::Predicate { args, .. } => args.iter().for_each(|a| term(a, out)),
        ProofExpr::Identity(a, b) => {
            term(a, out);
            term(b, out);
        }
        ProofExpr::And(l, r)
        | ProofExpr::Or(l, r)
        | ProofExpr::Implies(l, r)
        | ProofExpr::Iff(l, r) => {
            collect_constants(l, out);
            collect_constants(r, out);
        }
        ProofExpr::Not(x) => collect_constants(x, out),
        ProofExpr::ForAll { body, .. }
        | ProofExpr::Exists { body, .. }
        | ProofExpr::Temporal { body, .. } => collect_constants(body, out),
        ProofExpr::Term(t) => term(t, out),
        _ => {}
    }
}

/// Lower a `## Define` block (Rung 0a) to a proof-layer definition: the LHS
/// predicate supplies the (normalized) definiendum name and parameter symbols,
/// the RHS is the definiens. Both sides pass through the SAME
/// `logic_expr_to_proof_expr` + interner as theorem premises and goals, so the
/// definiendum name and parameters match the predicate occurrences the prover
/// sees. Returns `None` if the LHS is not a predicate application.
fn lower_definition(
    def: &logicaffeine_language::ast::DefinitionBlock,
    interner: &Interner,
) -> Option<logicaffeine_proof::verify::Definition> {
    use logicaffeine_language::proof_convert::logic_expr_to_proof_expr;
    let (name, params) = match logic_expr_to_proof_expr(def.definiendum, interner) {
        ProofExpr::Predicate { name, args, .. } => {
            let params = args
                .iter()
                .filter_map(|t| match t {
                    ProofTerm::Constant(n) | ProofTerm::Variable(n) => Some(n.clone()),
                    _ => None,
                })
                .collect();
            (name, params)
        }
        _ => return None,
    };
    let definiens = logic_expr_to_proof_expr(def.definiens, interner);
    Some(logicaffeine_proof::verify::Definition {
        name,
        params,
        definiens,
    })
}

/// [`theorem_proof_exprs`] with optional DEFEASIBLE conversion: premises keep
/// their generics/implicatures as abnormality-guarded defaults (returned for
/// the circumscription pass); the goal always converts strictly. Also returns
/// the document's `## Define` blocks lowered to proof-layer definitions (Rung 0a).
fn theorem_problem(
    input: &str,
    defeasible: bool,
) -> Result<
    (
        Vec<ProofExpr>,
        ProofExpr,
        Vec<logicaffeine_language::proof_convert::DefaultRule>,
        Vec<logicaffeine_proof::verify::Definition>,
    ),
    ParseError,
> {
    // === STEP 1: Parse ===
    let mut interner = Interner::new();
    let mut lexer = Lexer::new(input, &mut interner);
    let tokens = lexer.tokenize();

    let mwe_trie = mwe::build_mwe_trie();
    let tokens = mwe::apply_mwe_pipeline(tokens, &mwe_trie, &mut interner);

    let type_registry = {
        let mut discovery = DiscoveryPass::new(&tokens, &mut interner);
        discovery.run()
    };

    let expr_arena = Arena::new();
    let term_arena = Arena::new();
    let np_arena = Arena::new();
    let sym_arena = Arena::new();
    let role_arena = Arena::new();
    let pp_arena = Arena::new();

    let ctx = AstContext::new(
        &expr_arena,
        &term_arena,
        &np_arena,
        &sym_arena,
        &role_arena,
        &pp_arena,
    );

    let mut world_state = drs::WorldState::new();
    let mut parser = Parser::new(tokens, &mut world_state, &mut interner, ctx, type_registry);
    if defeasible {
        // The defeasible layer reasons over the pragmatic channel too:
        // scalar implicatures become guarded defaults.
        parser.set_pragmatic_mode(true);
    }
    let statements = parser.parse_program()?;

    let theorem = statements
        .iter()
        .find_map(|stmt| {
            if let Stmt::Theorem(t) = stmt {
                Some(t)
            } else {
                None
            }
        })
        .ok_or_else(|| ParseError {
            kind: logicaffeine_language::error::ParseErrorKind::Custom("No theorem block found in input".to_string()),
            span: logicaffeine_language::token::Span::default(),
        })?;

    // === STEP 2: Convert premises and goal to ProofExpr ===
    let mut defaults = Vec::new();
    let proof_exprs: Vec<ProofExpr> = theorem
        .premises
        .iter()
        .map(|premise| {
            if defeasible {
                logicaffeine_language::proof_convert::logic_expr_to_proof_expr_defeasible(
                    premise,
                    &interner,
                    &mut defaults,
                )
            } else {
                logic_expr_to_proof_expr(premise, &interner)
            }
        })
        .collect();
    let goal_expr = logic_expr_to_proof_expr(theorem.goal, &interner);

    // Rung 0a: collect every `## Define` block in the document, lowered to a
    // proof-layer definition the prover can δ-unfold.
    let definitions: Vec<logicaffeine_proof::verify::Definition> = statements
        .iter()
        .filter_map(|stmt| match stmt {
            Stmt::Definition(d) => lower_definition(d, &interner),
            _ => None,
        })
        .collect();

    Ok((proof_exprs, goal_expr, defaults, definitions))
}