big_code_analysis_ast/

checker.rs

// Per-language metric and AST modules deliberately consume the macro-
// generated tree-sitter token enums via `use crate::*` and `use Foo::*`
// inside match expressions — explicit imports would list dozens of
// variants per arm and obscure the per-language token sets that are the
// point of these files. Allowed at the module level rather than per
// function so the per-language impl blocks stay readable.
#![allow(clippy::wildcard_imports, clippy::enum_glob_use)]

//! The [`Checker`] classification predicates and their per-language impls.

use std::sync::OnceLock;

use aho_corasick::AhoCorasick;
use regex::bytes::Regex;

use crate::cfg_predicate::attribute_marks_test as rust_attribute_marks_test;
use crate::lang_helpers::{DocBinding, KeptDocs};
use crate::macros::csharp_invocation_expr_kinds;
use crate::*;

static AHO_CORASICK: OnceLock<AhoCorasick> = OnceLock::new();
static RE: OnceLock<Regex> = OnceLock::new();

// Shared ancestor-walk scaffold behind `check_if_func!` /
// `check_if_arrow_func!`. Both decide whether a JS `function_expression`
// / `arrow_function` is a *named function* (vs. an anonymous closure) by
// counting ancestors that bind the expression to a name (`$up` kinds:
// `var x = …`, `x = …`, `label:`, object `pair`) while stopping the walk
// at frames that prove the expression is used positionally (`$stop`
// kinds: a block, return, `new`, call/arguments). A positive count, or a
// per-call `$extra` adjacency check, marks it a named function.
//
// This must stay a `macro_rules!` (not a `fn`): the `$up` / `$stop`
// variant lists are matched against each JS-family language's own
// `kind_id` enum, brought into scope by a per-language `use $language::*`
// glob at the call site. A function could not name variants that only
// exist after that glob import.
//
// Kept token-identical to the two hand-written predicates it replaced:
// the bracketed variant lists expand straight into the `matches!`
// patterns and `$extra` into the trailing `|| …` disjunct.
macro_rules! js_ancestor_walk {
    (
        $node:ident,
        $ancestors:ident,
        [$($up:ident)|+],
        [$($stop:ident)|+],
        $extra:expr $(,)?
    ) => {
        // `Self` is the enclosing `impl Checker for <Lang>Code` type, so
        // the ancestor walk's `is_else_if` filter is bound directly to the
        // language's own `Checker` (no `ParserTrait` round-trip).
        //
        // `$ancestors` is the chain the caller descended through, so each
        // step is a slice index. The walk stops at the first frame that
        // proves positional use — an enclosing block, `return`, `new`, or
        // call — so on ordinary JS it takes one or two steps. That is
        // what made the pre-#1088 cost surprising: `Node::parent` is
        // `O(depth)` per step whatever the count, so even a two-step walk
        // was quadratic over a deeply nested file, which is what the
        // `nom/nested-arrow` probe measures.
        //
        // A chain of *expression*-bodied arrows (`a => b => c => …`) is
        // the one shape no chain can flatten: no `$stop` kind intervenes,
        // so the walk itself is `O(depth)` long. The chain takes it from
        // `O(depth^2)` per node to `O(depth)`; it stays quadratic overall
        // and would need a different rule, not a cheaper lookup, to fix.
        $node.count_specific_ancestors::<Self>(
            $ancestors,
            |node| matches!(node.kind_id().into(), $($up)|+),
            |node| matches!(node.kind_id().into(), $($stop)|+),
        ) > 0
            || $extra
    };
}

// Is this expression *bound* to a name, or used positionally?
//
// `check_if_func!` and `check_if_arrow_func!` answer that question for
// the two anonymous-function forms, and until #1188 they answered it
// with two different ancestor walks. Four divergences, each measured:
//
// 1. `$stop` was `Arguments` for one and `CallExpression` for the other
//    — different tree levels, since `call_expression > arguments >
//    function_expression`. An IIFE's chain is `parenthesized_expression
//    -> call_expression` with no `arguments` on it at all, so the func
//    walk ran past the call and reached the binding: `(function(c){})(1)`
//    was a closure while `const v = (function(c){})(1)` was a function.
//    The same construct, classified by what happened one level up past
//    the call.
// 2. `Pair` was in the func `$up` only, so `({ "k": function(){} })` was
//    a function and `({ "k": () => {} })` a closure. The arrow reached
//    the same verdict for a *bare* key only, and by a different
//    mechanism — `$extra`'s `property_identifier` sibling — which a
//    string, number or computed key does not provide.
// 3. `has_sibling(PropertyIdentifier)` was in the arrow `$extra` only,
//    the exact mirror: `class C { p = () => {}; }` was a function and
//    `class C { p = function(){}; }` a closure.
// 4. `is_child(Identifier)` is in the func `$extra` only, and **must
//    stay there**. For a function expression the only possible
//    `identifier` *child* is its optional name, so it means "carries its
//    own name" and `run(function g(){})` is rightly a function. An arrow
//    stores its un-parenthesised single parameter in exactly that
//    position, so unifying it would make `run(x => x)` — the commonest
//    callback shape in any JS corpus — a function.
//
// The first three are unified below, which makes an IIFE a closure in
// both spellings: that is what a reader sees, and it matches the
// behaviour the arrow form already had. `$up` and `$stop` are
// token-identical and only `$extra` differs, for the reason in (4).
//
// The field-definition kind is threaded in per language rather than
// named here because JS and TS spell it differently (`field_definition`
// vs `public_field_definition`). Both macros take it: `$extra`'s
// `property_identifier` sibling covers a class field only when its name
// *is* an identifier, so without the `$up` entry
// `class C { ["k"] = () => 1 }`, `{ "s" = … }` and `{ #p = … }` stayed
// closures while their `function` spellings were already functions —
// divergence (3) surviving in the shapes the first fix did not reach.
macro_rules! check_if_func {
    ($node: ident, $ancestors: ident, $field_definition: ident) => {
        js_ancestor_walk!(
            $node,
            $ancestors,
            [VariableDeclarator
                | AssignmentExpression
                | LabeledStatement
                | Pair
                | $field_definition],
            [StatementBlock | ReturnStatement | NewExpression | CallExpression | Arguments],
            $node.is_child(Identifier as u16),
        )
    };
}

macro_rules! check_if_arrow_func {
    ($node: ident, $ancestors: ident, $field_definition: ident) => {
        js_ancestor_walk!(
            $node,
            $ancestors,
            [VariableDeclarator
                | AssignmentExpression
                | LabeledStatement
                | Pair
                | $field_definition],
            [StatementBlock | ReturnStatement | NewExpression | CallExpression | Arguments],
            $node.has_sibling($ancestors, PropertyIdentifier as u16),
        )
    };
}

// A generator is a function, not a closure (#1186). `function* g() {}`
// is lexically a function *declaration* — the `*` is a modifier on the
// declaration, not a different binding form — and `get_space_kind`
// already calls both generator kinds `SpaceKind::Function`. The Checker
// disagreeing with the Getter is what made `nom` report a generator as a
// closure, bill its parameters to `closure_args`, and (reaching no
// cognitive boundary arm) let it inherit the enclosing nesting.
//
// The two kinds are **not** symmetrical, which the issue's "move both"
// framing misses. `generator_function_declaration` has a required `name`
// field and mirrors `function_declaration`, so it is unconditionally a
// function. `generator_function` has an *optional* one and mirrors
// `function_expression`, so it routes through the same
// binding-site walk: making it unconditional would classify
// `run(function*(){})` as a function while `run(function(){})` stays a
// closure.
macro_rules! is_js_func {
    ($node: ident, $ancestors: ident, $field_definition: ident) => {
        match $node.kind_id().into() {
            FunctionDeclaration | GeneratorFunctionDeclaration | MethodDefinition => true,
            FunctionExpression | GeneratorFunction => {
                check_if_func!($node, $ancestors, $field_definition)
            }
            ArrowFunction => check_if_arrow_func!($node, $ancestors, $field_definition),
            _ => false,
        }
    };
}

macro_rules! is_js_closure {
    ($node: ident, $ancestors: ident, $field_definition: ident) => {
        match $node.kind_id().into() {
            FunctionExpression | GeneratorFunction => {
                !check_if_func!($node, $ancestors, $field_definition)
            }
            ArrowFunction => !check_if_arrow_func!($node, $ancestors, $field_definition),
            _ => false,
        }
    };
}

macro_rules! is_js_func_and_closure_checker {
    ($language: ident, $field_definition: ident) => {
        #[inline]
        fn is_func<'a>(node: &Node<'a>, ancestors: Ancestors<'a, '_>) -> bool {
            use $language::*;
            is_js_func!(node, ancestors, $field_definition)
        }

        #[inline]
        fn is_closure<'a>(node: &Node<'a>, ancestors: Ancestors<'a, '_>) -> bool {
            use $language::*;
            is_js_closure!(node, ancestors, $field_definition)
        }
    };
}

// Generate an `is_string` impl for a JS-family `Checker` block: the
// named `String` literal, `TemplateString`, plus per-language
// anonymous *literal* aliases passed as extras. The alterator
// flattens the same kinds; the generic `string` filter must agree
// (issue #283). Type-keyword aliases (`: string`, which parses as an
// anonymous `"string"` token under a `predefined_type` wrapper) are
// not literals and stay out (#1261); per-call rationale sits above
// each invocation.
macro_rules! impl_js_family_is_string {
    ($lang:ident $(, $extra:ident)* $(,)?) => {
        fn is_string(node: &Node) -> bool {
            matches!(
                node.kind_id().into(),
                $lang::String | $lang::TemplateString
                    $(| $lang::$extra)*
            )
        }
    };
}

// Generate the documentation rule for a JS-family `Checker` block: a
// `/**` comment only — an Annex-B `<!-- -->` `HtmlComment` never — read by
// `$rule`, the block rule of the language's own tool, and kept inside a
// function body, because JSDoc documents inner symbols (`outer~inner`), so
// a `/** */` on a function or property declared in a body is documentation
// (#1597).
macro_rules! impl_js_family_doc_comment {
    ($lang:ident, $rule:ident) => {
        const DOC_COMMENTS_IN_FUNCTION_BODIES: bool = true;

        fn is_doc_comment(node: &Node, code: &[u8]) -> bool {
            node.kind_id() == $lang::Comment
                && $crate::lang_helpers::doc_comment::$rule(
                    $crate::lang_helpers::doc_comment::node_bytes(node, code),
                )
        }
    };
}

// Generate the documentation rule for a C-family `Checker` block: a
// Doxygen comment block, kept inside a function body. Doxygen appends
// every doc block in a function body to that function's documentation as
// its in-body description (`HIDE_IN_BODY_DOCS` defaults to NO), so an
// in-body `///` or `/** */` is documentation here, as it is in the
// published output (#1597).
macro_rules! impl_doxygen_doc_comment {
    () => {
        const DOC_COMMENTS_IN_FUNCTION_BODIES: bool = true;

        fn is_doc_comment(node: &Node, code: &[u8]) -> bool {
            Self::is_comment(node)
                && $crate::lang_helpers::doc_comment::doxygen(
                    $crate::lang_helpers::doc_comment::node_bytes(node, code),
                )
        }
    };
}

// Generate an `is_string` impl for languages whose `is_string`
// predicate is a flat `matches!` against one or more variant
// kinds. Reduces drift risk for new alias additions and gives a
// single table that answers "which kinds count as a string for
// `find string` / `count string`?" per language (issue #301).
//
// Languages whose `is_string` needs anything beyond a flat variant
// list (e.g. JS family's `String` + `String2` + `TemplateString`
// pattern) keep their own dedicated macros or impls.
macro_rules! impl_simple_is_string {
    ($lang:ident, $first:ident $(, $rest:ident)* $(,)?) => {
        fn is_string(node: &Node) -> bool {
            matches!(
                node.kind_id().into(),
                $lang::$first $(| $lang::$rest)*
            )
        }
    };
}

// Generate an `is_else_if` impl for the grammars that model `else if` as
// an `$if_kind` node nested directly inside an `else_clause` wrapper
// (`else_clause → if`). Covers C++/JS/TS/TSX (`IfStatement`) and Rust
// (`IfExpression`). Per-call rationale (e.g. the `IfExpression` choice)
// is hoisted above each invocation per `.claude/rules/macro-comments.md`.
macro_rules! impl_is_else_if_parent_clause {
    ($lang:ident, $if_kind:ident, $else_clause:ident) => {
        #[inline]
        fn is_else_if<'a>(node: &Node<'a>, ancestors: Ancestors<'a, '_>) -> bool {
            node.kind_id() == $lang::$if_kind
                && ancestors
                    .parent(node)
                    .is_some_and(|parent| parent.kind_id() == $lang::$else_clause)
        }
    };
}

// Generate an `is_else_if` impl for the grammars that emit a bare `else`
// keyword token followed by a nested `$if_kind` sibling, with no wrapping
// clause node. Covers Java/C#/Groovy (`IfStatement`) and Kotlin
// (`IfExpression`).
macro_rules! impl_is_else_if_prev_sibling {
    ($lang:ident, $if_kind:ident, $else_kw:ident) => {
        #[inline]
        fn is_else_if<'a>(node: &Node<'a>, ancestors: Ancestors<'a, '_>) -> bool {
            node.kind_id() == $lang::$if_kind
                && ancestors
                    .previous_sibling(node)
                    .is_some_and(|prev| prev.kind_id() == $lang::$else_kw)
        }
    };
}

// Generate an `is_else_if` impl for the grammars that expose a dedicated
// elseif/elsif clause node, so the clause node itself is the else-if
// (no `if`-nesting or sibling-token shape to inspect). Flat `matches!`
// against one or more variant kinds, analogous to `impl_simple_is_string!`;
// the multi-variant form covers PHP's `ElseIfClause | ElseIfClause2`.
macro_rules! impl_is_else_if_clause {
    ($lang:ident, $first:ident $(, $rest:ident)* $(,)?) => {
        #[inline]
        fn is_else_if(node: &Node, _ancestors: Ancestors<'_, '_>) -> bool {
            matches!(
                node.kind_id().into(),
                $lang::$first $(| $lang::$rest)*
            )
        }
    };
}

#[inline]
fn get_aho_corasick_match(code: &[u8]) -> bool {
    AHO_CORASICK
        .get_or_init(|| {
            AhoCorasick::new(vec![b"<div rustbindgen"])
                .expect("constant single-needle AhoCorasick automaton always compiles")
        })
        .is_match(code)
}

/// A [`Checker::should_skip_subtree`] verdict, plus the following
/// nodes it also answers for.
///
/// Most verdicts are about one node and carry no reach. The exception
/// is a verdict a classifier can only reach by reading a *run* of
/// siblings: Rust's `#[…]` rows all decorate the same item, so they all
/// get the same answer, and re-deriving it per row is `O(run)` work
/// done `run` times. A run of `3 * depth` attributes under a parent
/// wide enough to stay inside the forward attribute-scan budget made
/// that quadratic — 5.2 s at depth 2 000, against 0.05 s with
/// `exclude_tests` off (#1446). Reporting the reach lets the walker ask
/// the classifier once per run.
///
/// # Soundness
///
/// The reach is a byte offset, and the walk visits nodes in
/// non-decreasing `start_byte` order, so "starts before it" is a state
/// the walker can test in `O(1)` and can never re-enter. A claim is
/// sound when every node the walk *visits* inside the reach genuinely
/// has the reported verdict.
///
/// "Visits" is the load-bearing word, and it cuts the obligation in
/// two. A `SKIP` reach need only hold for the nodes it skips, since a
/// skipped node's descendants are never pushed. A `RETAIN` reach
/// carries the heavier claim: the members' descendants *are* walked, so
/// it must be right for them too.
#[derive(Clone, Copy, PartialEq, Eq, Debug)]
pub struct SubtreeSkip {
    skipped: bool,
    covers_through_byte: Option<usize>,
}

impl SubtreeSkip {
    /// Keep this node. Answers for nothing else.
    pub const RETAIN: Self = Self {
        skipped: false,
        covers_through_byte: None,
    };

    /// Elide this node and its descendants. Answers for nothing else.
    pub const SKIP: Self = Self {
        skipped: true,
        covers_through_byte: None,
    };

    /// [`Self::SKIP`] or [`Self::RETAIN`], from a predicate's answer.
    #[inline]
    #[must_use]
    pub const fn of(skipped: bool) -> Self {
        if skipped { Self::SKIP } else { Self::RETAIN }
    }

    /// Extend this verdict to every node that starts before `byte`.
    ///
    /// The caller is asserting that it read far enough to answer for
    /// all of them — see the soundness note on the type.
    #[inline]
    #[must_use]
    pub const fn covering_through(self, byte: usize) -> Self {
        Self {
            covers_through_byte: Some(byte),
            ..self
        }
    }

    /// Whether this node and its descendants are elided.
    #[inline]
    #[must_use]
    pub const fn is_skipped(self) -> bool {
        self.skipped
    }

    /// Whether this verdict already answers for a node starting at
    /// `start_byte`, so the classifier need not be asked again.
    #[inline]
    #[must_use]
    pub const fn answers_for(self, start_byte: usize) -> bool {
        match self.covers_through_byte {
            Some(through) => start_byte < through,
            None => false,
        }
    }
}

/// Per-language AST classification predicates the metric walkers use to
/// recognize comments, function spaces, calls, strings, branches, and so
/// on by node kind.
///
/// Every predicate defaults to `false` — "this node is never of that
/// category." A language implements only the predicates its grammar
/// actually expresses; categories it lacks fall through to the default,
/// so adding a language no longer requires copy-pasting `-> false` stubs.
/// The flip side is that the compiler cannot flag a forgotten override, so
/// each language's per-metric tests are the safety net for completeness.
///
/// `#[doc(hidden)]` for the reason the crate root gives: this is
/// plumbing, and the surface the README points a reader at is `LANG` /
/// `Node` / `SpaceKind`, not a 200-method classifier trait.
#[doc(hidden)]
pub trait Checker {
    /// Whether `node` is a comment.
    #[inline]
    #[must_use]
    fn is_comment(_: &Node) -> bool {
        false
    }
    /// Whether this comment must survive `strip_comments`.
    ///
    /// `ancestors` is the chain the comment-removal walk descended
    /// through. Rust's impl reads the parent to keep a comment that is
    /// a macro token; reaching it with [`Node::parent`] costs
    /// `O(depth)` per comment (#1096).
    #[inline]
    #[must_use]
    fn is_useful_comment<'a>(_: &Node<'a>, _: &[u8], _: Ancestors<'a, '_>) -> bool {
        false
    }
    /// Whether this comment node is documentation: a doc comment the
    /// grammar marks (Rust's `doc_comment` child, Groovy's
    /// `groovydoc_comment`, Perl's `pod_statement`) or one its
    /// language's own tool recognises by the opening marker in `code`
    /// (Doxygen, Javadoc, KDoc, JSDoc, PHPDoc, C# XML docs; see
    /// `lang_helpers::doc_comment`). Implies [`is_comment`]; the LOC
    /// walk bills such a row as `dloc` as well as `cloc`, unless it sits
    /// in a function body the language's tool does not read
    /// ([`DOC_COMMENTS_IN_FUNCTION_BODIES`]) (#1597).
    ///
    /// [`is_comment`]: Self::is_comment
    /// [`DOC_COMMENTS_IN_FUNCTION_BODIES`]: Self::DOC_COMMENTS_IN_FUNCTION_BODIES
    #[inline]
    #[must_use]
    fn is_doc_comment(_: &Node, _code: &[u8]) -> bool {
        false
    }
    /// Whether a doc comment inside a function or closure body is still
    /// documentation. `false` where it is better read as a working note,
    /// so a note spelled `/// …` stays a charged comment: rustc lints one
    /// on a statement as unused (`unused_doc_comments`), Roslyn raises
    /// CS1587, javadoc and Dokka document declarations only, and an
    /// in-body PHPDoc `/** @var */` is a type hint. `true` where the tool
    /// publishes what a body holds: Doxygen's in-body description (C, C++,
    /// Objective-C — `HIDE_IN_BODY_DOCS` defaults to NO), JSDoc's inner
    /// symbols, and Perl POD, which perldoc extracts wherever it sits
    /// (#1597). Which part of a function counts as its body is
    /// [`holds_doc_comment_in_body`](Self::holds_doc_comment_in_body).
    const DOC_COMMENTS_IN_FUNCTION_BODIES: bool = false;
    /// Whether `ancestor`, one of a doc comment's ancestors, holds it in a
    /// function or closure body, where
    /// [`DOC_COMMENTS_IN_FUNCTION_BODIES`](Self::DOC_COMMENTS_IN_FUNCTION_BODIES)
    /// decides whether it is documentation. `ancestors` is `ancestor`'s
    /// own chain, and `below` is the node directly under `ancestor` on the
    /// path to the comment, which is what a "does the body hold it" test
    /// reads. The answer is memoised per ancestor for the whole walk, so
    /// it must not depend on which comment asked; a comment-specific
    /// exception is
    /// [`doc_comment_opens_body`](Self::doc_comment_opens_body) (#1597).
    /// By default the whole function or closure node counts:
    /// javac binds a doc comment only before a declaration's first token,
    /// and the C# specification wants one before a member's attributes,
    /// so a doc comment after an annotation or on a parameter documents
    /// nothing there. A language whose tool reads doc comments in the
    /// signature narrows this to the body proper (#1597).
    #[inline]
    #[must_use]
    fn holds_doc_comment_in_body<'a>(
        ancestor: &Node<'a>,
        ancestors: Ancestors<'a, '_>,
        _below: &Node<'a>,
    ) -> bool {
        Self::is_func(ancestor, ancestors) || Self::is_closure(ancestor, ancestors)
    }
    /// Whether `comment`, held in a body by `ancestor` (one of its two
    /// nearest ancestors, with `below` the node between them), is that
    /// body's own documentation rather than a note in it, so `ancestor`
    /// does not count as holding it. `false` by default; Rust's inner doc
    /// comment opening a function body is its `#![doc]` (#1597, #1615).
    #[inline]
    #[must_use]
    fn doc_comment_opens_body(
        _comment: &Node,
        _ancestor: &Node,
        _below: &Node,
        _code: &[u8],
    ) -> bool {
        false
    }
    /// Whether a comment run before `node` documents it, so the LOC walk
    /// bills that run as `dloc`. A positional language asks it of the
    /// node on the row directly below the run — go/doc's lead comment,
    /// YARD's convention; see `lang_helpers::doc_comment::leading_doc_run`
    /// (#1597). A language with a [`DOC_BINDING`](Self::DOC_BINDING) asks
    /// it of the first sibling after the run's comments and the nodes its
    /// [`DocBinding::skips`] names, and answers for the declarations its
    /// tool attaches a doc comment to; see
    /// `lang_helpers::doc_comment::bound_doc_run` (#1613). `ancestors`
    /// is `node`'s chain, for a rule that depends on where it sits.
    #[inline]
    #[must_use]
    fn is_doc_target<'a>(_: &Node<'a>, _code: &[u8], _: Ancestors<'a, '_>) -> bool {
        false
    }
    /// How a doc comment attaches to the declaration after it (#1613).
    /// `None` by default: a doc comment documents wherever it sits outside
    /// a function body, attached to anything or not, which is every
    /// language not yet bound. A bound language's forward doc comments
    /// ([`DocBinding::forward`]) document only when their stretch ends at
    /// an [`is_doc_target`](Self::is_doc_target), and only the ones the
    /// binding keeps; see `lang_helpers::doc_comment::bound_doc_run`.
    const DOC_BINDING: Option<DocBinding> = None;
    /// The row a doc target begins on, which a comment run must end
    /// directly above. Normally `node`'s own start row; Ruby's `then`
    /// wrapper starts on the row of the comment the grammar parks ahead of
    /// it, so its first statement's row stands in for it (#1597).
    #[inline]
    #[must_use]
    fn doc_target_start_row(node: &Node) -> usize {
        node.start_row()
    }
    /// Whether the comment run that `first` opens is documentation by its
    /// own content rather than by a [`is_doc_target`](Self::is_doc_target)
    /// below it: an LDoc block, opened by `---` and continued over the `--`
    /// rows below it, read under LDoc's rule for what follows it; or
    /// RDoc's `##` and YARD's `@!` directive blocks, which document
    /// metaprogrammed methods with no `def` to sit above. `next` is the
    /// sibling on the row directly after the run, if any, and `ancestors`
    /// is the run's chain, for a rule that depends on where the run sits
    /// (#1597).
    #[inline]
    #[must_use]
    fn is_doc_run_opener<'a>(
        _first: &Node<'a>,
        _next: Option<&Node<'a>>,
        _code: &[u8],
        _: Ancestors<'a, '_>,
    ) -> bool {
        false
    }
    /// Whether a comment run cannot continue past `node` onto the next
    /// row: LDoc ends a block at a long comment (`--[[ … ]]`), because
    /// the comment's closing newline is not part of it (#1597). A
    /// comment later on the same row still joins.
    #[inline]
    #[must_use]
    fn ends_doc_run(_: &Node, _code: &[u8]) -> bool {
        false
    }
    /// A member of a documentation run that is not documentation text
    /// and is not billed as `dloc`: a go/doc directive such as
    /// `//go:generate`, a Ruby shebang or magic comment (#1597).
    #[inline]
    #[must_use]
    fn is_doc_run_directive(_: &Node, _code: &[u8]) -> bool {
        false
    }
    /// Whether `node` opens a space: a function, closure, or one of the
    /// containers (class, struct, trait, impl, namespace, interface).
    #[inline]
    #[must_use]
    fn is_func_space(_: &Node) -> bool {
        false
    }
    /// Whether `node` is a *named* function definition.
    ///
    /// `ancestors` is the chain the caller descended through. The
    /// JS-family grammars have no distinct production for a named
    /// function: `const f = () => …` and `g(() => …)` share one
    /// `arrow_function` node, and only what encloses it says which is
    /// which. Those steps come off the chain, because `Node::parent`
    /// costs `O(depth)` each however few of them a walk takes (#1088).
    /// Every other grammar answers `is_func` from the node's own kind
    /// and ignores the parameter.
    #[inline]
    #[must_use]
    fn is_func<'a>(_: &Node<'a>, _ancestors: Ancestors<'a, '_>) -> bool {
        false
    }
    /// Whether `node` is an anonymous function — the complement of
    /// [`is_func`](Self::is_func) on the grammars that express both. Takes `ancestors`
    /// for the same reason, plus one of its own: Ruby's `{ … }` block is
    /// a closure only when its parent is not the `Lambda` that already
    /// counted it.
    #[inline]
    #[must_use]
    fn is_closure<'a>(_: &Node<'a>, _ancestors: Ancestors<'a, '_>) -> bool {
        false
    }
    /// Whether `node` is a *call site* — a function, method or command
    /// invocation a reader would navigate to.
    ///
    /// This is the predicate behind the `call` filter that
    /// [`ParserTrait::filters`](crate::traits::ParserTrait::filters)
    /// builds, so it decides what `bca find --type call` and
    /// `bca count --type call` report.
    ///
    /// It is deliberately narrower than the ABC `branches` axis, which
    /// applies Fitzpatrick's *"function invocation or object creation"*
    /// rule. Object construction (`new T(…)`) and constructor
    /// delegation — C#'s `: base(x)` / `: this(a)` and a C# 12 primary
    /// constructor's `: Base(x)`, Java's `super(…)` / `this(…)`,
    /// Kotlin's `constructor_invocation` — are ABC branches and are
    /// **not** call sites here: a constructor is not a function the
    /// `find` result should point a reader at. The languages whose
    /// `branches` axis consequently exceeds their `is_call` set are C#,
    /// Java, Groovy, Kotlin, C++, Mozcpp, JavaScript, MozJS,
    /// TypeScript, TSX, PHP, Ruby (`super` / `yield` count as branches)
    /// and Rust (`?` does).
    ///
    /// The relation is not a plain superset everywhere. Three languages
    /// depart from it, each in its own direction:
    ///
    /// * Tcl and iRules bill a mutator command (`incr`, `append`,
    ///   `lappend`) as an ABC *assignment* rather than a branch while
    ///   `is_call` still counts it. `set` is not one of those: it has a
    ///   grammar production of its own (`Tcl::Set` / `Irules::Set`)
    ///   rather than being a `command`, so it is an ABC assignment that
    ///   `is_call` never matches either.
    /// * Elixir excludes its definition and directive `Call`s (`def`,
    ///   `defmodule`, `alias`, `import`, …) from `branches` while
    ///   counting each `|>` pipeline step.
    ///
    /// The two sets coincide in C, Objective-C, Go, Python, Lua and
    /// Bash, which have neither a separate construction syntax nor a
    /// gated branch arm, and in Perl, whose one gated arm ABC shares
    /// with [`is_call_with_code`](Self::is_call_with_code): a bareword
    /// is no call of its own as a wrapper's callee (`print` in
    /// `print shift;`) or as an auto-quoted key (`(name => 1)`). That
    /// gate needs the ancestor chain, so for Perl this byte-less form is
    /// only the kind table the override narrows.
    // `is_call` and each language's ABC branch arm classify the same
    // nodes through parallel `matches!()` tables, so they drift
    // silently; the divergence is pinned by
    // `groovy_is_call_excludes_constructors` (#430) and
    // `csharp_is_call_excludes_constructors` in this file's tests.
    #[inline]
    #[must_use]
    fn is_call(_: &Node) -> bool {
        false
    }
    /// [`is_call`](Self::is_call), given the source bytes and the
    /// chain the caller descended through, which the `call` filter
    /// hands it. The default forwards to [`is_call`](Self::is_call).
    ///
    /// Perl overrides it: the grammar nests a call's name inside the
    /// call as a bareword call of its own, and parses an auto-quoted
    /// hash key as one, so only the position says which words are
    /// calls (#1580). As with [`is_string_with_code`](Self::is_string_with_code),
    /// a call site reaching for the byte-less spelling reads as correct
    /// against every language without an override;
    /// `perl_bareword_is_a_call_only_to_the_byteless_spelling` pins that
    /// the two really do disagree.
    #[inline]
    #[must_use]
    fn is_call_with_code<'a>(node: &Node<'a>, _code: &[u8], _ancestors: Ancestors<'a, '_>) -> bool {
        Self::is_call(node)
    }
    /// Whether `node`, met among a parameter list's children, is
    /// punctuation or a keyword rather than a parameter (`nargs` skips it).
    #[inline]
    #[must_use]
    fn is_non_arg(_: &Node) -> bool {
        false
    }
    /// Whether `param` is a language's way of spelling *"this function
    /// takes nothing"* inside an otherwise ordinary parameter list.
    ///
    /// C's `int f(void)` declares zero parameters, but the grammar
    /// emits a real `parameter_declaration` for the `void`, so the
    /// negative filters in `big_code_analysis::nargs` counted it as one — `f` and
    /// `int f(int)` both reported 1 despite one of them taking no
    /// argument at all.
    ///
    /// It takes `code` because the question is not structural: a
    /// `parameter_declaration` holding a bare `primitive_type` and no
    /// declarator is *also* how an unnamed parameter is spelled, and
    /// `void f(int)` really does take one. Only the bytes tell them
    /// apart — `.claude/rules/grammar-dispatch.md` §10.
    #[inline]
    #[must_use]
    fn is_empty_param_marker(_param: &Node, _code: &[u8]) -> bool {
        false
    }
    /// Whether `node` — the value of a function's `parameters` field —
    /// is itself a *single formal parameter* rather than a parameter
    /// *list*.
    ///
    /// `compute_args` walks the field node's children, which is right
    /// for a list and yields zero for a lone parameter that has none.
    /// Java's `lambda_expression` spells its `parameters` field three
    /// ways — `formal_parameters` and `inferred_parameters` are lists,
    /// but the un-parenthesised `x -> …` form puts a bare, childless
    /// `identifier` there — and C#'s puts a childless
    /// `implicit_parameter`. Both reported `nargs = 0` for a lambda
    /// that plainly has one argument (#1185).
    ///
    /// Gated per language rather than inferred from `child_count() == 0`
    /// so a MISSING or zero-width ERROR node under `parameters` on a
    /// broken parse cannot silently conjure an argument.
    #[inline]
    #[must_use]
    fn is_bare_param(_: &Node) -> bool {
        false
    }
    /// Whether `node` is a string literal, under every aliased kind the
    /// grammar emits for one.
    ///
    /// This is the *kind table* half of the answer. Where a grammar
    /// spells a string literal and something else with one kind, the
    /// table cannot separate them and the walk must ask
    /// [`is_string_with_code`](Self::is_string_with_code) instead.
    #[inline]
    #[must_use]
    fn is_string(_: &Node) -> bool {
        false
    }
    /// Source-aware variant of [`is_string`](Self::is_string), and the
    /// spelling every walk calls.
    ///
    /// The default forwards to the byte-less predicate, so a language
    /// whose string literals are a closed set of kinds needs no
    /// override. The Tcl family is the case that does: `braced_word` is
    /// both the literal of `lappend x {a b}` and the *script* of a
    /// `proc` body or an iRules `when` handler, and only the enclosing
    /// command's leading word tells the two apart (#1318, #1381). So
    /// `bca find --type string` reported every script body in the file
    /// as a string literal.
    ///
    /// **A call site reaching for the byte-less spelling reads as
    /// correct against every language without an override, and is wrong
    /// only on the one that has one** —
    /// `.claude/rules/grammar-dispatch.md` §7. There is one call site,
    /// the `"string"` arm of [`ParserTrait::filters`], and it takes this
    /// spelling; `tcl_script_body_is_a_string_only_to_the_byteless_spelling`
    /// pins that the two really do disagree, so a future call site that
    /// picks the wrong one is not silently right.
    ///
    /// [`ParserTrait::filters`]: crate::traits::ParserTrait::filters
    #[inline]
    #[must_use]
    fn is_string_with_code<'a>(
        node: &Node<'a>,
        _code: &[u8],
        _ancestors: Ancestors<'a, '_>,
    ) -> bool {
        Self::is_string(node)
    }
    /// Whether `node` is the `if` that continues an `else if` chain,
    /// rather than a freshly nested branch.
    ///
    /// `ancestors` is the chain the caller descended through; the
    /// grammars that model `else if` as nesting need the parent (or the
    /// preceding `else` token) to tell the two apart, and resolving
    /// either from the node alone costs `O(depth)` (#1084). Languages
    /// with a dedicated `elsif` clause node ignore it.
    #[inline]
    #[must_use]
    fn is_else_if(_: &Node, _: Ancestors<'_, '_>) -> bool {
        false
    }
    /// Whether `node` is a primitive type name (`int`, `double`, …),
    /// which Halstead keys by text so each distinct primitive counts once.
    ///
    /// Precondition: call this only for a node
    /// [`Getter::get_op_type_with_code`] has already classified as
    /// [`TokenRole::Operator`]. An impl that cannot read the source may
    /// answer by kind alone: Groovy's returns `true` for every
    /// `identifier` and `type_identifier`, which is right only because
    /// its getter bills those kinds as operators for a primitive
    /// spelling and nothing else.
    #[inline]
    #[must_use]
    fn is_primitive(_node: &Node) -> bool {
        false
    }

    /// Whether `node` is, or contains, a tree-sitter `ERROR` node.
    #[must_use]
    fn is_error(node: &Node) -> bool {
        node.has_error()
    }

    /// Answer [`SubtreeSkip::SKIP`] to elide this node and all its
    /// descendants from every metric. Used by language modules to
    /// filter test-only / generated / preprocessor-disabled subtrees.
    ///
    /// The default retains every node, preserving the pre-#182
    /// behavior. Language overrides drive opt-in skips (currently:
    /// `RustCode` filters `#[cfg(test)]` items, gated by the runtime
    /// `MetricsOptions::exclude_tests` flag).
    ///
    /// `ancestors` is the chain the caller descended through. Rust's
    /// override needs the parent to read the run of `#[…]` siblings
    /// before an item, and resolving siblings from the node alone
    /// costs `O(depth)` per step (#1100).
    ///
    /// A verdict that took a *sibling-run* to reach should say so with
    /// [`SubtreeSkip::covering_through`], so the walker asks once per
    /// run rather than once per member. See that method for what makes
    /// such a claim sound.
    #[inline]
    #[must_use]
    fn should_skip_subtree<'a>(
        _node: &Node<'a>,
        _code: &[u8],
        _ancestors: Ancestors<'a, '_>,
    ) -> SubtreeSkip {
        SubtreeSkip::RETAIN
    }

    /// Source-aware variant of [`is_func_space`](Self::is_func_space). The default forwards
    /// to the byte-less predicate so languages whose function-space
    /// classification is encoded in distinct grammar productions (Java,
    /// Rust, Python, …) need no override. Languages whose function
    /// boundaries are macro-shaped — Elixir's `def` / `defp` /
    /// `defmacro` / `defmacrop` / `defmodule` — override this to
    /// disambiguate `Call` nodes by their target identifier text
    /// (#275).
    ///
    /// `ancestors` is the chain the caller descended through, so a
    /// language whose classification depends on an enclosing construct
    /// (Elixir's `quote` template) can answer without paying
    /// `Node::parent`'s `O(depth)` per step (#1084).
    #[inline]
    #[must_use]
    fn is_func_space_with_code<'a>(
        node: &Node<'a>,
        _code: &[u8],
        _ancestors: Ancestors<'a, '_>,
    ) -> bool {
        Self::is_func_space(node)
    }

    /// Source-aware variant of [`is_func`](Self::is_func). Same rationale as
    /// [`is_func_space_with_code`](Self::is_func_space_with_code) (#275).
    #[inline]
    #[must_use]
    fn is_func_with_code<'a>(node: &Node<'a>, _code: &[u8], ancestors: Ancestors<'a, '_>) -> bool {
        Self::is_func(node, ancestors)
    }

    /// Combined predicate the walker uses to decide whether to promote
    /// `node` to a new function-space frame. Default forwards to
    /// `is_func_with_code || is_func_space_with_code` so existing
    /// languages need no override — each kept the freedom to expose
    /// `is_func` and `is_func_space` as disjoint sets (Rust includes
    /// closures in `is_func_space`, Java keeps lambdas out, …) and
    /// this method preserves that flexibility while letting Elixir
    /// halve its per-`Call` source-text lookups: a single
    /// `elixir_call_keyword` call answers both halves at once
    /// (#310 follow-on perf).
    #[inline]
    #[must_use]
    fn promotes_to_func_space_with_code<'a>(
        node: &Node<'a>,
        code: &[u8],
        ancestors: Ancestors<'a, '_>,
    ) -> bool {
        Self::is_func_with_code(node, code, ancestors)
            || Self::is_func_space_with_code(node, code, ancestors)
    }

    /// Whether the function space `node` opens belongs to something
    /// other than the container that syntactically encloses it.
    ///
    /// The space tree is built by a single-pass stack walk, so a space
    /// is a child of whichever space was open when the walk reached it
    /// — syntactic containment, never membership. The two normally
    /// coincide. C++ `friend` is the case where they do not: a friend
    /// with an inline body is written inside the class braces but is a
    /// free function the class merely grants access to, so `npm` does
    /// not count it as a method while `wmc` weighted it anyway (#1301).
    /// Objective-C has the same shape in a different spelling: a plain
    /// C `function_definition` inside `@implementation` / `@interface`
    /// / `@protocol` is a file-static helper, never a method (#1356).
    ///
    /// `wmc` is the sole consumer: WMC is defined as the sum over a
    /// class's *methods*, so a `true` here keeps the space out of the
    /// enclosing class's WMC while leaving it a `Function` space like
    /// any other — its own metrics, its `nom` contribution, and the
    /// file-level cyclomatic roll-up are all unaffected.
    ///
    /// A `true` for a function no container encloses is inert rather
    /// than wrong, since `merge` credits a `Function` child only to a
    /// `Class` or `Interface` parent. So a language whose grammar makes
    /// the answer unconditional — Objective-C, where a method is always
    /// a `method_definition` — may answer on node kind alone, which is
    /// total where a parent-kind list is not.
    ///
    /// Deliberately one method rather than a byte-less / `_with_code`
    /// pair: the pair forwards by default, so a call site reaching for
    /// the wrong spelling reads as correct against every language
    /// without an override (`.claude/rules/grammar-dispatch.md` §7).
    /// One spelling has nothing to disagree with.
    #[inline]
    #[must_use]
    fn is_non_member_function<'a>(
        _node: &Node<'a>,
        _code: &[u8],
        _ancestors: Ancestors<'a, '_>,
    ) -> bool {
        false
    }
}

// The doc-comment binding macros of the C and JS families (#1613),
// declared before the language modules that invoke them.
#[macro_use]
mod doxygen_doc_binding;
#[macro_use]
mod js_doc_binding;
mod bash;
mod c;
mod ccomment;
mod cpp;
mod csharp;
mod elixir;
mod go;
mod groovy;
mod irules;
mod java;
mod javascript;
mod kotlin;
mod lua;
mod mozcpp;
mod mozjs;
mod objc;
mod perl;
mod php;
mod preproc;
mod python;
pub mod ruby;
mod rust;
mod tcl;
mod tsx;
mod typescript;

/// Returns the `class_body` child of a Java `object_creation_expression`
/// when the node is an anonymous class (`new Runnable() { ... }`), or
/// `None` for a plain constructor call (`new Foo()`). Shared by
/// `JavaCode::is_func_space` and `JavaCode::get_space_kind` so both agree
/// on exactly which `object_creation_expression` nodes open a Class space
/// (#463); a lambda is a distinct `lambda_expression` node and never
/// reaches this path.
#[must_use]
pub fn java_anonymous_class_body<'a>(node: &Node<'a>) -> Option<Node<'a>> {
    node.first_child(|id| id == Java::ClassBody as u16)
}

/// Counts the `accessor_declaration` children (`get` / `set` / `init`) of a C#
/// `property_declaration` / `indexer_declaration` by walking its
/// `accessor_list`. Returns `0` for the expression-bodied form
/// (`this[int i] => _d[i];` / `int W => _w;`), which has no accessor list.
///
/// Used by the npm reference `csharp_count_member`, which keeps its own
/// `.max(1)` fallback so an accessor-less expression-bodied member still
/// counts as one method (#464). `csharp_member_has_accessors` asks only
/// whether one exists and stops at the first (#1612).
#[must_use]
pub fn csharp_accessor_count(node: &Node) -> usize {
    csharp_accessors(node).count()
}

/// The `accessor_declaration`s in a C# member's `accessor_list`.
fn csharp_accessors<'a>(node: &Node<'a>) -> impl Iterator<Item = Node<'a>> + use<'a> {
    node.children()
        .filter(|c| c.kind_id() == Csharp::AccessorList as u16)
        .flat_map(|list| list.children())
        .filter(|c| c.kind_id() == Csharp::AccessorDeclaration as u16)
}

/// Returns `true` when a C# `indexer_declaration` / `property_declaration`
/// carries bodied accessors — an `accessor_list` containing at least one
/// `accessor_declaration` (`get` / `set` / `init`). Returns `false` for the
/// expression-bodied form (`this[int i] => _d[i];` / `int W => _w;`), which
/// has no accessor list and defines a single implicit getter.
///
/// Shared by `CsharpCode::is_func` / `is_func_space` and
/// `CsharpCode::get_space_kind` so all three agree on when an indexer or
/// property opens its own Function space versus deferring to its accessor
/// children. This mirrors the npm reference (`csharp_count_member`): the
/// member counts as its accessor count, falling back to 1 for the
/// accessor-less expression-bodied form (#464 indexer, #472 property).
#[must_use]
pub fn csharp_member_has_accessors(node: &Node) -> bool {
    csharp_accessors(node).next().is_some()
}

/// Returns `true` when a C++ `struct_specifier` / `class_specifier`
/// defines a type body (`struct S { … }`) rather than only naming one.
/// A forward declaration (`struct S;`) and an elaborated type specifier
/// (`struct S *p;`, the parameter type in `void f(struct S *q)`) carry no
/// `body` field.
///
/// Shared by the C++ and Mozcpp `is_func_space` and `get_space_kind`, so
/// a bodiless specifier neither opens a space nor is labelled one (#1555).
#[must_use]
pub(crate) fn cpp_specifier_has_body(node: &Node) -> bool {
    node.child_by_field_name("body").is_some()
}

/// Whether `node` is a bare keyword token the C# grammar *aliased* to
/// `modifier`, rather than the `modifier` rule itself.
///
/// The rule is `choice('public', 'static', 'async', 'readonly', …)`, so
/// a real `modifier` wraps one keyword leaf that the getter already
/// classifies on its own. Two other rules alias a bare token onto the
/// same named kind, and those nodes are **childless**:
///
/// - `_parameter_type_with_modifiers` — `this`, `scoped`, `ref`, `out`,
///   `in`, `readonly` in parameter position, the six #1418 bills.
/// - `_lambda_expression_init` and `anonymous_method_expression` —
///   `static` and `async` on a lambda or `delegate { … }`, billed as
///   operators like their declaration spelling (#1482).
///
/// Child-presence is therefore what separates an alias from the rule,
/// and it is the grammar's shape rather than an inference about a broken
/// parse — contrast [`Checker::is_bare_param`], which refuses the same
/// test precisely because there it would be a guess.
///
/// Shared by [`CsharpCode::get_op_type_with_code`], which decides the
/// keyword's Halstead role from its text, and
/// [`CsharpCode::is_primitive`], which routes the operator half through
/// the lexeme-keyed map so the alias and the bare token are one
/// operator (#1418).
///
/// [`CsharpCode::get_op_type_with_code`]: crate::Getter::get_op_type_with_code
/// [`CsharpCode::is_primitive`]: Checker::is_primitive
#[must_use]
pub(crate) fn csharp_is_aliased_modifier(node: &Node) -> bool {
    node.kind_id() == Csharp::Modifier as u16 && node.child_count() == 0
}

/// Whether `param` is C's `(void)` marker — the spelling for an empty
/// parameter list — rather than a parameter.
///
/// Shared by C, C++, Mozcpp and Objective-C, which all inherit the
/// idiom and all emit a real `parameter_declaration` for it.
///
/// The declarator check is what separates the marker from `void *p`,
/// which is an ordinary parameter of type `void *`, and the byte
/// comparison is what separates it from an unnamed parameter of any
/// other type: `f(int)` and `f(void)` are the same shape and different
/// arities.
#[must_use]
pub fn c_family_void_parameter(param: &Node, code: &[u8]) -> bool {
    param.child_by_field_name("declarator").is_none()
        && param
            .child_by_field_name("type")
            .and_then(|ty| code.get(ty.start_byte()..ty.end_byte()))
            == Some(b"void".as_slice())
}

/// Strip the `#` / `#!` marker plus the `[...]` brackets from a
/// Rust `AttributeItem` / `InnerAttributeItem` token's raw text,
/// returning the inner body. Returns `None` if the input shape is
/// unexpected — callers skip silently rather than feed the matcher
/// the literal `#[...]` form.
fn rust_attribute_body<'a>(text: &'a str, marker: &str) -> Option<&'a str> {
    text.trim()
        .strip_prefix(marker)
        .and_then(|t| t.trim_start().strip_prefix('['))
        .and_then(|t| t.trim().strip_suffix(']'))
}

/// Classifies one attribute node's body, given the `#` / `#!` marker
/// its kind carries. Returns `false` for a node whose text is not
/// valid UTF-8 or does not have the `marker[...]` shape.
fn rust_attribute_node_marks_test(node: &Node, code: &[u8], marker: &str) -> bool {
    node.utf8_text(code)
        .and_then(|text| rust_attribute_body(text, marker))
        .is_some_and(rust_attribute_marks_test)
}

fn rust_item_is_test_only<'a>(node: &Node<'a>, code: &[u8], ancestors: Ancestors<'a, '_>) -> bool {
    rust_outer_attr_marks_test(node, code, ancestors) || rust_inner_attr_marks_test(node, code)
}

/// The item kinds the `exclude_tests` prune removes, given a test-only
/// verdict.
///
/// `RustCode::should_skip_subtree` and
/// [`rust_attribute_run_marks_test_item`] both consult this, so an
/// attribute run and the item it decorates can never disagree about
/// whether that item is prunable: `#[cfg(test)] use foo;` keeps both
/// rows, because a `use_declaration` is not on this list (#1431).
fn rust_prunable_item(node: &Node) -> bool {
    matches!(
        node.kind_id().into(),
        Rust::ModItem
            | Rust::FunctionItem
            | Rust::ImplItem
            | Rust::TraitItem
            | Rust::ConstItem
            | Rust::StaticItem
    )
}

/// Whether the outer-attribute run `node` belongs to decorates an item
/// the prune removes — which makes `node`'s own row test code too.
///
/// The whole run is asked about the *item*, not about itself, so every
/// row of a stacked `#[cfg(test)]\n#[allow(dead_code)]\nmod t` goes:
/// the second attribute says nothing about tests, and
/// [`rust_item_is_test_only`] reads the run from the item end either
/// way (#1431).
///
/// That is also why the answer is reported as covering the rest of the
/// run. Every attribute from `node` up to the item resolves the same
/// item and therefore the same verdict, so a walker that re-asks per
/// row pays the run twice over per row: once in
/// [`rust_attributed_item`]'s lookahead and once in the backward
/// reading [`rust_item_is_test_only`] takes. Both are `O(run)` under a
/// parent inside [`forward_attribute_scan_budget`], and a run of
/// `3 * depth` attributes is inside it by construction — the diagonal
/// #1446 measured at 5.2 s / depth 2 000. One ask per run makes it one
/// pass, not `run` passes.
fn rust_attribute_run_verdict<'a>(
    node: &Node<'a>,
    code: &[u8],
    ancestors: Ancestors<'a, '_>,
) -> SubtreeSkip {
    let Some(item) = rust_attributed_item(node, ancestors) else {
        // Attributes all the way to the end of the parent: nothing is
        // decorated, so nothing is pruned — and no later member of this
        // run will find an item either, so say so once. Without a
        // parent to bound the claim there is nothing to say, which
        // costs only a re-ask.
        return match ancestors.parent(node) {
            Some(parent) => SubtreeSkip::RETAIN.covering_through(parent.end_byte()),
            None => SubtreeSkip::RETAIN,
        };
    };
    // `item` is `node`'s sibling, so the chain that describes one
    // describes the other — parent and depth alike.
    let skipped = rust_prunable_item(&item) && rust_item_is_test_only(&item, code, ancestors);
    // Everything strictly between `node` and `item` is an
    // `AttributeItem` — that is what `rust_attributed_item` searched
    // for — so the reach holds for the run's remaining members. It
    // holds for *their* descendants too, which the walk visits whenever
    // the run is retained: an attribute's children are the `#`, `[`,
    // `]` tokens and the `attribute` body, none of them a
    // `rust_prunable_item` or an `AttributeItem`, so `RETAIN` is their
    // answer as well.
    SubtreeSkip::of(skipped).covering_through(item.start_byte())
}

/// The item an outer-attribute run decorates: the first sibling after
/// `node` that is not itself an `AttributeItem`.
///
/// Dispatched on [`forward_attribute_scan_budget`] for the reason the
/// backward reading of the same run is (#1100), with the directions
/// swapped: a cursor pass over a narrow parent's children is `O(width)`
/// and depth-free, while [`Node::next_sibling`] descends from the root
/// per step. Reading forward unconditionally would be `O(width)` *per
/// attribute*, which is quadratic on the flat `#[derive]`-per-item file
/// `bindgen` emits — the exact shape `nom/wide-attributed-fn` prices.
///
/// A `parent` that is not `node`'s parent leaves the first pass with
/// nothing to skip to and answers `None`, i.e. "nothing is pruned" —
/// the same conservative reading `rust_attribute_run_under` takes, and
/// a caller error either way (`Ancestors::checked`, #1122).
///
/// # What this costs
///
/// An attribute now pays a lookahead plus the run reading the item
/// pays, so a file of nothing but attributed items resolves about 2.5x
/// the siblings it did. Measured in release: a generated 20 000-item
/// `#[inline] fn f() {}` file went 245 ms to 305 ms and a 3 000-deep
/// nest 33 ms to 35 ms, while this repository's own 90 kloc of Rust did
/// not move off 59 ms — real code is a few percent attributed, not all
/// of it. Every scaling probe kept its exponent (`nom/wide-attributed-
/// fn` 1.02 to 1.04, `nom/nested-attributed-fn` 1.00 to 1.02), which is
/// the property that matters: the constant is the price of asking, and
/// the class is what #1100 was about.
///
/// Reusing [`rust_item_is_test_only`] rather than fusing the two walks
/// into one pass is deliberate at that price. A fused scan would save
/// one of the three sibling resolutions and would be a third reading of
/// the same run, free to drift from the other two — the failure mode
/// `.claude/rules/grammar-dispatch.md` §7 is about. #1446 kept that
/// choice: what was quadratic there was asking `run` times, not the
/// number of passes each ask makes, and
/// [`rust_attribute_run_verdict`]'s reach removes the repetition
/// without adding a reading.
///
/// # Why the same budget as the backward reading
///
/// The dispatch here weighs the same two costs #1100 did, and #1446 —
/// which found the diagonal `width ~= 3 * depth` shape this branch is
/// slowest on — did not move it. `parent.child_count() <= 3 * depth`
/// picks the forward pass exactly when `width * 35 ns` beats
/// `depth * 120 ns`, and on that diagonal it is still the cheaper of
/// the two: the sibling walk would be `O(run)` *and* `O(depth)` per
/// step. The budget was never what made the shape quadratic — asking
/// once per attribute was.
fn rust_attributed_item<'a>(node: &Node<'a>, ancestors: Ancestors<'a, '_>) -> Option<Node<'a>> {
    match ancestors.parent(node) {
        Some(parent) if parent.child_count() <= forward_attribute_scan_budget(ancestors) => parent
            .children()
            .skip_while(|child| child.id() != node.id())
            .find(|child| child.kind_id() != Rust::AttributeItem),
        _ => {
            let mut sibling = node.next_sibling();
            while let Some(candidate) = sibling {
                if candidate.kind_id() != Rust::AttributeItem {
                    return Some(candidate);
                }
                sibling = candidate.next_sibling();
            }
            None
        }
    }
}

/// Children a parent may have before reading its child list costs more
/// than resolving the sibling from `node` itself.
///
/// The two scans below grow on different axes and neither dominates.
/// [`Node::previous_sibling`] resolves the parent by descending from
/// the root, so its cost tracks the node's *depth*; a cursor pass over
/// the parent's children tracks the parent's *width*. Measured on this
/// workspace's grammar build, one cursor sibling step costs ~68 ns
/// against ~420 ns for a whole `previous_sibling` on a shallow node —
/// and that 420 ns did not move with the sibling index, so the
/// break-even sits near six children.
///
/// This is the budget at depth 1 only; deeper nodes get
/// [`FORWARD_ATTRIBUTE_SCAN_CHILDREN_PER_DEPTH`] more per level. It
/// still bounds the shallow-and-wide shape — a generated file of
/// thousands of top-level items stays on exactly the walk it had
/// before.
const MAX_FORWARD_ATTRIBUTE_SCAN_CHILDREN: usize = 6;

/// Children the budget above gains per level of depth.
///
/// Dispatching on width alone (#1100) reads the break-even as a
/// constant, but only the forward scan's price is: a
/// [`Node::previous_sibling`] descends from the root, so every extra
/// level makes the backward walk dearer while a cursor step stays flat.
/// One child too many therefore returned the whole `O(depth²)` walk —
/// measured on nested `mod` bodies of `#[inline] fn`s, a 7-child body
/// nested 3_200 deep took 2.67 s against 0.045 s for the same shape one
/// child narrower.
///
/// Sized from a crossover sweep of a forward-only against a
/// backward-only build over that fixture family: the *marginal* cost of
/// one more level of depth measured ~120 ns against ~35 ns for a cursor
/// step, and the observed break-even tracked that ratio — near 6
/// children at depth 1, near 13 at depth 3, and past 161 at depth 100.
/// (Marginal end-to-end figures, so they run below the isolated ~68 ns
/// / ~420 ns above; it is the ratio the budget uses, not either
/// absolute.) Three per level rounds the ratio down, keeping the
/// backward walk for the shallow-but-wide shape it is genuinely
/// cheaper on.
const FORWARD_ATTRIBUTE_SCAN_CHILDREN_PER_DEPTH: usize = 3;

/// The widest parent worth reading forward from, for a node at
/// `ancestors`' depth. A node reached without a chain has no depth to
/// scale by, so the shallow cut stands.
fn forward_attribute_scan_budget(ancestors: Ancestors<'_, '_>) -> usize {
    ancestors
        .depth()
        .map_or(0, |depth| {
            depth.saturating_mul(FORWARD_ATTRIBUTE_SCAN_CHILDREN_PER_DEPTH)
        })
        .max(MAX_FORWARD_ATTRIBUTE_SCAN_CHILDREN)
}

// The tree-sitter Rust grammar exposes outer attributes (`#[...]`) as
// `AttributeItem` siblings *before* the decorated item, so what marks
// `node` test-only is the run of attributes immediately preceding it —
// a non-attribute sibling (another item, a comment) ends the run. This
// scan covers every item kind, including `mod_item`, so
// `#[cfg(test)] mod tests` (an outer attribute on the module) is caught
// here while `mod tests { #![cfg(test)] }` is caught by the inner scan.
//
// The two readings agree on every node — `rust_outer_attr_scans_agree`
// pins that — so this only picks the cheaper one.
fn rust_outer_attr_marks_test<'a>(
    node: &Node<'a>,
    code: &[u8],
    ancestors: Ancestors<'a, '_>,
) -> bool {
    match ancestors.parent(node) {
        Some(parent) if parent.child_count() <= forward_attribute_scan_budget(ancestors) => {
            rust_attribute_run_under(&parent, node, code)
        }
        _ => rust_attribute_run_before(node, code),
    }
}

/// Reads the attribute run ending at `node` forward from `parent`,
/// which the walker already holds. `O(parent.child_count())` — two
/// cursor passes over the children when a preceding run exists, one
/// otherwise — and independent of how deep `node` sits.
///
/// Two passes and not one: classifying each attribute as the first pass
/// goes by would classify every attribute in the parent rather than
/// only the adjacent run, and classification is itself `O(len)` in the
/// attribute body (#1105). Folding them into a single-pass boolean
/// would reclassify a huge `#[cfg(all(all(…)))]` once per subsequent
/// sibling and put that quadratic straight back; the second
/// [`Node::children`] cursor is the cheaper half of that trade.
fn rust_attribute_run_under(parent: &Node, node: &Node, code: &[u8]) -> bool {
    let mut run_start = None;
    for child in parent.children() {
        if child.id() == node.id() {
            break;
        }
        if child.kind_id() == Rust::AttributeItem {
            if run_start.is_none() {
                run_start = Some(child);
            }
        } else {
            run_start = None;
        }
    }
    // No preceding attribute. A `parent` that is not `node`'s parent is
    // a caller error, not a shape answered for here: the first pass then
    // runs to the end of `parent`'s children, so the answer depends on
    // whether those end in an attribute run and is meaningless either
    // way. The walker constructs its chain through `Ancestors::checked`,
    // which catches a mismatch that shows in the spans in any debug build
    // and every mismatch under `make chain-audit` (#1122).
    let Some(run_start) = run_start else {
        return false;
    };
    parent
        .children()
        .skip_while(|child| child.id() != run_start.id())
        .take_while(|child| child.id() != node.id() && child.kind_id() == Rust::AttributeItem)
        .any(|attribute| rust_attribute_node_marks_test(&attribute, code, "#"))
}

/// Reads the same run backward from `node`. Each step resolves the
/// parent by descending from the root, so this is `O(attributes x
/// depth)` — but it does not touch the siblings it does not need,
/// which is what makes it the cheaper reading under a wide parent.
fn rust_attribute_run_before(node: &Node, code: &[u8]) -> bool {
    let mut sibling = node.previous_sibling();
    while let Some(attribute) = sibling {
        if attribute.kind_id() != Rust::AttributeItem {
            break;
        }
        if rust_attribute_node_marks_test(&attribute, code, "#") {
            return true;
        }
        sibling = attribute.previous_sibling();
    }
    false
}

// `mod_item` additionally accepts inner attributes (`#![cfg(test)]`).
// The grammar nests these inside the module's `declaration_list` body,
// not as direct `mod_item` children, so descend one level via the
// `body` field before scanning. Non-module items have no inner-attribute
// test form, so this returns `false` for them immediately.
fn rust_inner_attr_marks_test(node: &Node, code: &[u8]) -> bool {
    node.kind_id() == Rust::ModItem
        && node.child_by_field_name("body").is_some_and(|body| {
            body.children()
                .filter(|child| child.kind_id() == Rust::InnerAttributeItem)
                .any(|attribute| rust_attribute_node_marks_test(&attribute, code, "#!"))
        })
}

#[cfg(test)]
#[allow(
    clippy::float_cmp,
    clippy::cast_precision_loss,
    clippy::cast_possible_truncation,
    clippy::cast_sign_loss,
    clippy::similar_names,
    clippy::doc_markdown,
    clippy::needless_raw_string_hashes,
    clippy::too_many_lines
)]
mod tests {
    use super::*;
    use crate::count::count;
    use crate::langs::{
        BashParser, JavascriptCode, JavascriptParser, MozjsParser, PhpParser, TsxParser,
        TypescriptParser,
    };
    use crate::test_support::{ast_has_kind_id, for_each_node_with_chain};
    use std::fmt::Write as _;
    use std::path::PathBuf;

    #[cfg(feature = "bash")]
    fn parse(source: &str) -> BashParser {
        BashParser::new(source.as_bytes().to_vec(), &PathBuf::from("test.sh"), None)
    }

    #[cfg(feature = "bash")]
    fn count_strings(source: &str) -> usize {
        count(&parse(source), &["string".to_string()]).0
    }

    // `count`'s filter parser accepts a numeric string as a `kind_id` match
    // (parser.rs `filters`), so `has_kind` reuses the same primitive.
    #[cfg(feature = "bash")]
    fn has_kind(source: &str, kind_id: u16) -> bool {
        count(&parse(source), &[kind_id.to_string()]).0 > 0
    }

    #[cfg(feature = "bash")]
    #[test]
    fn bash_is_string_excludes_word_tokens() {
        // `echo hello world` produces three Word nodes — none of them are
        // string literals. Regression for #44 (Word must not match
        // is_string).
        assert_eq!(count_strings("echo hello world\n"), 0);
        assert_eq!(
            count_strings("if [ -f file.txt ]; then cat file.txt; fi\n"),
            0
        );
    }

    #[cfg(feature = "bash")]
    #[test]
    fn bash_is_string_matches_quoted_literals() {
        // Regular double-quoted string -> `string` (Bash::String).
        assert_eq!(count_strings("echo \"double\"\n"), 1);
        // Single-quoted string -> `raw_string` (Bash::RawString).
        assert_eq!(count_strings("echo 'single'\n"), 1);
        // ANSI-C quoting -> `ansi_c_string` (Bash::AnsiCString).
        assert_eq!(count_strings("echo $'ansi-c'\n"), 1);
    }

    #[cfg(feature = "bash")]
    #[test]
    fn bash_is_string_matches_translated_string() {
        // tree-sitter-bash only emits a visible `translated_string` node
        // in assignment-style contexts; in command arguments the `$"..."`
        // tokenizes as `$` plus a regular `string`. Use an assignment so
        // the wrapper actually appears in the AST.
        let src = "x=$\"translated\"\n";
        assert!(
            has_kind(src, Bash::TranslatedString as u16),
            "expected a translated_string node in {src:?}"
        );
        // The wrapper plus its inner `string` child both match is_string,
        // so count is 2.
        assert_eq!(count_strings(src), 2);
    }

    #[cfg(feature = "bash")]
    #[test]
    fn bash_is_string_matches_heredoc_bodies() {
        // Plain heredoc body.
        assert_eq!(
            count_strings("cat <<EOF\nhello world\nEOF\n"),
            1,
            "heredoc body should be counted as a string literal"
        );
        // Quoted-tag heredoc disables expansions but is still a string.
        assert_eq!(
            count_strings("cat <<'EOF'\nliteral $not_expanded\nEOF\n"),
            1
        );
        // Heredoc with an embedded expansion still yields exactly one
        // body node (parallel to a JS template string with `${x}`).
        assert_eq!(count_strings("cat <<EOF\nhi $name\nEOF\n"), 1);
    }

    // ===== PHP `is_string` regression tests (issue #288) =====

    #[cfg(feature = "php")]
    fn parse_php(source: &str) -> PhpParser {
        PhpParser::new(source.as_bytes().to_vec(), &PathBuf::from("test.php"), None)
    }

    #[cfg(feature = "php")]
    fn count_php_strings(source: &str) -> usize {
        count(&parse_php(source), &["string".to_string()]).0
    }

    #[cfg(feature = "php")]
    #[test]
    fn php_is_string_matches_single_quoted_literal() {
        // `Php::String` is the named single-quoted literal. Inert
        // single-quoted strings have always been matched; this anchors
        // the baseline before exercising the alias kinds.
        assert_eq!(count_php_strings("<?php $x = 'single';"), 1);
    }

    #[cfg(feature = "php")]
    #[test]
    fn php_is_string_matches_encapsed_heredoc_nowdoc_shell() {
        // `EncapsedString` (double-quoted), `Heredoc`, `Nowdoc`, and
        // `ShellCommandExpression` (backticks) must all match
        // `is_string`. Pre-#288 the alterator/checker arms were almost
        // aligned for these named literals — this test locks in the
        // shape.
        assert_eq!(count_php_strings("<?php $x = \"double\";"), 1);
        assert_eq!(
            count_php_strings("<?php $x = <<<EOT\nbody\nEOT;\n"),
            1,
            "heredoc should match is_string"
        );
        assert_eq!(
            count_php_strings("<?php $x = <<<'EOT'\nbody\nEOT;\n"),
            1,
            "nowdoc should match is_string"
        );
        assert_eq!(
            count_php_strings("<?php $x = `ls`;"),
            1,
            "shell command (backtick) should match is_string"
        );
    }

    #[cfg(feature = "php")]
    #[test]
    fn php_is_string_excludes_type_keyword_alias_1474() {
        // `Php::String2` (kind_id 25) is the `string` *type* keyword,
        // not a literal: `bca dump` finds it only as the sole child of
        // a `primitive_type` wrapper. #288 listed it in `is_string` for
        // kind-name identity and for parity with `get_op_type` and the
        // `Alterator`; #1293 moved it into `get_op_type`'s
        // `primitive_type`-suppression arm, so #1474 withdrew it here
        // too — the same reversal #1261 made for TS's `String2` and
        // TSX's `String3`. On the repo's PHP corpus this took
        // `find -t string` from 24 hits to 12 on `strings.php` and from
        // 27 to 8 on `classes.php`.
        //
        // The fixture carries every position the keyword appears in —
        // property type, plain parameter, `?string`, a `string|int`
        // union and the return type. Today's predicate is a `kind_id`
        // `matches!`, so it cannot tell those five apart; they are here
        // against a future ancestor-sensitive `is_string_with_code`
        // override of the kind Tcl already carries (#1381), which could
        // suppress one position and not another. The literals anchor
        // the other direction: without them this test would also pass
        // if `is_string` were emptied out entirely.
        let src = concat!(
            "<?php\n",
            "class C {\n",
            "    private string $prop = 'p';\n",
            "    public function f(string $a, ?string $b, string|int $c): string {\n",
            "        $d = \"y\";\n",
            "        $e = <<<EOT\n",
            "        body\n",
            "        EOT;\n",
            "        $g = <<<'EOT'\n",
            "        lit\n",
            "        EOT;\n",
            "        return (string) $a;\n",
            "    }\n",
            "}\n",
        );
        let parser = parse_php(src);
        // Pin the keyword spellings by *count*, not with a boolean
        // `ast_has_kind_id`. The zero below holds for one spelling
        // exactly as for five, so a fixture trimmed to a single
        // `string` keyword would keep every assertion here green while
        // the paragraph above claimed full coverage — measured, by
        // stripping four of the five. The count also subsumes the
        // "is the keyword in the parse at all" guard the boolean gave.
        let keyword_nodes = parser
            .root()
            .preorder()
            .filter(|n| n.kind_id() == Php::String2 as u16)
            .count();
        assert_eq!(
            keyword_nodes, 5,
            "fixture must keep all five `string` type-keyword positions: property, \
             parameter, `?string`, `string|int` union, return type",
        );
        assert_eq!(
            count_string_matches_for_kind(&parser, Php::String2 as u16, PhpCode::is_string),
            0,
            "Php::String2 (type keyword) must not match is_string",
        );
        // The `(string)` cast is the negative control: a childless
        // `cast_type`, a different kind_id, so the four below would be
        // five if anything ever classified a cast as a literal. Pinned
        // so the control cannot be trimmed out of the fixture silently.
        assert_eq!(
            parser
                .root()
                .preorder()
                .filter(|n| n.kind_id() == Php::CastType as u16)
                .count(),
            1,
            "fixture must keep the `(string)` cast negative control",
        );
        assert!(
            count_string_matches_for_kind(&parser, Php::String as u16, PhpCode::is_string) > 0,
            "the `'p'` literal must still match is_string",
        );
        // End-to-end through the filter chain `bca find` / `bca count`
        // actually use (`is_string_with_code`, `parser.rs`), not just
        // the predicate: `'p'`, `"y"`, the heredoc and the nowdoc — the
        // four literals, and none of the five keywords.
        assert_eq!(count_php_strings(src), 4);
    }

    // ===== JS-family `is_string` regression tests (issue #283) =====

    // Count nodes whose kind_id is exactly `target` *and* match
    // `is_string`. It answers both directions: a non-zero count proves
    // the alias reaches the predicate (the #283/#288 positive checks),
    // and a zero paired with an `ast_has_kind_id` presence check proves
    // a kind present in the parse is rejected (Go runes, C-family char
    // literals, PHP's `string` type keyword in #1474).
    #[cfg(any(
        feature = "bash",
        feature = "c",
        feature = "c-family-helpers",
        feature = "cpp",
        feature = "csharp",
        feature = "elixir",
        feature = "go",
        feature = "groovy",
        feature = "irules",
        feature = "java",
        feature = "javascript",
        feature = "kotlin",
        feature = "lua",
        feature = "mozcpp",
        feature = "mozjs",
        feature = "objc",
        feature = "perl",
        feature = "php",
        feature = "python",
        feature = "ruby",
        feature = "rust",
        feature = "tcl",
        feature = "typescript",
    ))]
    fn count_string_matches_for_kind<P: ParserTrait, F: Fn(&Node) -> bool>(
        parser: &P,
        target: u16,
        is_string: F,
    ) -> usize {
        parser
            .root()
            .preorder()
            .filter(|node| node.kind_id() == target && is_string(node))
            .count()
    }

    #[cfg(feature = "javascript")]
    #[test]
    fn javascript_is_string_matches_string2_alias() {
        // `Javascript::String2` (kind_id 221) aliases to `"string"`
        // (see `language_javascript.rs`). The alterator already
        // flattens it (#119); the generic `string` filter must agree
        // (#283). Use a source mix that exercises both the primary
        // `String` and the anonymous `String2` productions.
        let src = "const a = 'single';\nconst b = \"double\";\nimport \"m\";\n";
        let parser = JavascriptParser::new(src.as_bytes().to_vec(), &PathBuf::from("t.js"), None);
        // First confirm String2 actually surfaces in this parse —
        // otherwise the assertion below would be vacuously true.
        assert!(
            ast_has_kind_id(&parser, Javascript::String2 as u16),
            "expected Javascript::String2 to appear in the parse",
        );
        // Then assert every String2 node matches is_string.
        assert!(
            count_string_matches_for_kind(
                &parser,
                Javascript::String2 as u16,
                JavascriptCode::is_string,
            ) > 0,
            "Javascript::String2 nodes must match is_string",
        );
    }

    #[cfg(feature = "mozjs")]
    #[test]
    fn mozjs_is_string_matches_string2_alias() {
        // Parallel coverage for the MozJS dialect; same `String2`
        // alias as upstream JavaScript (kind_id 220 here).
        let src = "const a = 'single';\nconst b = \"double\";\nimport \"m\";\n";
        let parser = MozjsParser::new(src.as_bytes().to_vec(), &PathBuf::from("t.js"), None);
        assert!(
            ast_has_kind_id(&parser, Mozjs::String2 as u16),
            "expected Mozjs::String2 to appear in the parse",
        );
        assert!(
            count_string_matches_for_kind(&parser, Mozjs::String2 as u16, MozjsCode::is_string) > 0,
            "Mozjs::String2 nodes must match is_string",
        );
    }

    #[cfg(feature = "typescript")]
    #[test]
    fn typescript_is_string_excludes_type_keyword_alias_1261() {
        // TypeScript's only anonymous `"string"` alias, `String2`
        // (kind_id 135), is the `: string` type-annotation keyword —
        // not a literal. #313 made it match `is_string` for parity
        // with the operand classification; #1261 dropped it from both,
        // so `find string` / `count string` report literals only.
        // The `String` literal kinds must keep matching (the fixture's
        // `'x'` / `'y'`), otherwise this test would pass by
        // over-narrowing.
        let src = "const a: string = 'x';\nfunction f(): string { return 'y'; }\n";
        let parser = TypescriptParser::new(src.as_bytes().to_vec(), &PathBuf::from("t.ts"), None);
        assert!(
            ast_has_kind_id(&parser, Typescript::String2 as u16),
            "expected Typescript::String2 to appear in the parse",
        );
        assert_eq!(
            count_string_matches_for_kind(
                &parser,
                Typescript::String2 as u16,
                TypescriptCode::is_string,
            ),
            0,
            "Typescript::String2 (type keyword) must not match is_string",
        );
        assert!(
            count_string_matches_for_kind(
                &parser,
                Typescript::String as u16,
                TypescriptCode::is_string,
            ) > 0,
            "Typescript::String literals must still match is_string",
        );
    }

    #[cfg(feature = "typescript")]
    #[test]
    fn tsx_is_string_matches_literal_alias_not_type_keyword_1261() {
        // TSX uniquely carries two anonymous `"string"` aliases:
        // `String3` (kind_id 141, the type-annotation keyword) and
        // `String2` (kind_id 261, the string-literal alias). Both must
        // appear in this fixture: the `: string` annotation produces
        // `String3`, and the `'x'` / `"y"` / `"m"` / `"c"` literals
        // produce `String2`. Asserting presence of both *before*
        // checking `is_string` ensures a future grammar bump that
        // stops emitting either alias fails loudly here rather than
        // silently dropping coverage. The literal alias matches
        // (#283); the type keyword must not (#1261).
        let src = "const a: string = 'x';\n\
                   const b = \"y\";\n\
                   import \"m\";\n\
                   const el = <div className=\"c\">{\"t\"}</div>;\n";
        let parser = TsxParser::new(src.as_bytes().to_vec(), &PathBuf::from("t.tsx"), None);
        assert!(
            ast_has_kind_id(&parser, Tsx::String3 as u16),
            "expected Tsx::String3 (type-keyword `string`) in the parse",
        );
        assert!(
            ast_has_kind_id(&parser, Tsx::String2 as u16),
            "expected Tsx::String2 (string-literal alias) in the parse",
        );
        assert_eq!(
            count_string_matches_for_kind(&parser, Tsx::String3 as u16, TsxCode::is_string),
            0,
            "Tsx::String3 (type keyword) must not match is_string",
        );
        assert!(
            count_string_matches_for_kind(&parser, Tsx::String2 as u16, TsxCode::is_string) > 0,
            "Tsx::String2 (string-literal alias) nodes must match is_string",
        );
    }

    // The first node, in document order, whose `kind_id` equals `target`.
    #[cfg(any(
        feature = "c",
        feature = "cpp",
        feature = "csharp",
        feature = "groovy",
        feature = "mozcpp",
        feature = "python",
        feature = "rust",
    ))]
    fn find_first_kind<P: ParserTrait>(parser: &P, target: u16) -> Option<Node<'_>> {
        parser
            .root()
            .preorder()
            .find(|node| node.kind_id() == target)
    }

    /// `#[cfg(test)] mod tests { … }` carries the test marker as an
    /// *outer* attribute sibling before the `mod_item`. The outer scan
    /// must run for `mod_item` nodes too, so this case is caught by
    /// `rust_outer_attr_marks_test` — not the inner scan. Pins the key
    /// invariant of the helper split: a `mod_item`'s outer attributes are
    /// never skipped in favour of only its inner attributes.
    #[cfg(feature = "rust")]
    #[test]
    fn rust_outer_attr_on_mod_is_test_only() {
        let src = "#[cfg(test)]\nmod tests {\n    fn t() {}\n}\n";
        let parser = RustParser::new(src.as_bytes().to_vec(), &PathBuf::from("test.rs"), None);
        let code = parser.code();
        let node = find_first_kind(&parser, Rust::ModItem as u16).expect("mod_item");

        assert!(rust_item_is_test_only(&node, code, Ancestors::unknown()));
        // The marker lives on the outer scan; the inner scan sees no
        // `#![cfg(test)]` and must report false.
        assert!(rust_outer_attr_marks_test(
            &node,
            code,
            Ancestors::unknown()
        ));
        assert!(!rust_inner_attr_marks_test(&node, code));
    }

    /// `mod tests { #![cfg(test)] … }` carries the marker as an *inner*
    /// attribute nested in the module body. The outer sibling scan finds
    /// nothing; `rust_inner_attr_marks_test` descends via the `body`
    /// field and catches it.
    #[cfg(feature = "rust")]
    #[test]
    fn rust_inner_attr_in_mod_is_test_only() {
        let src = "mod tests {\n    #![cfg(test)]\n    fn t() {}\n}\n";
        let parser = RustParser::new(src.as_bytes().to_vec(), &PathBuf::from("test.rs"), None);
        let code = parser.code();
        let node = find_first_kind(&parser, Rust::ModItem as u16).expect("mod_item");

        assert!(rust_item_is_test_only(&node, code, Ancestors::unknown()));
        assert!(!rust_outer_attr_marks_test(
            &node,
            code,
            Ancestors::unknown()
        ));
        assert!(rust_inner_attr_marks_test(&node, code));
    }

    /// A plain, unattributed item is not test-only — neither scan matches.
    #[cfg(feature = "rust")]
    #[test]
    fn rust_plain_item_is_not_test_only() {
        let src = "fn foo() {}\n";
        let parser = RustParser::new(src.as_bytes().to_vec(), &PathBuf::from("test.rs"), None);
        let code = parser.code();
        let node = find_first_kind(&parser, Rust::FunctionItem as u16).expect("function_item");

        assert!(!rust_item_is_test_only(&node, code, Ancestors::unknown()));
        assert!(!rust_outer_attr_marks_test(
            &node,
            code,
            Ancestors::unknown()
        ));
        assert!(!rust_inner_attr_marks_test(&node, code));
    }

    /// Reading the `#[…]` run forward from the parent must answer
    /// exactly what walking backward from the item answers, on every
    /// node of every attribute shape.
    ///
    /// This is the acceptance criterion for #1100, not a nicety:
    /// `rust_outer_attr_marks_test` picks between the two by the
    /// parent's width against the node's depth, so a disagreement would
    /// make `exclude_tests` prune a different set of subtrees on one
    /// shape than on another and move published metric values —
    /// including as the budget is retuned. The comparison is against the
    /// backward reading directly rather than against the dispatcher,
    /// which would compare the backward walk with itself on every wide
    /// parent.
    ///
    /// The counted `marked` / `unmarked` totals are what stop it from
    /// passing vacuously — two readings that both answered `false`
    /// everywhere would agree perfectly.
    #[cfg(feature = "rust")]
    #[test]
    fn rust_outer_attr_scans_agree() {
        // One attribute run longer than the file's actual code, so the
        // run length cannot be confused with the child count.
        let long_run = format!("{}fn f() {{}}\n", "#[allow(dead_code)]\n".repeat(64));
        // A parent with hundreds of children: the shallow-but-wide
        // shape a per-step sibling scan would have been worst on.
        let mut wide = String::new();
        for i in 0..200 {
            let _ = writeln!(wide, "#[inline] fn f{i}() {{}}");
        }
        let sources = [
            // Attribute run opening the file, so the marked item is
            // its parent's first non-attribute child.
            "#[cfg(test)]\nfn a() {}\n",
            // Every recognised marker form, plus the `not(test)` form
            // that must *not* mark (#278).
            "#[test]\nfn a() {}\n#[tokio::test]\nfn b() {}\n\
             #[cfg(all(test, feature = \"x\"))]\nfn c() {}\n\
             #[cfg(not(test))]\nfn d() {}\n",
            // A run whose marker is not the attribute adjacent to the
            // item, the same run reversed, and one where an
            // intervening comment ends the run. Both orders matter:
            // only the second distinguishes a scan that stops at the
            // node it was asked about from one that reads the whole
            // run, since the query node here is itself an attribute.
            "#[cfg(test)]\n#[allow(dead_code)]\nfn a() {}\n\
             #[allow(dead_code)]\n#[cfg(test)]\nfn b() {}\n\
             #[cfg(test)]\n// break\nfn c() {}\n",
            // Nested modules, inner attributes, and an attributed item
            // inside a function body — parents other than the root.
            "mod outer {\n#![cfg(test)]\nmod inner {\n#[test]\nfn a() {}\n}\n}\n\
             fn b() {\n#[cfg(test)]\nfn c() {}\nlet x = 1;\n}\n",
            // A bare item: no parent for the root, no preceding
            // sibling for the item.
            "fn a() {}\n",
            // The dispatch boundary at depth 1, where the budget is
            // `MAX_FORWARD_ATTRIBUTE_SCAN_CHILDREN`: three attributed
            // items make a `source_file` exactly six children wide, and
            // a fourth bare item makes seven. Neither width occurred in
            // any fixture above, so `<=` against `<` went untested.
            "#[cfg(test)]\nfn a() {}\n#[inline]\nfn b() {}\n#[cfg(test)]\nfn c() {}\n",
            "#[cfg(test)]\nfn a() {}\n#[inline]\nfn b() {}\n#[cfg(test)]\nfn c() {}\nfn d() {}\n",
            // The same seven items one level down: two braces make the
            // `declaration_list` nine wide, which is exactly the budget
            // at depth 3, so the reading flips back to forward.
            "mod m {\n#[cfg(test)]\nfn a() {}\n#[inline]\nfn b() {}\n\
             #[cfg(test)]\nfn c() {}\nfn d() {}\n}\n",
            &long_run,
            &wide,
        ];

        let (mut marked, mut unmarked) = (0_usize, 0_usize);
        let (mut read_forward, mut read_backward) = (0_usize, 0_usize);
        for source in sources {
            let code = source.as_bytes();
            let visited = for_each_node_with_chain::<RustCode>(code, |node, chain| {
                let Some(parent) = chain.last() else {
                    return; // The root has no parent to read forward from.
                };
                let backward = rust_attribute_run_before(node, code);
                assert_eq!(
                    rust_attribute_run_under(parent, node, code),
                    backward,
                    "outer-attribute readings disagree on {} at row {}",
                    node.kind(),
                    node.start_row(),
                );
                assert_eq!(
                    rust_outer_attr_marks_test(node, code, Ancestors::known(chain)),
                    backward,
                    "the width dispatch changed the answer on {} at row {}",
                    node.kind(),
                    node.start_row(),
                );
                if backward {
                    marked += 1;
                } else {
                    unmarked += 1;
                }
                if parent.child_count() > forward_attribute_scan_budget(Ancestors::known(chain)) {
                    read_backward += 1;
                } else {
                    read_forward += 1;
                }
            });
            assert!(visited > 0, "fixture must have nodes to compare");
        }
        assert!(
            marked > 0 && unmarked > 0,
            "the comparison saw only one answer ({marked} marked, {unmarked} unmarked), \
             so agreeing proves nothing"
        );
        // Both arms have to fire: the dispatcher assertion above
        // compares the backward walk with itself on every node that
        // took the backward arm, so without forward readings it proves
        // nothing, and without backward ones the `wide` shape is unmet.
        assert!(
            read_forward > 0 && read_backward > 0,
            "one dispatch arm went unexercised ({read_forward} forward, \
             {read_backward} backward)"
        );
    }

    /// The budget is a width *per level of depth*, not a flat width.
    ///
    /// Pinned as a table because the shape of the formula is the whole
    /// point of #1100's follow-up: a flat cut returned the `O(depth²)`
    /// walk to any parent one child too wide, however deep it sat.
    #[cfg(feature = "rust")]
    #[test]
    fn the_forward_attribute_scan_budget_grows_with_depth() {
        assert_eq!(
            forward_attribute_scan_budget(Ancestors::unknown()),
            MAX_FORWARD_ATTRIBUTE_SCAN_CHILDREN,
            "no chain means no depth to scale by"
        );

        // `depth` reads the chain's length and nothing else, so one
        // real node repeated stands in for any ancestry — which is what
        // lets the table reach depths a fixture could not.
        let mut checked = 0_usize;
        for_each_node_with_chain::<RustCode>(b"fn a() {}\n", |node, _| {
            checked += 1;
            // The shallow cut floors the scaled one, so it governs
            // until the depth term overtakes it at depth 3.
            for (depth, want) in [(0, 6), (1, 6), (2, 6), (3, 9), (10, 30), (1_000, 3_000)] {
                let chain = vec![*node; depth];
                assert_eq!(
                    forward_attribute_scan_budget(Ancestors::known(&chain)),
                    want,
                    "budget at depth {depth}"
                );
            }
        });
        assert!(
            checked > 0,
            "fixture must yield a node to build chains from"
        );
    }

    /// The same equivalence one level up, on the hook the walker calls.
    ///
    /// `rust_item_is_test_only` folds the inner-attribute scan in, and
    /// `should_skip_subtree` adds the item-kind filter and the #1431
    /// attribute arm, so pinning it here is what carries the
    /// predicate-level agreement above through to the set of subtrees
    /// `exclude_tests` actually prunes.
    ///
    /// The oracle reads every run backward and every lookahead forward
    /// with the raw sibling accessors — no depth-against-width budget —
    /// so it disagrees with production the moment either dispatch arm
    /// answers differently from the other. The fixture spans both:
    /// `source_file` is wide enough at depth 1 to take the sibling
    /// arm, and `mod narrow`'s body is not.
    ///
    /// It also pins the [`SubtreeSkip`] reach, by replaying exactly what
    /// the walker does with it: carry the last covering verdict, and on
    /// every node it claims to answer for, check the claim against what
    /// the classifier says when asked directly. An over-broad reach —
    /// one byte past the item, say, or a run start taken as the run end
    /// — is a *silently wrong prune* in production, because the walker
    /// never asks again (#1446).
    #[cfg(feature = "rust")]
    #[test]
    fn rust_should_skip_subtree_matches_the_backward_reading() {
        let source = "#[cfg(test)]\nmod tests {\nfn a() {}\n}\n\
                      #[allow(dead_code)]\nstatic S: i32 = 1;\n\
                      mod inner {\n#![cfg(test)]\nconst C: i32 = 1;\n}\n\
                      #[rstest]\nfn b() {}\nfn c() {}\n\
                      #[cfg(test)]\nuse std::fmt;\n\
                      #[cfg(test)]\n#[allow(dead_code)]\nfn d() {}\n\
                      #[inline]\n#[allow(dead_code)]\n#[must_use]\n\
                      fn f() -> i32 {\n#[cfg(test)]\nfn inner() {}\n1\n}\n\
                      mod narrow {\n#[cfg(test)]\nfn e() {}\n}\n\
                      #[inline]\n#[allow(dead_code)]\n";
        let code = source.as_bytes();
        // Spelled out rather than reused from the production predicate:
        // the point is to pin which kinds the prune considers, so
        // borrowing `rust_prunable_item` would assert nothing.
        let is_item = |node: &Node| {
            matches!(
                node.kind_id().into(),
                Rust::ModItem
                    | Rust::FunctionItem
                    | Rust::ImplItem
                    | Rust::TraitItem
                    | Rust::ConstItem
                    | Rust::StaticItem
            )
        };
        let is_test_only = |node: &Node| {
            rust_attribute_run_before(node, code) || rust_inner_attr_marks_test(node, code)
        };
        let mut pruned = 0_usize;
        // The walker's own state, replayed: the last verdict that
        // claimed to answer for the nodes after it, and the end of the
        // deepest subtree pruned so far. The second is what makes the
        // replay a replay — the walker never pushes a skipped node's
        // children, so a reach owes them nothing and asserting over
        // them would fail on correct code.
        let mut reach = SubtreeSkip::RETAIN;
        let (mut pruned_through, mut reused) = (0_usize, 0_usize);
        let visited = for_each_node_with_chain::<RustCode>(code, |node, chain| {
            let reference = if node.kind_id() == Rust::AttributeItem {
                // Walk to the item this run decorates with the raw
                // forward accessor, independent of the budget dispatch.
                let mut following = node.next_sibling();
                while following.is_some_and(|s| s.kind_id() == Rust::AttributeItem) {
                    following = following.and_then(|s| s.next_sibling());
                }
                following.is_some_and(|item| is_item(&item) && is_test_only(&item))
            } else {
                is_item(node) && is_test_only(node)
            };
            let verdict = RustCode::should_skip_subtree(node, code, Ancestors::known(chain));
            let skipped = verdict.is_skipped();
            assert_eq!(
                skipped,
                reference,
                "prune decision moved on {} at row {}",
                node.kind(),
                node.start_row(),
            );
            if node.start_byte() >= pruned_through {
                if reach.answers_for(node.start_byte()) {
                    reused += 1;
                    assert_eq!(
                        reach.is_skipped(),
                        reference,
                        "a reach answered {} for {} at row {}",
                        reach.is_skipped(),
                        node.kind(),
                        node.start_row(),
                    );
                } else {
                    reach = verdict;
                }
                if skipped {
                    pruned_through = pruned_through.max(node.end_byte());
                }
            }
            pruned += usize::from(skipped);
        });
        assert!(visited > 0, "fixture must have nodes to compare");
        // Both halves of the reach have to fire, or the assertion above
        // is decoration: a run whose members are pruned (`fn d`'s two
        // rows) and one whose members are kept (`fn f`'s three, plus
        // the trailing pair that decorates nothing at all), the latter
        // also carrying each attribute's `#` / `[` / `]` / `attribute`
        // children through the same reach.
        //
        // `fn f`'s body is what makes an over-broad reach fail rather
        // than agree by luck. A reach is *always* right about the item
        // it stops at — the run's verdict is that item's verdict, by
        // construction — so a reach extended one node too far reads as
        // correct. Extended over the item's *body*, it swallows the
        // nested `#[cfg(test)] fn inner`, and the two answers part
        // company (verified by perturbation: `item.start_byte()` to
        // `item.end_byte()` fails this assertion, and without `fn f`'s
        // body it fails nothing).
        assert!(
            reused > 4,
            "only {reused} nodes were answered by a reach; the fixture \
             must carry runs long enough for the walker to reuse one"
        );
        // Seven test-only items — `mod tests`, `mod inner`, `fn b`,
        // `fn d`, `fn inner`, `fn e`, and `fn a` inside `mod tests` —
        // plus the six `#[…]` rows that mark five of them (`fn d`
        // carries two, `mod inner` an inner attribute that its own
        // subtree covers). Neither `#[allow(dead_code)]` on the
        // production `static`, nor `#[cfg(test)]` on the `use`, nor the
        // trailing run that decorates nothing, is the prune's to take.
        // A hook that pruned everything would satisfy the equality
        // above only if the oracle agreed; this pins the count the
        // fixture was written for.
        assert_eq!(
            pruned, 12,
            "expected the test-only items and the attributes marking them"
        );
    }

    /// A run that decorates nothing still answers for the rest of
    /// itself.
    ///
    /// `#[inline]` with no item after it parses cleanly — no ERROR node,
    /// just a `source_file` whose last children are attributes — so the
    /// lookahead runs off the end and there is no item to take the reach
    /// from. Answering per row there is *correct*, which is exactly why
    /// [`rust_should_skip_subtree_matches_the_backward_reading`] cannot
    /// see it: dropping this reach costs only time, and a trailing run
    /// is unbounded in length like any other.
    #[cfg(feature = "rust")]
    #[test]
    fn a_trailing_attribute_run_answers_for_its_whole_run() {
        let source = "fn a() {}\n#[inline]\n#[allow(dead_code)]\n";
        let code = source.as_bytes();
        let mut attributes = Vec::new();
        for_each_node_with_chain::<RustCode>(code, |node, chain| {
            if node.kind_id() == Rust::AttributeItem {
                attributes.push((
                    node.start_byte(),
                    RustCode::should_skip_subtree(node, code, Ancestors::known(chain)),
                ));
            }
        });
        let [(_, first), (second_start, second)] = attributes[..] else {
            panic!("fixture must hold exactly two attributes, got {attributes:?}");
        };
        assert!(
            !first.is_skipped() && !second.is_skipped(),
            "a run decorating nothing prunes nothing"
        );
        assert!(
            first.answers_for(second_start),
            "the first row's verdict must answer for the rest of the run"
        );
    }

    #[cfg(feature = "groovy")]
    #[test]
    fn groovy_is_else_if_recognises_else_followed_by_if() {
        // Direct assertion that `GroovyCode::is_else_if` returns true
        // for an `if_statement` whose previous sibling is the `else`
        // token. Defends the sibling-token strategy against accidental
        // regression to a `false` stub (lesson 10, #115 / #239).
        let src = "if (x) { } else if (y) { } else { }";
        let parser =
            GroovyParser::new(src.as_bytes().to_vec(), &PathBuf::from("test.groovy"), None);
        let outer =
            find_first_kind(&parser, Groovy::IfStatement as u16).expect("outer if_statement");
        // Locate the inner if_statement (in the `alternative` slot of
        // the outer if, after the `else` token).
        let mut inner: Option<Node> = None;
        for i in 0..outer.child_count() {
            if let Some(c) = outer.child(i)
                && c.kind_id() == Groovy::IfStatement as u16
            {
                inner = Some(c);
                break;
            }
        }
        let inner = inner.expect("expected an inner if_statement");
        assert!(
            GroovyCode::is_else_if(&inner, Ancestors::unknown()),
            "inner if_statement after `else` must be recognised as else-if"
        );
        assert!(
            !GroovyCode::is_else_if(&outer, Ancestors::unknown()),
            "outer if_statement must not be recognised as else-if"
        );
    }

    #[cfg(feature = "groovy")]
    #[test]
    fn groovy_is_else_if_false_for_standalone_if() {
        // A bare `if` (no `else` preceding it) must NOT register as
        // an else-if.
        let src = "if (x) { println(x) }";
        let parser =
            GroovyParser::new(src.as_bytes().to_vec(), &PathBuf::from("test.groovy"), None);
        let node = find_first_kind(&parser, Groovy::IfStatement as u16).expect("if_statement");
        assert!(!GroovyCode::is_else_if(&node, Ancestors::unknown()));
    }

    #[cfg(feature = "groovy")]
    #[test]
    fn groovy_is_call_excludes_constructors() {
        // Regression for #430. `GroovyCode::is_call` previously matched
        // `ObjectCreationExpression`, so `new Foo()` was counted as a
        // call in Groovy but not in Java/C# (whose `is_call` is method
        // invocation only). The Java-family convention is that `is_call`
        // (the `--ops`/`call` filter) counts method/function call sites
        // only; constructors are an ABC concern. This test mixes all
        // three Groovy call-shaped forms and pins the count at 2:
        //   * `new Foo()`        -> object_creation_expression (NOT a call)
        //   * `a.bar()`          -> method_invocation           (a call)
        //   * `println "hi"`     -> command_chain               (a call)
        // Pre-fix this count was 3 (the constructor was miscounted).
        let src = "def m() {\n  def a = new Foo()\n  a.bar()\n  println \"hi\"\n}\n";
        let parser =
            GroovyParser::new(src.as_bytes().to_vec(), &PathBuf::from("test.groovy"), None);
        assert_eq!(
            count(&parser, &["call".to_string()]).0,
            2,
            "is_call must count method_invocation + command_chain only, not the constructor"
        );

        // Direct predicate assertions: the constructor node must be
        // rejected while both genuine call forms are accepted.
        let ctor = find_first_kind(&parser, Groovy::ObjectCreationExpression as u16)
            .expect("object_creation_expression");
        assert!(
            !GroovyCode::is_call(&ctor),
            "object_creation_expression must not be a call"
        );
        let method =
            find_first_kind(&parser, Groovy::MethodInvocation as u16).expect("method_invocation");
        assert!(
            GroovyCode::is_call(&method),
            "method_invocation must be a call"
        );
        let chain = find_first_kind(&parser, Groovy::CommandChain as u16).expect("command_chain");
        assert!(GroovyCode::is_call(&chain), "command_chain must be a call");
    }

    #[cfg(feature = "csharp")]
    #[test]
    fn csharp_is_call_excludes_constructors() {
        // The C# half of the `is_call` contract documented on the
        // `Checker::is_call` trait method (#1456), mirroring the #430 test
        // above. C# spells four call-shaped constructs that ABC counts
        // as branches and this filter must not count as call sites:
        //
        //   * `new Foo()`                  -> object_creation_expression
        //   * `: this(a)`                  -> constructor_initializer
        //   * `class Sub(int x) : Base(x)` -> base_list > argument_list
        //   * `record R(int x) : Base(x)`  -> base_list >
        //                                     primary_constructor_base_type
        //
        // paired with one genuine call, `Helper(f)`. The fixture scores
        // `abc.branches` 5 and `call` 1, so widening `is_call` to any of
        // the four moves the count asserted here.
        let src = "class Sub(int x) : Base(x) {
                       public Sub(int a, int b) : this(a) { }
                       void M() { var f = new Foo(); Helper(f); }
                   }
                   record R(int x) : Base(x);";
        let parser = CsharpParser::new(src.as_bytes().to_vec(), &PathBuf::from("test.cs"), None);
        assert_eq!(
            count(&parser, &["call".to_string()]).0,
            1,
            "is_call must count `Helper(f)` only, not the four constructions"
        );

        let call = find_first_kind(&parser, Csharp::InvocationExpression3 as u16)
            .expect("invocation_expression");
        assert!(
            CsharpCode::is_call(&call),
            "an invocation_expression is a call"
        );
        // §2 drift markers for the other two members of
        // `csharp_invocation_expr_kinds!()`. All three render to
        // `"invocation_expression"`, and at the 0.23.5 pin every
        // spelling probed with `bca dump` — plain, `this.`-qualified,
        // `base.`-qualified, inside a query clause, a lambda body, an
        // attribute argument, an interpolation, a `when` guard and a
        // constant pattern — emits only the aliased id. The claim these
        // pin is therefore about *this* fixture; a grammar bump that
        // starts spelling an ordinary call with either id fails the
        // count above as well, since neither is `find_first_kind`'s
        // target here.
        for (id, name) in [
            (Csharp::InvocationExpression, "InvocationExpression"),
            (Csharp::InvocationExpression2, "InvocationExpression2"),
        ] {
            assert!(
                !ast_has_kind_id(&parser, id as u16),
                "{name} is unobserved at this grammar pin; its arm is defensive"
            );
        }

        for (id, what) in [
            (Csharp::ObjectCreationExpression as u16, "`new Foo()`"),
            (Csharp::ConstructorInitializer as u16, "`: this(a)`"),
            (
                Csharp::PrimaryConstructorBaseType as u16,
                "a record's `: Base(x)`",
            ),
        ] {
            let node = find_first_kind(&parser, id)
                .unwrap_or_else(|| panic!("fixture must contain {what}"));
            assert!(
                !CsharpCode::is_call(&node),
                "{what} is an ABC branch, not a call site"
            );
        }

        // The `class` spelling of the primary-constructor base call is a
        // bare `argument_list` under the `base_list`, and preorder reaches
        // `Sub`'s before the four that belong to `: this(a)`, `new Foo()`,
        // `Helper(f)` and the record's `: Base(x)` — the parent assertion
        // is what keeps this pointing at the construction rather than at
        // one of those.
        let base_args =
            find_first_kind(&parser, Csharp::ArgumentList as u16).expect("argument_list");
        assert_eq!(
            base_args.parent().map(|p| p.kind_id()),
            Some(Csharp::BaseList2 as u16),
            "the first argument_list must be `class Sub(int x) : Base(x)`'s"
        );
        assert!(
            !CsharpCode::is_call(&base_args),
            "a class's primary-constructor base call is an ABC branch, not a call site"
        );
    }

    // ===== C-family `is_call` regression tests (issue #1254) =====

    // Every C-family grammar here references `preproc_call_expression`
    // only under an `alias(…, $.call_expression)` — zero bare references
    // in `grammar.json`, and no entry in `node-types.json` — so the
    // unsuffixed `CallExpression` never reaches `kind_id()` while the
    // aliased `CallExpression2` carries every real call.
    //
    // The two halves do different jobs. The first keeps the fixture an
    // honest witness: a grammar bump that stops emitting the alias must
    // fail here rather than silently weaken the counts below. The second
    // cannot fail against the current parse tables — `ts_symbol_map`
    // rewrites the unsuffixed symbol onto the aliased one for *every*
    // input — so it is a drift marker, not a live guard: it fires only if
    // a bump renumbers the enum or stops aliasing the rule, which is
    // exactly what justifies keeping the unsuffixed arm defensively
    // (the grammar-dispatch §2 technique, applied to an alias-mapped
    // rule rather than a hidden one).
    #[cfg(any(feature = "c", feature = "cpp", feature = "mozcpp"))]
    #[track_caller]
    fn assert_call_kinds_are_aliased<P: ParserTrait>(
        parser: &P,
        lang: &str,
        aliased: u16,
        unsuffixed: u16,
    ) {
        assert!(
            ast_has_kind_id(parser, aliased),
            "{lang}: fixture must emit the aliased call_expression kind"
        );
        assert!(
            !ast_has_kind_id(parser, unsuffixed),
            "{lang}: the unsuffixed CallExpression is the aliased-away \
             preproc_call_expression and must not surface"
        );
    }

    // Six calls — statement position, initializer, `if` condition, both
    // binary operands, and the `return` operand. The other parenthesised
    // shapes are deliberate negatives: the `int f(int);` prototype's
    // `function_declarator`, the `sizeof(int)` operand, and the `(int)x`
    // cast. Six differs from the C++ fixture's five, so a test that
    // reaches for the wrong parser cannot pass on the right number.
    #[cfg(feature = "c")]
    const C_CALL_SHAPES: &str = "int f(int);\n\
                                 int main(void) {\n\
                                 f(1);\n\
                                 int x = g(2);\n\
                                 if (h(3)) { }\n\
                                 int y = a(4) + b(5);\n\
                                 int z = sizeof(int) + (int)x + y;\n\
                                 return c(z);\n\
                                 }\n";

    // Five calls — member (`t.m()`), arrow member (`p->m()`), qualified
    // (`ns::f()`), and both function-pointer spellings (`fp(2)`,
    // `(*fp)(3)`). `T t(1)` and `new T(4)` are the negatives that matter:
    // object creation is an ABC concern, not a call site, matching Groovy /
    // Java / C# (#430). Functional-style `T(4)` is deliberately absent
    // from the fixture — the grammar emits it as a plain `call_expression`,
    // so it *is* counted, and including it here would blur what the five
    // is pinning.
    #[cfg(any(feature = "cpp", feature = "mozcpp"))]
    const CPP_CALL_SHAPES: &str = "struct T { T(int); int m(); };\n\
                                   namespace ns { int f(); }\n\
                                   int main() {\n\
                                   T t(1);\n\
                                   t.m();\n\
                                   T *p = &t;\n\
                                   p->m();\n\
                                   ns::f();\n\
                                   int (*fp)(int) = nullptr;\n\
                                   fp(2);\n\
                                   (*fp)(3);\n\
                                   T *q = new T(4);\n\
                                   delete q;\n\
                                   return sizeof(T);\n\
                                   }\n";

    /// Every `call_expression` the pinned C grammar emits carries the
    /// *aliased* `CallExpression2` kind_id. The unsuffixed variant is the
    /// grammar's `preproc_call_expression`, referenced only under an
    /// `alias(…, $.call_expression)` — it is absent from `node-types.json`
    /// and never reaches `kind_id()`. Matching it alone left `is_call`
    /// dead: `bca count -t call` reported 0 on ordinary C (#1254).
    #[cfg(feature = "c")]
    #[test]
    fn c_is_call_matches_the_aliased_call_expression() {
        let parser = CParser::new(
            C_CALL_SHAPES.as_bytes().to_vec(),
            &PathBuf::from("test.c"),
            None,
        );

        assert_call_kinds_are_aliased(
            &parser,
            "C",
            C::CallExpression2 as u16,
            C::CallExpression as u16,
        );

        assert_eq!(
            count(&parser, &["call".to_string()]).0,
            6,
            "is_call must match every call_expression the C grammar emits"
        );

        // Parenthesised non-call shapes stay out. The exact count above is
        // the sharper guard — these rows document the intended boundary
        // and catch a future edit that widens the match to them.
        for (kind, label) in [
            (C::SizeofExpression as u16, "sizeof_expression"),
            (C::CastExpression as u16, "cast_expression"),
            (C::FunctionDeclarator as u16, "function_declarator"),
        ] {
            let node = find_first_kind(&parser, kind).expect(label);
            assert!(!CCode::is_call(&node), "{label} must not be a call");
        }
    }

    /// C++ counterpart of [`c_is_call_matches_the_aliased_call_expression`].
    /// Member, qualified and function-pointer calls all arrive as the same
    /// aliased `CallExpression2`; object construction does not (#1254).
    #[cfg(feature = "cpp")]
    #[test]
    fn cpp_is_call_matches_the_aliased_call_expression() {
        let parser = CppParser::new(
            CPP_CALL_SHAPES.as_bytes().to_vec(),
            &PathBuf::from("test.cpp"),
            None,
        );

        assert_call_kinds_are_aliased(
            &parser,
            "Cpp",
            Cpp::CallExpression2 as u16,
            Cpp::CallExpression as u16,
        );

        assert_eq!(
            count(&parser, &["call".to_string()]).0,
            5,
            "is_call must match member, qualified and function-pointer \
             calls, and only those"
        );

        // `new T(4)` and `sizeof(T)` are call-shaped but not call sites.
        for (kind, label) in [
            (Cpp::NewExpression as u16, "new_expression"),
            (Cpp::SizeofExpression as u16, "sizeof_expression"),
            (Cpp::InitDeclarator as u16, "init_declarator"),
        ] {
            let node = find_first_kind(&parser, kind).expect(label);
            assert!(!CppCode::is_call(&node), "{label} must not be a call");
        }
    }

    /// `Mozcpp` owns no file extension, so nothing routes to it and it
    /// gets no integration-snapshot coverage. Pin it against its
    /// extension-owning sibling instead: the Mozilla fork inherits
    /// upstream C++'s aliasing, so the same source must yield the same
    /// calls (grammar-dispatch, "sweep the rest").
    #[cfg(all(feature = "cpp", feature = "mozcpp"))]
    #[test]
    fn mozcpp_is_call_agrees_with_cpp() {
        let source = CPP_CALL_SHAPES.as_bytes().to_vec();
        let mozcpp = MozcppParser::new(source.clone(), &PathBuf::from("test.cpp"), None);
        let cpp = CppParser::new(source, &PathBuf::from("test.cpp"), None);

        assert_call_kinds_are_aliased(
            &mozcpp,
            "Mozcpp",
            Mozcpp::CallExpression2 as u16,
            Mozcpp::CallExpression as u16,
        );

        let mozcpp_calls = count(&mozcpp, &["call".to_string()]).0;
        assert_eq!(
            mozcpp_calls, 5,
            "Mozcpp must count the same five calls as C++"
        );
        assert_eq!(
            mozcpp_calls,
            count(&cpp, &["call".to_string()]).0,
            "Mozcpp is the Mozilla C++ fork; is_call must not diverge \
             from CppCode on shared source"
        );

        let new_expr =
            find_first_kind(&mozcpp, Mozcpp::NewExpression as u16).expect("new_expression");
        assert!(
            !MozcppCode::is_call(&new_expr),
            "new_expression must not be a call"
        );
    }

    #[cfg(feature = "python")]
    fn parse_python(src: &str) -> PythonParser {
        PythonParser::new(src.as_bytes().to_vec(), &PathBuf::from("test.py"), None)
    }

    // Walk the AST and return every node whose `kind_id` equals `target`,
    // in DFS pre-order. Used by the Python `is_else_if` tests below to
    // distinguish the outer if from the inner one in an `else: if` chain.
    #[cfg(feature = "python")]
    fn find_all_kinds<P: ParserTrait>(parser: &P, target: u16) -> Vec<Node<'_>> {
        let mut out = Vec::new();
        let mut stack = vec![parser.root()];
        while let Some(node) = stack.pop() {
            if node.kind_id() == target {
                out.push(node);
            }
            for i in (0..node.child_count()).rev() {
                if let Some(c) = node.child(i) {
                    stack.push(c);
                }
            }
        }
        out
    }

    #[cfg(feature = "python")]
    #[test]
    fn python_is_else_if_recognises_if_inside_else_clause() {
        // `else: if b:` chains parse as `else_clause → block → if_statement`
        // (the `block` wrapper is Python-specific). `is_else_if` must
        // walk through that wrapper. Regression for #276 (the stub
        // returned `false` unconditionally).
        let src = "if a:\n    pass\nelse:\n    if b:\n        pass\n";
        let parser = parse_python(src);
        let outer =
            find_first_kind(&parser, Python::IfStatement as u16).expect("outer if_statement");
        let inner =
            find_python_if_inside_else_block(&parser).expect("inner if_statement under else");
        assert!(
            PythonCode::is_else_if(&inner, Ancestors::unknown()),
            "if_statement inside else_clause's block must be recognised as else-if"
        );
        assert!(
            !PythonCode::is_else_if(&outer, Ancestors::unknown()),
            "outer if_statement must not be recognised as else-if"
        );
    }

    #[cfg(feature = "python")]
    #[test]
    fn python_is_else_if_false_for_standalone_if() {
        // A bare `if` whose parent is the module / function body must
        // NOT register as an else-if.
        let src = "if a:\n    pass\n";
        let parser = parse_python(src);
        let node = find_first_kind(&parser, Python::IfStatement as u16).expect("if_statement");
        assert!(!PythonCode::is_else_if(&node, Ancestors::unknown()));
    }

    #[cfg(feature = "python")]
    #[test]
    fn python_is_else_if_false_for_outer_if_with_elif_alternative() {
        // `elif` parses as an `ElifClause`, not an `IfStatement`, so the
        // only `IfStatement` in `if … elif …` is the outer one. Its
        // parent is the module / function body, not an `else_clause`.
        // Pins that the presence of `elif` in the AST does not trip
        // `is_else_if` for the outer `if`.
        let src = "if a:\n    pass\nelif b:\n    pass\n";
        let parser = parse_python(src);
        let outer = find_first_kind(&parser, Python::IfStatement as u16).expect("if_statement");
        assert!(!PythonCode::is_else_if(&outer, Ancestors::unknown()));
    }

    // Return the inner `if_statement` that sits directly inside an
    // `else_clause`'s `block` wrapper, or `None` if no such node exists.
    // Used by tests below instead of relying on `find_all_kinds`'s DFS
    // pre-order to land at `ifs[1]`.
    #[cfg(feature = "python")]
    fn find_python_if_inside_else_block(parser: &PythonParser) -> Option<Node<'_>> {
        find_all_kinds(parser, Python::IfStatement as u16)
            .into_iter()
            .find(|n| {
                n.parent().is_some_and(|p| {
                    matches!(p.kind_id().into(), Python::Block | Python::Block2)
                        && p.parent()
                            .is_some_and(|gp| gp.kind_id() == Python::ElseClause)
                })
            })
    }

    #[cfg(feature = "python")]
    #[test]
    fn python_is_else_if_false_when_else_body_has_siblings() {
        // `else: if b:` followed by another statement at the same indent
        // is a real nested-if, not a chain. The block has 2 named
        // children, so `is_else_if` must return false.
        let src = "if a:\n    pass\nelse:\n    if b:\n        pass\n    pass\n";
        let parser = parse_python(src);
        let inner =
            find_python_if_inside_else_block(&parser).expect("inner if_statement under else");
        assert!(
            !PythonCode::is_else_if(&inner, Ancestors::unknown()),
            "inner if must NOT be recognised as else-if when its block has siblings"
        );
    }

    // Drift guard for #419: tree-sitter-python lists hidden aliases for
    // both `block` and `lambda` that are NOT emitted today —
    // `Block` (135, the hidden `_block` supertype; only `Block2` 160 is
    // emitted) and `Lambda2` (197, an unseen lambda alias; only
    // `Lambda` 196 and the `lambda` keyword token `Lambda3` 73 appear).
    // Several metric sites (`is_closure` — which now routes through
    // `cognitive::python_is_lambda`, the three cognitive lambda-scope
    // sites it feeds (#422), `is_else_if`, `python_is_block` in npa,
    // npm's class-body lookup, loc's no-op arm) already enumerate these
    // aliases defensively so a future grammar bump that promotes either
    // supertype to a concrete node cannot silently undercount.
    // This test pins their current absence across representative Python
    // (function, class, if/for bodies, lambda); if a bump ever emits one,
    // the guard flips red and forces a positive assertion to be added —
    // mirroring the `Php::String3` hidden-supertype guard above.
    #[cfg(feature = "python")]
    #[test]
    fn python_hidden_block_and_lambda_aliases_stay_unseen() {
        let src = "def f(a, b):\n    if a:\n        return b\n    for x in b:\n        print(x)\n\nclass C:\n    def m(self):\n        pass\n\ng = lambda x: x + 1\n";
        let parser = parse_python(src);

        // Confirm the LIVE aliases actually surface, so this fixture is
        // a meaningful witness (a guard that never parses a block or a
        // lambda would pass vacuously).
        assert!(
            ast_has_kind_id(&parser, Python::Block2 as u16),
            "expected Python::Block2 (160, the emitted `block`) in the parse",
        );
        assert!(
            ast_has_kind_id(&parser, Python::Lambda as u16),
            "expected Python::Lambda (196, the emitted `lambda`) in the parse",
        );

        // The hidden supertypes must NOT surface. If either now appears,
        // add a positive assertion routing it through the relevant
        // predicate (see #419).
        assert!(
            !ast_has_kind_id(&parser, Python::Block as u16),
            "Python::Block (135) is the hidden `_block` supertype; if it now appears, route it through python_is_block and assert positively (#419)",
        );
        assert!(
            !ast_has_kind_id(&parser, Python::Lambda2 as u16),
            "Python::Lambda2 (197) is an unseen lambda alias; if it now appears, cognitive::python_is_lambda (reused by is_closure and the three cognitive lambda-scope sites) must detect it and a positive closure assertion is required (#419/#422)",
        );
    }

    // #422: the cognitive lambda-alias chokepoint and `is_closure` must
    // recognise the *same* live `lambda` node. Before #422, cognitive's
    // three lambda sites compared against `Lambda` (196) only while
    // `is_closure` accepted `Lambda | Lambda2`; both now route through
    // `cognitive::python_is_lambda`, so this pins that they agree on the
    // emitted `Lambda` and that the predicate is not vacuously true (it
    // rejects the enclosing `FunctionDefinition`). The unseen `Lambda2`
    // half of the set is covered by the drift guard above.
    #[cfg(feature = "python")]
    #[test]
    fn python_is_lambda_matches_live_lambda_and_agrees_with_is_closure() {
        use crate::lang_helpers::python::python_is_lambda;

        let parser = parse_python("def f():\n    return lambda x: x and x\n");
        let lambda = find_first_kind(&parser, Python::Lambda as u16)
            .expect("the lambda expression must parse as Python::Lambda (196)");

        assert!(
            python_is_lambda(&lambda),
            "python_is_lambda must accept the emitted Lambda node",
        );
        assert!(
            PythonCode::is_closure(&lambda, Ancestors::unknown()),
            "is_closure must agree with python_is_lambda on the same lambda node",
        );

        // Not vacuously true: a non-lambda node (the enclosing function)
        // must be rejected, so the predicate is discriminating.
        let func = find_first_kind(&parser, Python::FunctionDefinition as u16)
            .expect("the def must parse as Python::FunctionDefinition");
        assert!(
            !python_is_lambda(&func),
            "python_is_lambda must reject a non-lambda node",
        );
    }

    /// The two `is_else_if` implementations that answer a flat `false`
    /// are a behavioural claim, not a placeholder: the language has no
    /// `else if` *chain* to collapse, so every `if` it contains is a
    /// freshly nested branch and pays cognitive nesting.
    ///
    /// Neither is reachable from a metric walk — `Preproc` and
    /// `Ccomment` (which inherit the `Checker` default) have no `if`
    /// construct at all, and Elixir's branching is `cond do` or nested
    /// `if/else` — so nothing else in the suite executes them. #1084
    /// changed both signatures to take `Ancestors`, which is exactly the
    /// kind of edit that can quietly invert a one-line predicate, so pin
    /// the contract directly.
    #[cfg(all(feature = "c-family-helpers", feature = "elixir"))]
    #[test]
    fn languages_without_else_if_chains_answer_false() {
        // The `Checker` default, via the two grammars that do not
        // override it.
        let preproc = PreprocParser::new(b"#define A 1\n".to_vec(), &PathBuf::from("a.h"), None);
        let ccomment = CcommentParser::new(b"/* c */\n".to_vec(), &PathBuf::from("a.c"), None);
        assert!(
            !PreprocCode::is_else_if(&preproc.root(), Ancestors::unknown()),
            "the Checker default must answer false"
        );
        assert!(
            !CcommentCode::is_else_if(&ccomment.root(), Ancestors::unknown()),
            "the Checker default must answer false"
        );

        // Elixir's explicit override, asserted on a real `if` node so the
        // test would catch an impl that started consulting the ancestor
        // chain and got the answer wrong.
        let parser = ElixirParser::new(
            b"defmodule M do\n  def f(a) do\n    if a do\n      if a do\n        :ok\n      end\n    end\n  end\nend\n".to_vec(),
            std::path::Path::new("m.ex"),
            None,
        );
        let ifs: Vec<_> = parser
            .root()
            .preorder()
            .filter(|n| {
                n.kind() == "call"
                    && n.utf8_text(parser.code())
                        .is_some_and(|t| t.starts_with("if"))
            })
            .collect();
        assert!(
            ifs.len() >= 2,
            "fixture must contain a nested if, else the assertion below is vacuous"
        );
        for node in &ifs {
            assert!(
                !ElixirCode::is_else_if(node, Ancestors::unknown()),
                "Elixir has no else-if chain; every if is a fresh branch"
            );
        }
    }

    // Regression for #301: every language consolidated under
    // `impl_simple_is_string!` must still recognise its canonical
    // string literal via the `"string"` filter (which routes through
    // `Checker::is_string`). The positive test now drills down to
    // every individual variant of every multi-variant language so a
    // future macro invocation that drops a variant (e.g. forgetting
    // `Cpp::ConcatenatedString` after a grammar bump) fails loudly.
    //
    // JS-family languages (Mozjs/Javascript/Typescript/Tsx) keep
    // their dedicated `impl_js_family_is_string!` macro and have
    // their own alias-aware tests above; they are intentionally
    // not duplicated here.
    #[cfg(any(
        feature = "bash",
        feature = "c-family-helpers",
        feature = "cpp",
        feature = "csharp",
        feature = "elixir",
        feature = "go",
        feature = "groovy",
        feature = "irules",
        feature = "java",
        feature = "kotlin",
        feature = "lua",
        feature = "perl",
        feature = "php",
        feature = "python",
        feature = "ruby",
        feature = "rust",
        feature = "tcl",
    ))]
    fn count_with_parser<P: ParserTrait>(parser: &P) -> usize {
        count(parser, &["string".to_string()]).0
    }

    // Assert that `target` kind_id appears in the parse and every
    // such node matches `is_string`. The two-step check makes test
    // failures unambiguous: a presence failure means the fixture no
    // longer produces the variant (likely grammar drift); a match
    // failure means a macro invocation dropped the variant.
    #[cfg(any(
        feature = "bash",
        feature = "c-family-helpers",
        feature = "cpp",
        feature = "csharp",
        feature = "elixir",
        feature = "go",
        feature = "groovy",
        feature = "irules",
        feature = "java",
        feature = "kotlin",
        feature = "lua",
        feature = "perl",
        feature = "php",
        feature = "python",
        feature = "ruby",
        feature = "rust",
        feature = "tcl",
    ))]
    fn assert_variant_is_string<P: ParserTrait, F: Fn(&Node) -> bool>(
        parser: &P,
        target: u16,
        is_string: F,
        lang: &str,
        variant: &str,
    ) {
        assert!(
            ast_has_kind_id(parser, target),
            "{lang}::{variant} (kind_id {target}) did not appear in the parse — fixture broken",
        );
        assert!(
            count_string_matches_for_kind(parser, target, is_string) > 0,
            "{lang}::{variant} must route through is_string",
        );
    }

    // Collapses the 6-line-per-variant `assert_variant_is_string`
    // call into a single token per variant. `$lang` is the language
    // enum (e.g. `Cpp`); `$code` is the `Checker`-implementing type
    // (e.g. `CppCode`); the trailing list names the enum variants to
    // exercise. The macro feeds `stringify!` for both the language
    // and variant labels so test failures keep the same "Lang::Variant"
    // wording the helper already emits.
    #[cfg(all(
        feature = "bash",
        feature = "c-family-helpers",
        feature = "cpp",
        feature = "csharp",
        feature = "elixir",
        feature = "go",
        feature = "groovy",
        feature = "irules",
        feature = "java",
        feature = "kotlin",
        feature = "lua",
        feature = "perl",
        feature = "php",
        feature = "python",
        feature = "ruby",
        feature = "rust",
        feature = "tcl",
    ))]
    macro_rules! assert_variants_is_string {
        ($parser:expr, $lang:ident, $code:ident, [$($variant:ident),+ $(,)?]) => {
            $(
                assert_variant_is_string(
                    $parser,
                    $lang::$variant as u16,
                    $code::is_string,
                    stringify!($lang),
                    stringify!($variant),
                );
            )+
        };
    }

    // Parse `$src` with `$parser_ty` and assert the generic `"string"`
    // filter yields zero matches. Used by the negative test, which
    // walks every language consolidated under `impl_simple_is_string!`
    // with identical per-language shape (parse → count → assert_eq! 0).
    #[cfg(all(
        feature = "bash",
        feature = "c-family-helpers",
        feature = "cpp",
        feature = "csharp",
        feature = "elixir",
        feature = "go",
        feature = "groovy",
        feature = "irules",
        feature = "java",
        feature = "kotlin",
        feature = "lua",
        feature = "perl",
        feature = "php",
        feature = "python",
        feature = "ruby",
        feature = "rust",
        feature = "tcl",
    ))]
    macro_rules! assert_no_string_matches {
        ($parser_ty:ident, $path:expr, $src:expr, $lang:literal $(,)?) => {{
            let parser = $parser_ty::new($src.to_vec(), $path, None);
            assert_eq!(count_with_parser(&parser), 0, $lang);
        }};
    }

    #[cfg(all(
        feature = "bash",
        feature = "c-family-helpers",
        feature = "cpp",
        feature = "csharp",
        feature = "elixir",
        feature = "go",
        feature = "groovy",
        feature = "irules",
        feature = "java",
        feature = "kotlin",
        feature = "lua",
        feature = "perl",
        feature = "php",
        feature = "python",
        feature = "ruby",
        feature = "rust",
        feature = "tcl",
    ))]
    #[test]
    fn simple_is_string_macro_recognises_each_language() {
        use crate::langs::{
            CcommentParser, CppParser, CsharpParser, ElixirParser, GoParser, GroovyParser,
            IrulesParser, JavaParser, KotlinParser, LuaParser, PerlParser, PreprocParser,
            PythonParser, RubyParser, RustParser, TclParser,
        };

        let path = PathBuf::from("test");

        // ---- Preproc (2 variants): StringLiteral, RawStringLiteral ----
        let src = b"#include \"foo.h\"\nR\"(raw)\"\n".to_vec();
        let parser = PreprocParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Preproc,
            PreprocCode,
            [StringLiteral, RawStringLiteral]
        );

        // ---- Ccomment (2 variants): StringLiteral, RawStringLiteral ----
        // The Ccomment grammar is a stub that only emits Comment /
        // StringLiteral / RawStringLiteral; feed it both forms.
        let src = b"\"hello\"\nR\"(raw)\"\n".to_vec();
        let parser = CcommentParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Ccomment,
            CcommentCode,
            [StringLiteral, RawStringLiteral]
        );

        // ---- Cpp (3 variants): StringLiteral, ConcatenatedString, RawStringLiteral ----
        // C++ string concatenation (`"a" "b"`) produces a
        // `concatenated_string` node wrapping the literals.
        let src =
            b"const char* a = \"hi\";\nconst char* b = \"a\" \"b\";\nconst char* c = R\"(raw)\";\n"
                .to_vec();
        let parser = CppParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Cpp,
            CppCode,
            [StringLiteral, ConcatenatedString, RawStringLiteral]
        );

        // ---- Python (2 variants): String, ConcatenatedString ----
        // Python concatenates adjacent string literals into a
        // `concatenated_string` node.
        let src = b"a = \"hi\"\nb = \"a\" \"b\"\n".to_vec();
        let parser = PythonParser::new(src, &path, None);
        assert_variants_is_string!(&parser, Python, PythonCode, [String, ConcatenatedString]);

        // ---- Java (2 variants): StringLiteral, MultilineStringLiteral ----
        // `Java::MultilineStringLiteral` maps to `_multiline_string_literal`
        // (leading-underscore hidden supertype) and does NOT surface as a
        // concrete kind_id in observed parses. Triple-quoted text blocks
        // instead produce regular `StringLiteral` nodes. The variant is
        // intentionally listed in the macro so a future grammar revision
        // that promotes it can't bypass `is_string`; presence is asserted
        // below to flag drift.
        let src = b"class C { String a = \"hi\"; String b = \"\"\"\nmulti\n\"\"\"; }\n".to_vec();
        let parser = JavaParser::new(src, &path, None);
        assert_variants_is_string!(&parser, Java, JavaCode, [StringLiteral]);
        assert!(
            !ast_has_kind_id(&parser, Java::MultilineStringLiteral as u16),
            "Java::MultilineStringLiteral is documented as the hidden _multiline_string_literal supertype; if it now appears in parses, replace this with a positive variant assertion",
        );

        // ---- Csharp (4 variants): StringLiteral, VerbatimStringLiteral,
        // RawStringLiteral, InterpolatedStringExpression ----
        // Verbatim: `@"..."`; raw: triple-double-quote; interpolated: `$"..."`.
        let src = b"class C { string a = \"hi\"; string b = @\"verb\"; string c = \"\"\"raw\"\"\"; string d = $\"int{1}\"; }\n".to_vec();
        let parser = CsharpParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Csharp,
            CsharpCode,
            [
                StringLiteral,
                VerbatimStringLiteral,
                RawStringLiteral,
                InterpolatedStringExpression,
            ]
        );

        // ---- Rust (2 variants): StringLiteral, RawStringLiteral ----
        let src = b"fn main() { let a = \"hi\"; let b = r\"raw\"; }\n".to_vec();
        let parser = RustParser::new(src, &path, None);
        assert_variants_is_string!(&parser, Rust, RustCode, [StringLiteral, RawStringLiteral]);

        // ---- Go (2 variants): InterpretedStringLiteral, RawStringLiteral ----
        // Backtick-delimited string is the raw form.
        let src = b"package main\nfunc main() { _ = \"hi\"; _ = `raw` }\n".to_vec();
        let parser = GoParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Go,
            GoCode,
            [InterpretedStringLiteral, RawStringLiteral]
        );

        // ---- Kotlin (2 variants): StringLiteral, MultilineStringLiteral ----
        let src = b"fun main() { val a = \"hi\"; val b = \"\"\"multi\"\"\" }\n".to_vec();
        let parser = KotlinParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Kotlin,
            KotlinCode,
            [StringLiteral, MultilineStringLiteral]
        );

        // ---- Lua (1 variant): String ----
        let src = b"local a = \"hi\"\nlocal b = [[long]]\n".to_vec();
        let parser = LuaParser::new(src, &path, None);
        assert_variants_is_string!(&parser, Lua, LuaCode, [String]);

        // ---- Perl (7 variants): StringSingleQuoted, StringDoubleQuoted,
        // StringQQuoted, StringQqQuoted, BacktickQuoted, CommandQxQuoted,
        // HeredocBodyStatement ----
        let src = b"my $a = 'single';\nmy $b = \"double\";\nmy $c = q(qquoted);\nmy $d = qq(qqquoted);\nmy $e = `cmd`;\nmy $f = qx(qxcmd);\nmy $g = <<EOT;\nbody\nEOT\n".to_vec();
        let parser = PerlParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Perl,
            PerlCode,
            [
                StringSingleQuoted,
                StringDoubleQuoted,
                StringQQuoted,
                StringQqQuoted,
                BacktickQuoted,
                CommandQxQuoted,
                HeredocBodyStatement,
            ]
        );

        // ---- Bash (5 variants): String, RawString, AnsiCString,
        // TranslatedString, HeredocBody2 ----
        // TranslatedString surfaces as a wrapper node only in
        // assignment-style contexts (see `bash_is_string_matches_translated_string`).
        let src = b"a=\"d\"\nb='r'\nc=$'ansi'\nd=$\"t\"\ncat <<EOF\nbody\nEOF\n".to_vec();
        let parser = crate::langs::BashParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Bash,
            BashCode,
            [
                String,
                RawString,
                AnsiCString,
                TranslatedString,
                HeredocBody2
            ]
        );

        // ---- Tcl (3 variants): QuotedWord, BracedWord, BracedWordSimple ----
        // Tcl uses two `braced_word` rules: the named rule `braced_word`
        // (an argument-position braced expression that admits commands
        // and substitutions, e.g. `proc ... { body }`'s `body` arg) and
        // `braced_word_simple` (a plain inert braced word like
        // `{braced}` in `set v {braced}`). The fixture below mixes a
        // `proc ... {body}` (BracedWord), a simple `set` assignment
        // (BracedWordSimple), and a `set` to a quoted literal
        // (QuotedWord).
        let src = b"set a \"quoted\"\nset b {braced}\nproc p {x y} { return $x }\n".to_vec();
        let parser = TclParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Tcl,
            TclCode,
            [QuotedWord, BracedWordSimple, BracedWord]
        );

        // ---- iRules (3 variants): QuotedWord, BracedWord, BracedWordSimple ----
        // Same Tcl-family braced-word split as Tcl above: the `proc`
        // body is a named `braced_word` (BracedWord), the simple `set`
        // assignment is a `braced_word_simple` (BracedWordSimple), and
        // the quoted literal is a `QuotedWord`.
        let src = b"set a \"quoted\"\nset b {braced}\nproc p {x y} { return $x }\n".to_vec();
        let parser = IrulesParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Irules,
            IrulesCode,
            [QuotedWord, BracedWordSimple, BracedWord]
        );

        // ---- Php (6 variants): String, String3, EncapsedString,
        // Heredoc, Nowdoc, ShellCommandExpression ----
        // The `: string` return type in the fixture produces a
        // `String2` — the `string` *type* keyword, which is NOT a
        // variant here: #1474 withdrew it, and
        // `php_is_string_excludes_type_keyword_alias_1474` pins its
        // absence. String3 is the hidden `_string` supertype (kind_id
        // => "_string" — name starts with `_`), which tree-sitter does
        // NOT emit as a concrete node — see its empirical absence
        // asserted below.
        let src = b"<?php function f(): string { $a = 'single'; $b = \"double\"; $c = <<<EOT\nbody\nEOT;\n$d = <<<'EOT'\nnow\nEOT;\n$e = `ls`; return $a; }\n".to_vec();
        let parser = PhpParser::new(src, &path, None);
        assert_variants_is_string!(&parser, Php, PhpCode, [String]);
        // `Php::String3` is the hidden `_string` supertype — never
        // surfaces as a concrete kind_id in observed parses; the
        // checker still lists it so future grammar revisions that
        // promote it cannot silently bypass `is_string`. Verified
        // unreachable empirically (assertion below proves the
        // fixture does not produce it; the variant is intentionally
        // unverifiable through a positive test until then).
        assert!(
            !ast_has_kind_id(&parser, Php::String3 as u16),
            "Php::String3 is documented as the hidden _string supertype; if it now appears in parses, add a positive variant assertion",
        );
        assert_variants_is_string!(
            &parser,
            Php,
            PhpCode,
            [EncapsedString, Heredoc, Nowdoc, ShellCommandExpression]
        );

        // ---- Elixir (3 variants): String, Charlist, Sigil ----
        // Charlists use single quotes; sigils use `~s(...)` etc.
        let src = b"a = \"hi\"\nb = 'charlist'\nc = ~s(sigil)\n".to_vec();
        let parser = ElixirParser::new(src, &path, None);
        assert_variants_is_string!(&parser, Elixir, ElixirCode, [String, Charlist, Sigil]);

        // ---- Ruby (11 variants): String, ChainedString, BareString,
        // Subshell, Regex, HeredocBody, DelimitedSymbol, SimpleSymbol,
        // StringArray, SymbolArray, Character ----
        // ChainedString: two adjacent string literals (`"a" "b"`).
        // BareString: an unquoted element inside `%w[...]` / inside
        // a string array context. `%w[bare1 bare2]` emits a
        // StringArray whose children are BareString nodes — both
        // surface in the parse. Use a script that produces every
        // form so the per-variant assertion can hit each.
        let src = b"a = \"hi\"\nb = \"x\" \"y\"\nc = `cmd`\nd = /re/\ne = <<EOT\nbody\nEOT\nf = :sym\ng = :\"dsym\"\nh = %w[bare1 bare2]\ni = %i[s1 s2]\nj = ?A\n".to_vec();
        let parser = RubyParser::new(src, &path, None);
        assert_variants_is_string!(
            &parser,
            Ruby,
            RubyCode,
            [
                String,
                ChainedString,
                BareString,
                Subshell,
                Regex,
                HeredocBody,
                DelimitedSymbol,
                SimpleSymbol,
                StringArray,
                SymbolArray,
                Character,
            ]
        );

        // ---- Groovy (1 variant): StringLiteral ----
        // The dekobon Groovy grammar consolidates every string shape
        // (single / double / triple-quoted, slashy `/.../`, dollar-
        // slashy `$/.../$`, GString-interpolated) under one
        // `string_literal` rule, so a single variant suffices —
        // unlike Java, character literals are not a separate node.
        let src =
            b"def m() { def a = \"hi\"; def b = \"\"\"multi\"\"\"; def c = /pat/ }\n".to_vec();
        let parser = GroovyParser::new(src, &path, None);
        assert_variants_is_string!(&parser, Groovy, GroovyCode, [StringLiteral]);
    }

    #[cfg(all(
        feature = "bash",
        feature = "c-family-helpers",
        feature = "cpp",
        feature = "csharp",
        feature = "elixir",
        feature = "go",
        feature = "groovy",
        feature = "irules",
        feature = "java",
        feature = "kotlin",
        feature = "lua",
        feature = "perl",
        feature = "php",
        feature = "python",
        feature = "ruby",
        feature = "rust",
        feature = "tcl",
    ))]
    #[test]
    fn simple_is_string_macro_rejects_non_string_nodes() {
        // Pure-identifier source must produce zero string matches.
        // Catches a regression where a macro invocation accidentally
        // included a too-broad variant (e.g. `Identifier`). Covers
        // every language consolidated under `impl_simple_is_string!`.
        use crate::langs::{
            BashParser, CcommentParser, CppParser, CsharpParser, ElixirParser, GoParser,
            GroovyParser, IrulesParser, JavaParser, KotlinParser, LuaParser, PerlParser,
            PreprocParser, PythonParser, RubyParser, RustParser, TclParser,
        };

        let path = PathBuf::from("test");

        // Each fixture is the most minimal identifier-only input that
        // still parses for the target language. Per-language comments
        // below flag the few cases where the input choice is load-
        // bearing (Bash `s=$y` to avoid `string`-kind expansion nodes,
        // Tcl `set x $y` to keep the bareword as `Word`, etc.).
        assert_no_string_matches!(PreprocParser, &path, b"#define FOO 1\n", "Preproc");
        // Ccomment: input that lexes as a single line comment only.
        assert_no_string_matches!(CcommentParser, &path, b"// just a comment\n", "Ccomment");
        assert_no_string_matches!(CppParser, &path, b"int main() { return x; }\n", "Cpp");
        assert_no_string_matches!(PythonParser, &path, b"x = y\n", "Python");
        assert_no_string_matches!(JavaParser, &path, b"class C { int x = y; }\n", "Java");
        // Csharp: identifier-only field initializer.
        assert_no_string_matches!(CsharpParser, &path, b"class C { int x = y; }\n", "Csharp");
        assert_no_string_matches!(RustParser, &path, b"fn main() { let x = y; }\n", "Rust");
        assert_no_string_matches!(
            GoParser,
            &path,
            b"package main\nfunc main() { _ = x }\n",
            "Go"
        );
        assert_no_string_matches!(KotlinParser, &path, b"fun main() { val x = y }\n", "Kotlin");
        // Perl: identifier-only assignment with no quoted forms.
        assert_no_string_matches!(PerlParser, &path, b"my $x = $y;\n", "Perl");
        assert_no_string_matches!(LuaParser, &path, b"local x = y\n", "Lua");
        // Bash: assignment of one variable to another, no literals.
        // `s=$y` produces only Variable/SimpleExpansion nodes; no
        // string-kind node should appear.
        assert_no_string_matches!(BashParser, &path, b"s=$y\n", "Bash");
        // Tcl: `set` of an unquoted identifier word. The unquoted
        // bareword surfaces as `Word`, not any of the three string
        // kinds.
        assert_no_string_matches!(TclParser, &path, b"set x $y\n", "Tcl");
        // iRules: same Tcl-family `set` of an unquoted identifier word —
        // the bareword surfaces as `Word`, not any string kind.
        assert_no_string_matches!(IrulesParser, &path, b"set x $y\n", "Irules");
        // Php: identifier-only assignment.
        assert_no_string_matches!(PhpParser, &path, b"<?php $x = $y;\n", "Php");
        // Elixir: integer assignment, no string/charlist/sigil.
        assert_no_string_matches!(ElixirParser, &path, b"x = 1\n", "Elixir");
        assert_no_string_matches!(RubyParser, &path, b"x = y\n", "Ruby");
        // Groovy: identifier-only method body.
        assert_no_string_matches!(GroovyParser, &path, b"def m() { def x = y }\n", "Groovy");
    }

    #[cfg(feature = "mozjs")]
    #[test]
    fn mozjs_parses_using_declaration() {
        // Drift marker for the JS-base-grammar bump 0.23.1 -> 0.25.0
        // (#407): the `using` / `await using` explicit-resource-
        // management declaration is a 0.25.0 grammar feature. The
        // bundled mozjs parser was previously stale at 0.23.1 while
        // its marker claimed 0.25.0 (the #400 baseline lie), so this
        // node could not appear. Asserting it surfaces pins the regen
        // against reversion — it fails against the pre-#407 parser.
        let src = "function f() {\n  using r = acquire();\n  return r;\n}\n";
        let parser = MozjsParser::new(src.as_bytes().to_vec(), &PathBuf::from("t.js"), None);
        assert!(
            ast_has_kind_id(&parser, Mozjs::UsingDeclaration as u16),
            "expected Mozjs::UsingDeclaration to appear in the parse",
        );
    }

    #[cfg(feature = "go")]
    #[test]
    fn go_rune_literal_is_not_a_string() {
        // #699 verdict: Go `RuneLiteral` is an operand in `get_op_type`
        // and is flattened by the alterator, but it is deliberately
        // excluded from `is_string` — a rune is a character (`int32`),
        // not a string, exactly as Rust/Cpp `CharLiteral` are operand +
        // flattened yet absent from their `is_string`. This pins the
        // exclusion: the fixture must produce a `rune_literal` (proving
        // the surrounding grammar still emits it) while `GoCode::is_string`
        // returns false for it. The flattening half lives in
        // `alterator.rs::go_rune_literal_flattened_but_not_a_string_kind`.
        use crate::langs::GoParser;
        let path = PathBuf::from("test.go");
        let src = b"package main\nfunc main() { r := 'x'; _ = r }\n".to_vec();
        let parser = GoParser::new(src, &path, None);
        let rune_id = Go::RuneLiteral as u16;
        assert!(
            ast_has_kind_id(&parser, rune_id),
            "fixture should produce a rune_literal node",
        );
        assert_eq!(
            count_string_matches_for_kind(&parser, rune_id, GoCode::is_string),
            0,
            "Go rune_literal must not match is_string (a rune is a char, not a string)",
        );
    }

    #[cfg(all(feature = "c", feature = "cpp", feature = "mozcpp", feature = "objc"))]
    #[test]
    fn c_family_char_literal_is_not_a_string() {
        // The C-family half of the #699 verdict `go_rune_literal_is_not_a_string`
        // cites: a `char_literal` is flattened by the alterator and — since
        // #1316 — is a Halstead operand, yet it is deliberately absent from
        // `is_string`, because a character is not a string.
        //
        // #1316 made that citation true rather than aspirational (the operand
        // half had never been implemented), so the exclusion is now a live
        // decision and wants a pin of its own.
        //
        // Each row is two-sided. The `char_literal` count must be zero, and
        // the `string_literal` in the same fixture must be one — otherwise a
        // predicate that had stopped matching anything at all would read as a
        // correctly-rejected character. The `ast_has_kind_id` guard covers the
        // third way to get a zero: a fixture that no longer parses to a
        // `char_literal` at all.
        use crate::langs::{CParser, CppParser, MozcppParser, ObjcParser};

        #[cfg(any(feature = "c", feature = "cpp", feature = "mozcpp", feature = "objc"))]
        fn check<P: ParserTrait, F: Fn(&Node) -> bool + Copy>(
            source: &[u8],
            file: &str,
            (char_literal, string_literal): (u16, u16),
            is_string: F,
            label: &str,
        ) {
            let parser = P::new(source.to_vec(), &PathBuf::from(file), None);
            assert!(
                ast_has_kind_id(&parser, char_literal),
                "{label}: fixture should produce a char_literal node",
            );
            assert_eq!(
                count_string_matches_for_kind(&parser, char_literal, is_string),
                0,
                "{label}: char_literal must not match is_string (a char is not a string)",
            );
            assert_eq!(
                count_string_matches_for_kind(&parser, string_literal, is_string),
                1,
                "{label}: is_string must still match the fixture's string_literal, or the \
                 zero above is a dead predicate rather than a rejected character",
            );
        }

        let src = b"char a = 'x';\nconst char *b = \"s\";\n";
        check::<CParser, _>(
            src,
            "t.c",
            (C::CharLiteral as u16, C::StringLiteral as u16),
            CCode::is_string,
            "c",
        );
        check::<CppParser, _>(
            src,
            "t.cpp",
            (Cpp::CharLiteral as u16, Cpp::StringLiteral as u16),
            CppCode::is_string,
            "cpp",
        );
        check::<MozcppParser, _>(
            src,
            "t.cpp",
            (Mozcpp::CharLiteral as u16, Mozcpp::StringLiteral as u16),
            MozcppCode::is_string,
            "mozcpp",
        );
        check::<ObjcParser, _>(
            src,
            "t.m",
            (Objc::CharLiteral as u16, Objc::StringLiteral as u16),
            ObjcCode::is_string,
            "objc",
        );
    }

    /// Asserts `is_func` / `is_closure` answer the same off a known
    /// chain as off a `Node::parent` climb, for every node of `code`.
    /// Returns the `(functions, closures)` the chain path counted, so a
    /// caller can additionally pin what the predicates actually decided.
    ///
    /// Both predicates fold their ancestor lookups into one boolean, so
    /// a wrong ancestor does not fail — it silently moves a function
    /// into the closure column. The counters keep a fixture honest: both
    /// answers have to occur, for both predicates, or the parity holds
    /// only over nodes the predicates never classify.
    #[cfg(any(
        feature = "javascript",
        feature = "mozjs",
        feature = "ruby",
        feature = "typescript"
    ))]
    fn assert_func_parity<L: crate::traits::LanguageInfo + Checker>(
        label: &str,
        code: &[u8],
    ) -> (usize, usize) {
        let mut funcs = 0;
        let mut closures = 0;
        let visited = for_each_node_with_chain::<L>(code, |node, chain| {
            let known = Ancestors::known(chain);
            let climbing = Ancestors::unknown();
            assert_eq!(
                L::is_func(node, known),
                L::is_func(node, climbing),
                "{label}: is_func disagrees on {} at row {}",
                node.kind(),
                node.start_row()
            );
            assert_eq!(
                L::is_closure(node, known),
                L::is_closure(node, climbing),
                "{label}: is_closure disagrees on {} at row {}",
                node.kind(),
                node.start_row()
            );
            funcs += usize::from(L::is_func(node, known));
            closures += usize::from(L::is_closure(node, known));
        });
        assert!(visited > 20, "{label}: fixture is too small to prove much");
        assert!(
            funcs > 0 && closures > 0,
            "{label}: fixture must contain both a named function and a closure, \
             else parity holds over nodes the walk never classifies: \
             {funcs} functions, {closures} closures"
        );
        (funcs, closures)
    }

    /// The JS-family `is_func` / `is_closure` must classify a node the
    /// same whether they read the walker's ancestor chain or climb with
    /// `Node::parent` (#1088).
    ///
    /// These are the predicates with the longest ancestor lookup: an
    /// upward walk, its `is_else_if` filter, and a `has_sibling`
    /// adjacency check.
    #[cfg(all(feature = "javascript", feature = "mozjs", feature = "typescript"))]
    #[test]
    fn js_func_and_closure_agree_between_known_and_climbing() {
        // One of each shape the walk distinguishes: an arrow bound to a
        // name, an arrow passed positionally, a `function` expression
        // named by an object `pair`, and a nested arrow whose only
        // enclosing frame is a `statement_block`.
        //
        // The returned object literal is load-bearing. It is the one
        // shape here where the trailing `has_sibling` adjacency check
        // decides the answer: the upward walk stops at the `return`
        // having counted no binding, so whether the arrow is a *named*
        // function rests entirely on the `property_identifier` beside
        // it. Everywhere else a binding is found first and the adjacency
        // check is never reached, which would leave a wrong parent
        // lookup invisible.
        let code = concat!(
            "const f = a => { g(() => 1); a => 2; };\n",
            "const o = { m: function () { return 1; } };\n",
            "function h() { return { k: (a) => a + 1 }; }\n",
            "function i() { return function () { return 2; }; }\n",
        )
        .as_bytes();
        // `is_js_func_and_closure_checker!` expands separately against
        // each grammar's own `kind_id` enum, so the four impls are four
        // distinct variant lists rather than one shared body. Checking
        // only JavaScript would leave a drifted id in any of the other
        // three uncovered.
        assert_func_parity::<crate::langs::JavascriptCode>("javascript", code);
        assert_func_parity::<crate::langs::MozjsCode>("mozjs", code);
        assert_func_parity::<crate::langs::TypescriptCode>("typescript", code);
        assert_func_parity::<crate::langs::TsxCode>("tsx", code);
    }

    /// Ruby's `is_closure` is the one non-JS predicate that needs an
    /// ancestor: a `{ … }` block is a closure unless its parent is the
    /// `Lambda` that already counted it (#465), so it reads the chain
    /// for the same reason the JS family does (#1088).
    ///
    /// The stabby lambda is the discriminating shape. Its body block is
    /// the only node here whose answer depends on the parent lookup —
    /// every other block's parent is a `Call`, which the arm accepts
    /// whether the lookup is right or wrong.
    #[cfg(feature = "ruby")]
    #[test]
    fn ruby_block_closure_agrees_between_known_and_climbing() {
        let code = concat!(
            "def m\n",
            "  f = ->(z) { z + 1 }\n",
            "  [1].each { |x| x }\n",
            "  lambda { 2 }\n",
            "end\n",
        )
        .as_bytes();
        let (funcs, closures) = assert_func_parity::<crate::langs::RubyCode>("ruby", code);
        assert_eq!(funcs, 1, "only `def m` is a named function");
        assert_eq!(
            closures, 3,
            "the stabby lambda, the `each` block and the `lambda` block \
             are three closures — the stabby lambda's own body block must \
             not add a fourth, which is the answer the parent lookup decides"
        );
    }

    /// The verdicts the byte-less and the source-aware `is_string` give
    /// the `braced_word`-kind nodes of `code`, as their source texts,
    /// plus how many string-kind nodes of *any other* kind were seen.
    ///
    /// Returns `(kept_literal, withdrawn, other_strings)`. Texts rather
    /// than counts because a count is polarity-blind: `(1, 1)` holds
    /// just as well when the predicate calls the `proc` body a literal
    /// and `{a b}` a script, which is the claim inverted. `other_strings`
    /// anchors the control the third fixture line supplies — without it
    /// the every-other-kind assertion below can be deleted from the
    /// fixture and the kind-gate perturbation stops failing anything.
    ///
    /// Also asserts the two spellings agree off a known chain and off a
    /// `Node::parent` climb: the walk now threads a chain, and the
    /// answer must not depend on which it gets.
    #[cfg(any(feature = "irules", feature = "tcl"))]
    fn braced_word_string_verdicts<L: crate::traits::LanguageInfo + Checker>(
        label: &str,
        code: &[u8],
        script_kind: u16,
    ) -> (Vec<String>, Vec<String>, usize) {
        let (mut kept, mut withdrawn, mut other) = (Vec::new(), Vec::new(), 0);
        for_each_node_with_chain::<L>(code, |node, chain| {
            let known = L::is_string_with_code(node, code, Ancestors::known(chain));
            assert_eq!(
                known,
                L::is_string_with_code(node, code, Ancestors::unknown()),
                "{label}: is_string_with_code disagrees between a known chain and a \
                 parent climb on {} at row {}",
                node.kind(),
                node.start_row()
            );
            if node.kind_id() != script_kind {
                // Every other kind must answer identically, or the
                // override has widened past the one ambiguous kind.
                assert_eq!(
                    known,
                    L::is_string(node),
                    "{label}: the override moved a {} node at row {}",
                    node.kind(),
                    node.start_row()
                );
                other += usize::from(known);
                return;
            }
            assert!(
                L::is_string(node),
                "{label}: the kind table must still list the script kind, or this \
                 test is measuring its absence rather than the override"
            );
            let text = String::from_utf8_lossy(&code[node.start_byte()..node.end_byte()]);
            if known {
                kept.push(text.into_owned());
            } else {
                withdrawn.push(text.into_owned());
            }
        });
        (kept, withdrawn, other)
    }

    /// A Tcl-family script body is a string literal to
    /// [`Checker::is_string`] and not to
    /// [`Checker::is_string_with_code`] (#1381).
    ///
    /// This is the test the end-to-end `bca find` ones cannot be: they
    /// observe the *result*, which is equally consistent with the
    /// override never firing and the byte-less default happening to be
    /// right. Asserting the disagreement directly is what makes a future
    /// call site that reaches for the byte-less spelling a wrong answer
    /// rather than an indistinguishable one
    /// (`.claude/rules/grammar-dispatch.md` §7).
    ///
    /// Each fixture holds a braced value, a script body reached through
    /// a *modelled* construct, a script body reached through the
    /// **command-name** list (`eval`), and a non-braced string literal —
    /// four rows that no constant and no inverted polarity satisfies.
    /// The two script routes are structurally independent
    /// (grammar-dispatch §11): `is_value_braced_word` answers "script"
    /// either because the word fills a modelled slot or because the
    /// enclosing command is in `SCRIPT_TAKING_COMMANDS`, and a fixture
    /// exercising only the first leaves the second dead.
    #[test]
    #[cfg(any(feature = "tcl", feature = "irules"))]
    fn tcl_script_body_is_a_string_only_to_the_byteless_spelling() {
        let mut ran = 0;
        #[cfg(feature = "tcl")]
        {
            ran += 1;
            // `{x}` is the parameter list, which the grammar spells
            // `arguments` rather than a braced word, so it is neither.
            // `set s "q"` is the control for the kind gate: a
            // `quoted_word` under a *modelled* command is a node the
            // role predicate answers "not a value" for, so an override
            // that dropped its `kind_id` test would unstring it — which
            // the `other` count below is what notices.
            let (kept, withdrawn, other) = braced_word_string_verdicts::<crate::langs::TclCode>(
                "tcl",
                b"proc p {x} { puts $x }\nlappend l {a b}\neval {puts hi}\nset s \"q\"\n",
                Tcl::BracedWord as u16,
            );
            assert_eq!(kept, ["{a b}"], "tcl: only the braced value is a literal");
            assert_eq!(
                withdrawn,
                ["{ puts $x }", "{puts hi}"],
                "tcl: the proc body (a modelled slot) and the `eval` argument \
                 (a named script-taking command) are both scripts"
            );
            assert_eq!(other, 1, "tcl: the `set` quoted word stays a string");
        }
        #[cfg(feature = "irules")]
        {
            ran += 1;
            // The `"hi"` inside the handler body is this dialect's
            // `other` row, so the control sits inside the construct
            // under test rather than beside it.
            let (kept, withdrawn, other) = braced_word_string_verdicts::<crate::langs::IrulesCode>(
                "irules",
                b"when HTTP_REQUEST { log local0. \"hi\" }\nlappend l {x y}\neval {puts hi}\n",
                Irules::BracedWord as u16,
            );
            assert_eq!(
                kept,
                ["{x y}"],
                "irules: only the braced value is a literal"
            );
            assert_eq!(
                withdrawn,
                ["{ log local0. \"hi\" }", "{puts hi}"],
                "irules: the `when` handler body and the `eval` argument are scripts"
            );
            assert_eq!(other, 1, "irules: the quoted word stays a string");
        }
        assert!(
            ran > 0,
            "neither tcl nor irules is enabled; this test asserted nothing"
        );
    }

    /// A Perl bareword is a call to [`Checker::is_call`] and, as a
    /// wrapper's callee or an auto-quoted key, not to
    /// [`Checker::is_call_with_code`] (#1580). The end-to-end
    /// `find --type call` tests observe the result, which is equally
    /// consistent with the override never firing; this asserts the
    /// disagreement, so a call site reaching for the byte-less spelling
    /// is wrong rather than indistinguishable (grammar-dispatch §7).
    ///
    /// Three structurally independent routes withdraw a bareword — key
    /// position, the spaced wrapper's callee, the bracketed wrapper's
    /// callee — and two keep one: a bareword argument and a lone
    /// bareword (grammar-dispatch §11). The two wrappers are the control
    /// for the kind gate: both spellings must still call them.
    #[test]
    #[cfg(feature = "perl")]
    fn perl_bareword_is_a_call_only_to_the_byteless_spelling() {
        let code = b"my %h = (name => 1);\nprint shift;\nfoo(1);\nfoo;\n";
        let (mut kept, mut withdrawn, mut other) = (Vec::new(), Vec::new(), 0);
        for_each_node_with_chain::<crate::langs::PerlCode>(code, |node, chain| {
            let known = PerlCode::is_call_with_code(node, code, Ancestors::known(chain));
            assert_eq!(
                known,
                PerlCode::is_call_with_code(node, code, Ancestors::unknown()),
                "known chain and parent climb disagree on {} at row {}",
                node.kind(),
                node.start_row()
            );
            if node.kind_id() != Perl::CallExpressionWithBareword as u16 {
                assert_eq!(
                    known,
                    PerlCode::is_call(node),
                    "the override moved a {} node at row {}",
                    node.kind(),
                    node.start_row()
                );
                other += usize::from(known);
                return;
            }
            assert!(
                PerlCode::is_call(node),
                "the kind table must still list the bareword kind, or this test is \
                 measuring its absence rather than the override"
            );
            let text =
                String::from_utf8_lossy(&code[node.start_byte()..node.end_byte()]).into_owned();
            if known {
                kept.push(text);
            } else {
                withdrawn.push(text);
            }
        });
        assert_eq!(
            kept,
            ["shift", "foo"],
            "a bareword argument and a lone bareword are calls"
        );
        assert_eq!(
            withdrawn,
            ["name", "print", "foo"],
            "an auto-quoted key and the callee inside either wrapper are not"
        );
        assert_eq!(other, 2, "the two wrappers stay calls to both spellings");
    }

    /// The value slots of the three constructs whose argument lists hold
    /// both roles stay string literals, though `is_value_braced_word`
    /// classifies each construct whole as script-taking
    /// (`Getter::is_braced_literal_slot`).
    ///
    /// Every fixture line carries a literal *and* a script in the same
    /// construct, so each rule is pinned from both sides: a rule that
    /// withdrew the whole construct fails on the literal, and one that
    /// rescued it whole fails on the script. The two switch lines are the
    /// guards' own rows — an arm whose *pattern* is spelled `proc`, `on`
    /// or `trap` parses as that construct around arm bodies, which must
    /// stay scripts. They need a line each: `proc {…} on {…}` parses as
    /// one `procedure`, so only the second line puts an `on` *command*
    /// in front of the owner guard.
    ///
    /// The Tcl `trap` line stands alone because the walk helper refuses a
    /// tree with a parse error, and `try {…} trap …` on one line reaches
    /// the same generic `trap` command only through error recovery (a
    /// missing terminator the grammar inserts after the `try` body).
    #[test]
    #[cfg(any(feature = "tcl", feature = "irules"))]
    fn tcl_family_value_slots_of_script_takers_stay_strings() {
        let mut ran = 0;
        #[cfg(feature = "tcl")]
        {
            ran += 1;
            let (kept, withdrawn, _) = braced_word_string_verdicts::<crate::langs::TclCode>(
                "tcl",
                b"proc {my proc} {x} { puts $x }\n\
                  namespace export {a b}\n\
                  namespace eval ns { puts hi }\n\
                  namespace ensemble create -map {add ::a}\n\
                  trap {POSIX ENOENT} {msg} { puts $msg }\n\
                  switch $k { proc {puts p} on {puts o} }\n\
                  switch $k { on {puts o} trap {puts t} }\n",
                Tcl::BracedWord as u16,
            );
            assert_eq!(
                kept,
                ["{my proc}", "{a b}", "{add ::a}", "{POSIX ENOENT}", "{msg}"],
                "tcl: a proc name, a non-`eval` namespace argument, and a \
                 `trap` pattern and variable list are literals"
            );
            assert_eq!(
                withdrawn,
                [
                    "{ puts $x }",
                    "{ puts hi }",
                    "{ puts $msg }",
                    "{ proc {puts p} on {puts o} }",
                    "{puts p}",
                    "{puts o}",
                    "{ on {puts o} trap {puts t} }",
                    "{puts o}",
                    "{puts t}",
                ],
                "tcl: the proc body, the `namespace eval` body, the `trap` \
                 script, and the switch arm list and its bodies are scripts"
            );
        }
        #[cfg(feature = "irules")]
        {
            ran += 1;
            let (kept, withdrawn, other) = braced_word_string_verdicts::<crate::langs::IrulesCode>(
                "irules",
                b"proc {my proc} {x} { log local0. $x }\n\
                  namespace export {a b}\n\
                  when HTTP_REQUEST { log local0. \"hi\" }\n\
                  on error {m2} {drop}\n",
                Irules::BracedWord as u16,
            );
            assert_eq!(
                kept,
                ["{my proc}", "{a b}", "{m2}"],
                "irules: a proc name, a `namespace` value and an `on` variable \
                 list are literals"
            );
            assert_eq!(
                withdrawn,
                ["{ log local0. $x }", "{ log local0. \"hi\" }", "{drop}"],
                "irules: the proc and `when` bodies and the `on` handler are scripts"
            );
            assert_eq!(other, 1, "irules: the quoted word stays a string");
        }
        assert!(
            ran > 0,
            "neither tcl nor irules is enabled; this test asserted nothing"
        );
    }

    /// A command whose signature mixes the two roles keeps its *value*
    /// arguments string literals, one rule per row of
    /// `lang_helpers::tcl_family::SCRIPT_TAKING_COMMANDS` (#1381 review).
    ///
    /// `is_value_braced_word` answers by command name alone, so before
    /// the slot table every braced argument of a script-taking command
    /// was a script: `time {set x 1} {5}` lost the count, `after {100}`
    /// the delay, `switch {foo} {…}` the subject, and
    /// `namespace eval {foo bar} {…}` the namespace name.
    ///
    /// Every line carries a value *and* a script, so each rule is pinned
    /// from both sides: a rule that withdrew the whole construct fails on
    /// the value, and one that rescued it whole fails on the script.
    /// Flipping which index a rule calls the value moves a row from one
    /// list to the other, so neither list can be satisfied by a constant
    /// or by an inverted polarity.
    ///
    /// The two `uplevel` lines are the level gate's own pair, the last
    /// Tcl line is the `is_switch_arm` guard's — an arm whose *pattern*
    /// is spelled `after` builds an `after` command around what are
    /// really arm bodies, whose first would otherwise become the delay —
    /// and the `-matchvar` line pins the subject scan's option arity and
    /// its `--` terminator. iRules has no `switch` rows because its
    /// grammar models the construct and rejects a braced subject
    /// outright (the word lands under an `ERROR`, where the walk helper
    /// refuses to go).
    #[test]
    #[cfg(any(feature = "tcl", feature = "irules"))]
    fn tcl_family_mixed_signature_value_arguments_stay_strings() {
        let mut ran = 0;
        #[cfg(feature = "tcl")]
        {
            ran += 1;
            let (kept, withdrawn, _) = braced_word_string_verdicts::<crate::langs::TclCode>(
                "tcl",
                b"after {100} {puts a}\n\
                  time {puts b} {5}\n\
                  uplevel {1} {puts c}\n\
                  uplevel {puts d}\n\
                  switch {sub} {p {puts e}}\n\
                  switch -exact -- {sub2} {q {puts f}}\n\
                  switch -- {sub4} {t} {puts m}\n\
                  switch -nocase -matchvar {mv} -regexp {sub3} {r {puts g}}\n\
                  switch $v {^s} {puts h}\n\
                  namespace eval {my ns} {puts i}\n\
                  namespace inscope {my ns2} {puts j}\n\
                  namespace code {puts k}\n\
                  switch $w { after {puts x} time {puts y} }\n",
                Tcl::BracedWord as u16,
            );
            assert_eq!(
                kept,
                [
                    "{100}", "{5}", "{1}", "{sub}", "{sub2}", "{sub4}", "{t}", "{mv}", "{sub3}",
                    "{^s}", "{my ns}", "{my ns2}",
                ],
                "tcl: an `after` delay, a `time` count, an `uplevel` level, a \
                 `switch` subject reached past three option spellings, a \
                 `-matchvar` operand, two flat-form patterns and the two \
                 `namespace` names are literals"
            );
            assert_eq!(
                withdrawn,
                [
                    "{puts a}",
                    "{puts b}",
                    "{puts c}",
                    "{puts d}",
                    "{p {puts e}}",
                    "{puts e}",
                    "{q {puts f}}",
                    "{puts f}",
                    "{puts m}",
                    "{r {puts g}}",
                    "{puts g}",
                    "{puts h}",
                    "{puts i}",
                    "{puts j}",
                    "{puts k}",
                    "{ after {puts x} time {puts y} }",
                    "{puts x}",
                    "{puts y}",
                ],
                "tcl: every arm list, arm body and evaluated argument is a \
                 script — including the sole argument of the second \
                 `uplevel`, which does not read as a level, and the two \
                 bodies of an arm whose pattern is spelled `after`"
            );
        }
        #[cfg(feature = "irules")]
        {
            ran += 1;
            let (kept, withdrawn, _) = braced_word_string_verdicts::<crate::langs::IrulesCode>(
                "irules",
                b"after {100} {log a}\n\
                  time {log b} {5}\n\
                  uplevel {1} {log c}\n\
                  uplevel {log d}\n\
                  namespace eval {my ns} {log e}\n\
                  namespace inscope {my ns2} {log f}\n\
                  namespace code {log g}\n",
                Irules::BracedWord as u16,
            );
            assert_eq!(
                kept,
                ["{100}", "{5}", "{1}", "{my ns}", "{my ns2}"],
                "irules: the same delay, count, level and namespace names are \
                 literals — the two dialects share one predicate and must not \
                 drift"
            );
            assert_eq!(
                withdrawn,
                [
                    "{log a}", "{log b}", "{log c}", "{log d}", "{log e}", "{log f}", "{log g}",
                ],
                "irules: every evaluated argument is a script"
            );
        }
        assert!(
            ran > 0,
            "neither tcl nor irules is enabled; this test asserted nothing"
        );
    }

    /// A mixed-signature command whose *layout* cannot be resolved keeps
    /// the construct-wide script answer, and a resolvable one beside it
    /// still yields its value.
    ///
    /// Three shapes reach a fallback in
    /// `lang_helpers::tcl_family` that no other fixture does:
    ///
    /// - `namespace $sub {…}` — the subcommand is a substitution, not a
    ///   plain word, so no `ScriptSlots` row can be looked up at all.
    /// - `switch -matchvar {m} -indexvar {i}` — both options consume the
    ///   word after them, so the scan runs off the end without finding a
    ///   subject. *Two* pairs, because one leaves the scan at a list of
    ///   two whose second word an invented subject 0 would call the arm
    ///   list — the same answer, so a single pair proves nothing.
    /// - `switch bar {p} {puts c} {q} {puts d}` — the subject is a bare
    ///   `simple_word`, the one spelling neither the braced nor the `$v`
    ///   fixtures above exercise. *Two* arms, for the same reason: with
    ///   one, the trailing argument is the braced arm list either way, so
    ///   the subject's index does not change any answer.
    ///
    /// Pairing each with a resolvable line is what makes the claim
    /// two-sided: the exact vectors below say `{my ns}`, `{p}` and `{q}`
    /// are values *and* that `{puts b}`, `{m}` and `{i}` are not, so a
    /// fallback answering "value" instead moves rows between the lists.
    ///
    /// The `switch` rows are Tcl-only by construction. The iRules grammar
    /// models `switch`, so a well-formed one is never the generic
    /// `command` this rule reads, and a malformed one lands under an
    /// `ERROR` that `is_braced_literal_slot` rejects before the slot
    /// table — the same asymmetry `SCRIPT_TAKING_COMMANDS` records for
    /// its `for` and `switch` rows.
    #[test]
    #[cfg(any(feature = "tcl", feature = "irules"))]
    fn tcl_family_unresolvable_slot_layouts_keep_the_script_answer() {
        let mut ran = 0;
        #[cfg(feature = "tcl")]
        {
            ran += 1;
            let (kept, withdrawn, _) = braced_word_string_verdicts::<crate::langs::TclCode>(
                "tcl",
                b"namespace eval {my ns} {puts a}\n\
                  namespace $sub {puts b}\n\
                  switch -matchvar {m} -indexvar {i}\n\
                  switch bar {p} {puts c} {q} {puts d}\n",
                Tcl::BracedWord as u16,
            );
            assert_eq!(
                kept,
                ["{my ns}", "{p}", "{q}"],
                "tcl: a resolvable `namespace eval` name and both flat-form \
                 patterns behind a bare-word subject are literals"
            );
            assert_eq!(
                withdrawn,
                ["{puts a}", "{puts b}", "{m}", "{i}", "{puts c}", "{puts d}"],
                "tcl: an unresolvable `namespace` subcommand and a `switch` \
                 with no subject keep every braced argument a script"
            );
        }
        #[cfg(feature = "irules")]
        {
            ran += 1;
            let (kept, withdrawn, _) = braced_word_string_verdicts::<crate::langs::IrulesCode>(
                "irules",
                b"namespace eval {my ns} {log a}\n\
                  namespace $sub {log b}\n",
                Irules::BracedWord as u16,
            );
            assert_eq!(
                kept,
                ["{my ns}"],
                "irules: the resolvable `namespace eval` name is a literal"
            );
            assert_eq!(
                withdrawn,
                ["{log a}", "{log b}"],
                "irules: the unresolvable subcommand keeps its argument a script"
            );
        }
        assert!(
            ran > 0,
            "neither tcl nor irules is enabled; this test asserted nothing"
        );
    }
}