big_code_analysis_ast/

getter.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 [`Getter`] accessors — names, space kinds, Halstead classes — and
//! their per-language impls.

use crate::token_role::TokenRole;

use crate::lang_helpers::tcl_family::{
    ArgumentRoles, Dialect, SWITCH_COMMAND, argument_slots_are_readable, fills_script_slot,
    namespace_script_slots, script_slots,
};
use crate::space_kind::SpaceKind;
use crate::traits::Search;
use generic_angles::{GenericAngleKinds, never_misparsed};
use js_family::impl_js_family_get_op_type;

use crate::*;

/// Bounds- and UTF-8-checked text extraction for a node's byte span.
///
/// `code` is `&[u8]`, so slicing it carries no char-boundary
/// precondition; the two guards here cover two unrelated failure modes:
///
/// * `std::str::from_utf8` rejects non-UTF-8 bytes. This one is
///   reachable from ordinary use — `big_code_analysis::Ast::parse` accepts
///   arbitrary bytes, so a node span in a partially-binary source need
///   not be valid UTF-8.
/// * `code.get` bounds-checks the range. This one is reachable only by
///   violating the same-parse precondition documented on `Getter`, i.e.
///   `big_code_analysis::Ast::from_tree_sitter` adopting a tree built from longer
///   source than the `code` passed alongside it.
///
/// Both degrade to `None`. The walker stores a space's name as
/// `Option<String>` (`spaces.rs`), so that records an unnamed space
/// rather than crashing — note this is *not* the same path as a node
/// with no `name` field, which `get_func_space_name` reports as
/// `Some("<anonymous>")` without reaching here. The unguarded sibling
/// slice sites return infallible
/// types feeding metric arithmetic or rendered output; there the only
/// available fallback would be a fabricated empty value that silently
/// corrupts a count, so they rely on the precondition instead. The
/// asymmetry is deliberate: guarding is free here because `Option` is
/// already part of this signature's contract (#1059).
#[inline]
pub(crate) fn node_text<'a>(code: &'a [u8], node: &Node) -> Option<&'a str> {
    code.get(node.start_byte()..node.end_byte())
        .and_then(|bytes| std::str::from_utf8(bytes).ok())
}

macro_rules! get_operator {
    ($language:ident) => {
        #[inline]
        fn get_operator_id_as_str(id: u16) -> &'static str {
            let typ = id.into();
            match typ {
                $language::LPAREN => "()",
                $language::LBRACK => "[]",
                $language::LBRACE => "{}",
                _ => typ.into(),
            }
        }
    };
}

/// The default space name: the node's `name` field, else `<anonymous>`.
///
/// A free function as well as the trait default so a language that needs
/// to name a *few* kinds specially can delegate the rest rather than
/// restate the rule — `<get>` / `<set>` / `<init>` / `<static-init>` in
/// Kotlin, Java and Groovy all do (#1184). Calling `Self::…` there would
/// recurse.
#[must_use]
pub fn default_func_space_name<'a, 'tree>(
    node: &Node<'tree>,
    code: &'a [u8],
    _ancestors: Ancestors<'tree, '_>,
) -> Option<&'a str> {
    // we're in a function or in a class
    if let Some(name) = node.child_by_field_name("name") {
        node_text(code, &name)
    } else {
        Some("<anonymous>")
    }
}

/// Per-language accessors the space walker and the Halstead
/// operator/operand classification dispatch through.
///
/// # Precondition
///
/// Every method taking a `code: &[u8]` next to a `&Node` slices `code`
/// by that node's byte range. `code` must be the exact buffer `node` was
/// parsed from — `big_code_analysis::Ast::source` for a node obtained from the
/// same `big_code_analysis::Ast`. Pairing a node with any other buffer reads the
/// wrong bytes at best and panics on an out-of-bounds index at worst;
/// the same precondition is documented on [`crate::dump_node`] (#795).
/// `node_text` — reached from the default `get_func_space_name` and
/// from the per-language `get_func_name` overrides — is the sole
/// bounds-checked slice; see its docs for why the rest deliberately
/// are not.
#[doc(hidden)]
/// The kinds one Tcl-family dialect spells the braced-word construct
/// with, for [`Getter::is_subsumed_braced_word`] and
/// [`Getter::braced_word_op_type`]. A struct rather than a row of `u16`
/// parameters because the ids are same-typed and positional: a
/// transposed pair compiles and silently inverts the rule for that
/// dialect.
pub struct BracedWordKinds {
    /// `braced_word_simple`, the literal *value* form.
    pub value: u16,
    /// `braced_word`, the *script* form.
    pub script: u16,
    /// `comment`, the one named child of a script that is not a
    /// command.
    pub comment: u16,
    /// `command`, the generic command node. Its `name` field carries
    /// the leading word [`Getter::is_value_braced_word`] recognises the
    /// construct by, and it is the only parent a *generic* argument
    /// list hangs from. The hidden `_command` supertype (`Command2` in
    /// both enums) is deliberately absent: the parser never emits it
    /// (grammar-dispatch §2).
    pub command: u16,
    /// `word_list`, a command's `arguments` field — and, in both
    /// grammars, also the argument list of the modelled `namespace`
    /// construct, which is why the rule reads the grandparent rather
    /// than stopping here.
    pub word_list: u16,
    /// `simple_word`, the only spelling of a command name this rule
    /// resolves. A computed name (`$cmd {…}`, `[pick] {…}`) parses as
    /// `variable_substitution` / `command_substitution` and is not
    /// statically resolvable at all.
    pub simple_word: u16,
    /// `argument`, one entry of a `proc` parameter list. A modelled
    /// slot that holds a *value* rather than a script: a defaulted
    /// parameter (`proc p {a {b {x y}}}`) spells its default as a
    /// `braced_word`, and a default is never evaluated as code.
    pub argument: u16,
    /// `procedure`, the modelled `proc` construct. Its `name` field is a
    /// `braced_word` when the name holds a space (`proc {my proc} …`),
    /// and a name is a literal although the construct's other braced
    /// slot, the body, is a script
    /// ([`Getter::is_braced_literal_slot`]).
    pub procedure: u16,
    /// `namespace`, the modelled construct — not the keyword token that
    /// shares its name. Its `word_list` holds a subcommand followed by
    /// that subcommand's arguments, and only three subcommands take a
    /// script ([`Getter::is_braced_literal_slot`]).
    pub namespace: u16,
    /// `{`, the brace opener — the *only* node
    /// [`Getter::braced_word_op_type`] may revise. A braced word's
    /// operator children are not the opener alone: `_terminator` is a
    /// hidden rule, so a `;` separating two commands is inlined as a
    /// direct child of the `braced_word` and both dialects classify it
    /// as an operator. Keying the revision on the parent kind alone
    /// therefore swallowed those separators too, taking
    /// `lappend x {puts a ; puts b}` to `n1` 0 / `N1` 0 and
    /// `halstead.effort` — a gated threshold metric — to `0.0`.
    pub open_brace: u16,
}

/// Per-language accessors that *name* and *classify* what a node is:
/// the function or space name, the [`SpaceKind`] a node opens, and the
/// [`TokenRole`] of a leaf.
///
/// Every method has a default that answers "nothing" (`None`,
/// [`SpaceKind::Unknown`], [`TokenRole::Unknown`]); a language
/// overrides only the ones its grammar expresses.
pub trait Getter {
    /// Names the function `node` declares. Defaults to
    /// [`get_func_space_name`](Self::get_func_space_name).
    #[must_use]
    fn get_func_name<'a, 'tree>(
        node: &Node<'tree>,
        code: &'a [u8],
        ancestors: Ancestors<'tree, '_>,
    ) -> Option<&'a str> {
        Self::get_func_space_name(node, code, ancestors)
    }

    /// Names the space `node` opens.
    ///
    /// `ancestors` is the chain the caller descended through. Elixir
    /// needs it: its `def` / `defmodule` heads are ordinary `Call`
    /// nodes, and one inside a `quote` template names no space at all,
    /// which is a question about what encloses the call (#1088).
    #[must_use]
    fn get_func_space_name<'a, 'tree>(
        node: &Node<'tree>,
        code: &'a [u8],
        ancestors: Ancestors<'tree, '_>,
    ) -> Option<&'a str> {
        default_func_space_name(node, code, ancestors)
    }

    /// The kind of space `node` opens, or [`SpaceKind::Unknown`] when it
    /// opens none.
    #[must_use]
    fn get_space_kind(_node: &Node) -> SpaceKind {
        SpaceKind::Unknown
    }

    /// Source-aware variant of [`get_space_kind`](Self::get_space_kind). The default
    /// forwards to the byte-less classifier; languages whose space
    /// kinds are encoded in macro identifier text (Elixir's
    /// `defmodule` / `def` / `defp` / `defmacro` / `defmacrop` Calls)
    /// override this so the walker can attribute the correct
    /// `SpaceKind` to each promoted func space (#275).
    ///
    /// `ancestors` is the chain the caller descended through; Elixir
    /// needs it to see whether the `Call` sits inside a `quote`
    /// template without paying `Node::parent`'s `O(depth)` (#1084).
    #[inline]
    #[must_use]
    fn get_space_kind_with_code<'a>(
        node: &Node<'a>,
        _code: &[u8],
        _ancestors: Ancestors<'a, '_>,
    ) -> SpaceKind {
        Self::get_space_kind(node)
    }

    /// Classifies `node` as a Halstead operator, operand, or neither.
    ///
    /// **Punctuation and delimiters are vocabulary, so they classify as
    /// operators whatever grammatical role they serve.** Halstead
    /// counts the alphabet a program is written in and has no notion of
    /// a decision, so a `<` is the same vocabulary entry whether it
    /// opens a JSX tag, brackets a Lua `<const>` attribute, names a C#
    /// `operator <` overload, disambiguates a Kotlin `super<A>`, or
    /// delimits a Perl `<FH>` readline. Every impl that classifies
    /// punctuation at all bills `(`, `{`, `[`, `,`, `;`, `.` and `:` on
    /// exactly that reading, for each glyph its language spells. Python
    /// and Rust were the last two exceptions, closed by #1486.
    ///
    /// A *decision* metric gates the same token by role, and is
    /// supposed to disagree: ABC's `conditions` counts a `<` only under
    /// a `binary_expression` (#1274, #1275, #1280, #1297), so each of
    /// the five constructs above scores zero there while `bca ops`
    /// still reports the bracket. #1395 settled that the two answers
    /// are both correct rather than a drift to be reconciled;
    /// `tests/parity/abc_halstead_bracket_parity.rs` fails if a later
    /// consistency pass flips either side.
    ///
    /// The one suppression that reading admits is **a literal's own
    /// delimiter**. There the punctuation is not vocabulary in its own
    /// right — the literal it wraps is already counted, so billing its
    /// quotes fabricates an arithmetic or comparison operator the
    /// source never spelled, and the score moves with the author's
    /// choice of delimiter. It takes two shapes, and only the first is
    /// an arm of this method.
    ///
    /// Nine arms suppress a delimiter whose enclosing literal node is
    /// itself the operand, each guarded on the delimiter's *parent*
    /// being that literal: Elixir's sigil (#1256), the regex `/` of the
    /// JS family (#1314), of Perl and of Ruby (#1312), Ruby's subshell
    /// backtick (#1360), Groovy's slashy string and the raw-string `(`
    /// of both C++ grammars (#1314), and Objective-C's `@` before an
    /// `NSString` literal.
    ///
    /// The tenth is [`braced_word_op_type`], which Tcl and iRules route
    /// their `{` through. A braced *value* is a literal to
    /// [`Checker::is_string_with_code`] and to the dump, so its opener
    /// is withdrawn for the same reason — `puts {c d}` and
    /// `puts "c d"` have to agree on the operator column — but the
    /// operands are the words *inside* it rather than the braced word
    /// itself, so the parent test above does not describe it. #1318
    /// established that arm and #1382 widened it to the value slots of
    /// a script-taking command.
    ///
    /// A bracket delimiting *syntax* fits neither shape: no literal
    /// would absorb its contribution, so it stays an operator. A
    /// generic list's `<` / `>` is billed like the other bracket pairs,
    /// once as `<>` on the opener ([`Self::get_operator_spelling`], #1559),
    /// which leaves the comparison `<` and `>` their own entries.
    ///
    /// `ancestors` is the chain the walker descended through. Most
    /// impls read a parent from it to disambiguate a token whose role
    /// depends on what encloses it — the nine literal-delimiter arms
    /// above, plus, among others, Python's `not` / `in` / `is` inside
    /// the compound `not in` / `is not`, Rust's `||` and `!` inside a
    /// binary expression rather than a doc-comment marker, the
    /// namespace identifier in both C++ grammars, Bash's `$name`, the
    /// name inside a Tcl or iRules variable substitution, and Groovy's
    /// `super` in a wildcard bound. Reaching those parents with
    /// [`Node::parent`] instead costs `O(depth)` per node (#1096).
    ///
    /// [`braced_word_op_type`]: Self::braced_word_op_type
    /// [`Checker::is_string_with_code`]: crate::Checker::is_string_with_code
    #[must_use]
    fn get_op_type<'a>(_node: &Node<'a>, _ancestors: Ancestors<'a, '_>) -> TokenRole {
        TokenRole::Unknown
    }

    /// Source-aware variant of [`get_op_type`]. The default forwards
    /// to the byte-less classifier; languages whose Halstead operand
    /// classification depends on token text override this. Kotlin uses
    /// it to recover the variable in a short-form string template
    /// (`"Hi $name"`), which the grammar emits as bare `string_content`
    /// tokens with no structured interpolation node — the distinction
    /// between an interpolated `$name` and a literal `$5` is only
    /// visible in the source bytes (#454).
    ///
    /// [`get_op_type`]: Self::get_op_type
    #[inline]
    #[must_use]
    fn get_op_type_with_code<'a>(
        node: &Node<'a>,
        _code: &[u8],
        ancestors: Ancestors<'a, '_>,
    ) -> TokenRole {
        Self::get_op_type(node, ancestors)
    }

    /// The fixed spelling an operator `node` is keyed under in place of
    /// its kind id, for a token whose kind the grammar shares with a
    /// different operator; `None` keys it by kind. A generic list's `<`
    /// is the case (#1559): it has the comparison's kind but is keyed as
    /// the `"<>"` pair, as `()`, `[]` and `{}` are. Consulted only for a
    /// node [`get_op_type_with_code`](Self::get_op_type_with_code) calls
    /// an operator.
    #[inline]
    #[must_use]
    fn get_operator_spelling<'a>(
        _node: &Node<'a>,
        _ancestors: Ancestors<'a, '_>,
    ) -> Option<&'static str> {
        None
    }

    /// Returns the source-byte slice used to key a Halstead *operand*.
    /// The default keys on the operand node's full byte range. Kotlin
    /// overrides this to narrow a short-interpolation name token
    /// (`$name`) to its leading identifier prefix, because the grammar
    /// glues trailing inter-segment text onto the name token
    /// (`"$a $b"` → `"a "`); keying the raw bytes would record a
    /// distinct `"a "` operand and break parity with the long `${a}`
    /// form (#454).
    #[inline]
    #[must_use]
    fn get_operand_id<'a>(
        node: &Node<'a>,
        code: &'a [u8],
        _ancestors: Ancestors<'a, '_>,
    ) -> &'a [u8] {
        &code[node.start_byte()..node.end_byte()]
    }

    /// Classifies a string-literal `node` as a single Halstead
    /// operand, *unless* it wraps an interpolation child drawn from
    /// `interp_kinds` — in which case the wrapper yields
    /// [`TokenRole::Unknown`] because the inner expressions are
    /// walked and counted separately. Counting the wrapper too would
    /// double-count their contribution to `N2`.
    ///
    /// This declares the per-language interpolation skip once (issue
    /// #420), replacing nine independently-added regression fixes
    /// (#183 / #184 / #191 / #192 / #199 / #277, …). Each language
    /// supplies only its own grammar's interpolation child-kind ids;
    /// the per-call rationale lives at each call site.
    #[must_use]
    fn string_operand_type(node: &Node, interp_kinds: &[u16]) -> TokenRole {
        if node.wraps_any(interp_kinds) {
            TokenRole::Unknown
        } else {
            TokenRole::Operand
        }
    }

    /// Whether `node` is a fragment of a Tcl-family braced word rather
    /// than a token of its own, and so contributes no operand (#1354).
    ///
    /// The two dialects are deliberate clones, so the rule is declared
    /// once and instantiated with each grammar's [`BracedWordKinds`];
    /// the per-call rationale lives at each call site. It answers yes
    /// in two cases:
    ///
    /// - `node` sits directly inside a braced *value*
    ///   (`braced_word_simple`). Tcl evaluates nothing between braces,
    ///   so the value is its whole span and the parts inside it are
    ///   letters — the wrapper carries the one operand. This subsumes
    ///   #1314's narrower guard on the opening `{`, which is one more
    ///   such child.
    /// - `node` is a braced *script* (`braced_word`) holding at least
    ///   one command — a named child other than a comment. Commands are
    ///   what the walk descends into and counts in their own right, so
    ///   billing the block as well would count its whole text a second
    ///   time. A comment is named too, but no arm classifies it, so a
    ///   body holding only comments is billed like an empty one rather
    ///   than like nothing at all.
    ///
    /// A `script` with no command is deliberately not subsumed: the
    /// same kind serves as the value slot of every command the grammar
    /// does not special-case, where `lappend l {}` is an empty list and
    /// the brace pair is its only carrier (grammar-dispatch §6). An
    /// empty or comment-only `proc` body is spelled identically and so
    /// also scores one operand, its whole text. #1318 can now tell the
    /// two apart by the enclosing command, but it bills them alike, so
    /// this arm needs no help from it.
    ///
    /// This is the *byte-less* half of the braced-word rule and it is
    /// not the whole of it. [`braced_word_op_type`] suppresses the `{`
    /// of a script-kind word the enclosing command shows to be a plain
    /// value, and both Tcl-family getters reach it through
    /// [`get_op_type_with_code`], which is what the walk calls
    /// (grammar-dispatch §7). It revises no operand, so this arm's
    /// answers stand unchanged.
    ///
    /// [`braced_word_op_type`]: Self::braced_word_op_type
    /// [`get_op_type_with_code`]: Self::get_op_type_with_code
    #[must_use]
    fn is_subsumed_braced_word<'a>(
        node: &Node<'a>,
        ancestors: Ancestors<'a, '_>,
        kinds: &BracedWordKinds,
    ) -> bool {
        ancestors.parent_has_kind(node, kinds.value)
            || (node.kind_id() == kinds.script
                && node
                    .children()
                    .any(|child| child.is_named() && child.kind_id() != kinds.comment))
    }

    /// The generic `command` node `word` is an argument of, paired with
    /// that command's own ancestry — or `None` when `word` instead
    /// fills a slot of a construct the grammar models, and so is a
    /// script by construction.
    ///
    /// Both dialects hang a *generic* command's arguments off a
    /// `word_list` and a *modelled* construct's slots off that
    /// construct's own node (`procedure`, `if`, `else`, `elseif`,
    /// `while`, `foreach`, `catch`, `try`, `finally`, `argument`, plus
    /// iRules' `when_event`, `for`, `switch_arm`, `on_handler`,
    /// `trap_handler` and the three `dict` loops). Asking the parent
    /// kind therefore answers "is this a generic argument" without
    /// enumerating the modelled parents — a list that would be a
    /// coverage claim to re-derive on every grammar bump
    /// (grammar-dispatch §1), and that a grammar gaining one more
    /// modelled construct would silently invalidate.
    ///
    /// A `word_list` whose own parent is *not* a `command` is
    /// `namespace`'s argument list — the one modelled construct in
    /// either grammar that reaches its body through a `word_list`. The
    /// grandparent test answers `None` there, so `namespace eval ns
    /// {…}` keeps its body a script.
    ///
    /// A braced word directly under a `command` is the command *name*
    /// (`{puts} hi`), not an argument, and also answers `None`: its
    /// caller decides that case before asking, because a name is
    /// always a literal and must not pick up the `switch`-arm rescue
    /// below.
    #[must_use]
    fn generic_argument_command<'tree, 'chain>(
        word: &Node<'tree>,
        ancestors: Ancestors<'tree, 'chain>,
        kinds: &BracedWordKinds,
    ) -> Option<(Node<'tree>, Ancestors<'tree, 'chain>)> {
        let mut chain = ancestors.iter(word);
        let (parent, _) = chain.next()?;
        if parent.kind_id() != kinds.word_list {
            return None;
        }
        let (grandparent, grandparent_ancestors) = chain.next()?;
        (grandparent.kind_id() == kinds.command).then_some((grandparent, grandparent_ancestors))
    }

    /// Whether a braced *script*-kind word is really a plain value —
    /// the literal `{a b}` of `lappend x {a b}` rather than the block
    /// of `eval {…}` (#1318).
    ///
    /// The two roles share one kind, so no kind-scoped arm can separate
    /// them; recognition is out-of-band, by the enclosing command's
    /// leading word (grammar-dispatch §9). A word that fills a modelled
    /// construct's slot is a script; a word passed to a command in
    /// `SCRIPT_TAKING_COMMANDS` is a script; **everything else is a
    /// value**.
    ///
    /// That default is the load-bearing choice, and it is deliberately
    /// the opposite of "keep today's answer for anything unrecognised":
    ///
    /// - The script-taking set is *closed* and small — Tcl defines no
    ///   user-extensible control structures, so a command that
    ///   evaluates a braced argument is either a core command listed
    ///   above or a proc that forwards to `eval` / `uplevel`. The
    ///   value-taking set is *open*: every user proc and every package
    ///   command that takes a list or a pattern is in it. An
    ///   unrecognised name is therefore far likelier to be value-taking.
    /// - Defaulting to script *fabricates* — it reports a `{}` operator
    ///   for a block the source does not contain. Defaulting to value
    ///   can only *omit*: one `N1` occurrence per script handed to an
    ///   unlisted command, and — when no modelled construct in the same
    ///   space opens a block — the `{}` vocabulary entry with them.
    ///   That last case is where the cost shows: a top-level
    ///   `dict for {k v} $d { puts $k }` with nothing around it has no
    ///   operator left, and Halstead's difficulty is a product with
    ///   `n1` in it, so `effort` reads `0.0` rather than a little low.
    ///   Inside a `proc` the `proc` keyword and body brace keep the
    ///   space non-zero, which is the scope `halstead.effort` is gated
    ///   at; and the same `0.0` was already the answer for a file of
    ///   `set x {a b}` lines, so it is Halstead's shape on an
    ///   operator-free space, not a new failure mode. The omission
    ///   stays bounded only because the answer is scoped to the brace:
    ///   see [`braced_word_op_type`] for the measurement that decided
    ///   it, and for why suppressing the *contents* of a value would
    ///   have made it unbounded instead.
    /// - It stops the score moving with the author's choice of
    ///   delimiter, which is what #695, #1312 and #1314 each restored
    ///   elsewhere: `puts {c d}` and `puts "c d"` now agree on the
    ///   operator column.
    ///
    /// The cost is a script passed to an unlisted command — a Tk
    /// `-command {…}` callback, `trace add variable v w {…}`, a
    /// subcommand dispatch (`dict for`, `interp eval`), a user-defined
    /// `with_lock {…}` — losing the `{}` its block deserves. No
    /// structural signal distinguishes those cases: the grammar parses
    /// `{a b}` and `{puts hi}` into the same shape. The code inside
    /// them is still counted, so the loss is one operator occurrence
    /// per such block, plus the vocabulary entry in the case above.
    ///
    /// [`braced_word_op_type`]: Self::braced_word_op_type
    #[must_use]
    fn is_value_braced_word<'a>(
        word: &Node<'a>,
        code: &[u8],
        ancestors: Ancestors<'a, '_>,
        kinds: &BracedWordKinds,
    ) -> bool {
        // Two positions answer "value" without consulting any command
        // name, and both must be asked before the lookup below,
        // because neither hangs off a `command` the way an argument
        // does — and the second must not reach the `switch`-arm
        // rescue.
        //
        // A defaulted `proc` parameter (`proc p {a {b {x y}}}`) is the
        // one modelled slot holding a value: a default is data the
        // interpreter assigns, never a script it evaluates.
        //
        // A braced word directly under a `command` is that command's
        // *name* (`{puts} hi` invokes `puts`), so it is a literal word.
        // Inside a `switch` arm list the name is the arm's *pattern*,
        // and a braced pattern (`switch -regexp $v { {^a.*b$} {…} }`)
        // is idiomatic — rescuing it as an arm body would fabricate a
        // block around a regex.
        if ancestors
            .parent(word)
            .is_some_and(|parent| [kinds.argument, kinds.command].contains(&parent.kind_id()))
        {
            return true;
        }
        let Some((command, command_ancestors)) =
            Self::generic_argument_command(word, ancestors, kinds)
        else {
            return false;
        };
        if Self::command_leading_word(&command, code, kinds)
            .and_then(script_slots)
            .is_some()
        {
            return false;
        }
        !(Self::is_switch_arm(&command, code, command_ancestors, kinds)
            && Self::is_switch_arm_body(word, &command))
    }

    /// Whether `node` is a Tcl-family braced *script* — a `proc` or `if`
    /// body, an iRules `when` handler — rather than a braced literal
    /// (#1381).
    ///
    /// The positive form of [`is_value_braced_word`], narrowed to the
    /// script kind so it answers `false` for every node of every other
    /// kind and for the two literal spellings, and narrowed again by the
    /// value slots [`is_braced_literal_slot`] recognises. That makes it
    /// the question the *non-Halstead* classifiers ask:
    /// `Checker::is_string_with_code` must not call a `proc` body a
    /// string literal, and `Alterator::keeps_children` must stop the dump
    /// flattening one into a leaf, which dropped the body from it. Both
    /// once listed `braced_word` beside `quoted_word` and
    /// `braced_word_simple`, which is right for the literal role and
    /// wrong for the script role the same kind also serves.
    ///
    /// Stated here rather than in each of the five call sites so the
    /// string, dump and Halstead classifiers cannot drift apart on the
    /// same bytes (grammar-dispatch §7). Halstead was the deliberate
    /// exception between #1318 and #1382: `braced_word_op_type` asked
    /// [`is_value_braced_word`] alone, so the braces of the value slots
    /// [`is_braced_literal_slot`] adds still billed a `{}` operator
    /// while the other two called the same bytes a literal. #1382
    /// closed it — that call site now asks this predicate — so a braced
    /// value is a value to all three, and `after {100} {puts hi}` bills
    /// the one `{}` its `{puts hi}` block earns rather than two.
    ///
    /// The remaining §7 note is one of scope, not of disagreement: the
    /// three differ only where Halstead's *operand* half keys on
    /// whether a body holds a command, which makes an empty or
    /// comment-only `proc` body an operand there and not a string here.
    ///
    /// [`is_value_braced_word`]: Self::is_value_braced_word
    /// [`is_braced_literal_slot`]: Self::is_braced_literal_slot
    #[must_use]
    fn is_braced_script_word<'a>(
        node: &Node<'a>,
        code: &[u8],
        ancestors: Ancestors<'a, '_>,
        kinds: &BracedWordKinds,
    ) -> bool {
        node.kind_id() == kinds.script
            && !Self::is_value_braced_word(node, code, ancestors, kinds)
            && !Self::is_braced_literal_slot(node, code, ancestors, kinds)
    }

    /// Whether `word`, a braced word [`is_value_braced_word`] calls a
    /// script, fills a slot whose documented syntax takes a *value*.
    ///
    /// [`is_value_braced_word`] answers by command name alone. Most of the
    /// names it recognises have a mixed signature, though, so that answer
    /// made a *value* — a millisecond count, a namespace name, a `switch`
    /// subject — a script (#1381 review). This
    /// is the per-argument half: `proc` has one value slot the grammar
    /// names as a field, and every other construct declares its layout in
    /// one of the two tables in `lang_helpers::tcl_family`, which
    /// `fills_script_slot` applies.
    ///
    /// Each value here was a string and a flat dump leaf before #1381, and
    /// would otherwise have become a script under it: the dump rendered
    /// `{my proc}` as a command named `my` and `{100}` as one named `100`,
    /// and `namespace export {…}` and `namespace ensemble create -map {…}`
    /// both occur in the Tcl 8.6 standard library. Since #1382 Halstead
    /// reads the same answer through [`is_braced_script_word`], so the
    /// braces of these values quote rather than open (`after {100} {…}`
    /// scores the `{}` of its script and no other).
    ///
    /// Two guards keep the construct-wide answer. A switch arm list parses
    /// as commands, so an arm whose *pattern* is spelled `proc`,
    /// `namespace`, `after` or `switch` builds one of these shapes around
    /// what are really arm bodies — [`is_switch_arm`] recognises it first.
    /// And an argument list whose own slot sequence is unreadable has no
    /// positions worth trusting.
    ///
    /// That second guard is `tcl_family::argument_slots_are_readable`,
    /// which asks the *slots* rather than the owner — asking
    /// [`Node::has_error`] of the command withdrew every value slot of a
    /// command whose only error sat inside one of its script arguments.
    /// Its doc carries the measurement. A list *longer* than the
    /// signature admits is the second way to be unreadable, and
    /// `tcl_family::ScriptSlots::LastOf` is where that one lives.
    ///
    /// A `try … trap` clause is out of reach only where the grammar
    /// wraps it: inside a `proc` or a `namespace eval` body the Tcl
    /// grammar recovers into an `ERROR` node and no role signal
    /// survives, but at the top level the same clause parses as a plain
    /// `trap` command whose slots this does read. An earlier revision
    /// of this paragraph said "the multi-line `try … trap` clause" was
    /// out of reach outright, which is where the `finally` defect
    /// `LastOf` now guards hid (#1382 review).
    ///
    /// [`Node::has_error`]: crate::Node::has_error
    /// [`is_braced_script_word`]: Self::is_braced_script_word
    /// [`is_value_braced_word`]: Self::is_value_braced_word
    /// [`is_switch_arm`]: Self::is_switch_arm
    #[must_use]
    fn is_braced_literal_slot<'a>(
        word: &Node<'a>,
        code: &[u8],
        ancestors: Ancestors<'a, '_>,
        kinds: &BracedWordKinds,
    ) -> bool {
        let mut chain = ancestors.iter(word);
        // A value slot is an *argument*, so it sits at least two levels
        // below the root; neither grammar puts a braced word at the root
        // or directly under it (a statement-position braced word is
        // wrapped in a `command`, so `{a b}` alone on a line is still two
        // deep). Both exits are therefore unreachable from a walk and are
        // here because this is a `pub` trait method any node may be
        // handed — `braced_slot_tests` calls it at both depths.
        let Some((parent, above_parent)) = chain.next() else {
            return false;
        };
        if parent.kind_id() == kinds.procedure {
            let is_name =
                matches!(parent.child_by_field_name("name"), Some(name) if name.id() == word.id());
            return is_name && !Self::is_switch_arm(&parent, code, above_parent, kinds);
        }
        let Some((owner, above_owner)) = chain.next() else {
            return false;
        };
        if parent.kind_id() != kinds.word_list
            || !argument_slots_are_readable(&owner, &parent)
            || Self::is_switch_arm(&owner, code, above_owner, kinds)
        {
            return false;
        }
        let dialect = Dialect { code, kinds };
        // At both pinned grammars a `word_list` has exactly two possible
        // owners — `command`'s `arguments` field and `namespace`'s child
        // — so the third arm is unreachable and stays as the answer a
        // grammar that grew a third owner should get: no layout, hence
        // the construct-wide script answer.
        // `a_word_list_is_owned_only_by_a_command_or_a_namespace` fails
        // when that stops being true, which is the event this arm exists
        // for.
        let roles = if owner.kind_id() == kinds.namespace {
            namespace_script_slots(&parent, dialect).map(|slots| ArgumentRoles { slots, first: 1 })
        } else if owner.kind_id() == kinds.command {
            Self::command_leading_word(&owner, code, kinds)
                .and_then(script_slots)
                .map(|slots| ArgumentRoles { slots, first: 0 })
        } else {
            None
        };
        roles.is_some_and(|roles| !fills_script_slot(word, &parent, roles, dialect))
    }

    /// Whether `word`, an argument of a `switch` arm command
    /// ([`is_switch_arm`]), sits in a *body* position rather than a
    /// *pattern* position.
    ///
    /// The arm list is a flat run of `pattern body pattern body …`,
    /// and the grammar breaks it into commands at newlines. One arm per
    /// line gives each command a pattern for its name and a body for
    /// its sole argument; several arms on one line give the first
    /// pattern's command the whole run as arguments, so the arguments
    /// alternate `body pattern body …`. The command-name test above
    /// keeps the *first* braced pattern a literal, but without this
    /// gate every later one was rescued as a body, and a one-line
    /// `switch -regexp $v { {^a} {…} {^b} {…} }` fabricated a `{}`
    /// around `{^b}` that the same arms written one per line did not —
    /// the score moving with layout, which is what the rule exists to
    /// stop.
    ///
    /// Position is the only signal because it is what Tcl itself uses:
    /// `switch` pairs the list up by index, with no marker on either
    /// half. The parity holds across the words that can interpose — a
    /// `-` fall-through body and a `default` pattern are both
    /// `simple_word`s and take a slot each — so an even index is a body
    /// and an odd one a pattern. This is the one place the rule needs a
    /// sibling *index* rather than a parent kind, and it runs once per
    /// braced argument of such a command.
    ///
    /// The index comes from [`Cursor::goto_first_child_for_byte`], which
    /// is `O(log n)` in the argument count. A `children().position(..)`
    /// scan answers identically in `O(n)`, and since every braced word
    /// of a one-line arm list asks, that made the Halstead walk — and,
    /// once #1381 routed them through this rule, `find` / `count
    /// --type string` and the `Ast` dump — quadratic in the width of
    /// one line: seconds per request at tens of KB (#1381 review). The
    /// id check keeps the scan's answer for a word that is not one of
    /// the command's arguments at all.
    ///
    /// [`is_switch_arm`]: Self::is_switch_arm
    /// [`Cursor::goto_first_child_for_byte`]: crate::node::Cursor::goto_first_child_for_byte
    #[must_use]
    fn is_switch_arm_body(word: &Node<'_>, command: &Node<'_>) -> bool {
        // `is_value_braced_word` asks only for a word it has already
        // placed inside this command's `arguments`, so the field is
        // always there on that path. It is not always there on the
        // method's own contract — an argument-less `switch` is a
        // `command` with a `name` and nothing else — and such a command
        // has no body position at all.
        let Some(arguments) = command.child_by_field_name("arguments") else {
            return false;
        };
        let mut cursor = arguments.cursor();
        let index = cursor.goto_first_child_for_byte(word.start_byte());
        cursor.node().id() == word.id() && matches!(index, Some(index) if index.is_multiple_of(2))
    }

    /// A command's leading word, when it is a statically resolvable
    /// `simple_word`.
    ///
    /// Located by field rather than by index (grammar-dispatch §3): the
    /// name is `command`'s `name` field in both grammars. A computed
    /// name (`$cmd {…}`, `[pick] {…}`) parses as a substitution node and
    /// is not resolvable at all, and a name spelled in quotes or braces
    /// (`"eval" {…}`) is legal Tcl that this deliberately leaves
    /// unresolved — the same limitation `tcl_command_name` records for
    /// the Cognitive and Cyclomatic walkers.
    #[must_use]
    fn command_leading_word<'c>(
        command: &Node<'_>,
        code: &'c [u8],
        kinds: &BracedWordKinds,
    ) -> Option<&'c str> {
        let name = command.child_by_field_name("name")?;
        if name.kind_id() != kinds.simple_word {
            return None;
        }
        let text = node_text(code, &name)?;
        // `::eval` *is* `eval`, and without the strip the qualified form
        // fell to the value default and lost its block. The rule and the
        // reason it stops at the *leading* qualifier live with the
        // helper, which the four metrics that resolve a leading word
        // without this table share (#1381 review).
        Some(crate::lang_helpers::strip_global_qualifier(text))
    }

    /// Whether `command` is really one `pattern body` pair of a Tcl
    /// `switch` arm list rather than a command of its own.
    ///
    /// Tcl models no `switch`, so `switch $v {a {…} b {…}}` parses the
    /// arm list as a braced word whose interior is a `command` named
    /// after the *first pattern* — `a` here — with the arm bodies as
    /// its arguments (grammar-dispatch §9; the Cognitive walker reads
    /// the same shape through `tcl_switch_arm_list`). Without this
    /// test the bodies would read as literals passed to a command
    /// called `a`, which is exactly what the value default is for
    /// everywhere else. iRules models `switch` with `switch_arm`
    /// children, so its arm bodies never reach here.
    #[must_use]
    fn is_switch_arm<'a>(
        command: &Node<'a>,
        code: &[u8],
        ancestors: Ancestors<'a, '_>,
        kinds: &BracedWordKinds,
    ) -> bool {
        ancestors
            .iter(command)
            .next()
            .is_some_and(|(arm_list, above)| {
                arm_list.kind_id() == kinds.script
                    && Self::generic_argument_command(&arm_list, above, kinds)
                        .and_then(|(switch, _)| Self::command_leading_word(&switch, code, kinds))
                        == Some(SWITCH_COMMAND)
            })
    }

    /// Whether the `{` of `node`'s parent braced word opens a block,
    /// or merely quotes a literal (#1318).
    ///
    /// This is #1314's guard with its kind test replaced by a role
    /// test. #1314 suppressed the opener of a `braced_word_simple`,
    /// the literal form the grammars emit only in the value slots they
    /// special-case; everywhere else a literal is a `braced_word`, the
    /// same kind a block uses, and its `{` still reported a `{}`
    /// operator for a block the source does not contain —
    /// `lappend x {a b}`, `puts {c d}`, and every user proc taking a
    /// list.
    ///
    /// The role is [`is_braced_script_word`], the same predicate
    /// `Checker::is_string_with_code` and `Alterator::keeps_children`
    /// ask, so the three classifiers answer one question about one set
    /// of bytes (grammar-dispatch §7). That was not so between #1318
    /// and #1382: this call site asked [`is_value_braced_word`] alone
    /// and billed a `{}` for the value slots of a script-taking
    /// command, which the other two had called literals since #1381.
    /// The exception, and what closing it moved, is recorded on
    /// [`is_braced_script_word`], which owns the §7 story for all five
    /// call sites.
    ///
    /// **It revises the operator only, and deliberately leaves the
    /// words inside a value alone.** Suppressing them too would make
    /// `lappend x {a b}` score like its `set x {a b}` synonym, which
    /// is the tidier answer for that line — and the wrong one in
    /// general, because the same braced argument of an unrecognised
    /// command is just as often real code: an `oo::class create C {…}`
    /// body, a `tcltest` `-body {…}`, an `apply {{x} {…}}` lambda.
    /// Suppressing contents was built and measured on a file holding
    /// one of each; it took n1 5 → 2, N1 8 → 2, n2 25 → 15 and
    /// N2 32 → 15, the class body, the test body and the lambda each
    /// collapsing into a single operand, and `halstead.effort` — a
    /// gated threshold metric — collapsing with them. So the rule only
    /// ever withdraws a claim the classifier cannot support; it never
    /// discards code the walk has already read.
    ///
    /// The residual is the asymmetry #1382 opens with, and it is the
    /// **contract**, not an outstanding defect: a braced value scores
    /// one operand where the grammar names the command (`set x {a b}`)
    /// and one per word where only the command name would
    /// (`lappend x {a b}`). #1382 weighed the two ways of closing it
    /// and took neither. Reading the *contents* — no substitution, no
    /// nested block, therefore list-like — is the measurement above,
    /// already taken and already rejected. Widening the one-operand
    /// treatment to a *value*-taking command list needs the mirror of
    /// `SCRIPT_TAKING_COMMANDS`, and that set is open where this one is
    /// closed: every user proc is in it, so the list is incomplete by
    /// construction and would leave three behaviours where there are
    /// two. Both trade a bounded over-report of the operand vocabulary
    /// for an unbounded loss of real code, which is the trade the
    /// paragraph above declines.
    ///
    /// Keeping to the operator also keeps the whole thing cheap: only
    /// a braced word's own opener can change answer, so the test is one
    /// kind comparison and one parent lookup, the same scope #1354 and
    /// #1314 use. Two lookups are `O(log n)` rather than `O(1)`, both
    /// cursor seeks by byte and both reached only from a *command whose
    /// name resolves*: [`is_switch_arm_body`] needs a `switch` arm
    /// word's index among its command's arguments, and — since #1382 —
    /// [`is_braced_literal_slot`] needs the same index to place a word
    /// against the layout `lang_helpers::tcl_family` records. An
    /// ancestor scan would have been `O(depth)` per node and quadratic
    /// on a deeply nested `expr`, the shape #1122 warns about.
    ///
    /// [`is_braced_literal_slot`]: Self::is_braced_literal_slot
    /// [`is_braced_script_word`]: Self::is_braced_script_word
    /// [`is_switch_arm_body`]: Self::is_switch_arm_body
    /// [`is_value_braced_word`]: Self::is_value_braced_word
    ///
    /// [`get_op_type`]: Self::get_op_type
    #[must_use]
    fn braced_word_op_type<'a>(
        node: &Node<'a>,
        code: &[u8],
        ancestors: Ancestors<'a, '_>,
        kinds: &BracedWordKinds,
    ) -> TokenRole {
        let base = Self::get_op_type(node, ancestors);
        // Only the opener can change answer, and the test has to say so
        // rather than infer it from the parent kind. A braced word's
        // operator children are the `{` *and* every `;` separating two
        // of its commands — `_terminator` is a hidden rule, so the
        // separator is inlined as a direct child of the `braced_word`
        // and both dialects list `SEMI` as an operator. The closer is
        // unclassified and a `\n` terminator is not an operator, so
        // those two are the whole set.
        if node.kind_id() != kinds.open_brace || !matches!(base, TokenRole::Operator) {
            return base;
        }
        let quotes_a_literal = ancestors.iter(node).next().is_some_and(|(parent, above)| {
            parent.kind_id() == kinds.script
                && !Self::is_braced_script_word(&parent, code, above, kinds)
        });
        if quotes_a_literal {
            TokenRole::Unknown
        } else {
            base
        }
    }

    /// The grammar's name for the operator kind `id`, used to render a
    /// Halstead operator; empty for languages that do not report them.
    #[must_use]
    fn get_operator_id_as_str(_id: u16) -> &'static str {
        ""
    }
}

mod bash;
mod c;
mod ccomment;
mod cpp;
mod csharp;
mod elixir;
mod generic_angles;
mod go;
mod groovy;
mod irules;
mod java;
mod javascript;
mod js_family;
mod kotlin;
mod lua;
mod mozcpp;
mod mozjs;
mod objc;
mod perl;
mod php;
mod preproc;
mod python;
mod ruby;
mod rust;
mod tcl;
mod tsx;
mod typescript;

#[cfg(test)]
mod node_text_tests {
    use super::node_text;
    use crate::langs::RustParser;
    use crate::traits::ParserTrait;
    use std::path::PathBuf;

    /// A node whose span lies inside the buffer it was parsed from
    /// yields its exact source text.
    #[cfg(feature = "rust")]
    #[test]
    fn in_bounds_span_returns_text() {
        let src = "fn x() {}";
        let parser = RustParser::new(src.as_bytes().to_vec(), &PathBuf::from("t.rs"), None);
        let root = parser.root();
        assert_eq!(node_text(parser.code(), &root), Some(src));
    }

    /// Reslicing a node against a *shorter* buffer (the stale-span hazard
    /// the guard exists for) must degrade to `None`, not panic. A direct
    /// `&code[start..end]` would panic here — this is the revert check.
    #[cfg(feature = "rust")]
    #[test]
    fn out_of_bounds_span_returns_none_not_panic() {
        let src = "fn x() {}";
        let parser = RustParser::new(src.as_bytes().to_vec(), &PathBuf::from("t.rs"), None);
        let root = parser.root();
        assert!(root.end_byte() > 2);
        let truncated = &src.as_bytes()[..2];
        assert_eq!(node_text(truncated, &root), None);
    }

    /// The UTF-8 guard is the *other* failure mode, and unlike the range
    /// guard it is reachable without violating the same-parse
    /// precondition: `Ast::parse` accepts arbitrary bytes. A span whose
    /// bytes are not valid UTF-8 must yield `None`, not a panic and not
    /// lossy replacement characters.
    #[cfg(feature = "rust")]
    #[test]
    fn non_utf8_span_returns_none() {
        let mut src = b"fn ".to_vec();
        src.extend_from_slice(&[0xF0, 0x9F]);
        src.extend_from_slice(b"() {}");
        let parser = RustParser::new(src.clone(), &PathBuf::from("t.rs"), None);
        let root = parser.root();
        assert_eq!(root.end_byte(), src.len());
        assert_eq!(node_text(parser.code(), &root), None);
    }
}

#[cfg(test)]
mod ancestor_tests {
    use super::Getter;
    use crate::node::{Ancestors, Node};
    use crate::test_support::for_each_node_with_chain;
    use crate::traits::LanguageInfo;

    /// `get_func_space_name` must name a space the same whether it reads
    /// the walker's ancestor chain or climbs with `Node::parent`.
    ///
    /// Two grammars consult an ancestor here, for different reasons:
    /// the JS family names an anonymous `function` / arrow from the
    /// `pair` or `variable_declarator` holding it, and Elixir skips
    /// naming a `def` that sits inside a `quote` template. #1088 moved
    /// both onto the chain.
    #[cfg(any(
        feature = "elixir",
        feature = "javascript",
        feature = "mozjs",
        feature = "typescript",
    ))]
    fn assert_name_parity<L: LanguageInfo + Getter>(
        label: &str,
        code: &[u8],
        expect_named: &[&str],
    ) {
        let mut seen: Vec<&str> = Vec::new();
        let visited = for_each_node_with_chain::<L>(code, |node: &Node<'_>, chain| {
            let known = L::get_func_space_name(node, code, Ancestors::known(chain));
            let climbing = L::get_func_space_name(node, code, Ancestors::unknown());
            assert_eq!(
                known,
                climbing,
                "{label}: name of {} at row {} disagrees",
                node.kind(),
                node.start_row()
            );
            if let Some(name) = known
                && expect_named.contains(&name)
                && !seen.contains(&name)
            {
                seen.push(name);
            }
        });
        assert!(visited > 20, "{label}: fixture is too small to prove much");
        for name in expect_named {
            assert!(
                seen.contains(name),
                "{label}: no node resolved to {name:?}, so the fixture no longer \
                 exercises the ancestor-derived naming it was added for"
            );
        }
    }

    #[cfg(all(
        feature = "elixir",
        feature = "javascript",
        feature = "mozjs",
        feature = "typescript",
    ))]
    #[test]
    fn func_space_name_agrees_between_known_and_climbing() {
        // `outer` and `keyed` are only reachable through the parent:
        // the function expressions themselves carry no `name` field.
        //
        // All four JS-family grammars are exercised, not just
        // JavaScript: their `get_func_space_name` impls are separate
        // copies of the same body against four distinct `kind_id`
        // enums, so a `Pair` / `VariableDeclarator` id that drifted in
        // one grammar would be invisible here if only one were checked.
        let js_source =
            b"var outer = function () { return 1; };\nvar o = { keyed: function () { return 2; } };\n";
        assert_name_parity::<crate::langs::JavascriptCode>(
            "javascript",
            js_source,
            &["outer", "keyed"],
        );
        assert_name_parity::<crate::langs::MozjsCode>("mozjs", js_source, &["outer", "keyed"]);
        assert_name_parity::<crate::langs::TypescriptCode>(
            "typescript",
            js_source,
            &["outer", "keyed"],
        );
        assert_name_parity::<crate::langs::TsxCode>("tsx", js_source, &["outer", "keyed"]);
        // `multi` is named from its `Call` head; the `def a` inside the
        // `quote` template is not a definition, so it falls through to
        // the field-less default.
        assert_name_parity::<crate::langs::ElixirCode>(
            "elixir",
            b"defmodule Foo do\n  defmacro multi do\n    quote do\n      def a, do: 1\n    end\n  end\nend\n",
            &["Foo", "multi"],
        );
    }
}

/// The `is_braced_literal_slot` / `is_switch_arm_body` guards that no
/// Tcl-family walk can reach, and the grammar shape each rests on.
#[cfg(test)]
#[cfg(any(feature = "tcl", feature = "irules"))]
mod braced_slot_tests {
    use super::{BracedWordKinds, Getter};
    // Gated with the rows that name it: `Tcl` is read only by the
    // `feature = "tcl"` helpers below, so the module's wider
    // `any(tcl, irules)` gate would leave this unused — and `ci.yml` sets
    // `RUSTFLAGS: "-D warnings"`, which turns that into a hard failure on
    // an `irules`-without-`tcl` build.
    #[cfg(feature = "tcl")]
    use crate::Tcl;
    use crate::node::{Ancestors, Node};
    use crate::test_support::for_each_node_with_chain;
    use crate::traits::LanguageInfo;

    /// No node shallower than an argument can fill a value slot.
    ///
    /// A value slot is an argument, and both grammars hang an argument at
    /// least two levels below the root — under a `word_list` under a
    /// `command`, or under a construct's own node. A statement-position
    /// braced word is no exception: `{a b}` alone on a line is a `command`
    /// wrapping the braced word, never the braced word itself. So the root
    /// and its direct children are exactly the depths
    /// `is_braced_literal_slot`'s two chain exits answer for, and no walk
    /// reaches either.
    ///
    /// They are still the method's contract — it is `pub`, takes any node,
    /// and must not read a slot into something with no argument list above
    /// it. Flipping either exit to `true` makes this fail.
    fn assert_no_shallow_value_slot<L: LanguageInfo + Getter>(
        label: &str,
        code: &[u8],
        kinds: &BracedWordKinds,
    ) {
        let (mut roots, mut children) = (0, 0);
        for_each_node_with_chain::<L>(code, |node: &Node<'_>, chain| {
            if chain.len() > 1 {
                return;
            }
            for ancestors in [Ancestors::known(chain), Ancestors::unknown()] {
                assert!(
                    !L::is_braced_literal_slot(node, code, ancestors, kinds),
                    "{label}: a {} at depth {} filled a value slot",
                    node.kind(),
                    chain.len()
                );
            }
            if chain.is_empty() {
                roots += 1;
            } else {
                children += 1;
            }
        });
        assert_eq!(roots, 1, "{label}: exactly one root");
        assert!(
            children > 1,
            "{label}: the fixture must put several nodes directly under the \
             root, or the depth-1 exit goes unexercised"
        );
    }

    #[test]
    #[cfg(feature = "tcl")]
    fn tcl_no_node_above_an_argument_fills_a_value_slot() {
        assert_no_shallow_value_slot::<crate::langs::TclCode>(
            "tcl",
            b"{a b}\nproc {my proc} {} {}\nnamespace export {c d}\n",
            &crate::lang_helpers::tcl::BRACED_WORD_KINDS,
        );
    }

    #[test]
    #[cfg(feature = "irules")]
    fn irules_no_node_above_an_argument_fills_a_value_slot() {
        assert_no_shallow_value_slot::<crate::langs::IrulesCode>(
            "irules",
            b"{a b}\nproc {my proc} {} {}\nnamespace export {c d}\n",
            &crate::lang_helpers::irules::BRACED_WORD_KINDS,
        );
    }

    /// Every braced word of `code` that fills a value slot, in source
    /// order. Tcl only: `is_braced_literal_slot` is a provided method, so
    /// both dialects run the identical body over their own kind table.
    ///
    /// Walks with [`Ancestors::unknown`] rather than through
    /// `for_each_node_with_chain`, whose fixtures must parse cleanly so a
    /// walk cannot cover error recovery by accident — which is exactly
    /// what the callers below are about. The two spellings answer
    /// identically here; `assert_no_shallow_value_slot` above asserts
    /// that on every node it visits.
    #[cfg(feature = "tcl")]
    fn value_slot_texts(code: &[u8]) -> Vec<String> {
        let tree = crate::node::Tree::new::<crate::langs::TclCode>(code);
        tree.get_root()
            .preorder()
            .filter(|node| {
                <crate::langs::TclCode as Getter>::is_braced_literal_slot(
                    node,
                    code,
                    Ancestors::unknown(),
                    &crate::lang_helpers::tcl::BRACED_WORD_KINDS,
                )
            })
            .map(|node| node.utf8_text(code).unwrap_or_default().to_owned())
            .collect()
    }

    /// A parse error nested inside one argument's *script* leaves its
    /// sibling *value* slots alone (#1381 review).
    ///
    /// `Node::has_error` is transitive, so asking it of the command
    /// withdrew the rescue from every argument: `trap {pat} {v} {puts ]}`
    /// lost both value slots, and the dump then rendered each as a
    /// fabricated `command` named after its own text. What decides
    /// whether the positions are readable is the argument list's own slot
    /// sequence, and an error *inside* an argument moves no sibling's
    /// index. Verified by reverting the guard to `owner.has_error()`:
    /// the error row then comes back empty and this fails.
    ///
    /// Asserted as the *same* list for both fixtures rather than as a
    /// non-empty one for the error row. The well-formed row is what pins
    /// that the fixture still spells two value slots at all, so trimming
    /// one out fails here instead of quietly reducing the claim to
    /// "nothing changed".
    #[test]
    #[cfg(feature = "tcl")]
    fn an_error_inside_a_script_argument_keeps_the_sibling_value_slots() {
        let well_formed = value_slot_texts(b"trap {pat} {v} {puts j}\n");
        assert_eq!(
            well_formed,
            ["{pat}", "{v}"],
            "the fixture must spell two value slots before an error is added"
        );
        assert_eq!(
            value_slot_texts(b"trap {pat} {v} {puts ]}\n"),
            well_formed,
            "an error inside the script argument moves no sibling's index"
        );
    }

    /// The other half of the same rule: an `ERROR` token occupying a slot
    /// of the argument list *does* shift every argument after it, so the
    /// construct-wide answer is the right one there.
    ///
    /// `on code varList script` reads its slots by position, and the
    /// stray `]` takes one of its own — so `{code}` is no longer
    /// followed by the variable list the signature puts there and the
    /// layout cannot be trusted. Without this row the guard could be
    /// deleted outright and the test above would still pass.
    ///
    /// The fixture keeps the clause's three arguments deliberately: the
    /// `]` replaces the variable list rather than joining it, so
    /// `ScriptSlots::LastOf`'s arity rule still admits the list and
    /// `argument_slots_are_readable` is the only guard that can reject
    /// it. A four-argument spelling would be rejected twice over and
    /// this row would stop naming which guard it tests (#1382 review).
    ///
    /// The well-formed row is what keeps the empty one from passing
    /// vacuously: `on` reaching the slot table at all is what makes the
    /// error row's emptiness mean something, and dropping that row from
    /// `SCRIPT_TAKING_COMMANDS` empties both. The siblings either side
    /// of this test anchor the same way.
    #[test]
    #[cfg(feature = "tcl")]
    fn an_error_occupying_a_slot_withdraws_the_layout() {
        assert_eq!(
            value_slot_texts(b"on {code} {v} {puts j}\n"),
            ["{code}", "{v}"],
            "the fixture must spell two value slots before the error \
             takes one of their places"
        );
        assert!(
            value_slot_texts(b"on {code} ] {puts j}\n").is_empty(),
            "an ERROR token in the argument list makes every position \
             unreadable"
        );
    }

    /// The third way a layout stops being readable: the grammar groups
    /// a *following* clause into this one, so the list is longer than
    /// the signature admits (#1382 review).
    ///
    /// Tcl models neither `on` nor `trap` outside a `try`, so a
    /// top-level `try … trap … finally` parses the whole tail as one
    /// generic `trap` command — `p v b finally c`. Reading "the script
    /// is the last argument" off that names the `finally` body and
    /// calls the handler body a value, which withdraws the `{}` of a
    /// real block and flattens it out of the dump. Both rows run the
    /// same clause, so the second cannot pass by the fixture having
    /// stopped spelling a value slot at all.
    #[test]
    #[cfg(feature = "tcl")]
    fn a_handler_clause_of_the_wrong_arity_withdraws_the_layout() {
        assert_eq!(
            value_slot_texts(b"try {puts a} trap {p} {v} {puts b}\n"),
            ["{p}", "{v}"],
            "a three-argument `trap` clause spells two value slots"
        );
        assert!(
            value_slot_texts(b"try {puts a} trap {p} {v} {puts b} finally {puts c}\n").is_empty(),
            "a trailing `finally` joins the same generic `trap` command, \
             so no position in its argument list is trustworthy"
        );
    }

    /// A command with no argument list has no `switch`-arm body position.
    ///
    /// `is_value_braced_word` only ever asks about a word it has already
    /// placed inside the command's `arguments`, so that field is always
    /// present on the walk's path. It is absent for an argument-less
    /// command — `switch` on a line by itself is a `command` carrying only
    /// its `name` — and the method has to answer for one, because a
    /// bodiless command has no even-indexed slot to be in.
    ///
    /// The second half is the same method's positive answer on a real arm
    /// list, so the first cannot be satisfied by a constant.
    #[test]
    #[cfg(feature = "tcl")]
    fn a_command_without_arguments_has_no_switch_arm_body() {
        use crate::node::Tree;

        let code = b"switch\nswitch $v { p {puts a} }\n";
        let tree = Tree::new::<crate::langs::TclCode>(code);
        let root = tree.get_root();
        let bare = root.child(0).expect("the bare `switch` command");
        let name = bare
            .child_by_field_name("name")
            .expect("even a bare command names itself");
        assert!(
            bare.child_by_field_name("arguments").is_none(),
            "the fixture's first command must carry no argument list"
        );
        assert!(
            !<crate::langs::TclCode as Getter>::is_switch_arm_body(&name, &bare),
            "a command with no argument list has no body position"
        );

        let arm = find_arm_named_p(&root, code).expect("the arm command named `p`");
        let body = arm
            .child_by_field_name("arguments")
            .and_then(|args| args.child(0))
            .expect("the arm's body");
        assert!(
            <crate::langs::TclCode as Getter>::is_switch_arm_body(&body, &arm),
            "the arm's sole argument is its body, or the assertion above is \
             satisfied by a constant"
        );
    }

    /// The `command` a Tcl `switch` arm list spells its first pattern as.
    #[cfg(feature = "tcl")]
    fn find_arm_named_p<'a>(node: &Node<'a>, code: &[u8]) -> Option<Node<'a>> {
        let named_p = node.kind_id() == Tcl::Command as u16
            && node
                .child_by_field_name("name")
                .is_some_and(|name| name.utf8_text(code) == Some("p"));
        if named_p {
            return Some(*node);
        }
        node.children()
            .find_map(|child| find_arm_named_p(&child, code))
    }

    /// Only a `command` and a `namespace` own a `word_list`, which is what
    /// makes `is_braced_literal_slot`'s third owner arm unreachable.
    ///
    /// Checked against both pinned grammars' `node-types.json` when that
    /// arm was written; this is the executable half, so a grammar bump
    /// handing a `word_list` to a third construct fails here rather than
    /// silently routing it to the no-layout answer.
    fn assert_word_list_owners<L: LanguageInfo>(label: &str, code: &[u8], kinds: &BracedWordKinds) {
        let mut seen = 0;
        for_each_node_with_chain::<L>(code, |node: &Node<'_>, chain| {
            if node.kind_id() != kinds.word_list {
                return;
            }
            let owner = chain.last().expect("a word_list is never the root");
            assert!(
                [kinds.command, kinds.namespace].contains(&owner.kind_id()),
                "{label}: a word_list owned by {}",
                owner.kind()
            );
            seen += 1;
        });
        assert!(seen > 2, "{label}: fixture reached only {seen} word_lists");
    }

    /// Both owners plus the modelled constructs either grammar spells, so
    /// a new owner among them is what the assertion catches.
    const WORD_LIST_OWNER_FIXTURE: &[u8] = b"proc p {x} { puts $x }\n\
        namespace eval ns { puts hi }\n\
        namespace export {a b}\n\
        if {$x} { puts a } else { puts b }\n\
        while {$x} { puts a }\n\
        foreach i {1 2} { puts $i }\n\
        catch { puts a } msg\n\
        eval {puts hi}\n\
        switch $v { p {puts a} }\n\
        lappend l {c d}\n";

    #[test]
    #[cfg(feature = "tcl")]
    fn tcl_a_word_list_is_owned_only_by_a_command_or_a_namespace() {
        assert_word_list_owners::<crate::langs::TclCode>(
            "tcl",
            WORD_LIST_OWNER_FIXTURE,
            &crate::lang_helpers::tcl::BRACED_WORD_KINDS,
        );
    }

    #[test]
    #[cfg(feature = "irules")]
    fn irules_a_word_list_is_owned_only_by_a_command_or_a_namespace() {
        assert_word_list_owners::<crate::langs::IrulesCode>(
            "irules",
            WORD_LIST_OWNER_FIXTURE,
            &crate::lang_helpers::irules::BRACED_WORD_KINDS,
        );
    }
}