git_xcrypt/git/

config.rs

//! Reading and writing git configuration through a library.
//!
//! Spawning `git config` is not an option: git starts a filter process per
//! operation, so it would be N process spawns on the hot path — most expensive
//! on exactly the platform where it hurts most. Writing only ever touches the
//! repository-local file, which is the only one this tool has business changing.

use std::fmt;
use std::path::{Path, PathBuf};

use bstr::ByteSlice as _;
use gix_config::File;
use gix_config::file::Metadata;

use crate::{Error, Result};

/// The global attributes file, resolved the way git resolves it.
///
/// Reading `core.attributesFile` verbatim is not enough, and the gap is not
/// cosmetic: it is a source in git's attribute stack, so a line in it can put
/// `text` back on a path this tool encrypts — and git then converts the
/// ciphertext. Measured on git 2.55, 2 MB, the line living in
/// `~/.config/git/attributes` while the same line in the tree is refused: `git
/// add` exited **0**, 27 `CR` bytes were eaten out of the blob, the commit
/// succeeded and the checkout left **no file at all**. The refusal in
/// [`crate::commands::filter`] and the gate in `status` both resolve the stack correctly;
/// they simply were not being handed this file, so both reported a healthy
/// repository over a destroyed one.
///
/// Git's rule, measured on 2.55 rather than read from the documentation — the
/// five shapes are a table test in this module:
///
/// | `core.attributesFile` | what git reads |
/// | --- | --- |
/// | unset | `$XDG_CONFIG_HOME/git/attributes`, else `$HOME/.config/git/attributes` |
/// | `~/name`, `~user/name` | expanded, exactly as `core.excludesFile` is |
/// | an absolute path | that path |
/// | empty | **nothing** — and no XDG fallback either |
///
/// Returns the path whether or not it exists; a missing file is an empty source
/// to the resolver, which is what git does with one too.
#[must_use]
pub fn global_attributes_file(config: &File) -> Option<PathBuf> {
    global_attributes_file_for(
        config,
        // `HOME` first, then the platform's own answer — git's order, and the
        // reason a Windows user can keep a linux-style home somewhere else.
        gix_path::env::home_dir().as_deref(),
        std::env::var_os("XDG_CONFIG_HOME").as_deref(),
    )
}

/// [`global_attributes_file`], with the two environment variables as arguments.
///
/// Split out so the table below can be a test rather than a hope. `HOME` and
/// `XDG_CONFIG_HOME` belong to the *process*, so setting them to exercise a row
/// would need `unsafe` — which `unsafe_code = "forbid"` refuses, and that
/// refusal is worth more than the convenience. Passing the difference in as an
/// argument is the same shape `eol::apply_where` and `repo::with_separator`
/// already use for a platform the test is not running on.
#[must_use]
fn global_attributes_file_for(
    config: &File,
    home: Option<&Path>,
    xdg_config_home: Option<&std::ffi::OsStr>,
) -> Option<PathBuf> {
    let Ok(value) = config.raw_value("core.attributesFile") else {
        // Git treats an empty `XDG_CONFIG_HOME` as unset, so `is_empty` is part
        // of the rule and not a defensive extra.
        if let Some(xdg) = xdg_config_home
            && !xdg.is_empty()
        {
            return Some(PathBuf::from(xdg).join("git").join("attributes"));
        }
        return Some(home?.join(".config").join("git").join("attributes"));
    };
    // Set but empty turns the file off; it does **not** fall back to XDG.
    if value.is_empty() {
        return None;
    }
    // `~/`, `~user/` and `%(prefix)/`, through the same crate that parsed the
    // value. Hand-rolling the expansion would be a second spelling of a rule
    // git already has one of.
    gix_config::Path::from(value)
        .interpolate(gix_config::path::interpolate::Context {
            home_dir: home,
            ..Default::default()
        })
        .ok()
}

/// The repository-local configuration, loaded for editing.
///
/// Includes are deliberately not followed: we are about to write this file back,
/// and following includes would fold someone else's file into ours.
///
/// # Errors
///
/// [`Error::Config`] when the file exists but cannot be parsed.
pub fn open_local(path: &Path) -> Result<File> {
    read_optional(path, gix_config::Source::Local)
}

/// One configuration file, where a missing one is empty and a broken one is not.
///
/// The asymmetry is the point: git creates these files lazily, so absence is an
/// ordinary state and saying so costs nothing. A file that is *there* and does
/// not parse is a different answer, and swallowing it would let a repository
/// whose `filter.git-xcrypt.required` cannot be read pass for one that has it.
fn read_optional(path: &Path, source: gix_config::Source) -> Result<File> {
    if !path.exists() {
        return Ok(File::new(Metadata::from(source)));
    }
    File::from_path_no_includes(path.to_path_buf(), source)
        .map_err(|err| Error::Config(format!("could not read {}: {err}", path.display())))
}

/// The configuration git itself would see, for reading only.
///
/// Full precedence: git installation, system, global, repository-local,
/// worktree and `GIT_CONFIG_*` overrides, with `include`/`includeIf` followed.
/// The smudge path needs this rather than `.git/config` alone, because
/// `core.autocrlf` and `core.eol` are almost always set globally — on Windows
/// the installer does it — and reading only the local file would leave the
/// measured line-ending table unreachable on exactly the platform it exists for.
///
/// This assembles the cascade the way `gix_config::File::from_git_dir` does,
/// with **one** difference, and it is the reason it is spelled out here rather
/// than called: that function reads `config.worktree` whenever
/// `extensions.worktreeConfig` is true, and treats the file's absence as an
/// error. Git treats it as empty — the extension is permission to look, not a
/// promise the file exists, and git creates it lazily on the first
/// `git config --worktree`.
///
/// Measured on git 2.55, 2026-08-05: in a repository where
/// `extensions.worktreeConfig` was set and no `config.worktree` had been written
/// yet, git ran `add`, `commit` and `status` at exit 0 while this build could
/// not start its own filter — so with `required = true`, **every git operation
/// in the repository failed**, `git add` exiting 128 with
/// `could not read git configuration`. The trigger is the command git's own
/// documentation gives for enabling per-worktree configuration, and the window
/// is however long it takes to run the next one.
///
/// Fail-closed, so nothing was ever stored in the clear over it — but an outage
/// is exactly what `required = true` turns a configuration read into, and the
/// rest of this crate is careful about that.
///
/// **The two directories are different on purpose, and that is the second
/// difference.** `config` is shared, so it comes from `common_dir`;
/// `config.worktree` is per-checkout and comes from `git_dir`, which for a
/// linked worktree is `…/.git/worktrees/<name>`. Reading both from the common
/// directory — which is what this did, and what `gix-config` does when handed
/// one path — gives a linked worktree the *main* checkout's per-worktree
/// configuration, which is nobody's configuration.
///
/// Measured on git 2.55, 2026-08-05, 2 MB in a linked worktree whose
/// `config.worktree` set `core.attributesFile` to a file declaring `vault/**
/// text`: `git check-attr text` answered `set` there and `unspecified` in the
/// main checkout, `git add` exited **0** because the refusal never saw the file,
/// 40 bytes were eaten out of the blob, and the checkout left **no file at
/// all**. Same shape as an unresolved `~/` in [`global_attributes_file`], one
/// directory further out.
///
/// # Errors
///
/// [`Error::Config`] when a file in the cascade exists and cannot be parsed.
pub fn open_full(git_dir: &Path, common_dir: &Path) -> Result<File> {
    let broken =
        |err: &dyn fmt::Display| Error::Config(format!("could not read git configuration: {err}"));

    let mut local = read_optional(&common_dir.join("config"), gix_config::Source::Local)?;
    // The one file git looks for only conditionally, and the one whose absence
    // must not be an error. `Source::Worktree` rather than `Local`, so it keeps
    // the precedence git gives it: above the local file, below the environment.
    let worktree = get(&local, "extensions.worktreeConfig")
        .is_some_and(|value| is_true(&value))
        .then(|| {
            read_optional(
                &git_dir.join("config.worktree"),
                gix_config::Source::Worktree,
            )
        })
        .transpose()?;

    let home = gix_path::env::home_dir();
    let options = gix_config::file::init::Options {
        includes: gix_config::file::includes::Options::follow(
            gix_config::path::interpolate::Context {
                home_dir: home.as_deref(),
                ..Default::default()
            },
            gix_config::file::includes::conditional::Context {
                git_dir: Some(git_dir),
                branch_name: None,
            },
        ),
        ..Default::default()
    };

    let mut config = File::from_globals().map_err(|err| broken(&err))?;
    config
        .resolve_includes(options)
        .map_err(|err| broken(&err))?;
    local
        .resolve_includes(options)
        .map_err(|err| broken(&err))?;
    config.append(local).map_err(|err| broken(&err))?;
    if let Some(mut worktree) = worktree {
        worktree
            .resolve_includes(options)
            .map_err(|err| broken(&err))?;
        config.append(worktree).map_err(|err| broken(&err))?;
    }
    config
        .append(File::from_environment_overrides().map_err(|err| broken(&err))?)
        .map_err(|err| broken(&err))?;
    // Last, so `-c` outranks everything — including the `GIT_CONFIG_COUNT` set
    // just above, which is the order git applies them in.
    if let Some(overrides) = cli_overrides(std::env::var_os(CLI_OVERRIDE_ENV).as_deref()) {
        config.append(overrides).map_err(|err| broken(&err))?;
    }
    Ok(config)
}

/// Where git puts the overrides given as `git -c key=value`.
const CLI_OVERRIDE_ENV: &str = "GIT_CONFIG_PARAMETERS";

/// The `git -c key=value` overrides, as a configuration source.
///
/// **Why this is hand-rolled rather than left to `gix-config`.** That crate
/// reads the *other* mechanism — `GIT_CONFIG_COUNT` with `GIT_CONFIG_KEY_n` and
/// `GIT_CONFIG_VALUE_n` — and git 2.55 does not populate it for `-c`; measured,
/// by asking an alias to print its own environment. Without this, `git -c
/// core.autocrlf=true checkout` converted the paths git owns and left ours
/// alone: one command, two answers, on paths sitting next to each other.
///
/// **The format is measured, not assumed.** git 2.55 writes shell-quoted words,
/// key and value quoted separately, and escapes `!` outside the quotes for csh:
///
/// ```text
/// 'core.autocrlf'='true' 'core.eol'='lf' 'user.name'='a b'\''c'
/// 'alias.x'=''\!'printenv FOO'
/// ```
///
/// So a word is a concatenation of quoted runs, backslash escapes and bare
/// characters, and it is only after unquoting that `key=value` can be split —
/// a value may legally contain spaces, quotes and `=`. Older git quoted the
/// whole pair as `'key=value'`, which this reads identically.
///
/// **Fail-open throughout, and that direction is deliberate.** Everything here
/// runs on the clean path, where `required = true` turns a wrong answer into a
/// blocked repository rather than a diagnostic. A value nobody typed is worse
/// than a value we did not notice: a bogus `core.attributesFile` invented by a
/// misparse would refuse `git add` across the whole repository. So an unterminated
/// quote drops the entire variable, a word that does not name a dotted key is
/// skipped, and a value `gix-config` will not accept is skipped — in every case
/// leaving the configuration files to speak for themselves, exactly as before
/// this function existed.
fn cli_overrides(raw: Option<&std::ffi::OsStr>) -> Option<File> {
    let words = split_quoted(raw?.to_str()?)?;
    let mut file = File::new(Metadata::from(gix_config::Source::Cli));
    let mut any = false;

    for word in words {
        // No `=` at all is git's boolean shorthand: `git -c core.autocrlf` means
        // true. `get` already spells a value-less key that way, so this stays
        // the one spelling every caller tests for.
        let (key, value) = match word.split_once('=') {
            Some((key, value)) => (key.to_string(), value.to_string()),
            None => (word, "true".to_string()),
        };
        // Asked with `gix-config`'s own parser rather than by looking for a dot,
        // so this and `get` cannot drift apart — and so `set_raw_value`, which
        // panics on a key it will not parse, is never handed one.
        if gix_config::AsKey::try_as_key(&key.as_str()).is_none() {
            continue;
        }
        if file.set_raw_value(key.as_str(), value.as_str()).is_ok() {
            any = true;
        }
    }

    any.then_some(file)
}

/// Splits one shell-quoted line into words, or gives up on the whole line.
///
/// Only the three constructs git's own quoting produces: a single-quoted run is
/// literal, a backslash outside quotes takes the next character literally, and
/// unquoted whitespace ends a word. Returning `None` for an unterminated quote
/// is the fail-open half of [`cli_overrides`] — half a word could name a key
/// nobody asked for.
fn split_quoted(line: &str) -> Option<Vec<String>> {
    let mut words = Vec::new();
    let mut current = String::new();
    let mut started = false;
    let mut quoted = false;
    let mut chars = line.chars();

    while let Some(character) = chars.next() {
        match character {
            '\'' => {
                quoted = !quoted;
                started = true;
            }
            '\\' if !quoted => {
                current.push(chars.next()?);
                started = true;
            }
            character if character.is_whitespace() && !quoted => {
                if started {
                    words.push(std::mem::take(&mut current));
                    started = false;
                }
            }
            character => {
                current.push(character);
                started = true;
            }
        }
    }

    if quoted {
        return None;
    }
    if started {
        words.push(current);
    }
    Some(words)
}

/// Writes a configuration file back to disk, replacing it in one step.
///
/// This file carries the driver registration, so a half-written one leaves git
/// with no filter and the next `git add` storing plaintext with exit code 0.
///
/// # Errors
///
/// [`Error::Io`] when the file cannot be written.
pub fn save_local(path: &Path, config: &File) -> Result<()> {
    crate::util::atomic::write(path, &config.to_bstring())
}

/// Sets a dotted key such as `filter.git-xcrypt.required`, creating what is missing.
///
/// # Errors
///
/// [`Error::Config`] when the key cannot be set.
pub fn set(config: &mut File, key: &str, value: &str) -> Result<()> {
    config
        .set_raw_value(key, value)
        .map(|_| ())
        .map_err(|err| Error::Config(format!("could not set {key}: {err}")))
}

/// Removes a dotted key, if it is there at all.
///
/// # Errors
///
/// [`Error::Config`] when the key names a section that cannot be addressed.
pub fn unset(config: &mut File, key: &str) -> Result<()> {
    let (section_key, name) = key
        .rsplit_once('.')
        .ok_or_else(|| Error::Config(format!("`{key}` is not a dotted configuration key")))?;

    if let Ok(mut section) = config.section_mut_by_key(section_key) {
        while section.remove(name).is_some() {}
    }
    Ok(())
}

/// Reads a dotted key, if present.
///
/// A key written with no value at all — `[core]\n\tautocrlf` — is `true` to git,
/// but has no raw value to return, so it comes back as `Some("true")`: git's own
/// reading of that line, spelled the way every caller already tests for.
///
/// **`Some("true")` rather than `Some("")`, and the difference is a security
/// one.** `gix-config` reports `key` (no `=`) and `key =` (an empty value)
/// identically — the first as `Err(KeyMissing)` from `raw_value`, the second as
/// `Ok("")` — and git does not: measured on git 2.55, `git config --type=bool`
/// reads the first as `true` and the second as **`false`**. Flattening both to
/// the empty string and calling that true made `filter.git-xcrypt.required = `
/// read as enabled, while git ignored the failing filter and stored the
/// plaintext with `git add` exiting 0 — and `status`, the gate that exists to
/// catch exactly that, reported no gap.
///
/// **A key `gix-config` will not parse gives `None` rather than a panic.** Its
/// `raw_value` takes the key through `AsKey::as_key`, which panics on anything
/// it cannot split — `notdotted` is enough. No caller passes such a key today,
/// and the `-c` reader filters them out before they get here, so this is not a
/// fix for a live bug; it is a fix for the *shape* of one. `get` is on the
/// filter path, and with `required = true` a panic there does not fail one
/// command, it aborts every git operation in the repository until someone
/// unregisters the driver by hand. A lookup that cannot name a section has no
/// answer, and saying so is the same thing this function already does two lines
/// further down for the same key.
#[must_use]
pub fn get(config: &File, key: &str) -> Option<String> {
    // `gix-config`'s own rule, not a second spelling of it: a guard that drifted
    // narrower than the one that panics would leave the panic reachable.
    gix_config::AsKey::try_as_key(&key)?;

    if let Ok(value) = config.raw_value(key) {
        // An explicit value, the empty string included. Git reads `key =` as
        // false, so it must not be turned into a spelling of true below.
        return Some(value.to_string());
    }

    let (section_key, name) = key.rsplit_once('.')?;
    let (section, subsection) = match section_key.split_once('.') {
        Some((section, subsection)) => (section, Some(subsection.as_bytes().as_bstr())),
        None => (section_key, None),
    };

    let present = config
        .sections_by_name(section)?
        .filter(|section| section.header().subsection_name() == subsection)
        .any(|section| section.value_names().any(|value_name| value_name == name));

    present.then(|| "true".to_string())
}

/// Whether a value is one of git's spellings of true.
///
/// Git accepts `1`, `yes` and `on` beside `true`, case insensitively. Every
/// caller that branches on a git boolean has to accept the same set, or a
/// perfectly ordinary `required = 1` reads as "off".
///
/// The empty string is **not** in the set. Git reads `key =` as `false`
/// (measured with `git config --type=bool` on 2.55), and the value-less
/// `key` that git does read as true never arrives here as empty — [`get`]
/// returns it as `"true"`.
#[must_use]
pub fn is_true(value: &str) -> bool {
    matches!(
        value.to_ascii_lowercase().as_str(),
        "true" | "yes" | "on" | "1"
    )
}

#[cfg(test)]
mod tests {
    use std::ffi::OsStr;

    use super::*;
    use tempfile::TempDir;

    /// The five shapes of `core.attributesFile`, as measured on git 2.55.
    ///
    /// The scenario in `tests/attributes.rs` proves the two that cost a file;
    /// this proves the whole table, including the two that must resolve to
    /// **nothing**. Over-eager resolution is its own failure mode: with
    /// `required = true`, a global file we invent and the user does not have
    /// would refuse operations in every repository on the machine.
    #[test]
    fn the_global_attributes_file_resolves_where_git_resolves_it() {
        let home = Path::new("/home/user");
        let resolve = |contents: &str, xdg: Option<&str>| {
            let file = File::try_from(contents).expect("the fixture is valid configuration");
            global_attributes_file_for(&file, Some(home), xdg.map(std::ffi::OsStr::new))
        };

        assert_eq!(
            resolve("[core]\n", None),
            Some(home.join(".config").join("git").join("attributes")),
            "unset must fall back to the XDG default, which is where git looks"
        );
        assert_eq!(
            resolve("[core]\n", Some("/xdg")),
            Some(Path::new("/xdg").join("git").join("attributes")),
            "XDG_CONFIG_HOME must win over the $HOME/.config default"
        );
        assert_eq!(
            resolve("[core]\n", Some("")),
            Some(home.join(".config").join("git").join("attributes")),
            "git treats an empty XDG_CONFIG_HOME as unset, so this must too"
        );
        assert_eq!(
            resolve("[core]\n\tattributesFile = ~/attrs\n", None),
            Some(home.join("attrs")),
            "`~/` must be expanded, exactly as git expands `core.excludesFile`"
        );
        assert_eq!(
            resolve(
                "[core]\n\tattributesFile = /elsewhere/attrs\n",
                Some("/xdg")
            ),
            Some(PathBuf::from("/elsewhere/attrs")),
            "an absolute path must be taken as written, XDG or no XDG"
        );
        assert_eq!(
            resolve("[core]\n\tattributesFile = \n", Some("/xdg")),
            None,
            "an empty value turns the file off — and does **not** fall back to XDG"
        );
    }

    #[test]
    fn an_empty_value_is_false_to_git_and_must_be_false_here() {
        // Measured on git 2.55: `git config --type=bool` reads `key` (no `=`) as
        // `true` and `key = ` as `false`. `gix-config` reports the two
        // identically once they are flattened to a string, so this is the one
        // place the difference can be kept. Getting it wrong let
        // `filter.git-xcrypt.required = ` read as enabled while git ignored the
        // failing filter and stored the plaintext, and `status` saw no gap.
        let dir = TempDir::new().expect("temporary directory");
        let path = dir.path().join("config");
        std::fs::write(
            &path,
            "[filter \"git-xcrypt\"]\n\trequired = \n[core]\n\tautocrlf =   \n",
        )
        .expect("writing must succeed");

        let config = open_local(&path).expect("valid config");
        for key in ["filter.git-xcrypt.required", "core.autocrlf"] {
            let value = get(&config, key).unwrap_or_else(|| panic!("{key} must read as present"));
            assert!(
                !is_true(&value),
                "{key} = `{value}` was taken for true, which git does not"
            );
        }
    }

    /// `git -c` overrides, in the shape git 2.55 actually hands them over.
    ///
    /// The line below is not invented: it is what
    /// `git -c alias.showenv='!printenv GIT_CONFIG_PARAMETERS' -c core.autocrlf=true
    /// -c core.eol=lf -c "user.name=a b'c" showenv` printed on 2.55, copied
    /// verbatim. Three constructs have to survive it — key and value quoted
    /// separately, a `!` escaped outside the quotes for csh, and a value holding
    /// both a space and a quote.
    #[test]
    fn overrides_from_the_command_line_are_read_the_way_git_writes_them() {
        let measured = r"'alias.showenv'=''\!'printenv GIT_CONFIG_PARAMETERS' 'core.autocrlf'='true' 'core.eol'='lf' 'user.name'='a b'\''c'";
        let config = cli_overrides(Some(OsStr::new(measured))).expect("the line names four keys");

        assert_eq!(get(&config, "core.autocrlf").as_deref(), Some("true"));
        assert_eq!(get(&config, "core.eol").as_deref(), Some("lf"));
        assert_eq!(get(&config, "user.name").as_deref(), Some("a b'c"));
        assert_eq!(
            get(&config, "alias.showenv").as_deref(),
            Some("!printenv GIT_CONFIG_PARAMETERS"),
            "a value may hold spaces and an escaped bang; splitting on either \
             would invent a key nobody typed"
        );

        // Older git quoted the whole pair instead. Same words, same answer.
        let old = cli_overrides(Some(OsStr::new("'core.autocrlf=input'"))).expect("one key");
        assert_eq!(get(&old, "core.autocrlf").as_deref(), Some("input"));

        // `git -c core.autocrlf` with no `=` is git's boolean shorthand, and it
        // has to arrive spelled the way a value-less line in a file arrives.
        let bare = cli_overrides(Some(OsStr::new("'core.autocrlf'"))).expect("one key");
        assert_eq!(get(&bare, "core.autocrlf").as_deref(), Some("true"));

        // An explicitly empty value is false to git, and must not be promoted.
        let empty = cli_overrides(Some(OsStr::new("'core.autocrlf'=''"))).expect("one key");
        let value = get(&empty, "core.autocrlf").expect("present");
        assert!(!is_true(&value), "`-c core.autocrlf=` is false to git");
    }

    /// Everything this parser cannot read has to leave the files in charge.
    ///
    /// This is the half that decides whether the feature is safe to have at all.
    /// It runs on the clean path, where `required = true` turns a wrong answer
    /// into a repository that refuses every git operation — so a value nobody
    /// typed is strictly worse than a value we failed to notice. Each row below
    /// is a shape that must produce *no* override rather than a guessed one.
    #[test]
    fn anything_unreadable_leaves_the_configuration_files_to_speak() {
        assert!(cli_overrides(None).is_none(), "unset means no overrides");
        assert!(cli_overrides(Some(OsStr::new(""))).is_none());
        assert!(cli_overrides(Some(OsStr::new("   "))).is_none());
        assert!(
            cli_overrides(Some(OsStr::new("'core.autocrlf'='true"))).is_none(),
            "an unterminated quote drops the whole variable: half a word could \
             name a key nobody asked for"
        );
        assert!(
            cli_overrides(Some(OsStr::new("notdotted=1"))).is_none(),
            "a word that is not a dotted key is skipped, not guessed at"
        );

        // A skipped word must not take its neighbours with it.
        let mixed =
            cli_overrides(Some(OsStr::new("notdotted=1 'core.eol'='crlf'"))).expect("one key");
        assert_eq!(get(&mixed, "core.eol").as_deref(), Some("crlf"));
        // Asked of the rendered file rather than through `get`, which asserts on
        // a key with no dot in it — the very shape being skipped here.
        assert!(
            !mixed.to_bstring().to_str_lossy().contains("notdotted"),
            "a word that names no section must not reach the configuration"
        );
    }

    /// A key `gix-config` will not parse must answer `None`, not abort the process.
    ///
    /// `raw_value` panics on one — `'notdotted' is not a valid configuration key`
    /// — and this function is on the filter path, where with `required = true` a
    /// panic does not fail one command: it aborts every git operation in the
    /// repository until the driver is unregistered by hand. No caller passes such
    /// a key today, which is exactly why this is worth pinning; nothing else
    /// would notice if one started to.
    #[test]
    fn a_key_that_names_no_section_has_no_value_and_does_not_panic() {
        let config = File::try_from("[core]\n\tautocrlf = true\n")
            .expect("the fixture is valid configuration");

        for key in ["notdotted", "", "."] {
            assert_eq!(
                get(&config, key),
                None,
                "`{key}` cannot name a value, and answering that must not cost a panic"
            );
        }

        // The rule is borrowed from `gix-config` rather than restated, so the
        // ordinary key beside it still has to work.
        assert_eq!(get(&config, "core.autocrlf").as_deref(), Some("true"));
    }
}