razeri-plugin-contract 0.21.0

// Copyright (C) 2026 Gabriel Shea
// Licensed under either of the Apache License, Version 2.0 or the MIT
// license, at your option.

//! **The one derivation of a surrogate key**, composed from the vocabulary a
//! dialect states.
//!
//! The key is **XXH64, seed 0, of the UTF-8 bytes of the value as the table
//! stores it**, read as a signed 64-bit integer. Nothing trims, lowercases or
//! replaces anything on the way in: a column that wants its value cleaned says
//! so in the design, the load that writes the table cleans it, and the key is
//! the hash of what the table then holds.
//!
//! [`SidSpelling`] says how one engine joins strings, casts a value to text,
//! writes a replacement and spells a 64-bit hash. This module turns those
//! answers into the expression a warehouse runs:
//!
//! ```text
//! guard( hash( part₁ ) )                                  one column
//! guard( hash( escape(part₁) ‖ '|' ‖ … ‖ escape(partₙ) ) )  two or more
//! ```
//!
//! It is published so that a dialect author can render their own key
//! expression and run it. Without it, *"the same key on every engine"* is a
//! claim the person making it cannot check — and a key that differs by one
//! character keys a warehouse to numbers nothing else reproduces.
//!
//! # What a dialect decides, and what it does not
//!
//! The **vocabulary** is the dialect's, and it is the only thing that varies.
//! The **derivation** is fixed: there is no table of engine names here and no
//! way to add one, so a third party spells their own engine's key by stating a
//! [`SidSpelling`] and nothing else.
//!
//! # The parts of the derivation
//!
//! - each **part** is the key column cast to text, and nothing more;
//! - a key of **one** column is hashed as that text, with nothing done to it;
//! - a key of **two or more** columns **escapes** each part — every `\`
//!   doubled, then a `\` put before every `|` ([`escape_part`]) — and joins
//!   the parts with the literal `|`. Without the escape `('a|b', 'c')` and
//!   `('a', 'b|c')` are one text and one key. A part holding neither character
//!   is its own text, so its key is the number it always was;
//! - **guard** is the [`NULL_SID`] sentinel, and it fires when **any** part is
//!   NULL: any NULL component keys the whole member to the sentinel;
//! - **hash** is either this engine's own spelling or a call to the
//!   [`HELPER_FUNCTION_NAME`] function, whose body is that same spelling over
//!   its argument as given. The two sites are one derivation — see
//!   [`HashSite`].

/// **The sentinel a member with any NULL key part is keyed to**, `(-2^63) + 2`.
///
/// One number, on every engine, for one reason: a NULL natural key is an
/// absence rather than a value, and two absences are not the same member. It
/// sits two above the minimum so that a warehouse can still reserve the
/// minimum itself.
pub const NULL_SID: i64 = -9_223_372_036_854_775_806_i64;

/// **The name of the surrogate-key function — the only one.**
///
/// One public name everywhere, whoever defines it. On engines where the hash is
/// an ordinary SQL expression, Razeri emits a function of this name whose body
/// is the derivation below. On engines where the hash lives outside SQL — a
/// compiled library, an assembly, a registered handler — the dialect's
/// prerequisites install an artifact under this same name and
/// [`SidSpelling::hash`] names it directly, so nothing emits a second
/// definition.
///
/// A second name is a second contract. There is no alias and no transitional
/// spelling.
pub const HELPER_FUNCTION_NAME: &str = "generate_sid";

use crate::ExpressionRenderer;
use crate::sid_spelling::{Concat, SidSpelling, TextCast, Trim};

/// Where the hash itself is spelled.
#[derive(Debug, Clone, Copy)]
pub enum HashSite<'a> {
    /// Inline this engine's own hash spelling, from [`SidSpelling::hash`].
    Native,
    /// Inline this engine's own hash spelling, in the schema Razeri installed
    /// it in — [`SidPrerequisiteHome::Schema`](crate::SidPrerequisiteHome).
    ///
    /// The schema, a dot, and then [`SidSpelling::hash`] exactly as [`Native`]
    /// writes it. So a dialect that names a home owes a `hash` whose opening
    /// text BEGINS with the function's own name: `razeri_xxh64(` qualifies to
    /// `razeri.razeri_xxh64(`, while a spelling that opened with a cast —
    /// `CAST(xxHash64(` — would put the schema in front of the cast and name
    /// nothing. Every engine that opens with a cast has a server-global
    /// function and names no home, which is why the two facts have never met.
    ///
    /// [`Native`]: Self::Native
    NativeIn(&'a str),
    /// Call the [`HELPER_FUNCTION_NAME`] function by this — possibly
    /// schema-qualified — name. Its body is the [`HashSite::Native`] spelling,
    /// so the two sites are the same derivation.
    ///
    /// **The qualifier is the function's home, not the caller's schema.** One
    /// function is created per deployment unit, so a call from any other schema
    /// of that unit has to name the schema that has it. A caller that qualifies
    /// with its own schema names a function nothing created.
    Helper(&'a str),
}

/// **`TRIM(expr)` in this engine's spelling** — the space around a value
/// taken off, and nothing else.
///
/// Public because taking the space off a value the source wrote is asked in
/// more than one place — reading an amount the source spelled its own way,
/// and a column's `trim` cleaning step — and two spellings of `TRIM` would be
/// two answers to one question. The key itself trims nothing.
///
/// Parameters:
/// - `spelling`: this engine's vocabulary;
/// - `expr`: an already-rendered value expression.
#[must_use]
pub fn normalize_trim(spelling: SidSpelling, expr: &str) -> String {
    match spelling.trim {
        Trim::Function(name) => format!("{name}({expr})"),
        Trim::LtrimRtrim => format!("LTRIM(RTRIM({expr}))"),
        Trim::FunctionWithSpaceSet(name) => format!("{name}({expr}, ' ')"),
    }
}

/// **`CAST(expr AS text)` in this engine's spelling**, total over NULL.
///
/// Public for the reason [`normalize_trim`] is public: turning a value into
/// text is not something only the key derivation does. A caller that has to
/// read a value back out of an engine as text — because the transport it
/// arrives over encodes the engine's own types in a form the same engine will
/// not accept back — needs this engine's cast and not a second opinion about
/// it. Two spellings of one cast are two answers to one question.
///
/// Parameters:
/// - `spelling`: this engine's vocabulary;
/// - `expr`: an already-rendered value expression.
#[must_use]
pub fn cast_to_text(spelling: SidSpelling, expr: &str) -> String {
    match spelling.cast_type {
        TextCast::Direct(ty) => format!("CAST({expr} AS {ty})"),
        TextCast::ClickHouseNullTolerant(ty) => {
            format!("assumeNotNull(CAST({expr} AS Nullable({ty})))")
        }
    }
}

/// **How an engine writes a backslash between quotes** — the one thing about
/// its string literals the escape inside a composite key has to know.
///
/// The escape writes `'\'`, `'\\'` and `'\|'` as literals, and engines
/// disagree about what those are. Standard SQL reads a backslash in a literal
/// as itself; some engines read it as the start of an escape, so one
/// backslash is written as two there. A literal written the wrong way is a
/// different string, or no string at all, and the key it makes is a number
/// no other engine computes.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Backslash {
    /// `'\'` is one backslash: standard SQL, and the default.
    AsWritten,
    /// `'\\'` is one backslash: the engine reads a backslash in a literal
    /// as an escape.
    Doubled,
}

impl Backslash {
    /// **This engine's answer, read from the engine's own words.**
    ///
    /// The one place every dialect already says how it writes a backslash
    /// between quotes is the list of words it reads as a missing value in a
    /// file ([`ExpressionRenderer::render_missing_value`]): an engine whose
    /// literals read a backslash as an escape writes each one twice there, so
    /// that a word holding one stays that word. So the dialect is asked for a
    /// word that is one backslash, and the answer is read off what it writes.
    /// A dialect that writes the word as standard SQL does is
    /// [`Backslash::AsWritten`].
    ///
    /// Parameters:
    /// - `dialect`: the engine the key is written for.
    #[must_use]
    pub fn of(dialect: &dyn ExpressionRenderer) -> Self {
        if dialect.render_missing_value("v", &["\\"]).contains("'\\\\'") {
            Self::Doubled
        } else {
            Self::AsWritten
        }
    }

    /// **`text` as a string literal on this engine**: between single quotes,
    /// each backslash written the way this engine reads one, behind `prefix`,
    /// the engine's Unicode-literal prefix
    /// ([`Lower::pattern_literal_prefix`](crate::Lower::pattern_literal_prefix)).
    /// A quote in `text` is doubled.
    ///
    /// Parameters:
    /// - `prefix`: what this engine puts in front of a Unicode literal.
    /// - `text`: the literal's characters.
    #[must_use]
    pub fn literal(self, prefix: &str, text: &str) -> String {
        let quoted = text.replace('\'', "''");
        match self {
            Self::AsWritten => format!("{prefix}'{quoted}'"),
            Self::Doubled => format!("{prefix}'{}'", quoted.replace('\\', "\\\\")),
        }
    }
}

/// **One part of a composite key, escaped in this engine's words**: every
/// `\` doubled, then a `\` put before every `|` — two replacements, the
/// backslash first, so the second one's own backslashes are not doubled
/// again. The subject is pinned to the engine's binary collation where its
/// key spelling names one
/// ([`Lower::match_collation`](crate::Lower::match_collation)), so the
/// replacement matches by the character.
///
/// Parameters:
/// - `part`: the part, already cast to text.
/// - `spelling`: this engine's vocabulary — its `REPLACE`, its literal prefix
///   and its binary collation.
/// - `backslash`: how this engine writes a backslash between quotes.
#[must_use]
pub fn escape_part(part: &str, spelling: SidSpelling, backslash: Backslash) -> String {
    let words = spelling.lower;
    let literal = |text: &str| backslash.literal(words.pattern_literal_prefix, text);
    let subject = match words.match_collation {
        Some(collation) => format!("{part} COLLATE {collation}"),
        None => part.to_owned(),
    };
    let doubled = format!("{}({subject}, {}, {})", words.replace, literal("\\"), literal("\\\\"));
    format!("{}({doubled}, {}, {})", words.replace, literal("|"), literal("\\|"))
}

/// **The join, written for an engine whose `CONCAT` takes exactly two
/// arguments**: the parts and the separators between them, folded left to
/// right, one call per pair. The same items in the same order as the
/// variadic call, so the joined string is the same string.
///
/// An empty key renders the bare call, as the variadic form does.
fn nest_pairwise(name: &str, parts: &[String]) -> String {
    let mut rest = parts.iter();
    let Some(first) = rest.next() else { return format!("{name}()") };
    rest.fold(first.clone(), |left, next| format!("{name}({name}({left}, '|'), {next})"))
}

/// **The derivation**: one part hashed as given; two or more parts each
/// escaped ([`escape_part`]), joined on `'|'`, and hashed.
///
/// `text_args` are already-rendered value expressions, already cast to text
/// where the caller needs it — a column reference inside a load statement, a
/// parameter name inside the key function's own body. Callers name the value
/// differently, so composing the name here would mean knowing which caller was
/// asking.
///
/// The result carries **no NULL guard**; [`guard_null_members`] wraps this.
///
/// Parameters:
/// - `text_args`: the key's parts, already cast to text;
/// - `spelling`: this engine's vocabulary;
/// - `backslash`: how this engine writes a backslash between quotes
///   ([`Backslash::of`]);
/// - `site`: whether the hash is inlined or reached through the function.
#[must_use]
pub fn sid_derivation(
    text_args: &[String],
    spelling: SidSpelling,
    backslash: Backslash,
    site: HashSite<'_>,
) -> String {
    let joined = if text_args.len() == 1 {
        text_args[0].clone()
    } else {
        let parts: Vec<String> =
            text_args.iter().map(|part| escape_part(part, spelling, backslash)).collect();
        match spelling.concat {
            Concat::Function(name) => format!("{name}({})", parts.join(", '|', ")),
            Concat::FunctionBinary(name) => nest_pairwise(name, &parts),
            Concat::Plus => parts.join(" + '|' + "),
            Concat::Operator(operator) => parts.join(&format!(" {operator} '|' {operator} ")),
        }
    };
    match site {
        HashSite::Native => format!("{}{joined}{}", spelling.hash.0, spelling.hash.1),
        HashSite::NativeIn(home) => {
            format!("{home}.{}{joined}{}", spelling.hash.0, spelling.hash.1)
        }
        HashSite::Helper(name) => format!("{name}({joined})"),
    }
}

/// **The [`NULL_SID`] sentinel, wrapped around the derivation.**
///
/// `raw_args` are the arguments **before** any cast — a cast preserves
/// NULL-ness, and `CAST(x AS TEXT) IS NULL` only makes the guard harder to
/// read.
///
/// **Any NULL component of a composite keys the whole member to the sentinel.**
/// That closes a collision as well as an omission: on several engines
/// `CONCAT('x', '|', NULL)` is `'x|'`, byte-identical to
/// `CONCAT('x', '|', '')`, so `('boston', NULL)` and `('boston', '')` would be
/// one member. Testing the components before they are joined is the only place
/// the distinction still exists.
///
/// The guard is a `CASE WHEN` rather than a `COALESCE` because it tests the
/// **input**. A `COALESCE` over the hash's *output* reaches the sentinel only
/// where the hash propagates NULL, and on at least one engine the hash of NULL
/// is a perfectly ordinary number.
///
/// An empty key renders no guard rather than an empty one.
///
/// Parameters:
/// - `raw_args`: the key's parts as the caller wrote them, before any cast;
/// - `derivation`: the expression [`sid_derivation`] produced.
#[must_use]
pub fn guard_null_members(raw_args: &[String], derivation: String) -> String {
    if raw_args.is_empty() {
        return derivation;
    }
    let test = raw_args.iter().map(|arg| format!("{arg} IS NULL")).collect::<Vec<_>>().join(" OR ");
    format!("CASE WHEN {test} THEN {NULL_SID} ELSE {derivation} END")
}

/// **The surrogate key as one expression**: cast, derive, guard.
///
/// This is the whole composition, and it is what every producer of a key
/// expression emits — the load that keys a dimension, the body of the
/// per-entity key function, the formula a design can name. Rendering it for
/// your own engine and running it is how you check that your
/// [`SidSpelling`] reaches the number every other engine reaches.
///
/// ```
/// use razeri_plugin_contract::{ANSI_SID_SPELLING, sid};
///
/// let key = sid::sid_expression(
///     ANSI_SID_SPELLING,
///     sid::Backslash::AsWritten,
///     None,
///     &["t.city".to_owned()],
/// );
/// assert!(key.starts_with("CASE WHEN t.city IS NULL THEN "));
/// assert!(key.contains("generate_sid(CAST(t.city AS TEXT))"));
/// ```
///
/// # `helper_home` is the key function's schema, never the caller's
///
/// The only thing a schema does here is qualify the call to
/// [`HELPER_FUNCTION_NAME`] on the engines whose
/// [`SidSpelling::through_helper`] is true. Where the key function is an
/// installed artifact instead, its own name is rendered and this argument is
/// ignored entirely — on at least one engine a compiled function lives server
/// globally and a qualified call to it does not resolve at all.
///
/// So the schema this wants is the one the function was **created** in, which
/// is a property of the deployment unit rather than of whichever entity is
/// being keyed. `None` renders the call unqualified, which is correct wherever
/// nothing is qualified: an engine without schemas, or a design that names no
/// schema.
///
/// Parameters:
/// - `spelling`: this engine's vocabulary;
/// - `backslash`: how this engine writes a backslash between quotes
///   ([`Backslash::of`]);
/// - `helper_home`: the schema the key function was created in, where the call
///   needs qualifying;
/// - `parts`: the key's parts as raw value expressions, in order.
#[must_use]
pub fn sid_expression(
    spelling: SidSpelling,
    backslash: Backslash,
    helper_home: Option<&str>,
    parts: &[String],
) -> String {
    let qualified = helper_home
        .map_or_else(|| HELPER_FUNCTION_NAME.to_owned(), |s| format!("{s}.{HELPER_FUNCTION_NAME}"));
    sid_expression_through(spelling, backslash, &qualified, parts)
}

/// **The same expression, with the helper's name already spelled by the
/// target.**
///
/// [`sid_expression`] writes the helper's name the way most engines read it —
/// the schema, a dot, the name, all bare. That is a name on every engine that
/// leaves an unquoted word alone and a *different* name on one that folds it,
/// where the function was created under the spelling that engine's own DDL
/// wrote. A caller holding the dialect spells the name first and hands it in
/// here, so the call and the `CREATE FUNCTION` say the same word.
///
/// Parameters:
/// - `spelling`: this engine's vocabulary;
/// - `backslash`: how this engine writes a backslash between quotes;
/// - `helper`: the key helper's name, qualified and spelled;
/// - `parts`: the key's parts as raw value expressions, in order.
#[must_use]
pub fn sid_expression_through(
    spelling: SidSpelling,
    backslash: Backslash,
    helper: &str,
    parts: &[String],
) -> String {
    let cast: Vec<String> = parts.iter().map(|arg| cast_to_text(spelling, arg)).collect();
    let site = if spelling.through_helper { HashSite::Helper(helper) } else { HashSite::Native };
    guard_null_members(parts, sid_derivation(&cast, spelling, backslash, site))
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::sid_spelling::{ANSI_SID_SPELLING, Lower};

    /// The methods a test renderer has to answer, and nothing else.
    macro_rules! for_test_renderer {
        ($name:ident) => {
            impl crate::SqlRenderer for $name {
                fn render_data_type(&self, _: &crate::DataType) -> String {
                    String::new()
                }
                fn quote_identifier(&self, name: &str) -> String {
                    name.to_owned()
                }
                fn qualify_name(&self, _: Option<&str>, name: &str) -> String {
                    name.to_owned()
                }
                fn render_current_timestamp(&self) -> String {
                    String::new()
                }
            }
        };
    }

    /// A vocabulary for the composition tests: a fixture, not an engine.
    const HELPER_ROUTED: SidSpelling =
        SidSpelling { platform: "test", hash: ("pg_xxhash64(", ")"), ..ANSI_SID_SPELLING };

    /// Standard SQL's backslash, for the composition tests.
    const STD: Backslash = Backslash::AsWritten;

    /// `a` and `b`, each escaped as standard SQL writes it.
    const A_B: &str = "REPLACE(REPLACE(a, '\\', '\\\\'), '|', '\\|'), '|', \
                       REPLACE(REPLACE(b, '\\', '\\\\'), '|', '\\|')";

    /// **Nothing is done to a one-column key before it is hashed**: no
    /// `TRIM`, no `LOWER`, no `REPLACE` — the cast and the hash, and nothing
    /// between.
    #[test]
    fn a_part_is_hashed_as_it_is_cast() {
        assert_eq!(
            sid_expression(HELPER_ROUTED, STD, None, &["x".to_owned()]),
            format!("CASE WHEN x IS NULL THEN {NULL_SID} ELSE generate_sid(CAST(x AS TEXT)) END")
        );
        let native = SidSpelling { through_helper: false, ..HELPER_ROUTED };
        for spelling in [HELPER_ROUTED, native] {
            for backslash in [STD, Backslash::Doubled] {
                let rendered = sid_expression(spelling, backslash, Some("acme"), &["x".to_owned()]);
                for fold in ["TRIM(", "LOWER(", "REPLACE("] {
                    assert!(!rendered.contains(fold), "{fold} in {rendered}");
                }
            }
        }
    }

    /// **Each part of a composite is escaped before the join**: every `\`
    /// doubled, then a `\` put before every `|`, in this engine's `REPLACE`,
    /// its literal prefix, its binary collation and its way of writing a
    /// backslash.
    #[test]
    fn a_composite_escapes_each_part_before_the_join() {
        let native = SidSpelling { through_helper: false, ..HELPER_ROUTED };
        assert_eq!(
            sid_derivation(&["a".to_owned(), "b".to_owned()], native, STD, HashSite::Native),
            format!("pg_xxhash64(CONCAT({A_B}))")
        );
        assert_eq!(
            escape_part("a", ANSI_SID_SPELLING, Backslash::Doubled),
            "REPLACE(REPLACE(a, '\\\\', '\\\\\\\\'), '|', '\\\\|')"
        );
        let tsql_like = SidSpelling {
            lower: Lower {
                function: "LOWER",
                replace: "REPLACE",
                pattern_literal_prefix: "N",
                match_collation: Some("Latin1_General_100_BIN2"),
            },
            ..ANSI_SID_SPELLING
        };
        assert_eq!(
            escape_part("a", tsql_like, STD),
            "REPLACE(REPLACE(a COLLATE Latin1_General_100_BIN2, N'\\', N'\\\\'), N'|', N'\\|')"
        );
        let clickhouse_like = SidSpelling {
            lower: Lower { replace: "replaceAll", ..ANSI_SID_SPELLING.lower },
            ..ANSI_SID_SPELLING
        };
        assert!(
            escape_part("a", clickhouse_like, Backslash::Doubled)
                .starts_with("replaceAll(replaceAll(a, "),
            "the engine's own replacement function"
        );
    }

    /// **An engine's backslash is read from its own missing-value words.**
    #[test]
    fn the_backslash_is_read_from_the_dialects_own_literal() {
        struct Escaping;
        struct Standard;
        for_test_renderer!(Escaping);
        for_test_renderer!(Standard);
        impl ExpressionRenderer for Standard {}
        impl ExpressionRenderer for Escaping {
            fn render_missing_value(&self, text: &str, words: &[&str]) -> String {
                let listed: Vec<String> =
                    words.iter().map(|w| format!("'{}'", w.replace('\\', "\\\\"))).collect();
                format!("CASE WHEN {text} IN ({}) THEN NULL ELSE {text} END", listed.join(", "))
            }
        }
        assert_eq!(Backslash::of(&Escaping), Backslash::Doubled);
        assert_eq!(Backslash::of(&Standard), Backslash::AsWritten);
        assert_eq!(Backslash::Doubled.literal("", "\\|"), "'\\\\|'");
        assert_eq!(Backslash::AsWritten.literal("N", "\\|"), "N'\\|'");
        assert_eq!(Backslash::AsWritten.literal("", "O'Hare"), "'O''Hare'");
    }

    /// **The call through the key function and the inlined spelling are one
    /// derivation** over the same parts.
    #[test]
    fn the_helper_site_and_the_native_site_hash_the_same_text() {
        let args = ["a".to_owned(), "b".to_owned()];
        assert_eq!(
            sid_derivation(&args, HELPER_ROUTED, STD, HashSite::Native),
            format!("pg_xxhash64(CONCAT({A_B}))")
        );
        assert_eq!(
            sid_derivation(&args, HELPER_ROUTED, STD, HashSite::Helper(HELPER_FUNCTION_NAME)),
            format!("generate_sid(CONCAT({A_B}))")
        );
        assert_eq!(
            sid_derivation(&args, HELPER_ROUTED, STD, HashSite::NativeIn("razeri")),
            format!("razeri.pg_xxhash64(CONCAT({A_B}))")
        );
    }

    /// **The four join forms are four spellings of one string**: the same
    /// items, in the same order, with the same separators.
    #[test]
    fn the_four_join_forms_join_the_same_items() {
        let three = ["a".to_owned(), "b".to_owned(), "c".to_owned()];
        let with = |concat| {
            sid_derivation(&three, SidSpelling { concat, ..HELPER_ROUTED }, STD, HashSite::Native)
        };
        let [a, b, c] = three.clone().map(|part| escape_part(&part, HELPER_ROUTED, STD));
        assert_eq!(
            with(Concat::Function("CONCAT")),
            format!("pg_xxhash64(CONCAT({a}, '|', {b}, '|', {c}))")
        );
        assert_eq!(
            with(Concat::FunctionBinary("CONCAT")),
            format!("pg_xxhash64(CONCAT(CONCAT(CONCAT(CONCAT({a}, '|'), {b}), '|'), {c}))")
        );
        assert_eq!(with(Concat::Plus), format!("pg_xxhash64({a} + '|' + {b} + '|' + {c})"));
        assert_eq!(
            with(Concat::Operator("||")),
            format!("pg_xxhash64({a} || '|' || {b} || '|' || {c})")
        );
        let one = ["a".to_owned()];
        assert_eq!(
            sid_derivation(
                &one,
                SidSpelling { concat: Concat::Plus, ..HELPER_ROUTED },
                STD,
                HashSite::Native
            ),
            "pg_xxhash64(a)",
            "a one-part key is neither escaped nor joined"
        );
        assert_eq!(
            sid_derivation(
                &[],
                SidSpelling { concat: Concat::FunctionBinary("CONCAT"), ..HELPER_ROUTED },
                STD,
                HashSite::Native
            ),
            "pg_xxhash64(CONCAT())"
        );
    }

    /// **Any NULL part keys the whole member to the sentinel**, the guard
    /// testing the parts before the cast; an empty key renders no guard.
    #[test]
    fn any_null_part_reaches_the_sentinel() {
        let composite = sid_expression(HELPER_ROUTED, STD, None, &["a".to_owned(), "b".to_owned()]);
        assert!(composite.starts_with("CASE WHEN a IS NULL OR b IS NULL THEN "), "{composite}");
        assert!(!sid_expression(HELPER_ROUTED, STD, None, &[]).contains("CASE WHEN"));
    }

    /// **A spelling that does not route through the key function ignores the
    /// schema.** An engine whose key function is an installed artifact
    /// renders a bare call: the artifact is registered where that engine
    /// registers artifacts, not in a schema Razeri created.
    #[test]
    fn a_spelling_that_inlines_its_artifact_ignores_the_schema_it_is_passed() {
        let external =
            SidSpelling { hash: ("generate_sid(", ")"), through_helper: false, ..HELPER_ROUTED };
        let qualified = sid_expression(external, STD, Some("acme"), &["x".to_owned()]);
        let bare = sid_expression(external, STD, None, &["x".to_owned()]);
        assert_eq!(qualified, bare);
        assert!(!qualified.contains("acme."), "the emitted call qualifies an installed artifact");
    }

    /// **An engine that joins with an operator gets one expression**, the
    /// operator between each part and the separator, left to right; a
    /// one-part key is not joined on this form either.
    #[test]
    fn an_operator_join_writes_the_operator_between_each_part_and_the_separator() {
        let spelling = SidSpelling { concat: Concat::Operator("||"), ..HELPER_ROUTED };
        let three = ["a".to_owned(), "b".to_owned(), "c".to_owned()];
        let [a, b, c] = three.clone().map(|part| escape_part(&part, spelling, STD));
        assert_eq!(
            sid_derivation(&three, spelling, STD, HashSite::Native),
            format!("pg_xxhash64({a} || '|' || {b} || '|' || {c})")
        );
        assert_eq!(
            sid_derivation(&["a".to_owned()], spelling, STD, HashSite::Native),
            "pg_xxhash64(a)"
        );
    }

    /// The sentinel is two above the smallest 64-bit integer, and it is spelled
    /// once.
    #[test]
    fn the_sentinel_is_two_above_the_smallest_integer() {
        assert_eq!(NULL_SID, i64::MIN + 2);
    }
}