vyre_libs/scan/

regex_compile.rs

//! Regex AST → `NfaPlan` frontend.
//!
//! `nfa::compile` ships a literal-only NFA (one byte per state). This
//! module is its regex-aware counterpart: parse a regex string with
//! `regex-syntax`, lower the AST into a Thompson NFA over byte
//! transitions, emit the same `(NfaPlan, transition_table,
//! epsilon_table)` triple the literal compiler produces.
//!
//! # Why a separate module instead of widening `nfa::compile`
//!
//! The literal compiler is hot-path simple  -  every byte is a single
//! state. Bolting alternation / repetition / character classes onto it
//! would either bloat the literal path or fork the construction code.
//! The lego-block fix is a SECOND construction module that emits the
//! SAME output shape, so every downstream component (`nfa_scan`
//! Program, `mega_scan::build`, `RulePipeline`) works unmodified.
//!
//! # Supported regex subset
//!
//! Targets the ~85% of vyre's expected detector regex shapes:
//!
//!   - Concatenation (default)
//!   - Alternation `a|b`
//!   - Character classes `[abc]`, `[a-z]`, `[^abc]`
//!   - Builtin escapes `\d \D \w \W \s \S` (ASCII semantics)
//!   - Bounded repetition `*`, `+`, `?`, `{n}`, `{n,m}`
//!   - Text anchors `^` and `$`
//!   - Escape literals `\.`, `\\`, `\(`, `\[`
//!
//! Explicitly NOT supported (returns `RegexCompileError::Unsupported`):
//!
//!   - Backreferences `\1` (NFA cannot represent)
//!   - Word-boundary and line-boundary lookarounds
//!   - Unicode character classes outside the ASCII range

use regex_syntax::hir::{Class, Hir, HirKind, Look, Repetition};

use crate::scan::nfa::NfaPlan;

const LANES: usize = vyre_primitives::nfa::subgroup_nfa::LANES_PER_SUBGROUP;

/// Capture output mode a consumer requests for a regex scan, wiring
/// `docs/optimization/REGEX_CAPTURE_MODE_CONTRACTS.toml` to code.
///
/// A consumer routes on this instead of parsing the TOML: [`accelerator_eligible`]
/// says whether the GPU DFA/AC path can serve the request directly, and
/// [`verifier_required`] says whether a scalar (CPU-semantics) verifier must own
/// the output. The three whole-match modes run entirely on the accelerator; the
/// three group-extraction modes need the verifier (the byte-DFA has no capture
/// stack). Keeping this a typed enum with one `contract_row` owner means the
/// routing decision has one home and cannot silently disagree with the contract.
///
/// [`accelerator_eligible`]: CaptureMode::accelerator_eligible
/// [`verifier_required`]: CaptureMode::verifier_required
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum CaptureMode {
    /// Whole match only, no spans (`whole_match_only`). Accelerator path.
    NonCapture,
    /// Match count per pattern (`match_count_per_pattern`). Accelerator path.
    Count,
    /// Whole-match `(start, end)` span (`whole_match_span`). Accelerator path.
    Span,
    /// Named group span records (`named_group_span_records`). Verifier-bound;
    /// unmatched group → null.
    NamedCapture,
    /// Ordered list of spans for a repeated group (`ordered_group_span_list`).
    /// Verifier-bound; an empty repeat yields an empty list.
    RepeatedCapture,
    /// Row × group value table (`row_group_value_table`). Verifier-bound;
    /// unmatched group → null.
    GroupExtraction,
}

/// Static per-mode contract row mirroring one `[[mode]]` entry of
/// `REGEX_CAPTURE_MODE_CONTRACTS.toml`. The [`CaptureMode::contract_row`] table
/// is the single code-side owner; `regex_capture_mode_contracts.rs` locks it to
/// the TOML so the two cannot drift.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct CaptureModeContract {
    /// Stable `mode_id` string, identical to the TOML.
    pub mode_id: &'static str,
    /// `output_shape` string, identical to the TOML.
    pub output_shape: &'static str,
    /// Whether the GPU accelerator path can serve this mode directly.
    pub accelerator_eligible: bool,
    /// Whether a scalar verifier must own the output for this mode.
    pub verifier_required: bool,
    /// `null_policy` string, identical to the TOML.
    pub null_policy: &'static str,
}

impl CaptureMode {
    /// Every mode, in contract order. One owner for iteration + coherence checks.
    pub const ALL: [CaptureMode; 6] = [
        CaptureMode::NonCapture,
        CaptureMode::Count,
        CaptureMode::Span,
        CaptureMode::NamedCapture,
        CaptureMode::RepeatedCapture,
        CaptureMode::GroupExtraction,
    ];

    /// The contract row for this mode, the single code-side source of truth for
    /// its `mode_id`, `output_shape`, routing bits, and null policy.
    #[must_use]
    pub const fn contract_row(self) -> CaptureModeContract {
        match self {
            CaptureMode::NonCapture => CaptureModeContract {
                mode_id: "noncapture",
                output_shape: "whole_match_only",
                accelerator_eligible: true,
                verifier_required: false,
                null_policy: "not_applicable",
            },
            CaptureMode::Count => CaptureModeContract {
                mode_id: "count",
                output_shape: "match_count_per_pattern",
                accelerator_eligible: true,
                verifier_required: false,
                null_policy: "not_applicable",
            },
            CaptureMode::Span => CaptureModeContract {
                mode_id: "span",
                output_shape: "whole_match_span",
                accelerator_eligible: true,
                verifier_required: false,
                null_policy: "absent-match-has-no-span",
            },
            CaptureMode::NamedCapture => CaptureModeContract {
                mode_id: "named_capture",
                output_shape: "named_group_span_records",
                accelerator_eligible: false,
                verifier_required: true,
                null_policy: "unmatched-group-null",
            },
            CaptureMode::RepeatedCapture => CaptureModeContract {
                mode_id: "repeated_capture",
                output_shape: "ordered_group_span_list",
                accelerator_eligible: false,
                verifier_required: true,
                null_policy: "empty-repeat-yields-empty-list",
            },
            CaptureMode::GroupExtraction => CaptureModeContract {
                mode_id: "group_extraction",
                output_shape: "row_group_value_table",
                accelerator_eligible: false,
                verifier_required: true,
                null_policy: "unmatched-group-null",
            },
        }
    }

    /// Whether the GPU accelerator path can serve this mode directly (no
    /// verifier). The three whole-match modes are eligible; group-extraction is not.
    #[must_use]
    pub const fn accelerator_eligible(self) -> bool {
        self.contract_row().accelerator_eligible
    }

    /// Whether a scalar (CPU-semantics) verifier must own this mode's output.
    /// The exact complement of [`accelerator_eligible`](Self::accelerator_eligible)
    /// under this contract, but named separately because the two are independent
    /// contract fields, a future mode could be neither (unsupported) rather than
    /// exactly one.
    #[must_use]
    pub const fn verifier_required(self) -> bool {
        self.contract_row().verifier_required
    }

    /// Look up a mode by its stable `mode_id` string (the reverse of
    /// `contract_row().mode_id`), for consumers that receive the mode as config
    /// text. Returns `None` for an unknown id.
    #[must_use]
    pub fn from_mode_id(mode_id: &str) -> Option<CaptureMode> {
        CaptureMode::ALL
            .into_iter()
            .find(|mode| mode.contract_row().mode_id == mode_id)
    }
}

/// Failure modes for [`compile_regex_set`]. Variants are non-exhaustive
/// so future regex features can be added without a breaking change.
#[derive(Debug, Clone)]
#[non_exhaustive]
pub enum RegexCompileError {
    /// `regex-syntax` rejected the pattern. Carries the parser's own
    /// diagnostic so callers can forward it.
    Parse {
        /// Index into the input slice that failed to parse.
        pattern_index: usize,
        /// `regex-syntax`'s error message.
        message: String,
    },
    /// The pattern uses a regex feature this GPU NFA frontend does not
    /// support. Callers must reject or rewrite the detector into supported
    /// GPU-NFA rule data.
    Unsupported {
        /// Index into the input slice that uses the unsupported feature.
        pattern_index: usize,
        /// One-line description of what isn't supported (e.g. "anchors").
        feature: &'static str,
    },
    /// The compiled NFA exceeds `LANES * 32` states (the lane-major
    /// transition table addresses states with one bit per lane).
    /// Mitigation: split the pattern set across multiple pipelines.
    TooManyStates {
        /// Number of states the AST would have produced.
        states: usize,
        /// Per-pipeline maximum.
        cap: usize,
    },
    /// Pattern count does not fit the GPU ABI's `u32` pattern id field.
    PatternCountOverflow {
        /// Number of patterns supplied by the caller.
        count: usize,
    },
    /// A compiled regex match length does not fit the `u32` match ABI.
    MatchLengthOverflow {
        /// Index into the input slice that produced the oversized match.
        pattern_index: usize,
        /// Longest matched byte length for the pattern.
        len: usize,
    },
    /// Transition or epsilon table word count overflowed host `usize`.
    TableWordCountOverflow {
        /// Table being built.
        table: &'static str,
    },
    /// Compiler staging allocation failed.
    StorageReserveFailed {
        /// Scratch vector being reserved.
        field: &'static str,
        /// Requested target capacity.
        requested: usize,
        /// Allocator failure details.
        message: String,
    },
}

impl RegexCompileError {
    /// The canonical `REGEX_UNSUPPORTED_DIAGNOSTICS.toml` diagnostic code for
    /// this error, or `None` when the error does not correspond to a tracked
    /// unsupported-construct in that registry.
    ///
    /// A consumer routes on this code, e.g. a `*_REQUIRES_VERIFIER` code means
    /// "send this detector to the scalar verifier", while a `*_UNSUPPORTED_*`
    /// code means "reject or rewrite". It returns `Some` only for constructs the
    /// GPU-NFA frontend can distinctly identify AND that have a registry code
    /// today (ASCII lookaround assertions and over-budget Unicode classes); it
    /// invents no codes. `Parse` errors, state-budget overflow, and ABI-sizing
    /// failures return `None`: they are not registry constructs. As the frontend
    /// learns to distinguish more constructs (backreferences, captures, huge
    /// alternations, nested repeats), map them here against their registry codes.
    ///
    /// The `feature` strings matched below are this crate's own construction-site
    /// constants (not upstream parser text), so the mapping is stable; the
    /// `regex_compile_diagnostic_codes` test locks the real compile path to them.
    #[must_use]
    pub fn diagnostic_code(&self) -> Option<&'static str> {
        match self {
            // The `feature` strings are this crate's own construction-site
            // constants (below), so the feature→construct→code chain has ONE
            // owner: `regex_feature_construct` + `regex_construct_diagnostic_code`.
            Self::Unsupported { feature, .. } => {
                regex_feature_construct(feature).map(regex_construct_diagnostic_code)
            }
            _ => None,
        }
    }
}

/// A regex construct vyre's GPU-NFA frontend distinctly detects AND that has a
/// canonical `REGEX_UNSUPPORTED_DIAGNOSTICS.toml` diagnostic code.
///
/// This enum is the ONE owner of the construct→code mapping. Both
/// [`RegexCompileError::diagnostic_code`] (for the constructs that surface as a
/// compile error) and [`CompiledRegexSet::capture_extraction_diagnostic_code`]
/// (for the non-error capture case) route through
/// [`regex_construct_diagnostic_code`], so a code string is never written twice.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[non_exhaustive]
pub enum RegexConstruct {
    /// `\1` / `\k<name>` / `(?P=name)`: not a regular language; rejected.
    Backreference,
    /// A non-edge lookaround assertion (`\b`, `(?=…)`, …) (verifier-routed).
    Lookaround,
    /// A Unicode character class over the byte-mode GPU expansion budget.
    UnicodeClassesGpu,
    /// A capture group whose submatch spans a whole-match engine cannot prove.
    /// NOT a compile error (whole-match still accelerates; verifier-routed).
    CaptureExtraction,
    /// An alternation with more arms than the state budget can ever hold.
    HugeAlternation,
    /// Nested bounded repeats whose unroll product exceeds the state budget.
    NestedRepeats,
}

/// The canonical `REGEX_UNSUPPORTED_DIAGNOSTICS.toml` code for a construct, the
/// single source of truth for these strings.
#[must_use]
pub fn regex_construct_diagnostic_code(construct: RegexConstruct) -> &'static str {
    match construct {
        RegexConstruct::Backreference => "VYRE_SCAN_UNSUPPORTED_BACKREFERENCE",
        RegexConstruct::Lookaround => "VYRE_SCAN_APPROXIMATED_LOOKAROUND_REQUIRES_VERIFIER",
        RegexConstruct::UnicodeClassesGpu => "VYRE_SCAN_UNSUPPORTED_UNICODE_MODE_GPU",
        RegexConstruct::CaptureExtraction => "VYRE_SCAN_CAPTURE_EXTRACTION_REQUIRES_VERIFIER",
        RegexConstruct::HugeAlternation => "VYRE_SCAN_UNSUPPORTED_HUGE_ALTERNATION_BUDGET",
        RegexConstruct::NestedRepeats => "VYRE_SCAN_UNSUPPORTED_NESTED_REPEAT_BUDGET",
    }
}

// Feature strings carried by `RegexCompileError::Unsupported`. Defined ONCE here
// and used at every construction site AND by `regex_feature_construct`, so the
// error text and the diagnostic mapping cannot drift apart.
const FEATURE_LOOKAROUND: &str = "non-edge lookaround assertion";
const FEATURE_UNICODE_CLASS_CAP: &str = "unicode character class exceeded expansion cap";
const FEATURE_BACKREFERENCE: &str = "backreference";
const FEATURE_HUGE_ALTERNATION: &str = "huge alternation exceeds budget";
const FEATURE_NESTED_REPEATS: &str = "nested repeat exceeds budget";

/// Map an `Unsupported { feature }` string back to its construct. Returns `None`
/// for feature strings that are real GPU-NFA limits but have no registry code
/// (e.g. the empty/byte-class expansion caps), so `diagnostic_code` invents none.
fn regex_feature_construct(feature: &str) -> Option<RegexConstruct> {
    match feature {
        FEATURE_LOOKAROUND => Some(RegexConstruct::Lookaround),
        FEATURE_UNICODE_CLASS_CAP => Some(RegexConstruct::UnicodeClassesGpu),
        FEATURE_BACKREFERENCE => Some(RegexConstruct::Backreference),
        FEATURE_HUGE_ALTERNATION => Some(RegexConstruct::HugeAlternation),
        FEATURE_NESTED_REPEATS => Some(RegexConstruct::NestedRepeats),
        _ => None,
    }
}

impl std::fmt::Display for RegexCompileError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            Self::Parse {
                pattern_index,
                message,
            } => write!(
                f,
                "regex_compile: pattern {pattern_index} parse error: {message}. \
                 Fix: review the regex syntax."
            ),
            Self::Unsupported {
                pattern_index,
                feature,
            } => write!(
                f,
                "regex_compile: pattern {pattern_index} uses unsupported feature `{feature}`. \
                 Fix: rewrite the detector into supported GPU-NFA syntax or split it into GPU-compatible rules."
            ),
            Self::TooManyStates { states, cap } => write!(
                f,
                "regex_compile: NFA needs {states} states; per-pipeline cap is {cap}. \
                 Fix: split the pattern set across multiple pipelines."
            ),
            Self::PatternCountOverflow { count } => write!(
                f,
                "regex_compile: pattern count {count} exceeds u32 capacity. Fix: shard the pattern set before GPU regex compilation."
            ),
            Self::MatchLengthOverflow {
                pattern_index,
                len,
            } => write!(
                f,
                "regex_compile: pattern {pattern_index} match length {len} exceeds u32 capacity. Fix: bound or shard the regex before GPU compilation."
            ),
            Self::TableWordCountOverflow { table } => write!(
                f,
                "regex_compile: {table} table word count overflows host usize. Fix: shard the regex pattern set before table construction."
            ),
            Self::StorageReserveFailed {
                field,
                requested,
                message,
            } => write!(
                f,
                "regex_compile: could not reserve {requested} {field} slot(s): {message}. Fix: shard the regex pattern set before GPU compilation."
            ),
        }
    }
}

impl std::error::Error for RegexCompileError {}

/// Output of [`compile_regex_set`]  -  same triple shape as the literal
/// `nfa::compile` returns plus the GPU side-tables `nfa::nfa_scan`
/// expects, so consumers can plug this into `RulePipeline` without
/// changing the dispatch path.
#[derive(Debug, Clone)]
pub struct CompiledRegexSet {
    /// State graph + accept-state metadata.
    pub plan: NfaPlan,
    /// Lane-major byte→bitset transition table:
    /// `[num_states × 256 × LANES_PER_SUBGROUP]` u32s.
    pub transition_table: Vec<u32>,
    /// Lane-major epsilon (free) transition table:
    /// `[num_states × LANES_PER_SUBGROUP]` u32s.
    pub epsilon_table: Vec<u32>,
    /// `true` when at least one source pattern contained a capture group.
    ///
    /// The GPU NFA is a WHOLE-MATCH multimatch engine: it accelerates the
    /// match decision but does NOT prove submatch (capture) spans, capture
    /// groups are stripped during lowering (whole-match still compiles and
    /// runs correctly). A consumer that needs submatch offsets must route
    /// these patterns to the scalar verifier; this flag is the distinct signal
    /// for the `VYRE_SCAN_CAPTURE_EXTRACTION_REQUIRES_VERIFIER` diagnostic
    /// (see [`regex_construct_diagnostic_code`]) WITHOUT rejecting the pattern
    /// (making captures a compile error would regress whole-match acceleration).
    pub captures_present: bool,
}

impl CompiledRegexSet {
    /// The `REGEX_UNSUPPORTED_DIAGNOSTICS.toml` code a consumer routes on when
    /// it needs submatch (capture) spans this whole-match GPU engine does not
    /// prove, or `None` when the compiled set has no capture groups.
    ///
    /// This is NOT an error: the set compiled and scans correctly for the
    /// whole-match decision. The code tells a consumer that wants capture
    /// offsets to run the scalar capture verifier for these patterns.
    #[must_use]
    pub fn capture_extraction_diagnostic_code(&self) -> Option<&'static str> {
        self.captures_present
            .then_some(regex_construct_diagnostic_code(
                RegexConstruct::CaptureExtraction,
            ))
    }
}

const STATE_CAP: usize = LANES * 32;

/// An alternation with more arms than this can NEVER fit the state budget
/// (each arm needs ≥1 state, plus the fork + join), so it is distinctly
/// diagnosed as a huge alternation instead of collapsing into a generic
/// `TooManyStates`. Equal to `STATE_CAP` so the reclassification is SOUND: any
/// alternation this wide already overflowed and never compiled, no successful
/// compile is turned into an error.
const MAX_ALTERNATION_ARMS: usize = STATE_CAP;

/// Nested bounded repeats unroll to (product of the bounds) copies of the body,
/// and each copy is ≥1 state, so when that product exceeds this budget the NFA
/// provably cannot fit. Such patterns are distinctly diagnosed as a nested-repeat
/// blowup rather than a generic `TooManyStates`. Equal to `STATE_CAP` (the unroll
/// product lower-bounds the state count), so no currently-compiling nested
/// repeat regresses.
const NESTED_REPEAT_UNROLL_BUDGET: u64 = STATE_CAP as u64;

/// Non-error signals gathered by [`scan_constructs`] while it validates budgets.
struct ConstructScan {
    /// A capture group was seen (whole-match compiles; submatch spans are not
    /// proven (a verifier-routed signal, never a compile error)).
    captures_present: bool,
}

/// Walk a parsed HIR once to (a) reject over-budget constructs with a DISTINCT
/// diagnostic, huge alternations and nested bounded repeats. BEFORE lowering
/// collapses them into a generic `TooManyStates`, and (b) record whether any
/// capture group is present. Returns the worst-case bounded-repeat unroll
/// product of `hir`, so a parent repetition can detect multiplicative nesting.
fn scan_constructs(
    hir: &Hir,
    pid: usize,
    scan: &mut ConstructScan,
) -> Result<u64, RegexCompileError> {
    match hir.kind() {
        HirKind::Alternation(alts) => {
            if alts.len() > MAX_ALTERNATION_ARMS {
                return Err(RegexCompileError::Unsupported {
                    pattern_index: pid,
                    feature: FEATURE_HUGE_ALTERNATION,
                });
            }
            let mut worst = 1u64;
            for a in alts {
                worst = worst.max(scan_constructs(a, pid, scan)?);
            }
            Ok(worst)
        }
        HirKind::Concat(parts) => {
            let mut worst = 1u64;
            for p in parts {
                worst = worst.max(scan_constructs(p, pid, scan)?);
            }
            Ok(worst)
        }
        HirKind::Repetition(rep) => {
            let inner = scan_constructs(&rep.sub, pid, scan)?;
            match rep.max {
                Some(m) => {
                    let product = u64::from(m).saturating_mul(inner.max(1));
                    // `inner > 1` means a bounded repeat is NESTED inside this
                    // bounded repeat (the case that multiplicatively explodes).
                    // A flat `a{5000}` (inner == 1) is left to the per-repeat
                    // `TooManyStates` guard, unchanged.
                    if inner > 1 && product > NESTED_REPEAT_UNROLL_BUDGET {
                        return Err(RegexCompileError::Unsupported {
                            pattern_index: pid,
                            feature: FEATURE_NESTED_REPEATS,
                        });
                    }
                    Ok(product)
                }
                // Unbounded (`*` / `+`) lowers to an O(1) Kleene wrapper: it does
                // not multiply the nesting product.
                None => Ok(inner.max(1)),
            }
        }
        HirKind::Capture(c) => {
            scan.captures_present = true;
            scan_constructs(&c.sub, pid, scan)
        }
        _ => Ok(1),
    }
}

/// Structured scan for a backreference construct (`\1`..`\9`, `\k<name>` /
/// `\k'name'` / `\k{name}`, or `(?P=name)`). `regex-syntax` does not support
/// backreferences at all, they surface as a raw parse error, so this runs
/// ONLY on the parse-failure path, to CLASSIFY the failure as the distinct
/// unsupported construct rather than a generic syntax error. It respects
/// backslash escaping: an escaped backslash (`\\`) consumes both bytes, so the
/// following digit is read as a literal, not a backreference.
fn pattern_uses_backreference(pat: &str) -> bool {
    let bytes = pat.as_bytes();
    let mut i = 0;
    while i < bytes.len() {
        match bytes[i] {
            b'\\' => {
                if let Some(&c) = bytes.get(i + 1) {
                    // Numeric backreference `\1`..`\9` (`\0` is a NUL escape).
                    if c.is_ascii_digit() && c != b'0' {
                        return true;
                    }
                    // Named backreference `\k<name>` / `\k'name'` / `\k{name}`.
                    if c == b'k' && matches!(bytes.get(i + 2), Some(b'<' | b'\'' | b'{')) {
                        return true;
                    }
                }
                // Skip the escape AND the escaped byte so `\\` is not misread.
                i += 2;
            }
            // Python-style named backreference `(?P=name)`. `bytes[i]` is the
            // ASCII `(`, so `i` is a char boundary and the slice is safe.
            b'(' if pat[i..].starts_with("(?P=") => return true,
            _ => i += 1,
        }
    }
    false
}

/// Compile a list of regex strings into a single multimatch NFA.
///
/// # Errors
/// See [`RegexCompileError`].
pub fn compile_regex_set(patterns: &[&str]) -> Result<CompiledRegexSet, RegexCompileError> {
    let mut builder = NfaBuilder::new();
    let _pattern_count =
        u32::try_from(patterns.len()).map_err(|_| RegexCompileError::PatternCountOverflow {
            count: patterns.len(),
        })?;
    let mut accept_states = Vec::new();
    reserve_vec(&mut accept_states, patterns.len(), "accept state")?;
    let mut accept_state_ids = Vec::new();
    reserve_vec(&mut accept_state_ids, patterns.len(), "accept state id")?;
    let mut accept_start_anchored = Vec::new();
    reserve_vec(
        &mut accept_start_anchored,
        patterns.len(),
        "accept start-anchor flag",
    )?;
    let mut accept_end_anchored = Vec::new();
    reserve_vec(
        &mut accept_end_anchored,
        patterns.len(),
        "accept end-anchor flag",
    )?;
    let entry = builder.fresh_state()?; // shared entry state 0
    let mut captures_present = false;

    // Use the byte-oriented parser configuration: `unicode(false)` +
    // `utf8(false)` makes `\d` / `\w` / `\s` ASCII-only, which is what
    // this primitive's byte-state automaton can represent.
    // `regex_syntax::parse(pat)` defaults to Unicode classes that
    // explode into hundreds of byte ranges and trip our `> 0x7F` guard.
    for (pid, pat) in patterns.iter().enumerate() {
        // Two-phase parse: byte-mode first (keeps `\d`/`\w`/`\s` ASCII
        // so they don't explode into hundreds of Unicode codepoint
        // ranges), then unicode-mode as a fallback when the source
        // contains a non-ASCII codepoint inside a character class
        // (e.g. homoglyph-expanded `[hнһｈ]`). The unicode-mode HIR
        // gets the same `build_class` lowering - non-ASCII members
        // expand into UTF-8 byte-sequence alternations.
        let hir = match regex_syntax::ParserBuilder::new()
            .unicode(false)
            .utf8(false)
            .build()
            .parse(pat)
        {
            Ok(h) => h,
            Err(byte_mode_err) => match regex_syntax::ParserBuilder::new()
                .unicode(true)
                .utf8(false)
                .build()
                .parse(pat)
            {
                Ok(h) => h,
                Err(_unicode_err) => {
                    // Both grammars rejected it. Classify a backreference
                    // which `regex-syntax` never supports, as its DISTINCT
                    // unsupported construct instead of a generic parse error,
                    // so a consumer can route on the registry code. Everything
                    // else keeps the byte-mode diagnostic (the narrow grammar
                    // the kernel actually supports; the unicode retry only
                    // widens the character-class path).
                    if pattern_uses_backreference(pat) {
                        return Err(RegexCompileError::Unsupported {
                            pattern_index: pid,
                            feature: FEATURE_BACKREFERENCE,
                        });
                    }
                    return Err(RegexCompileError::Parse {
                        pattern_index: pid,
                        message: format!("{byte_mode_err}"),
                    });
                }
            },
        };
        // Validate construct budgets (huge alternation / nested repeats) with a
        // DISTINCT diagnostic before lowering collapses them into a generic
        // `TooManyStates`, and note capture presence (a non-error signal).
        let mut construct_scan = ConstructScan {
            captures_present: false,
        };
        scan_constructs(&hir, pid, &mut construct_scan)?;
        captures_present |= construct_scan.captures_present;
        let (frag, anchors) = build_pattern_hir(&mut builder, &hir, pid)?;
        // Connect the shared entry to this pattern's start via epsilon.
        builder.add_epsilon(entry, frag.start);
        let pid_u32 = u32::try_from(pid).map_err(|_| RegexCompileError::PatternCountOverflow {
            count: patterns.len(),
        })?;
        let match_len_u32 =
            u32::try_from(frag.match_len).map_err(|_| RegexCompileError::MatchLengthOverflow {
                pattern_index: pid,
                len: frag.match_len,
            })?;
        accept_states.push((pid_u32, match_len_u32));
        accept_state_ids.push(frag.end);
        accept_start_anchored.push(anchors.start);
        accept_end_anchored.push(anchors.end);
    }

    if builder.state_count() > STATE_CAP {
        return Err(RegexCompileError::TooManyStates {
            states: builder.state_count(),
            cap: STATE_CAP,
        });
    }

    let plan = NfaPlan {
        num_states: u32::try_from(builder.state_count()).map_err(|_| {
            RegexCompileError::TooManyStates {
                states: builder.state_count(),
                cap: STATE_CAP,
            }
        })?,
        input_len: 0,
        accept_states,
        accept_state_ids,
        accept_start_anchored,
        accept_end_anchored,
    };
    let (transition_table, epsilon_table) = builder.emit_lane_major_tables()?;
    Ok(CompiledRegexSet {
        plan,
        transition_table,
        epsilon_table,
        captures_present,
    })
}

/// Build a [`crate::scan::RulePipeline`] directly from regex
/// sources. Convenience for consumers who don't need the
/// `CompiledRegexSet` intermediate. `input_len` matches the contract
/// of `mega_scan::build` (haystack byte count the dispatch will scan).
///
/// # Errors
/// Forwards [`RegexCompileError`].
pub fn build_rule_pipeline_from_regex(
    patterns: &[&str],
    input_buf: &str,
    hit_buf: &str,
    input_len: u32,
) -> Result<crate::scan::RulePipeline, RegexCompileError> {
    let compiled = compile_regex_set(patterns)?;
    let has_epsilon = compiled.epsilon_table.iter().any(|word| *word != 0);
    let program = crate::scan::nfa::nfa_scan_with_plan(
        &compiled.plan,
        has_epsilon,
        input_buf,
        hit_buf,
        input_len,
    )
    .map_err(|_| RegexCompileError::TooManyStates {
        states: compiled.plan.num_states as usize,
        cap: STATE_CAP,
    })?;
    Ok(crate::scan::RulePipeline {
        program,
        transition_table: compiled.transition_table,
        epsilon_table: compiled.epsilon_table,
        plan: compiled.plan.for_input_len(input_len),
    })
}

// ---- Thompson NFA construction over byte transitions ----

#[derive(Debug)]
struct NfaBuilder {
    state_count: usize,
    /// Flat byte transitions. Emission consumes the stream directly,
    /// so construction does not need one allocation per NFA state.
    transitions: Vec<ByteTransition>,
    /// Flat epsilon (free) transitions.
    epsilons: Vec<(u32, u32)>,
}

#[derive(Debug, Clone)]
struct ByteTransition {
    src: u32,
    set: ByteSet,
    dst: u32,
}

#[derive(Debug, Clone)]
struct ByteSet {
    bits: [u64; 4], // 256 bits → 4 u64s
}

impl ByteSet {
    fn new() -> Self {
        Self { bits: [0; 4] }
    }
    fn insert(&mut self, b: u8) {
        self.bits[(b / 64) as usize] |= 1u64 << (b % 64);
    }
    fn from_byte(b: u8) -> Self {
        let mut s = Self::new();
        s.insert(b);
        s
    }
    fn from_range(lo: u8, hi: u8) -> Self {
        let mut s = Self::new();
        for b in lo..=hi {
            s.insert(b);
        }
        s
    }
    fn for_each_set_byte(&self, mut f: impl FnMut(u8)) {
        for (word_idx, &word) in self.bits.iter().enumerate() {
            let mut bits = word;
            while bits != 0 {
                let bit = bits.trailing_zeros() as usize;
                f((word_idx * 64 + bit) as u8);
                bits &= bits - 1;
            }
        }
    }
}

#[derive(Debug, Clone, Copy)]
struct Fragment {
    start: u32,
    end: u32,
    /// Sum of byte-steps along the longest path. Used as the
    /// `pattern_len` reported in `NfaPlan::accept_states`.
    match_len: usize,
}

#[derive(Debug, Clone, Copy, Default)]
struct PatternAnchors {
    start: bool,
    end: bool,
}

impl NfaBuilder {
    fn new() -> Self {
        Self {
            state_count: 0,
            transitions: Vec::new(),
            epsilons: Vec::new(),
        }
    }

    fn state_count(&self) -> usize {
        self.state_count
    }

    fn fresh_state(&mut self) -> Result<u32, RegexCompileError> {
        if self.state_count >= STATE_CAP {
            return Err(RegexCompileError::TooManyStates {
                states: self.state_count.saturating_add(1),
                cap: STATE_CAP,
            });
        }
        let state =
            u32::try_from(self.state_count).map_err(|_| RegexCompileError::TooManyStates {
                states: self.state_count,
                cap: STATE_CAP,
            })?;
        self.state_count =
            self.state_count
                .checked_add(1)
                .ok_or(RegexCompileError::TooManyStates {
                    states: usize::MAX,
                    cap: STATE_CAP,
                })?;
        Ok(state)
    }

    fn add_byte_transition(&mut self, src: u32, set: ByteSet, dst: u32) {
        self.transitions.push(ByteTransition { src, set, dst });
    }

    fn add_epsilon(&mut self, src: u32, dst: u32) {
        self.epsilons.push((src, dst));
    }

    /// Lane-major emission, matching the contract of
    /// `nfa::build_transition_table` + `build_epsilon_table`.
    fn emit_lane_major_tables(&self) -> Result<(Vec<u32>, Vec<u32>), RegexCompileError> {
        let n = self.state_count();
        let mut transitions = zeroed_u32_table(
            table_word_count(n, 256, "transition")?,
            "transition table word",
        )?;
        let mut epsilons =
            zeroed_u32_table(table_word_count(n, 1, "epsilon")?, "epsilon table word")?;

        for edge in &self.transitions {
            let src = edge.src as usize;
            let dst_lane = (edge.dst / 32) as usize;
            let dst_bit = 1u32 << (edge.dst % 32);
            edge.set.for_each_set_byte(|byte| {
                let idx = src * 256 * LANES + (byte as usize) * LANES + dst_lane;
                transitions[idx] |= dst_bit;
            });
        }
        for &(src, dst) in &self.epsilons {
            let dst_lane = (dst / 32) as usize;
            let dst_bit = 1u32 << (dst % 32);
            let idx = src as usize * LANES + dst_lane;
            epsilons[idx] |= dst_bit;
        }
        Ok((transitions, epsilons))
    }
}

fn table_word_count(
    states: usize,
    byte_columns: usize,
    table: &'static str,
) -> Result<usize, RegexCompileError> {
    states
        .checked_mul(byte_columns)
        .and_then(|words| words.checked_mul(LANES))
        .ok_or(RegexCompileError::TableWordCountOverflow { table })
}

fn zeroed_u32_table(words: usize, field: &'static str) -> Result<Vec<u32>, RegexCompileError> {
    let mut table = Vec::new();
    reserve_vec(&mut table, words, field)?;
    table.resize(words, 0);
    Ok(table)
}

fn reserve_vec<T>(
    vec: &mut Vec<T>,
    requested: usize,
    field: &'static str,
) -> Result<(), RegexCompileError> {
    vyre_foundation::allocation::try_reserve_vec_to_capacity(vec, requested).map_err(|source| {
        RegexCompileError::StorageReserveFailed {
            field,
            requested,
            message: source.to_string(),
        }
    })
}

fn empty_fragment(b: &mut NfaBuilder) -> Result<Fragment, RegexCompileError> {
    let s = b.fresh_state()?;
    Ok(Fragment {
        start: s,
        end: s,
        match_len: 0,
    })
}

fn build_pattern_hir(
    b: &mut NfaBuilder,
    hir: &Hir,
    pid: usize,
) -> Result<(Fragment, PatternAnchors), RegexCompileError> {
    match hir.kind() {
        HirKind::Look(Look::Start) => Ok((
            empty_fragment(b)?,
            PatternAnchors {
                start: true,
                end: false,
            },
        )),
        HirKind::Look(Look::End) => Ok((
            empty_fragment(b)?,
            PatternAnchors {
                start: false,
                end: true,
            },
        )),
        HirKind::Concat(parts) => {
            let mut first = 0usize;
            let mut last = parts.len();
            let mut anchors = PatternAnchors::default();

            if first < last && is_text_start_look(&parts[first]) {
                anchors.start = true;
                first += 1;
            }
            if first < last && is_text_end_look(&parts[last - 1]) {
                anchors.end = true;
                last -= 1;
            }

            Ok((build_hir_slice(b, &parts[first..last], pid)?, anchors))
        }
        _ => Ok((build_hir(b, hir, pid)?, PatternAnchors::default())),
    }
}

fn is_text_start_look(hir: &Hir) -> bool {
    matches!(hir.kind(), HirKind::Look(Look::Start))
}

fn is_text_end_look(hir: &Hir) -> bool {
    matches!(hir.kind(), HirKind::Look(Look::End))
}

fn build_hir_slice(
    b: &mut NfaBuilder,
    parts: &[Hir],
    pid: usize,
) -> Result<Fragment, RegexCompileError> {
    let Some(first_part) = parts.first() else {
        return empty_fragment(b);
    };
    let mut acc = build_hir(b, first_part, pid)?;
    for child in &parts[1..] {
        let next = build_hir(b, child, pid)?;
        b.add_epsilon(acc.end, next.start);
        acc = Fragment {
            start: acc.start,
            end: next.end,
            match_len: acc.match_len + next.match_len,
        };
    }
    Ok(acc)
}

fn build_hir(b: &mut NfaBuilder, hir: &Hir, pid: usize) -> Result<Fragment, RegexCompileError> {
    match hir.kind() {
        HirKind::Empty => empty_fragment(b),
        HirKind::Literal(lit) => {
            // Each literal byte gets its own state.
            let start = b.fresh_state()?;
            let mut prev = start;
            for &byte in lit.0.iter() {
                let next = b.fresh_state()?;
                b.add_byte_transition(prev, ByteSet::from_byte(byte), next);
                prev = next;
            }
            Ok(Fragment {
                start,
                end: prev,
                match_len: lit.0.len(),
            })
        }
        HirKind::Class(cls) => build_class(b, cls, pid),
        HirKind::Repetition(rep) => build_repetition(b, rep, pid),
        HirKind::Concat(parts) => build_hir_slice(b, parts, pid),
        HirKind::Alternation(alts) => {
            // Diamond: shared fork → each branch → shared join.
            let fork = b.fresh_state()?;
            let join = b.fresh_state()?;
            let mut max_len = 0usize;
            for child in alts {
                let frag = build_hir(b, child, pid)?;
                b.add_epsilon(fork, frag.start);
                b.add_epsilon(frag.end, join);
                if frag.match_len > max_len {
                    max_len = frag.match_len;
                }
            }
            Ok(Fragment {
                start: fork,
                end: join,
                match_len: max_len,
            })
        }
        HirKind::Look(_) => Err(RegexCompileError::Unsupported {
            pattern_index: pid,
            feature: FEATURE_LOOKAROUND,
        }),
        HirKind::Capture(c) => {
            // We don't expose capture groups (NFA scan is multimatch,
            // not capture). Strip and recurse.
            build_hir(b, &c.sub, pid)
        }
    }
}

fn build_repetition(
    b: &mut NfaBuilder,
    rep: &Repetition,
    pid: usize,
) -> Result<Fragment, RegexCompileError> {
    let min = rep.min;
    let max = rep.max;

    // Keep pathological repetitions from materializing a giant transient NFA.
    // The final state cap is the source of truth, so oversized repetitions
    // report TooManyStates instead of pretending the syntax is unsupported.
    if let Some(m) = max {
        if m as usize > STATE_CAP {
            return Err(RegexCompileError::TooManyStates {
                states: m as usize,
                cap: STATE_CAP,
            });
        }
    }
    if min as usize > STATE_CAP {
        return Err(RegexCompileError::TooManyStates {
            states: min as usize,
            cap: STATE_CAP,
        });
    }

    // Build by unrolling: emit `min` copies, then either
    //   - a Kleene loop if max is None (`*` / `+`), OR
    //   - `max - min` optional copies if max is bounded.
    let start = b.fresh_state()?;
    let mut tail = start;
    let mut total_len = 0usize;

    for _ in 0..min {
        let frag = build_hir(b, &rep.sub, pid)?;
        b.add_epsilon(tail, frag.start);
        tail = frag.end;
        total_len += frag.match_len;
    }

    match max {
        None => {
            // Open-ended: insert a Kleene wrapper. tail → frag.start →
            // frag.end → tail (loop back) ; tail → join (skip).
            let join = b.fresh_state()?;
            let frag = build_hir(b, &rep.sub, pid)?;
            b.add_epsilon(tail, frag.start);
            b.add_epsilon(frag.end, frag.start); // loop
            b.add_epsilon(frag.end, join);
            b.add_epsilon(tail, join); // zero matches
            tail = join;
        }
        Some(m) => {
            for _ in min..m {
                let frag = build_hir(b, &rep.sub, pid)?;
                let join = b.fresh_state()?;
                b.add_epsilon(tail, frag.start);
                b.add_epsilon(frag.end, join);
                b.add_epsilon(tail, join); // skip this optional copy
                tail = join;
                // `match_len` is the MAXIMUM admissible match length (see
                // `build_class`: extraction uses it only to size the replay
                // window, so over-sizing is harmless but UNDER-sizing truncates
                // the walk before the longer accepts). A bounded repetition
                // `{n,m}` accepts every length in `n..=m` (the ε skip edges make
                // the fragment end reachable after each optional copy), so the
                // window must cover the MAX `m` copies, otherwise the anchored
                // windowed replay caps at `n` and never visits ends `n+1..=m`
                // (the root of BACKLOG items 18/27: `a{2,4}` surfaced only
                // length-2, and `{10,48}` under-scanned). Accumulate every
                // optional copy so `total_len` reaches `m * sub_len`.
                total_len += frag.match_len;
            }
        }
    }
    Ok(Fragment {
        start,
        end: tail,
        match_len: total_len,
    })
}

/// Lower a regex character class into an NFA fragment, taking the
/// single-byte fast path when the class fits in 0..=127 and the
/// UTF-8-alternation expansion path otherwise.
///
/// The single-byte path is identical to the original implementation:
/// one ByteSet, one transition, `match_len = 1`. The expansion path
/// emits one byte-chain fragment per codepoint (or per pre-existing
/// multi-byte range like `\u{0100}-\u{01FF}` enumerated codepoint-by-
/// codepoint) and ε-merges them via a shared end state.
///
/// `match_len` for the expansion case is the MAX byte length across
/// arms - anchored extraction uses `match_len` only to position
/// the post-process window, not to extract the credential text, and
/// over-sizing the window is harmless (the real regex re-extracts the
/// exact match inside it).
///
/// To keep state-budget worst case bounded, expansion is capped at
/// `MAX_CLASS_EXPANSION_CODEPOINTS = 256` enumerated codepoints (a
/// `[\u{0100}-\u{017F}]` Latin-Extended block sits at 128, which is
/// well within budget; a class spanning a full CJK block would refuse).
fn build_class(b: &mut NfaBuilder, cls: &Class, pid: usize) -> Result<Fragment, RegexCompileError> {
    if let Some(set) = try_class_as_ascii_byte_set(cls) {
        let start = b.fresh_state()?;
        let end = b.fresh_state()?;
        b.add_byte_transition(start, set, end);
        return Ok(Fragment {
            start,
            end,
            match_len: 1,
        });
    }
    let sequences = class_to_utf8_sequences(cls, pid)?;
    if sequences.is_empty() {
        return Err(RegexCompileError::Unsupported {
            pattern_index: pid,
            feature: "empty character class after Unicode expansion",
        });
    }
    let start = b.fresh_state()?;
    let end = b.fresh_state()?;
    let mut max_len = 1usize;
    for seq in &sequences {
        if seq.is_empty() {
            continue;
        }
        // Build a sequential chain start ε→ s0 -b0-> s1 -b1-> ... -bN-> end
        // for this UTF-8 byte sequence.
        let arm_start = b.fresh_state()?;
        b.add_epsilon(start, arm_start);
        let mut prev = arm_start;
        for &byte in seq {
            let next = b.fresh_state()?;
            b.add_byte_transition(prev, ByteSet::from_byte(byte), next);
            prev = next;
        }
        b.add_epsilon(prev, end);
        if seq.len() > max_len {
            max_len = seq.len();
        }
    }
    Ok(Fragment {
        start,
        end,
        match_len: max_len,
    })
}

/// Returns `Some(ByteSet)` when every member of the class fits in
/// 0..=127 (i.e. the original single-byte ASCII fast path). Otherwise
/// returns None so the caller takes the UTF-8 expansion path.
fn try_class_as_ascii_byte_set(cls: &Class) -> Option<ByteSet> {
    let mut out = ByteSet::new();
    match cls {
        Class::Bytes(byte_class) => {
            // Byte classes are already at the byte level - every member
            // is a u8, no codepoint expansion involved. The legacy fast
            // path always applies.
            for r in byte_class.iter() {
                let merged = ByteSet::from_range(r.start(), r.end());
                for w in 0..4 {
                    out.bits[w] |= merged.bits[w];
                }
            }
            Some(out)
        }
        Class::Unicode(uni) => {
            // ASCII-only fast path. The moment any range escapes
            // 0..=0x7F, fall through to UTF-8 expansion.
            for r in uni.iter() {
                if (r.end() as u32) > 0x7F {
                    return None;
                }
                let merged = ByteSet::from_range(r.start() as u8, r.end() as u8);
                for w in 0..4 {
                    out.bits[w] |= merged.bits[w];
                }
            }
            Some(out)
        }
    }
}

/// Cap on enumerated codepoints during UTF-8 expansion. A class like
/// `[\u{0100}-\u{017F}]` (Latin Extended-A) expands to 128 sequences,
/// well within the cap. A class spanning a full CJK block (~20 000
/// codepoints) would blow past it - the byte-state automaton can't
/// represent that cleanly, so the consumer should keep that pattern on
/// the CPU regex path.
const MAX_CLASS_EXPANSION_CODEPOINTS: usize = 256;

/// Enumerate every codepoint in `cls`, encode each into UTF-8, and
/// return the resulting `Vec<Vec<u8>>` so the caller can build an
/// alternation of byte-chain fragments. ASCII members come back as
/// 1-byte sequences; non-ASCII as 2-4 byte sequences.
fn class_to_utf8_sequences(cls: &Class, pid: usize) -> Result<Vec<Vec<u8>>, RegexCompileError> {
    let mut sequences: Vec<Vec<u8>> = Vec::new();
    let mut budget = MAX_CLASS_EXPANSION_CODEPOINTS;
    match cls {
        Class::Bytes(byte_class) => {
            for r in byte_class.iter() {
                for byte in r.start()..=r.end() {
                    if budget == 0 {
                        return Err(RegexCompileError::Unsupported {
                            pattern_index: pid,
                            feature: "byte character class exceeded expansion cap",
                        });
                    }
                    sequences.push(vec![byte]);
                    budget -= 1;
                }
            }
        }
        Class::Unicode(uni) => {
            for r in uni.iter() {
                let lo = r.start() as u32;
                let hi = r.end() as u32;
                for cp in lo..=hi {
                    if budget == 0 {
                        return Err(RegexCompileError::Unsupported {
                            pattern_index: pid,
                            feature: FEATURE_UNICODE_CLASS_CAP,
                        });
                    }
                    // Use a small buffer + `char::encode_utf8` to avoid
                    // pulling in a heavyweight UTF-8 dependency. Invalid
                    // codepoints (surrogates) are silently skipped -
                    // regex-syntax shouldn't emit them in a parsed HIR
                    // for character classes, but the `char::from_u32`
                    // guard catches the corner case if it ever does.
                    if let Some(c) = char::from_u32(cp) {
                        let mut buf = [0u8; 4];
                        let encoded = c.encode_utf8(&mut buf);
                        sequences.push(encoded.as_bytes().to_vec());
                        budget -= 1;
                    }
                }
            }
        }
    }
    Ok(sequences)
}

#[cfg(test)]
mod tests {
    use super::*;

    fn states_of(s: &str) -> u32 {
        compile_regex_set(&[s]).unwrap().plan.num_states
    }

    #[test]
    fn capture_mode_routing_splits_accelerator_from_verifier() {
        // Exactly the three whole-match modes run on the accelerator; the three
        // group-extraction modes require the verifier. Under this contract the
        // two bits are exact complements, assert both directions so a future
        // "neither" (unsupported) mode can't slip through as accelerator-eligible.
        for mode in CaptureMode::ALL {
            assert_eq!(
                mode.accelerator_eligible(),
                !mode.verifier_required(),
                "{mode:?}: accelerator_eligible must be the complement of verifier_required"
            );
        }
        let accel: Vec<CaptureMode> = CaptureMode::ALL
            .into_iter()
            .filter(|m| m.accelerator_eligible())
            .collect();
        assert_eq!(
            accel,
            vec![
                CaptureMode::NonCapture,
                CaptureMode::Count,
                CaptureMode::Span
            ],
            "only the whole-match modes are accelerator-eligible"
        );
    }

    #[test]
    fn capture_mode_id_round_trips_and_is_unique() {
        use std::collections::BTreeSet;
        let mut ids = BTreeSet::new();
        for mode in CaptureMode::ALL {
            let id = mode.contract_row().mode_id;
            assert!(ids.insert(id), "duplicate mode_id `{id}`");
            assert_eq!(
                CaptureMode::from_mode_id(id),
                Some(mode),
                "mode_id `{id}` must round-trip back to {mode:?}"
            );
        }
        assert_eq!(ids.len(), 6, "all six modes must have distinct ids");
        assert_eq!(CaptureMode::from_mode_id("no_such_mode"), None);
    }

    #[test]
    fn literal_compiles() {
        let r = compile_regex_set(&["abc"]).unwrap();
        // 1 entry + 1 literal-start + 3 letter states = 5
        assert_eq!(r.plan.num_states, 5);
        assert_eq!(r.plan.accept_states.len(), 1);
    }

    #[test]
    fn alternation_compiles() {
        let r = compile_regex_set(&["a|b"]).unwrap();
        // entry + fork + join + 2*(start + 1 byte) = 1+1+1+2+2 = 7
        // (exact count depends on builder; just sanity-check it's >0).
        assert!(r.plan.num_states > 0);
        assert_eq!(r.plan.accept_states.len(), 1);
    }

    #[test]
    fn class_compiles() {
        let r = compile_regex_set(&["[a-z]"]).unwrap();
        assert!(r.plan.num_states > 0);
        // Sanity: 26 lowercase bytes hit the same destination state.
        // We don't introspect the table here  -  just ensure it builds.
    }

    #[test]
    fn text_anchors_compile_to_accept_flags() {
        let r = compile_regex_set(&["^foo$"]).unwrap();
        assert_eq!(r.plan.accept_start_anchored, vec![true]);
        assert_eq!(r.plan.accept_end_anchored, vec![true]);
    }

    #[test]
    fn bounded_repetition_above_old_cap_compiles_under_state_cap() {
        let r = compile_regex_set(&["a{0,128}"]).unwrap();
        assert!(r.plan.num_states > 64);
        assert!(r.plan.num_states <= STATE_CAP as u32);
    }

    #[test]
    fn regex_compile_preserves_accept_metadata_through_checked_paths() {
        let r = compile_regex_set(&["a", "bc", "^de$"]).unwrap();

        assert_eq!(r.plan.accept_states, vec![(0, 1), (1, 2), (2, 2)]);
        assert_eq!(r.plan.accept_state_ids.len(), 3);
        assert_eq!(r.plan.accept_start_anchored, vec![false, false, true]);
        assert_eq!(r.plan.accept_end_anchored, vec![false, false, true]);
        assert_eq!(
            r.transition_table.len(),
            r.plan.num_states as usize * 256 * LANES
        );
        assert_eq!(r.epsilon_table.len(), r.plan.num_states as usize * LANES);
    }

    #[test]
    fn regex_compile_uses_checked_abi_and_table_allocation_paths() {
        let production = include_str!("regex_compile.rs")
            .split("#[cfg(test)]")
            .next()
            .expect("Fix: regex_compile.rs must contain production section");

        assert!(
            production.contains("u32::try_from(pid)")
                && production.contains("u32::try_from(frag.match_len)")
                && production.contains("u32::try_from(builder.state_count())")
                && production.contains("u32::try_from(self.state_count)")
                && production.contains("checked_add(1)")
                && production.contains("try_reserve_vec_to_capacity")
                && !production.contains("pid as u32")
                && !production.contains("frag.match_len as u32")
                && !production.contains("builder.state_count() as u32")
                && !production.contains("self.state_count as u32")
                && !production.contains("vec![0u32;")
                && !production.contains("Vec::with_capacity(patterns.len())"),
            "Fix: regex compilation must not truncate ids/counts or allocate NFA tables with infallible zero-vector construction."
        );
    }

    #[test]
    fn regex_pipeline_uses_compiled_plan_instead_of_literal_source_plan() {
        let compiled = compile_regex_set(&["a|bc"]).unwrap();
        let pipeline = build_rule_pipeline_from_regex(&["a|bc"], "input", "hits", 64).unwrap();

        assert_eq!(pipeline.plan.num_states, compiled.plan.num_states);
        assert_eq!(
            pipeline.plan.accept_state_ids,
            compiled.plan.accept_state_ids
        );
        assert_eq!(
            pipeline.epsilon_table.iter().any(|word| *word != 0),
            compiled.epsilon_table.iter().any(|word| *word != 0)
        );
        assert_ne!(
            pipeline.plan.num_states,
            crate::scan::nfa::compile(&["a|bc"]).num_states,
            "regex pipeline must not rebuild the scan program from literal regex source bytes"
        );
    }

    #[test]
    fn states_count_grows_with_concat() {
        let one = states_of("a");
        let two = states_of("ab");
        let three = states_of("abc");
        assert!(two > one);
        assert!(three > two);
    }

    #[test]
    fn state_cap_enforced() {
        // Build a regex that would exceed the per-pipeline state cap.
        // A literal of LANES*32+1 = 1025 chars exceeds the cap.
        let huge: String = (0..(STATE_CAP + 4)).map(|_| 'a').collect();
        let err = compile_regex_set(&[&huge]).unwrap_err();
        assert!(matches!(err, RegexCompileError::TooManyStates { .. }));
    }

    #[test]
    fn unsupported_regex_diagnostic_does_not_route_to_cpu_backend() {
        let err = compile_regex_set(&[r"\bsecret\b"]).unwrap_err();
        let message = err.to_string().to_ascii_lowercase();
        assert!(
            !message.contains("cpu"),
            "unsupported GPU-NFA regex diagnostics must not recommend host-side routing: {message}"
        );
        assert!(
            message.contains("gpu"),
            "unsupported GPU-NFA regex diagnostics must name the GPU-compatible rewrite contract: {message}"
        );
    }

    /// Contract: non-ASCII codepoints inside a character class no longer
    /// abort compile. They expand into a UTF-8 byte-sequence alternation
    /// the byte-NFA can represent. Mirrors the homoglyph-expanded
    /// detector patterns consumers feed in (e.g. openai `[hнһｈ]f_...`)
    /// that used to fall on the floor with "unicode character classes
    /// outside ASCII".
    #[test]
    fn unicode_class_outside_ascii_compiles_via_utf8_expansion() {
        // `н` (U+043D) and `һ` (U+04BB) are 2-byte UTF-8; `ｈ` (U+FF48)
        // is 3-byte UTF-8; `h` (U+0068) is 1-byte. All four must be
        // representable.
        let pat = "[hнһｈ]f_[a-zA-Z0-9]{4}";
        let result = compile_regex_set(&[pat]);
        let compiled = match result {
            Ok(c) => c,
            Err(e) => {
                panic!("unicode-extended character class must compile via UTF-8 expansion; got {e}")
            }
        };
        // 4 alternation arms (one per codepoint) × varying byte length
        // + chain states + literal `f_` chain + bounded repetition
        // states - the exact count is implementation-dependent, but
        // every successfully-compiled regex must produce >=2 accept-
        // state-ids worth of state graph.
        assert!(
            compiled.plan.num_states > 4,
            "expanded NFA must have non-trivial state count"
        );
        // accept_state_ids carries one entry per accept (one pattern,
        // so one accept) regardless of arm count; the load-bearing
        // assertion is that compile didn't error.
        assert_eq!(compiled.plan.accept_states.len(), 1);
    }

    /// Contract: classes containing ONLY ASCII still take the fast
    /// single-byte-transition path. Without this guarantee, every AC
    /// detector regex would pay the multi-state expansion cost.
    #[test]
    fn ascii_only_class_keeps_single_byte_transition_path() {
        // Single state for entry + 2 for `[ab]` (start + end) = 3.
        // Anything larger means we accidentally took the expansion arm.
        let r = compile_regex_set(&["[ab]"]).unwrap();
        assert_eq!(
            r.plan.num_states, 3,
            "[ab] must stay on the single-transition fast path (entry + 2 class states); got {} states",
            r.plan.num_states
        );
    }

    /// Contract: massive Unicode ranges that would blow past the
    /// expansion cap return a structured error instead of consuming
    /// unbounded memory.
    #[test]
    fn unicode_class_above_expansion_cap_errors_cleanly() {
        // 257 codepoints - one above MAX_CLASS_EXPANSION_CODEPOINTS = 256.
        let pat = "[\u{0100}-\u{0200}]";
        let err = compile_regex_set(&[pat]).unwrap_err();
        match err {
            RegexCompileError::Unsupported { feature, .. } => {
                assert!(
                    feature.contains("expansion cap"),
                    "over-cap expansion must name the cap in its diagnostic: {feature}"
                );
            }
            other => panic!("expected Unsupported expansion-cap error, got {other:?}"),
        }
    }

    /// The real compile path must emit the canonical registry diagnostic code for
    /// each construct the frontend distinctly identifies, so a consumer can route
    /// precisely (verifier vs reject) instead of parsing free-text `feature`.
    #[test]
    fn regex_compile_diagnostic_codes() {
        // A non-edge lookaround (word boundary) routes to the verifier.
        let look_err = compile_regex_set(&[r"a\bc"]).expect_err("word boundary is unsupported");
        assert_eq!(
            look_err.diagnostic_code(),
            Some("VYRE_SCAN_APPROXIMATED_LOOKAROUND_REQUIRES_VERIFIER"),
            "non-edge lookaround must map to its verifier diagnostic code; error was: {look_err}"
        );

        // An over-cap Unicode class routes to the Unicode-mode-GPU rejection.
        let uni_err =
            compile_regex_set(&["[\u{0100}-\u{0200}]"]).expect_err("over-cap unicode class");
        assert_eq!(
            uni_err.diagnostic_code(),
            Some("VYRE_SCAN_UNSUPPORTED_UNICODE_MODE_GPU"),
            "over-cap unicode class must map to its diagnostic code; error was: {uni_err}"
        );

        // Edge anchors are SUPPORTED (Look::Start/End), so no error at all.
        assert!(
            compile_regex_set(&["^abc$"]).is_ok(),
            "start/end anchors must compile, not be flagged as unsupported lookaround"
        );

        // A pure syntax error is not a registry construct -> no code.
        let parse_err = compile_regex_set(&["("]).expect_err("unbalanced group is a parse error");
        assert_eq!(
            parse_err.diagnostic_code(),
            None,
            "a parse error must not claim a registry diagnostic code"
        );

        // W2-3: a backreference is classified as its DISTINCT construct, not a
        // generic parse error, so a consumer can route on the registry code.
        let backref_err =
            compile_regex_set(&[r"(a)\1"]).expect_err("backreferences are unsupported");
        assert_eq!(
            backref_err.diagnostic_code(),
            Some("VYRE_SCAN_UNSUPPORTED_BACKREFERENCE"),
            "a backreference must map to its distinct code, not fall back to Parse; error was: {backref_err}"
        );

        // W2-3: a huge alternation gets its own budget code instead of collapsing
        // into a generic TooManyStates.
        let huge: String = (0..(MAX_ALTERNATION_ARMS + 8))
            .map(|i| format!("v{i}"))
            .collect::<Vec<_>>()
            .join("|");
        let alt_err = compile_regex_set(&[huge.as_str()]).expect_err("over-budget alternation");
        assert_eq!(
            alt_err.diagnostic_code(),
            Some("VYRE_SCAN_UNSUPPORTED_HUGE_ALTERNATION_BUDGET"),
            "a huge alternation must map to its budget code, not TooManyStates; error was: {alt_err}"
        );

        // W2-3: nested bounded repeats whose unroll product exceeds the budget get
        // their own code, distinct from a flat over-cap repeat.
        let nested_err =
            compile_regex_set(&[r"(?:a{40}){40}"]).expect_err("nested-repeat unroll blowup");
        assert_eq!(
            nested_err.diagnostic_code(),
            Some("VYRE_SCAN_UNSUPPORTED_NESTED_REPEAT_BUDGET"),
            "nested bounded repeats must map to their budget code; error was: {nested_err}"
        );
    }

    /// The backreference detector must respect backslash escaping and match every
    /// backreference syntax `regex-syntax` rejects, WITHOUT string-matching parser
    /// error text (a structured source scan. ONE PLACE, no parse-message hacks).
    #[test]
    fn backreference_detector_is_escaping_aware() {
        // Numeric backreferences \1..\9 (in any position).
        assert!(pattern_uses_backreference(r"\1"));
        assert!(pattern_uses_backreference(r"(a)\1"));
        assert!(pattern_uses_backreference(r"foo\9bar"));
        // Named backreferences.
        assert!(pattern_uses_backreference(r"\k<name>"));
        assert!(pattern_uses_backreference(r"\k'name'"));
        assert!(pattern_uses_backreference("(?P=name)"));

        // NOT backreferences: \0 is a NUL escape, an escaped backslash before a
        // digit is a literal backslash + literal digit, and ordinary escapes /
        // classes carry no backreference.
        assert!(!pattern_uses_backreference(r"\0"));
        assert!(
            !pattern_uses_backreference(r"\\1"),
            "an escaped backslash then a literal 1 is not a backreference"
        );
        assert!(!pattern_uses_backreference(r"\d+\w*"));
        assert!(!pattern_uses_backreference(r"[a-z]{3}"));
        assert!(!pattern_uses_backreference("plain text"));
        // `\\\1` = literal backslash, then a real backreference.
        assert!(pattern_uses_backreference(r"\\\1"));
    }

    /// Capture groups must NOT become a compile error (whole-match acceleration
    /// still works); instead the compiled set reports capture presence so a
    /// consumer that needs submatch spans can route to the verifier.
    #[test]
    fn captures_compile_and_surface_the_verifier_diagnostic() {
        // A pattern with a capture group compiles (whole-match works) and reports
        // its presence + the verifier diagnostic code.
        let with_cap = compile_regex_set(&[r"(abc)def"]).expect("captures compile for whole-match");
        assert!(with_cap.captures_present, "the capture group must be noted");
        assert_eq!(
            with_cap.capture_extraction_diagnostic_code(),
            Some("VYRE_SCAN_CAPTURE_EXTRACTION_REQUIRES_VERIFIER"),
            "a captured pattern must surface the capture-verifier code without erroring"
        );

        // A capture-free pattern compiles with no capture signal.
        let no_cap = compile_regex_set(&[r"abcdef"]).expect("plain pattern compiles");
        assert!(!no_cap.captures_present);
        assert_eq!(no_cap.capture_extraction_diagnostic_code(), None);

        // A non-capturing group is not a capture.
        let noncap = compile_regex_set(&[r"(?:abc)def"]).expect("non-capturing group compiles");
        assert!(
            !noncap.captures_present,
            "a (?:…) non-capturing group must not be flagged as a capture"
        );
    }

    /// SOUNDNESS / no-regression: the budget reclassification must only relabel
    /// patterns that ALREADY failed (state overflow). Patterns UNDER the budgets
    /// must still compile exactly as before.
    #[test]
    fn budget_reclassification_does_not_regress_compiling_patterns() {
        // A normal multi-arm alternation (well under both the arm budget AND the
        // state cap, each arm is a single byte) still compiles: the arm-count
        // check must not false-fire on ordinary alternations.
        let ok_alt: String = ('a'..='z')
            .chain('A'..='Z')
            .chain('0'..='9')
            .map(|c| c.to_string())
            .collect::<Vec<_>>()
            .join("|");
        let compiled = compile_regex_set(&[ok_alt.as_str()])
            .expect("a 62-arm single-byte alternation must still compile");
        // And it must NOT be misclassified as a huge alternation.
        assert!(compiled.plan.num_states > 0);

        // A nested bounded repeat whose unroll product is under the budget still
        // compiles (20*20 = 400 < 1024).
        assert!(
            compile_regex_set(&[r"(?:a{20}){20}"]).is_ok(),
            "a nested repeat under the unroll budget must still compile"
        );

        // The ONE-PLACE construct→code map round-trips every construct.
        assert_eq!(
            regex_construct_diagnostic_code(RegexConstruct::Backreference),
            "VYRE_SCAN_UNSUPPORTED_BACKREFERENCE"
        );
        assert_eq!(
            regex_construct_diagnostic_code(RegexConstruct::NestedRepeats),
            "VYRE_SCAN_UNSUPPORTED_NESTED_REPEAT_BUDGET"
        );
    }

    /// W8-2 (structured diagnostics quality): every capability refusal must carry
    /// the `regex_compile:` owner prefix AND a `Fix:` clause naming the remedy
    /// the engineering standard that error messages include context and the fix.
    /// The `variants` array below is enforced COMPLETE by the exhaustive match in
    /// `assert_covers_every_variant`: adding a `RegexCompileError` variant without
    /// listing it here fails to COMPILE (the refusal cannot ship un-audited), and
    /// the per-variant assertions fail if any Display drops its owner or fix path.
    #[test]
    fn every_compile_error_variant_names_its_owner_and_fix_path() {
        let variants = [
            RegexCompileError::Parse {
                pattern_index: 0,
                message: "unclosed group".to_string(),
            },
            RegexCompileError::Unsupported {
                pattern_index: 1,
                feature: "backreference",
            },
            RegexCompileError::TooManyStates {
                states: 5_000,
                cap: 1_024,
            },
            RegexCompileError::PatternCountOverflow { count: usize::MAX },
            RegexCompileError::MatchLengthOverflow {
                pattern_index: 2,
                len: usize::MAX,
            },
            RegexCompileError::TableWordCountOverflow {
                table: "transition",
            },
            RegexCompileError::StorageReserveFailed {
                field: "epsilon",
                requested: 9,
                message: "allocator refused".to_string(),
            },
        ];

        // Exhaustiveness guard: the match has no wildcard arm, so a new variant
        // breaks the build here until it is added to `variants` above and given a
        // fix path in `Display` (this test is in the defining crate, where a
        // `#[non_exhaustive]` enum can still be matched exhaustively).
        fn assert_covers_every_variant(error: &RegexCompileError) {
            match error {
                RegexCompileError::Parse { .. }
                | RegexCompileError::Unsupported { .. }
                | RegexCompileError::TooManyStates { .. }
                | RegexCompileError::PatternCountOverflow { .. }
                | RegexCompileError::MatchLengthOverflow { .. }
                | RegexCompileError::TableWordCountOverflow { .. }
                | RegexCompileError::StorageReserveFailed { .. } => {}
            }
        }

        for error in &variants {
            assert_covers_every_variant(error);
            let rendered = error.to_string();
            assert!(
                rendered.starts_with("regex_compile:"),
                "a RegexCompileError variant lacks the `regex_compile:` owner prefix: {rendered}"
            );
            assert!(
                rendered.contains("Fix:"),
                "a RegexCompileError variant lacks a `Fix:` remedy clause: {rendered}"
            );
        }
    }
}