ifc_lite_core/

decoder.rs

// This Source Code Form is subject to the terms of the Mozilla Public
// License, v. 2.0. If a copy of the MPL was not distributed with this
// file, You can obtain one at https://mozilla.org/MPL/2.0/.

//! Entity Decoder - On-demand entity parsing
//!
//! Lazily decode IFC entities from byte offsets without loading entire file into memory.

use crate::columnar_index::EntityIndexStore;
use crate::error::{Error, Result};
use crate::parser::{
    is_step_space, parse_entity, parse_step_numeric, report_scan_diagnostics, skip_step_trivia,
    EntityScanner,
};
use crate::schema_gen::{AttributeValue, DecodedEntity};
use rustc_hash::FxHashMap;
use std::sync::Arc;

#[path = "decoder/caches.rs"]
mod caches;
#[path = "decoder/fast_buffers.rs"]
mod fast_buffers;
#[path = "decoder/precision.rs"]
mod precision;
#[path = "decoder/schema_layout.rs"]
mod schema_layout;
#[path = "decoder/styled_items.rs"]
mod styled_items;
pub use schema_layout::GridAxisIndex;
pub(crate) type StyledItemIndexResult = std::result::Result<FxHashMap<u32, Vec<u32>>, String>;

/// Pre-built entity index type
pub type EntityIndex = FxHashMap<u32, (usize, usize)>;

/// Build an entity index from content.
///
/// This intentionally shares `EntityScanner`'s HEADER skipping, quoted-string
/// semantics, and its #3395 refusal of instance names wider than `u32` (reported
/// here, not handed back as a quietly short index), so scan and lookup agree.
#[inline]
pub fn build_entity_index<T>(content: &T) -> EntityIndex
where
    T: AsRef<[u8]> + ?Sized,
{
    let content = content.as_ref();
    let estimated_entities = content.len() / 50;
    let mut index = FxHashMap::with_capacity_and_hasher(estimated_entities, Default::default());
    let mut scanner = EntityScanner::new(content);
    while let Some((id, _type_name, start, end)) = scanner.next_entity() {
        index.insert(id, (start, end));
    }
    report_scan_diagnostics(scanner.skipped_oversized_ids(), scanner.malformed_record_start().is_some());
    index
}

/// Entity decoder for lazy parsing from raw IFC bytes.
///
/// String attributes are decoded lossily when tokens become `AttributeValue`s;
/// structural scanning and byte offsets always use the original source bytes.
pub struct EntityDecoder<'a> {
    content: &'a [u8],
    /// Cache of decoded entities (entity_id -> `Arc<DecodedEntity>`)
    /// Using Arc avoids expensive clones on cache hits
    cache: FxHashMap<u32, Arc<DecodedEntity>>,
    /// Entity offsets (entity_id -> (start, end)): legacy `FxHashMap` or compact
    /// columnar index. `pub(crate)` for `crate::columnar_index`'s constructors.
    pub(crate) entity_index: Option<EntityIndexStore>,
    /// Cache of cartesian point coordinates for FacetedBrep optimization
    /// Only populated when using get_polyloop_coords_cached
    point_cache: FxHashMap<u32, (f64, f64, f64)>,
    /// Number of `point_cache` hits served by [`Self::get_polyloop_coords_cached`].
    /// Pure instrumentation (never affects output); read via
    /// [`Self::point_cache_stats`] to prove cross-element memoization fires when
    /// the cache is hoisted across a worker's parts.
    point_cache_hits: u64,
    /// Number of `point_cache` misses (a CartesianPoint parsed for the first time)
    /// served by [`Self::get_polyloop_coords_cached`].
    point_cache_misses: u64,
    /// Lazy-cached multiplier converting file plane-angle units to radians.
    /// Populated on first call to [`Self::plane_angle_to_radians`]. Spec
    /// default (and Renga-style files) is 1.0 (RADIAN); degree-unit files
    /// resolve to π/180.
    plane_angle_to_radians_cache: Option<f64>,
    /// Lazy-cached multiplier converting file length units to metres.
    /// Populated on first call to [`Self::length_unit_scale`]. 1.0 for metre
    /// files, 0.001 for millimetre files, etc. Used to express absolute
    /// tolerances (e.g. curve-tessellation chord deviation) in file units.
    length_unit_scale_cache: Option<f64>,
    /// Smallest and largest declared geometric-context Precision in file units.
    /// Errors are cached too, so malformed input cannot trigger repeated scans.
    geometric_precision_cache: Option<std::result::Result<Option<(f64, f64)>, String>>,
    /// Per-worker memo of resolved placement world transforms, keyed by the
    /// IfcObjectPlacement entity id and stored as an opaque column-major
    /// `[f64; 16]` (core must not depend on nalgebra; the geometry crate owns
    /// the Matrix4 <-> array round-trip). Populated by the geometry router's
    /// placement resolver, which recursively composes `parent * local` down the
    /// IfcLocalPlacement chain. For a well-formed acyclic placement DAG the
    /// resolved transform is a pure function of the placement id, so reusing it
    /// across the elements a single worker meshes is byte-identical (speed
    /// only): storey/building placements shared by thousands of elements are
    /// composed once per worker instead of once per element. Hoisted across a
    /// worker's parts exactly like `point_cache`; see
    /// [`Self::take_placement_transform_cache`].
    placement_transform_cache: FxHashMap<u32, [f64; 16]>,
    styled_item_index: std::sync::OnceLock<StyledItemIndexResult>,
    /// Declared `FILE_SCHEMA`, resolved on first use (see `schema_layout`).
    schema_version_cache: Option<Option<crate::generated::schema_registry::SchemaVersion>>,
    grid_axis_index: Arc<GridAxisIndex>,
}

impl<'a> EntityDecoder<'a> {
    /// Create new decoder
    pub fn new<T>(content: &'a T) -> Self
    where
        T: AsRef<[u8]> + ?Sized,
    {
        let content = content.as_ref();
        Self {
            content,
            cache: FxHashMap::default(),
            entity_index: None,
            point_cache: FxHashMap::default(),
            point_cache_hits: 0,
            point_cache_misses: 0,
            plane_angle_to_radians_cache: None,
            length_unit_scale_cache: None,
            geometric_precision_cache: None,
            placement_transform_cache: FxHashMap::default(),
            styled_item_index: std::sync::OnceLock::new(),
            schema_version_cache: None,
            grid_axis_index: Arc::default(),
        }
    }

    /// Create decoder with pre-built index (faster for repeated lookups)
    pub fn with_index<T>(content: &'a T, index: EntityIndex) -> Self
    where
        T: AsRef<[u8]> + ?Sized,
    {
        let content = content.as_ref();
        Self {
            content,
            cache: FxHashMap::default(),
            entity_index: Some(EntityIndexStore::Hash(Arc::new(index))),
            point_cache: FxHashMap::default(),
            point_cache_hits: 0,
            point_cache_misses: 0,
            plane_angle_to_radians_cache: None,
            length_unit_scale_cache: None,
            geometric_precision_cache: None,
            placement_transform_cache: FxHashMap::default(),
            styled_item_index: std::sync::OnceLock::new(),
            schema_version_cache: None,
            grid_axis_index: Arc::default(),
        }
    }

    /// Create decoder with shared Arc index (for parallel processing)
    pub fn with_arc_index<T>(content: &'a T, index: Arc<EntityIndex>) -> Self
    where
        T: AsRef<[u8]> + ?Sized,
    {
        let content = content.as_ref();
        Self {
            content,
            cache: FxHashMap::default(),
            entity_index: Some(EntityIndexStore::Hash(index)),
            point_cache: FxHashMap::default(),
            point_cache_hits: 0,
            point_cache_misses: 0,
            plane_angle_to_radians_cache: None,
            length_unit_scale_cache: None,
            geometric_precision_cache: None,
            placement_transform_cache: FxHashMap::default(),
            styled_item_index: std::sync::OnceLock::new(),
            schema_version_cache: None,
            grid_axis_index: Arc::default(),
        }
    }

    /// Install a pre-built index into this decoder. Used when the caller already
    /// built the index during another pass (e.g. the processor scan loop) so the
    /// decoder does not lazily rebuild it via `build_index`. Set it before any
    /// `decode_by_id`/ref-resolving call; afterwards `build_index` no-ops.
    pub fn set_entity_index(&mut self, index: Arc<EntityIndex>) {
        self.entity_index = Some(EntityIndexStore::Hash(index));
    }

    /// Build entity index for O(1) lookups
    /// This scans the file once and maps entity IDs to byte offsets
    fn build_index(&mut self) {
        if self.entity_index.is_some() {
            return; // Already built
        }
        self.entity_index = Some(EntityIndexStore::Hash(Arc::new(build_entity_index(self.content))));
    }

    /// Decode entity at byte offset
    /// Returns cached entity if already decoded
    ///
    /// Validates the `(start, end)` span against `self.content.len()` before
    /// slicing. Out-of-range or inverted spans return `Error::parse` instead
    /// of panicking — callers (e.g. `decode_and_cache`, `decode_at_with_id`,
    /// the streaming pre-pass shard mergers) hand us spans derived from
    /// untrusted/streamed entity-index data, and a malformed span must not
    /// take down the whole worker.
    #[inline]
    pub fn decode_at(&mut self, start: usize, end: usize) -> Result<DecodedEntity> {
        let (id, ifc_type, tokens) = self.parse_at(start, end, "decode_at")?;

        // Check cache first - return clone of inner DecodedEntity
        if let Some(entity_arc) = self.cache.get(&id) {
            return Ok(entity_arc.as_ref().clone());
        }

        // Convert tokens to AttributeValues
        let attributes = tokens
            .iter()
            .map(|token| AttributeValue::from_token(token))
            .collect();

        let entity = DecodedEntity::new(id, ifc_type, attributes);
        self.cache.insert(id, Arc::new(entity.clone()));
        Ok(entity)
    }

    /// Decode the entity in `[start, end)` **without** touching the cache.
    ///
    /// [`decode_at`](Self::decode_at) memoizes every entity it parses, which is the
    /// right trade-off for geometry sub-tree walks (the same points/profiles are
    /// revisited many times). For a single linear pass over *every* entity in a
    /// large model — tens of millions of rows — that cache grows without bound and
    /// dominates memory. This variant parses and returns the entity but never
    /// inserts it, so a streaming walk stays O(1) in entity count. The caller owns
    /// the result; for identical bytes it yields an identical [`DecodedEntity`] to
    /// `decode_at`, only without the cache side effect.
    pub fn decode_at_uncached(&self, start: usize, end: usize) -> Result<DecodedEntity> {
        let (id, ifc_type, tokens) = self.parse_at(start, end, "decode_at_uncached")?;
        let attributes = tokens
            .iter()
            .map(|token| AttributeValue::from_token(token))
            .collect();
        Ok(DecodedEntity::new(id, ifc_type, attributes))
    }

    /// Shared full-record validation and existing error context. The returned
    /// tokens borrow immutable source bytes, not the decoder's entity cache.
    #[inline]
    fn parse_at(
        &self, start: usize, end: usize, method: &str,
    ) -> Result<(u32, crate::IfcType, Vec<crate::parser::Token<'a>>)> {
        let content_len = self.content.len();
        if start > end || end > content_len {
            return Err(Error::parse(
                0,
                format!(
                    "{method}: invalid byte span ({}, {}) for content length {}",
                    start, end, content_len,
                ),
            ));
        }
        let line = &self.content[start..end];
        parse_entity(line).map_err(|e| {
            // Add bounded, lossy debug info without requiring the source to be UTF-8.
            let cut = line.len().min(100);
            Error::parse(
                0,
                format!(
                    "Failed to parse entity: {:?}, input: {:?}",
                    e,
                    String::from_utf8_lossy(&line[..cut])
                ),
            )
        })
    }

    /// Decode entity at byte offset with known ID (faster - checks cache before parsing)
    /// Use this when the scanner provides the entity ID to avoid re-parsing cached entities
    #[inline]
    pub fn decode_at_with_id(
        &mut self,
        id: u32,
        start: usize,
        end: usize,
    ) -> Result<DecodedEntity> {
        // Check cache first - avoid parsing if already decoded
        if let Some(entity_arc) = self.cache.get(&id) {
            return Ok(entity_arc.as_ref().clone());
        }

        // Not in cache, parse and cache
        self.decode_at(start, end)
    }

    /// Decode entity by ID - O(1) lookup using entity index
    #[inline]
    pub fn decode_by_id(&mut self, entity_id: u32) -> Result<DecodedEntity> {
        // Check cache first - return clone of inner DecodedEntity
        if let Some(entity_arc) = self.cache.get(&entity_id) {
            return Ok(entity_arc.as_ref().clone());
        }

        // Build index if not already built
        self.build_index();

        // O(1) lookup in index
        let (start, end) = self
            .entity_index
            .as_ref()
            .and_then(|idx| idx.lookup(entity_id))
            .ok_or_else(|| Error::parse(0, format!("Entity #{} not found", entity_id)))?;

        self.decode_at(start, end)
    }

    /// Multiplier that converts file plane-angle units to radians.
    ///
    /// Lazy-resolved on first call by scanning for IFCPROJECT and reading
    /// its IFCUNITASSIGNMENT. Cached for subsequent calls. Returns `1.0`
    /// when no plane-angle unit is declared (IFC spec default = RADIAN).
    ///
    /// Use this at curve-sampling time wherever an `IfcParameterValue` is
    /// interpreted as an angle (IfcCircle / IfcEllipse trim parameters).
    /// Without it, `value.to_radians()` is correct only for DEGREE files
    /// and silently shrinks arcs on RADIAN files (issue #820).
    pub fn plane_angle_to_radians(&mut self) -> f64 {
        if let Some(cached) = self.plane_angle_to_radians_cache {
            return cached;
        }

        let mut scanner = crate::parser::EntityScanner::new(self.content);
        let mut project_id: Option<u32> = None;
        while let Some((id, type_name, _, _)) = scanner.next_entity() {
            // #4497: the keyword is the raw scanner slice; see `parser::keyword`.
            if crate::parser::keyword_eq(type_name, "IFCPROJECT") {
                project_id = Some(id);
                break;
            }
        }

        let scale = match project_id {
            Some(pid) => crate::units::extract_plane_angle_to_radians(self, pid).unwrap_or(1.0),
            None => 1.0,
        };
        self.plane_angle_to_radians_cache = Some(scale);
        scale
    }

    /// Multiplier that converts file length units to metres (1.0 for metre
    /// files, 0.001 for millimetre files, …). Lazy-resolved on first call by
    /// scanning for IFCPROJECT and reading its IFCUNITASSIGNMENT, then cached.
    /// Returns `1.0` when no length unit is declared.
    ///
    /// Use this to express an *absolute* metric tolerance in file units —
    /// e.g. a curve-tessellation chord-deviation budget that stays constant in
    /// millimetres whether the file is authored in mm or m.
    pub fn length_unit_scale(&mut self) -> f64 {
        if let Some(cached) = self.length_unit_scale_cache {
            return cached;
        }

        let mut scanner = crate::parser::EntityScanner::new(self.content);
        let mut project_id: Option<u32> = None;
        while let Some((id, type_name, _, _)) = scanner.next_entity() {
            if crate::parser::keyword_eq(type_name, "IFCPROJECT") {
                project_id = Some(id);
                break;
            }
        }

        let scale = match project_id {
            Some(pid) => crate::units::extract_length_unit_scale(self, pid).unwrap_or(1.0),
            None => 1.0,
        };
        self.length_unit_scale_cache = Some(scale);
        scale
    }

    /// Pre-seed the unit-scale caches so [`Self::length_unit_scale`] and
    /// [`Self::plane_angle_to_radians`] return immediately without the full-file
    /// `IFCPROJECT` scan.
    ///
    /// Both lazy resolvers walk the whole DATA section to locate the (singleton)
    /// `IFCPROJECT`. That scan is `O(file size)` and `IFCPROJECT` legally sits
    /// anywhere — IfcOpenShell emits it near the *end*, so on a large model the
    /// scan touches tens of MB. The cache is per-decoder, and the parallel
    /// geometry pipeline builds a fresh decoder per element, so without seeding
    /// every arc-bearing element re-pays the scan (≈135 ms each on a 75 MB
    /// file). The orchestrator resolves both scales once on a warm shared
    /// decoder and seeds each worker decoder here.
    pub fn seed_unit_scales(&mut self, length_unit_scale: f64, plane_angle_to_radians: f64) {
        self.length_unit_scale_cache = Some(length_unit_scale);
        self.plane_angle_to_radians_cache = Some(plane_angle_to_radians);
    }

    /// Resolve entity reference (follow #ID)
    /// Returns None for null/derived values
    #[inline]
    pub fn resolve_ref(&mut self, attr: &AttributeValue) -> Result<Option<DecodedEntity>> {
        match attr.as_entity_ref() {
            Some(id) => Ok(Some(self.decode_by_id(id)?)),
            None => Ok(None),
        }
    }

    /// Resolve list of entity references
    pub fn resolve_ref_list(&mut self, attr: &AttributeValue) -> Result<Vec<DecodedEntity>> {
        let list = attr
            .as_list()
            .ok_or_else(|| Error::parse(0, "Expected list".to_string()))?;

        let mut entities = Vec::with_capacity(list.len());
        for item in list {
            if let Some(id) = item.as_entity_ref() {
                entities.push(self.decode_by_id(id)?);
            }
        }
        Ok(entities)
    }

    /// Get cached entity (without decoding)
    pub fn get_cached(&self, entity_id: u32) -> Option<DecodedEntity> {
        self.cache.get(&entity_id).map(|arc| arc.as_ref().clone())
    }

    /// Reserve cache capacity to avoid HashMap resizing during processing.
    /// For a 487 MB file with 208 K building elements, the cache can grow to
    /// 300 K+ entries (elements + representation chains + placements).
    /// Pre-allocating avoids ~6 resize-and-rehash operations that each copy
    /// all entries, reducing both peak memory spikes and timing variance.
    pub fn reserve_cache(&mut self, additional: usize) {
        self.cache.reserve(additional);
    }

    /// Inject a pre-warmed Arc-shared cache into this decoder's local cache.
    ///
    /// Used by the de-normalized parallel path: a serial pre-pass builds a
    /// shared `Arc<FxHashMap<u32, Arc<DecodedEntity>>>` containing all
    /// entities reachable from the jobs. Each rayon task then injects
    /// that shared cache into its own decoder via this method, so the
    /// per-task hot path hits in-WASM-heap Arc handles instead of
    /// SAB-imported atomic memory.
    ///
    /// Cost: one Arc::clone per cached entry (atomic refcount bump).
    /// For a typical 100K-entry cache × 9 rayon tasks = 900K atomics
    /// total, ~90 ms wall (incurred ONCE at task setup; the parallel
    /// hot path then runs lock-free against the populated cache).
    pub fn inject_shared_cache(&mut self, shared: &FxHashMap<u32, Arc<DecodedEntity>>) {
        self.cache.reserve(shared.len());
        for (&id, entity) in shared.iter() {
            self.cache.insert(id, Arc::clone(entity));
        }
    }

    /// Decode + cache without returning. Used by the pre-warm pass to
    /// populate a shared cache. Returns the cached Arc so the caller
    /// can chase references without re-decoding.
    pub fn decode_and_cache(
        &mut self,
        id: u32,
        start: usize,
        end: usize,
    ) -> Result<Arc<DecodedEntity>> {
        if let Some(arc) = self.cache.get(&id) {
            return Ok(Arc::clone(arc));
        }
        let _ = self.decode_at(start, end)?;
        Ok(Arc::clone(self.cache.get(&id).ok_or_else(|| {
            Error::parse(0, "decode_at didn't populate cache".to_string())
        })?))
    }
    /// Get raw bytes for an entity (for direct/fast parsing)
    /// Returns the full entity line including type and attributes
    #[inline]
    pub fn get_raw_bytes(&mut self, entity_id: u32) -> Option<&'a [u8]> {
        self.build_index();
        let (start, end) = self.entity_index.as_ref()?.lookup(entity_id)?;
        // An installed index's spans are caller data never checked against
        // these bytes (#4697); a span outside them is absent, not a panic.
        self.content.get(start..end)
    }

    /// Fast extraction of first entity ref from raw bytes
    /// Useful for BREP -> shell ID, Face -> FaceBound, etc.
    /// Returns the first entity reference ID found in the entity
    #[inline]
    pub fn get_first_entity_ref_fast(&mut self, entity_id: u32) -> Option<u32> {
        let bytes = self.get_raw_bytes(entity_id)?;
        let len = bytes.len();
        let mut i = 0;

        // Skip to first '(' after '='
        while i < len && bytes[i] != b'(' {
            i += 1;
        }
        if i >= len {
            return None;
        }
        i += 1; // Skip first '('

        // Find first '#' which is the entity ref
        while i < len {
            // Skip whitespace
            while i < len && is_step_space(bytes[i]) {
                i += 1;
            }

            if i >= len {
                return None;
            }

            if bytes[i] == b'#' {
                i += 1;
                let start = i;
                while i < len && bytes[i].is_ascii_digit() {
                    i += 1;
                }
                if i > start {
                    // Shared checked accumulator (#3421): refuses rather than wraps.
                    return crate::express_id::parse_express_id(&bytes[start..i]);
                }
            }
            i += 1;
        }

        None
    }

    /// Fast extraction of PolyLoop point IDs directly from raw bytes
    /// Bypasses full entity decoding for BREP optimization
    /// Returns list of entity IDs for CartesianPoints
    ///
    /// `None` when ANY `#<digits>` in the list is unrepresentable, not just when
    /// none of them is. A reference above `u32::MAX` is refused rather than
    /// wrapped (#3421), and dropping only that vertex would hand the caller a
    /// polygon one corner SHORTER than the file's: a different face, meshed
    /// and rendered as if it were the authored one.
    ///
    /// This accessor sees ids, not entities, so a reference that fits `u32`
    /// but names no record in the file is returned as written. Whether that
    /// id RESOLVES is the caller's half of the same rule, and every caller
    /// must refuse the whole loop on a miss rather than skip the point
    /// (`processors::helpers::extract_loop_points_by_id` and
    /// `processors::surface` both do). The coordinate sibling
    /// [`Self::get_polyloop_coords_cached_into`] enforces both halves itself
    /// because it resolves the points too.
    #[inline]
    pub fn get_polyloop_point_ids_fast(&mut self, entity_id: u32) -> Option<Vec<u32>> {
        let bytes = self.get_raw_bytes(entity_id)?;

        // IFCPOLYLOOP((#id1,#id2,#id3,...));
        let mut i = 0;
        let len = bytes.len();

        // Skip to first '(' after '='
        while i < len && bytes[i] != b'(' {
            i += 1;
        }
        if i >= len {
            return None;
        }
        i += 1; // Skip first '('

        // Skip to second '(' for the point list
        while i < len && bytes[i] != b'(' {
            i += 1;
        }
        if i >= len {
            return None;
        }
        i += 1; // Skip second '('

        // Parse point IDs
        let mut point_ids = Vec::with_capacity(8); // Most faces have 3-8 vertices

        while i < len {
            // Skip whitespace and commas
            while i < len
                && (bytes[i] == b',' || is_step_space(bytes[i]))
            {
                i += 1;
            }

            if i >= len || bytes[i] == b')' {
                break;
            }

            // Expect '#' followed by number
            if bytes[i] == b'#' {
                i += 1;
                // One digit walk, shared with the scanner (#3395). A ref that
                // was met but does not resolve refuses the whole loop (#3421):
                // a shorter polygon is a different face, not a smaller error.
                let (digits, id) = crate::express_id::parse_express_id_prefix(&bytes[i..]);
                i += digits;
                if digits > 0 {
                    point_ids.push(id?);
                }
            } else {
                i += 1; // Skip unknown character
            }
        }

        if point_ids.is_empty() {
            None
        } else {
            Some(point_ids)
        }
    }

    /// Fast extraction of CartesianPoint coordinates directly from raw bytes
    /// Bypasses full entity decoding for ~3x speedup on BREP-heavy files
    /// Returns (x, y, z) as f64 tuple
    #[inline]
    pub fn get_cartesian_point_fast(&mut self, entity_id: u32) -> Option<(f64, f64, f64)> {
        parse_cartesian_point_inline(self.get_raw_bytes(entity_id)?)
    }

    /// Fast extraction of FaceBound info directly from raw bytes
    /// Returns (loop_id, orientation, is_outer_bound)
    /// Bypasses full entity decoding for BREP optimization
    #[inline]
    pub fn get_face_bound_fast(&mut self, entity_id: u32) -> Option<(u32, bool, bool)> {
        let bytes = self.get_raw_bytes(entity_id)?;
        let len = bytes.len();

        // Find '=' to locate start of type name, and '(' for end
        let mut eq_pos = 0;
        while eq_pos < len && bytes[eq_pos] != b'=' {
            eq_pos += 1;
        }
        if eq_pos >= len {
            return None;
        }

        // Check if this is an outer bound by looking for "OUTER" in the type name
        // IFCFACEOUTERBOUND vs IFCFACEBOUND
        // The type name is between '=' and '('
        let mut is_outer = false;
        let mut i = eq_pos + 1;
        // Look for "OUTER" pattern (must check for the full word, not just 'O')
        while i + 4 < len && bytes[i] != b'(' {
            if bytes[i] == b'O'
                && bytes[i + 1] == b'U'
                && bytes[i + 2] == b'T'
                && bytes[i + 3] == b'E'
                && bytes[i + 4] == b'R'
            {
                is_outer = true;
                break;
            }
            i += 1;
        }
        // Continue to find the '(' if we haven't already
        while i < len && bytes[i] != b'(' {
            i += 1;
        }
        if i >= len {
            return None;
        }

        i += 1; // Skip first '('

        // Skip whitespace
        while i < len && is_step_space(bytes[i]) {
            i += 1;
        }

        // Expect '#' for loop entity ref
        if i >= len || bytes[i] != b'#' {
            return None;
        }
        i += 1;

        // Parse loop ID
        let start = i;
        while i < len && bytes[i].is_ascii_digit() {
            i += 1;
        }
        if i <= start {
            return None;
        }
        // Shared checked accumulator (#3421): an oversized loop ref fails
        // this lookup outright, like any other unresolvable reference,
        // instead of resolving to the wrong loop.
        let loop_id = crate::express_id::parse_express_id(&bytes[start..i])?;

        // Find orientation after comma - default to true (.T.)
        // Skip to comma
        while i < len && bytes[i] != b',' {
            i += 1;
        }
        i += 1; // Skip comma

        // Skip whitespace
        while i < len && is_step_space(bytes[i]) {
            i += 1;
        }

        // Check for .F. (false) or .T. (true)
        let orientation = if i + 2 < len && bytes[i] == b'.' && bytes[i + 2] == b'.' {
            bytes[i + 1] != b'F'
        } else {
            true // Default to true
        };

        Some((loop_id, orientation, is_outer))
    }
}

/// Parse cartesian point coordinates inline from raw bytes
/// Used by the cached polyloop coordinate readers for maximum performance
#[inline]
fn parse_cartesian_point_inline(bytes: &[u8]) -> Option<(f64, f64, f64)> {
    let len = bytes.len();
    let mut i = 0;

    // Skip to first '(' after '='
    while i < len && bytes[i] != b'(' {
        i += 1;
    }
    if i >= len {
        return None;
    }
    i += 1; // Skip first '('

    // Skip to second '(' for the coordinate list
    while i < len && bytes[i] != b'(' {
        i += 1;
    }
    if i >= len {
        return None;
    }
    i += 1; // Skip second '('

    parse_point_coordinates(&bytes[i..])
}

/// `x,y)` or `x,y,z)`, starting just inside the coordinate list's `(`. Every
/// value goes through the shared STEP literal grammar and every separator is
/// checked, so a corrupted literal (`1.52.3`), a non-STEP spelling (`nan`) or a
/// corrupt last value (`3.x`) refuses the whole point instead of splitting it,
/// reading it as its prefix, or defaulting it to 0, which is what the full
/// tokenizer does with the same record (#5266). Comments are trivia here as
/// everywhere else (`3./* c */)` is a legal z). z defaults to 0 only for a
/// genuine 2D point, whose list closes right after y.
#[inline]
fn parse_point_coordinates(bytes: &[u8]) -> Option<(f64, f64, f64)> {
    let value_at = |i: usize| -> Option<(f64, usize)> {
        let start = skip_step_trivia(bytes, i)?;
        let (value, len) = parse_step_numeric::<f64>(&bytes[start..])?;
        Some((value, skip_step_trivia(bytes, start + len)?))
    };
    let (x, i) = value_at(0)?;
    if bytes.get(i) != Some(&b',') {
        return None;
    }
    let (y, i) = value_at(i + 1)?;
    match bytes.get(i)? {
        b')' => return Some((x, y, 0.0)),
        b',' => {}
        _ => return None,
    }
    let (z, i) = value_at(i + 1)?;
    (bytes.get(i) == Some(&b')')).then_some((x, y, z))
}

#[cfg(test)]
mod decoder_tests;