vole_document/materialize/

observation.rs

//! Partial materialization / observation views (Phase 7.3).
//!
//! [`materialize_observation`] serves a narrow byte range of the reconstructed
//! source from a parsed descriptor that carries an observation index. It is a
//! *view* API: the complete-materialization path
//! ([`crate::materialize::materialize`]) is unchanged and remains the authority
//! for whole-source exactness.
//!
//! ## Decline, never guess
//!
//! A descriptor without an index is declined with
//! `ErrorClass::UnsupportedFeature`; the function never silently falls back to
//! full materialization, so a reported cost is never a whole-file cost in
//! disguise.
//!
//! ## Op selection: linear skipping vs. a bounded prefix
//!
//! The op layout is re-derived from the program with
//! [`Program::analyze_ops`] (the index's op table is advisory; the program is
//! authoritative). Let the requested output range be `[a, b)`.
//!
//! * If every op is a *linear, independent block* producer — one of
//!   [`Op::EmitObject`], [`Op::Inline`], [`Op::DecodeChannel`],
//!   [`Op::DeflateReplay`], or [`Op::InterleaveChannels`] — then each op's output
//!   is a pure function of its own inputs and its absolute start offset. Only the
//!   ops whose output intersects `[a, b)` are evaluated; ops entirely before `a`
//!   are skipped (no bytes produced, no channel decoded).
//! * Otherwise (the program uses any position-dependent op: [`Op::RepeatLast`],
//!   [`Op::MarkOffset`], [`Op::EmitOffset`], [`Op::PackSegments`], or
//!   [`Op::PackedChannels`]) the ops `0..=last` are evaluated as a prefix. This
//!   is correct but can produce up to the whole document; it is the honest
//!   fallback. [`ObservationStats::work_amplification`] reports how much of the
//!   program was walked.
//!
//! ## Lazy entropy-channel decoding
//!
//! Only the entropy channels referenced by the evaluated ops are decoded; the
//! complete path in [`crate::materialize`] still decodes every channel. A
//! referenced channel is decoded in full, so the conversion-contract tail check
//! is preserved for exactly the channels that are used. Objects are already
//! resident in the descriptor, so "fetched" here means "read by an evaluated op".
//!
//! ## `descriptor_bytes_traversed` is an approximation
//!
//! This in-memory path parses the whole framed descriptor into memory, so it has
//! no true I/O seek accounting. [`ObservationStats::descriptor_bytes_traversed`]
//! reports the sum of the serialized record payload lengths the path *needed* —
//! the graph record, the referenced object payloads, the referenced channel
//! payloads, and the index record — as a documented CPU-side approximation, not
//! a byte-read figure. The referenced channel payloads are already included here,
//! so [`ObservationStats::entropy_bytes_decoded`] is a **subset** of
//! `descriptor_bytes_traversed` and the two fields must never be summed. This
//! path reports `bytes_read == 0` (it performs no I/O of its own); the
//! seek reader ([`crate::materialize::seek::materialize_observation_seeked`],
//! Phase 8) reports a real, instrumented `bytes_read` while sharing this path's
//! op selection and evaluation verbatim.
//!
//! A partial `view` — in-memory or seeked — is an *observation*: it serves bytes
//! consistent with the descriptor's own validated program/index/directory, but it
//! never recomputes the whole-source SHA-256, so `integrity_verified` is `false`.
//! Only `materialize`/`decode`/`verify` are the archival authority.

use crate::container::ParsedDescriptor;
use crate::container::observation::{
    ObservationIndex, ObservationSelector as IndexSelector, SECTION_PDF_SELECTORS, SELECTOR_OBJECT,
    SELECTOR_REVISION, SELECTOR_STREAM,
};
use crate::dra::{Op, Program};
use crate::entropy::codec::EntropyChannelDescriptor;
use crate::entropy::model::EntropyModel;
use crate::error::{Error, Result};
use crate::limits::Limits;

#[cfg(feature = "rans")]
use crate::entropy::rans::{Capsule, decode_channel};

/// A narrow observation of the reconstructed source.
///
/// Every selector resolves to a half-open output range `[a, b)` in the source
/// address space (output offset `O` is source offset `O`, by the exact profile).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ObservationSelector {
    /// Raw output byte range `[offset, offset + len)`.
    ByteRange {
        /// First output byte.
        offset: u64,
        /// Number of bytes; must be non-empty.
        len: u64,
    },
    /// The full physical byte extent of indirect object `object generation`
    /// (the `N G obj` introducer through `endobj`).
    PdfIndirectObject {
        /// PDF object number.
        object: u32,
        /// PDF generation number.
        generation: u16,
    },
    /// The physical encoded stream data span of the stream in indirect object
    /// `object generation`: the compressed bytes exactly as they appear in the
    /// source.
    PdfEncodedStream {
        /// PDF object number of the enclosing indirect object.
        object: u32,
        /// PDF generation number of the enclosing indirect object.
        generation: u16,
    },
    /// Every source byte belonging to incremental-update revision `index`.
    PdfRevision {
        /// Zero-based revision index.
        index: u32,
    },
}

/// Cost attribution for one observation.
///
/// All fields are measured from the evaluated path, not estimated.
#[derive(Debug, Clone, Copy, PartialEq)]
pub struct ObservationStats {
    /// Number of ops the observation actually evaluated.
    pub ops_evaluated: usize,
    /// Total ops in the descriptor's reconstruction program.
    pub ops_total: usize,
    /// Number of distinct raw objects read by evaluated ops.
    pub objects_fetched: usize,
    /// Total raw objects in the descriptor.
    pub objects_total: usize,
    /// Number of entropy channels decoded for this observation.
    pub channels_decoded: usize,
    /// Total entropy channels in the descriptor.
    pub channels_total: usize,
    /// Sum of the encoded renormalization payload bytes of the decoded channels.
    ///
    /// This is the descriptor-side entropy bytes consumed, not the decoded symbol
    /// count (which is the referenced channels' `decoded_length`). It is a
    /// **subset** of [`Self::descriptor_bytes_traversed`], which already counts
    /// the same referenced channel payloads, so it is a breakdown and must
    /// **never** be added to that field — the sum double-counts.
    pub entropy_bytes_decoded: u64,
    /// Approximate descriptor bytes the path needed (see the module docs).
    ///
    /// The honest decode-side bound: the graph record, the index record, the
    /// referenced object payloads, and the referenced channel payloads (which
    /// include [`Self::entropy_bytes_decoded`]).
    pub descriptor_bytes_traversed: u64,
    /// Number of output bytes served.
    pub output_bytes: u64,
    /// Real bytes read from the source by the seek reader.
    ///
    /// Zero for the in-memory Phase-7 path, which is handed a fully parsed
    /// descriptor and performs no I/O of its own. This is a *real* I/O figure and
    /// is distinct from [`Self::descriptor_bytes_traversed`], which is a CPU-side
    /// approximation kept for comparison.
    pub bytes_read: u64,
    /// Whether the whole-source archival digest was verified for this report.
    ///
    /// Always `false` for an observation view: a partial read cannot recompute the
    /// whole-source SHA-256, so a served slice is an *observation* consistent with
    /// the descriptor's own validated program/index/directory -- not a verified
    /// archival read. Only `materialize`/`decode`/`verify` check `INTEGRITY` and
    /// are the archival authority.
    pub integrity_verified: bool,
}

impl ObservationStats {
    /// Fraction of the program evaluated: `ops_evaluated / ops_total`.
    ///
    /// Returns `0.0` when the program has no ops, so the ratio is always finite.
    pub fn work_amplification(&self) -> f64 {
        if self.ops_total == 0 {
            0.0
        } else {
            self.ops_evaluated as f64 / self.ops_total as f64
        }
    }
}

/// A served observation: the exact requested bytes plus its cost attribution.
#[derive(Debug, Clone, PartialEq)]
pub struct ObservationReport {
    /// The resolved half-open output range `[a, b)` that was served. Reporting it
    /// lets a measurement harness compare the slice against the source and price
    /// the sequential baselines at the same output offset.
    pub range: (u64, u64),
    /// The exact requested output range.
    pub bytes: Vec<u8>,
    /// Measured cost attribution for serving that range.
    pub stats: ObservationStats,
}

/// The op window an observation range selects, shared by the in-memory
/// (Phase-7) and seek (Phase-8) readers so the two selection paths cannot
/// diverge.
#[derive(Debug, Clone)]
pub(crate) struct OpWindow {
    /// The ops to evaluate, in program order.
    pub ops: Vec<Op>,
    /// The absolute output offset the evaluated buffer begins at. Zero when the
    /// program was walked as a prefix (a position-dependent op forced a fallback).
    pub buf_start: u64,
    /// Total ops in the descriptor's program (for `work_amplification`).
    pub ops_total: usize,
}

/// Select the minimal op set serving output range `[a, b)`.
///
/// This is the exact Phase-7 selection, factored out so `materialize_observation`
/// and the seek reader share one implementation. For a program of linear,
/// independent block producers only the ops whose output intersects `[a, b)` are
/// selected (and `buf_start` is that window's absolute start); any
/// position-dependent op forces the honest bounded prefix `ops[..=last]` with
/// `buf_start == 0`.
pub(crate) fn select_ops(
    program: &Program,
    object_lens: &[u64],
    channel_lens: &[u64],
    a: u64,
    b: u64,
    limits: Limits,
) -> Result<OpWindow> {
    let per_op = program.analyze_ops(object_lens, channel_lens, limits)?;
    select_ops_from_lengths(program, &per_op, a, b)
}

/// Select the minimal op set serving output range `[a, b)` from an explicit
/// per-op output-length vector (program order).
///
/// This is the length-agnostic core of [`select_ops`], factored so a partial
/// reader that learns per-op lengths from a validated observation-index op table
/// can select a window with **exactly** the same linear-skipping vs. bounded-
/// prefix rule as the in-memory path. The caller is responsible for the lengths
/// being the authoritative ones (the program's own `analyze_ops` view).
pub(crate) fn select_ops_from_lengths(
    program: &Program,
    per_op: &[u64],
    a: u64,
    b: u64,
) -> Result<OpWindow> {
    let mut starts: Vec<u64> = Vec::with_capacity(per_op.len());
    let mut ends: Vec<u64> = Vec::with_capacity(per_op.len());
    let mut acc: u64 = 0;
    for &len in per_op {
        starts.push(acc);
        acc = acc
            .checked_add(len)
            .ok_or_else(|| Error::invalid_graph("observation op length overflow"))?;
        ends.push(acc);
    }

    let intersects = |i: usize| per_op[i] > 0 && starts[i] < b && ends[i] > a;
    let first = (0..per_op.len())
        .find(|&i| intersects(i))
        .ok_or_else(|| Error::invalid_graph("observation range is not covered by the program"))?;
    let last = (0..per_op.len())
        .rev()
        .find(|&i| intersects(i))
        .ok_or_else(|| Error::invalid_graph("observation range is not covered by the program"))?;

    let (ops, buf_start) = if is_linear_independent(program) {
        let ops = program
            .ops
            .iter()
            .enumerate()
            .filter(|&(i, _)| intersects(i))
            .map(|(_, op)| op.clone())
            .collect();
        (ops, starts[first])
    } else {
        (program.ops[..=last].to_vec(), 0)
    };

    Ok(OpWindow {
        ops,
        buf_start,
        ops_total: program.ops.len(),
    })
}

/// Mark the objects and channels a selected op set references.
///
/// The seek reader uses this to decide which `OBJECT`/`ENTROPY_CHANNEL`/`MODEL`
/// records it must read; the in-memory path uses it to decode lazily. Both paths
/// must agree, so it lives here.
pub(crate) fn selection_references(
    ops: &[Op],
    objects_len: usize,
    channels_len: usize,
) -> (Vec<bool>, Vec<bool>) {
    let mut objects_used = vec![false; objects_len];
    let mut channels_used = vec![false; channels_len];
    for op in ops {
        mark_references(op, &mut objects_used, &mut channels_used);
    }
    (objects_used, channels_used)
}

/// The served bytes plus the measured cost attribution of one evaluated
/// selection, before the path-specific `descriptor_bytes_traversed`/`bytes_read`
/// fields are attached.
pub(crate) struct ServedSelection {
    pub bytes: Vec<u8>,
    pub ops_evaluated: usize,
    pub ops_total: usize,
    pub objects_fetched: usize,
    pub referenced_object_bytes: u64,
    pub channels_decoded: usize,
    pub entropy_bytes_decoded: u64,
    pub referenced_channel_bytes: u64,
}

/// Evaluate `window` and slice the requested `[a, b)`, decoding only the
/// referenced channels.
///
/// This is the unchanged Phase-7 evaluation body: lazy channel decoding, the
/// sub-program `eval`, and the window slice. It is shared by both readers; the
/// seek reader passes partial `objects`/`channels`/`models` vectors where only the
/// referenced index positions are populated, and every unreferenced position is a
/// never-dereferenced placeholder.
#[allow(clippy::too_many_arguments)]
pub(crate) fn serve_selection(
    objects: &[Vec<u8>],
    channels_desc: &[EntropyChannelDescriptor],
    models: &[EntropyModel],
    window: OpWindow,
    objects_used: &[bool],
    channels_used: &[bool],
    a: u64,
    b: u64,
    limits: Limits,
) -> Result<ServedSelection> {
    let channels = decode_referenced_channels(channels_desc, models, channels_used, limits)?;

    let ops_evaluated = window.ops.len();
    let ops_total = window.ops_total;
    let buf_start = window.buf_start;
    let sub = Program::new(window.ops);
    let out = sub.eval(objects, &channels, limits)?;

    let lo = a
        .checked_sub(buf_start)
        .and_then(|v| usize::try_from(v).ok())
        .ok_or_else(|| Error::internal_invariant("observation window precedes evaluated buffer"))?;
    let hi = b
        .checked_sub(buf_start)
        .and_then(|v| usize::try_from(v).ok())
        .ok_or_else(|| Error::internal_invariant("observation window overflow"))?;
    if hi > out.len() {
        return Err(Error::internal_invariant(
            "evaluated buffer is shorter than the requested observation window",
        ));
    }
    let bytes = out[lo..hi].to_vec();

    let mut entropy_bytes_decoded: u64 = 0;
    let mut referenced_channel_bytes: u64 = 0;
    let mut channels_decoded: usize = 0;
    for (id, used) in channels_used.iter().enumerate() {
        if *used {
            channels_decoded += 1;
            let payload_len = channels_desc[id].payload.len() as u64;
            entropy_bytes_decoded += payload_len;
            referenced_channel_bytes += payload_len;
        }
    }
    let mut objects_fetched: usize = 0;
    let mut referenced_object_bytes: u64 = 0;
    for (id, used) in objects_used.iter().enumerate() {
        if *used {
            objects_fetched += 1;
            referenced_object_bytes += objects[id].len() as u64;
        }
    }

    Ok(ServedSelection {
        bytes,
        ops_evaluated,
        ops_total,
        objects_fetched,
        referenced_object_bytes,
        channels_decoded,
        entropy_bytes_decoded,
        referenced_channel_bytes,
    })
}

/// Serve one observation from a parsed descriptor carrying an observation index.
///
/// The returned bytes equal `materialize(parsed)[a..b]` for the resolved range;
/// a descriptor without an index is declined (never silently fully
/// materialized). See the module documentation for the op-selection and lazy
/// decoding rules, and for the `descriptor_bytes_traversed` caveat.
pub fn materialize_observation(
    parsed: &ParsedDescriptor,
    selector: ObservationSelector,
    limits: Limits,
) -> Result<ObservationReport> {
    let d = &parsed.descriptor;
    let index = d
        .observation_index
        .as_ref()
        .ok_or_else(|| Error::unsupported_feature("descriptor has no observation index"))?;

    // The in-memory observation lane takes no resolver; a descriptor carrying
    // external object references is declined rather than partially served.
    if d.objects
        .iter()
        .any(|o| matches!(o, crate::container::ObjectSource::External { .. }))
    {
        return Err(Error::unsupported_feature(
            "partial observation cannot resolve external objects (no resolver supplied)",
        ));
    }
    let objects: Vec<Vec<u8>> = d
        .objects
        .iter()
        .map(|o| o.as_inline().unwrap_or(&[]).to_vec())
        .collect();

    // 2. Resolve the selector to an output range `[a, b)`.
    let (a, b) = resolve_selector(index, selector, d.source_len)?;

    // 3. Per-op lengths and cumulative offsets. The program is authoritative; the
    //    index's op table was already cross-checked against it at parse time.
    let object_lens: Vec<u64> = objects.iter().map(|o| o.len() as u64).collect();
    let channel_lens: Vec<u64> = d.channels.iter().map(|c| c.decoded_length).collect();

    // 4. Selection and lazy evaluation, shared verbatim with the seek reader.
    let window = select_ops(&d.program, &object_lens, &channel_lens, a, b, limits)?;
    let (objects_used, channels_used) =
        selection_references(&window.ops, objects.len(), d.channels.len());
    let served = serve_selection(
        &objects,
        &d.channels,
        &d.models,
        window,
        &objects_used,
        &channels_used,
        a,
        b,
        limits,
    )?;

    let descriptor_bytes_traversed = parsed.cost.graph
        + parsed.cost.index
        + served.referenced_object_bytes
        + served.referenced_channel_bytes;

    let stats = ObservationStats {
        ops_evaluated: served.ops_evaluated,
        ops_total: served.ops_total,
        objects_fetched: served.objects_fetched,
        objects_total: objects.len(),
        channels_decoded: served.channels_decoded,
        channels_total: d.channels.len(),
        entropy_bytes_decoded: served.entropy_bytes_decoded,
        descriptor_bytes_traversed,
        output_bytes: served.bytes.len() as u64,
        bytes_read: 0,
        integrity_verified: false,
    };

    Ok(ObservationReport {
        range: (a, b),
        bytes: served.bytes,
        stats,
    })
}

/// Resolve a selector to a target output range `[a, b)`.
pub(crate) fn resolve_selector(
    index: &ObservationIndex,
    selector: ObservationSelector,
    source_len: u64,
) -> Result<(u64, u64)> {
    match selector {
        ObservationSelector::ByteRange { offset, len } => {
            if len == 0 {
                return Err(Error::usage("observation byte range must be non-empty"));
            }
            let end = offset
                .checked_add(len)
                .ok_or_else(|| Error::usage("observation byte range overflows"))?;
            if end > source_len {
                return Err(Error::usage(format!(
                    "observation byte range {offset}..{end} exceeds source length {source_len}"
                )));
            }
            Ok((offset, end))
        }
        ObservationSelector::PdfIndirectObject { object, generation } => resolve_pdf(
            index,
            SELECTOR_OBJECT,
            object,
            u32::from(generation),
            "indirect object",
        ),
        ObservationSelector::PdfEncodedStream { object, generation } => resolve_pdf(
            index,
            SELECTOR_STREAM,
            object,
            u32::from(generation),
            "encoded stream",
        ),
        ObservationSelector::PdfRevision { index: rev } => {
            if index.section_flags & SECTION_PDF_SELECTORS == 0 {
                return Err(Error::unsupported_feature(
                    "observation index has no PDF selector table",
                ));
            }
            let hits: Vec<&IndexSelector> = index
                .selectors
                .iter()
                .filter(|s| s.kind == SELECTOR_REVISION && s.number == rev)
                .collect();
            match hits.as_slice() {
                [one] => selector_range(one, "PDF revision", u64::from(rev)),
                [] => Err(Error::unsupported_feature(format!(
                    "observation index has no selector for PDF revision {rev}"
                ))),
                _ => Err(Error::invalid_container(format!(
                    "observation index has ambiguous selectors for PDF revision {rev}"
                ))),
            }
        }
    }
}

/// Look up a `(kind, number, generation)` PDF selector; absent or ambiguous is a
/// typed error (never a guess).
fn resolve_pdf(
    index: &ObservationIndex,
    kind: u8,
    number: u32,
    generation: u32,
    label: &str,
) -> Result<(u64, u64)> {
    if index.section_flags & SECTION_PDF_SELECTORS == 0 {
        return Err(Error::unsupported_feature(
            "observation index has no PDF selector table",
        ));
    }
    let hits: Vec<&IndexSelector> = index
        .selectors
        .iter()
        .filter(|s| s.kind == kind && s.number == number && s.generation == generation)
        .collect();
    match hits.as_slice() {
        [one] => selector_range(one, label, u64::from(number)),
        [] => Err(Error::unsupported_feature(format!(
            "observation index has no selector for {label} {number} {generation}"
        ))),
        _ => Err(Error::invalid_container(format!(
            "observation index has ambiguous selectors for {label} {number} {generation}"
        ))),
    }
}

/// Convert a validated index selector into a checked `[a, b)` range.
fn selector_range(selector: &IndexSelector, label: &str, id: u64) -> Result<(u64, u64)> {
    let end = selector
        .out_off
        .checked_add(selector.out_len)
        .ok_or_else(|| {
            Error::invalid_container(format!("{label} {id} selector range overflows"))
        })?;
    Ok((selector.out_off, end))
}

/// True when every op is a linear, independently evaluable block producer.
///
/// Position-dependent ops (`REPEAT_LAST`, `MARK_OFFSET`, `EMIT_OFFSET`,
/// `PACK_SEGMENTS`, `PACKED_CHANNELS`) disqualify skipping because their output
/// depends on the running output position or on a previous block's bytes.
fn is_linear_independent(program: &Program) -> bool {
    program.ops.iter().all(|op| {
        matches!(
            op,
            Op::EmitObject { .. }
                | Op::Inline { .. }
                | Op::DecodeChannel { .. }
                | Op::DeflateReplay { .. }
                | Op::InterleaveChannels { .. }
        )
    })
}

/// Record the objects and channels an op reads from.
fn mark_references(op: &Op, objects: &mut [bool], channels: &mut [bool]) {
    match op {
        Op::EmitObject { object_id } => mark(objects, *object_id),
        Op::Inline { .. }
        | Op::MarkOffset { .. }
        | Op::EmitOffset { .. }
        | Op::RepeatLast { .. } => {}
        Op::DecodeChannel { channel_id } => mark(channels, *channel_id),
        Op::InterleaveChannels {
            kinds_channel,
            lengths_channel,
            first_payload_channel,
            payload_channel_count,
        } => {
            mark(channels, *kinds_channel);
            mark(channels, *lengths_channel);
            for k in 0..u32::from(*payload_channel_count) {
                if let Some(id) = first_payload_channel.checked_add(k) {
                    mark(channels, id);
                }
            }
        }
        Op::PackSegments { data_object, .. } => mark(objects, *data_object),
        Op::PackedChannels {
            data_channel,
            plan_channel,
            ..
        } => {
            mark(channels, *data_channel);
            mark(channels, *plan_channel);
        }
        Op::DeflateReplay {
            source_kind,
            source_id,
            corrections_object,
            ..
        } => {
            match *source_kind {
                crate::dra::op::DEFLATE_SOURCE_OBJECT => mark(objects, *source_id),
                crate::dra::op::DEFLATE_SOURCE_CHANNEL => mark(channels, *source_id),
                _ => {}
            }
            mark(objects, *corrections_object);
        }
    }
}

/// Mark `id` in `flags` when it is in range; an out-of-range id is left to the
/// evaluator's own validation.
fn mark(flags: &mut [bool], id: u32) {
    if let Some(slot) = flags.get_mut(id as usize) {
        *slot = true;
    }
}

/// Decode only the entropy channels referenced by the evaluated ops.
#[cfg(feature = "rans")]
fn decode_referenced_channels(
    channels_desc: &[EntropyChannelDescriptor],
    models: &[EntropyModel],
    channels_used: &[bool],
    limits: Limits,
) -> Result<Vec<Vec<u8>>> {
    let mut channels: Vec<Vec<u8>> = vec![Vec::new(); channels_desc.len()];
    for (id, used) in channels_used.iter().enumerate() {
        if *used {
            channels[id] = decode_channel_by_id(channels_desc, models, id, limits)?;
        }
    }
    Ok(channels)
}

/// Without the `rans` feature a referenced channel cannot be decoded; decline
/// explicitly rather than silently producing wrong bytes.
#[cfg(not(feature = "rans"))]
fn decode_referenced_channels(
    channels_desc: &[EntropyChannelDescriptor],
    _models: &[EntropyModel],
    channels_used: &[bool],
    _limits: Limits,
) -> Result<Vec<Vec<u8>>> {
    if channels_used.iter().any(|&used| used) {
        return Err(Error::unsupported_feature(
            "this build was compiled without the `rans` feature",
        ));
    }
    Ok(vec![Vec::new(); channels_desc.len()])
}

/// Decode one referenced channel against the model it names.
#[cfg(feature = "rans")]
fn decode_channel_by_id(
    channels_desc: &[EntropyChannelDescriptor],
    models: &[EntropyModel],
    id: usize,
    limits: Limits,
) -> Result<Vec<u8>> {
    let channel = channels_desc.get(id).ok_or_else(|| {
        Error::invalid_model(format!(
            "observation references missing entropy channel {id}"
        ))
    })?;
    let model = models.get(channel.model_id as usize).ok_or_else(|| {
        Error::invalid_model(format!(
            "entropy channel {id} references missing model {}",
            channel.model_id
        ))
    })?;
    let capsule = Capsule {
        initial_state: channel.initial_state,
        payload: channel.payload.clone(),
        symbol_count: channel.symbol_count,
        decoded_length: channel.decoded_length,
    };
    decode_channel(model, &capsule, limits)
}