dbmd_core/

parser.rs

//! `parser` — read and write db.md markdown files.
//!
//! Parses the YAML frontmatter block, the markdown body, wiki-links, standard
//! markdown links, `##` sections, and the structured sections of the `DB.md`
//! config file. Also the atomic writer that round-trips a file while
//! preserving the operator-edited body verbatim and emitting frontmatter in
//! canonical key order.
//!
//! Strict on required fields, lenient on unknowns: any frontmatter key the
//! spec doesn't recognize is preserved in [`Frontmatter::extra`] as ambient
//! context and round-tripped untouched.

use std::collections::BTreeMap;
use std::path::{Path, PathBuf};

use chrono::{DateTime, FixedOffset};
use serde_norway::{Mapping, Value};

/// One db.md text file is bounded before allocation/YAML parsing. Interactive
/// records are intentionally small; a larger blob belongs in the asset layer.
/// This prevents a hostile path or imported store from making every read,
/// validate, or format operation allocate an unbounded string.
pub const MAX_DBMD_FILE_BYTES: u64 = 64 * 1024 * 1024;

/// The two canonical layer folder names. A path is "content" / a wiki-link is
/// "full-path" only when it resolves under one of these.
const LAYER_DIRS: [&str; 2] = ["sources", "records"];

/// Errors produced while parsing a markdown file or the `DB.md` config.
#[derive(Debug, thiserror::Error)]
pub enum ParseError {
    /// The frontmatter block was not valid YAML. Maps to validate code
    /// `FM_MALFORMED_YAML`.
    #[error("malformed YAML frontmatter in {file}: {source}")]
    MalformedYaml {
        /// The file whose frontmatter failed to parse.
        file: PathBuf,
        /// The underlying YAML error.
        source: serde_norway::Error,
    },

    /// The file has no `---`-delimited frontmatter block at its very start.
    #[error("missing frontmatter block in {file}")]
    MissingFrontmatter {
        /// The offending file.
        file: PathBuf,
    },

    /// A required field was absent. Maps to validate code `FM_MISSING_TYPE`
    /// (for `type`) and the per-type required-field codes.
    #[error("missing required field '{key}' in {file}")]
    MissingField {
        /// The file missing the field.
        file: PathBuf,
        /// The required key.
        key: String,
    },

    /// A timestamp field was not ISO-8601 / RFC3339. Maps to `FM_BAD_TIMESTAMP`.
    #[error("bad timestamp in field '{key}' of {file}: {value}")]
    BadTimestamp {
        /// The file.
        file: PathBuf,
        /// The frontmatter key.
        key: String,
        /// The unparseable value.
        value: String,
    },

    /// An I/O error reading the file.
    #[error(transparent)]
    Io(#[from] std::io::Error),
}

/// The parsed YAML frontmatter of a db.md file.
///
/// The universal-contract fields are typed accessors; everything else lands in
/// [`extra`](Frontmatter::extra) as ambient context (unknown-field passthrough)
/// and is round-tripped verbatim. The atomic writer re-emits keys in canonical
/// order: `type`, `id`, `created`, `updated`, `summary` first, then
/// type-specific fields, then `status` / `tags`.
#[derive(Debug, Clone, Default, PartialEq)]
pub struct Frontmatter {
    /// `type` — required on content files; the primary query key.
    pub type_: Option<String>,
    /// `meta-type` — records-only; the epistemic class `fact`/`operational`/
    /// `conclusion`. Absent ⇒ `fact` (the effective default is applied by the
    /// index/query layer for record-layer files; sources carry none).
    pub meta_type: Option<String>,
    /// `id` — optional; derived from the file path when absent.
    pub id: Option<String>,
    /// `created` — RFC3339; required and auto-set on content-file create.
    pub created: Option<DateTime<FixedOffset>>,
    /// `updated` — RFC3339; required and auto-maintained on content files.
    pub updated: Option<DateTime<FixedOffset>>,
    /// `summary` — the one-line catalog line; required on every content file.
    pub summary: Option<String>,
    /// `status` — optional lifecycle state.
    pub status: Option<String>,
    /// `tags` — optional flat list of short scalar labels.
    pub tags: Vec<String>,
    /// All other frontmatter keys (type-specific + custom), preserved verbatim
    /// in insertion-stable sorted order. Wiki-link-valued fields keep their raw
    /// YAML form here; [`Frontmatter::link_fields`] surfaces them as
    /// [`WikiLink`]s.
    pub extra: BTreeMap<String, Value>,
}

/// Does `s` contain a run of at least `min` consecutive ASCII digits? A cheap
/// guard so [`quote_oversized_integers`] only does real work when an oversized
/// literal is even possible (`i64::MAX` is 19 digits, `u64::MAX` is 20).
fn has_long_digit_run(s: &str, min: usize) -> bool {
    let mut run = 0usize;
    for b in s.bytes() {
        if b.is_ascii_digit() {
            run += 1;
            if run >= min {
                return true;
            }
        } else {
            run = 0;
        }
    }
    false
}

/// True if `s` is a bare decimal integer literal whose magnitude exceeds the
/// `i64`/`u64` range `serde_norway` can represent losslessly — exactly the
/// literals it either rejects (`(u64::MAX, u128::MAX]`) or silently truncates to
/// `f64` (`> u128::MAX`). A canonical (no leading zero) decimal only, so an
/// octal/leading-zero/typed scalar is never reinterpreted.
fn is_oversized_int_literal(s: &str) -> bool {
    let t = s.trim();
    if t.is_empty() {
        return false;
    }
    let (neg, body) = match t.strip_prefix('-') {
        Some(b) => (true, b),
        None => (false, t.strip_prefix('+').unwrap_or(t)),
    };
    if body.is_empty() || !body.bytes().all(|b| b.is_ascii_digit() || b == b'_') {
        return false;
    }
    let digits: String = body
        .bytes()
        .filter(|b| *b != b'_')
        .map(|b| b as char)
        .collect();
    if digits.is_empty() {
        return false; // all underscores
    }
    // Leading-zero decimals (`007`) are version-ambiguous (octal vs int vs
    // string); never touch them.
    if digits.len() > 1 && digits.starts_with('0') {
        return false;
    }
    let canon = if neg { format!("-{digits}") } else { digits };
    // Fits i64 / u64 → handled losslessly; leave untouched.
    if canon.parse::<i64>().is_ok() || (!neg && canon.parse::<u64>().is_ok()) {
        return false;
    }
    true
}

/// Byte index where the scalar VALUE begins on a simple block line
/// (`key: <value>`, `- <value>`, or `- key: <value>`), or `None` when the line
/// bears no inline value (a bare `key:` / lone `-` / indent-only line).
fn scalar_value_start(content: &str) -> Option<usize> {
    let mut base = content.len() - content.trim_start().len();
    let mut rest = &content[base..];
    // Consume leading `- ` block-sequence markers (possibly nested: `- - x`).
    while let Some(after) = rest.strip_prefix("- ") {
        base += rest.len() - after.len();
        let trimmed = after.trim_start_matches(' ');
        base += after.len() - trimmed.len();
        rest = trimmed;
    }
    if rest.is_empty() || rest == "-" {
        return None;
    }
    // `key: value` — first `:` followed by a space/tab introduces the value.
    if let Some(colon) = rest.find(':') {
        let after = &rest[colon + 1..];
        if after.starts_with(' ') || after.starts_with('\t') {
            let val = after.trim_start_matches([' ', '\t']);
            return Some(base + colon + 1 + (after.len() - val.len()));
        }
        if after.is_empty() {
            return None; // `key:` with the value on following (block) lines
        }
    }
    // A bare sequence-item scalar: the value is the whole remainder.
    Some(base)
}

/// True if `content` introduces a YAML block scalar (`key: |`, `- >2`, …): the
/// value region begins with a `|` or `>` indicator. Its body must be skipped by
/// [`quote_oversized_integers`] so a digit line inside literal text is untouched.
fn introduces_block_scalar(content: &str) -> bool {
    match scalar_value_start(content) {
        Some(start) => {
            let v = content[start..].trim_start();
            v.starts_with('|') || v.starts_with('>')
        }
        None => false,
    }
}

/// Quote an oversized bare-integer value on a single block line, returning the
/// rewritten line, or `None` if the line carries no such value.
///
/// Handles two value shapes:
/// - a bare scalar value (`key: <int>`, `- <int>`, `- key: <int>`), and
/// - a single-line flow collection value (`key: [ … ]` / `key: { … }`) holding
///   one or more oversized integer literals (possibly mixed with in-range ints,
///   strings, and nested flow collections) — see [`quote_oversized_ints_in_flow`].
///
/// In both cases only the offending integer scalar(s) are single-quoted; every
/// other byte is preserved exactly.
fn quote_int_value_in_line(content: &str) -> Option<String> {
    let value_start = scalar_value_start(content)?;
    let region = &content[value_start..];
    let value = region.trim_end();

    // Single-line flow collection: scan inside it for oversized int literals.
    // (A bare scalar never starts with `[`/`{`, so these arms are disjoint.)
    if value.starts_with('[') || value.starts_with('{') {
        let trailing = &region[value.len()..];
        let rewritten = quote_oversized_ints_in_flow(value)?;
        return Some(format!(
            "{}{}{}",
            &content[..value_start],
            rewritten,
            trailing
        ));
    }

    if !is_oversized_int_literal(value) {
        return None;
    }
    // A pure digit literal contains no `'`, so single-quoting needs no escaping.
    let trailing = &region[value.len()..];
    Some(format!(
        "{}'{}'{}",
        &content[..value_start],
        value,
        trailing
    ))
}

/// Scan a single-line YAML flow collection (`[ … ]` / `{ … }`) and single-quote
/// each oversized bare-integer literal it contains, returning the rewritten flow
/// text, or `None` when it holds no such literal (so the caller can leave the
/// line untouched and `changed` stays false for an unaffected file).
///
/// The flow grammar is tokenized by its structural characters — `[ ] { } , :` —
/// at the top level: text between two structural characters (and outside any
/// single/double quoted scalar) is one plain scalar. A plain scalar whose trimmed
/// form is an oversized canonical decimal integer (per [`is_oversized_int_literal`])
/// is wrapped in single quotes; everything else — in-range ints, quoted strings,
/// floats, booleans, nested collections, the structural punctuation and all
/// surrounding whitespace — is emitted verbatim. Nested collections and multiple
/// literals on one line are handled by the same single left-to-right pass.
fn quote_oversized_ints_in_flow(flow: &str) -> Option<String> {
    let mut out = String::with_capacity(flow.len() + 2);
    let mut changed = false;
    // Byte offset where the current plain-scalar token began (None ⇒ not inside
    // a plain scalar, e.g. just after a structural char or inside a quote).
    let mut scalar_start: Option<usize> = None;
    let bytes = flow.as_bytes();
    let mut i = 0usize;

    // Flush the plain scalar spanning `[start, end)`: quote it iff it is an
    // oversized integer literal, otherwise copy it through verbatim.
    fn flush(out: &mut String, flow: &str, start: usize, end: usize, changed: &mut bool) {
        let raw = &flow[start..end];
        let trimmed = raw.trim();
        if !trimmed.is_empty() && is_oversized_int_literal(trimmed) {
            // Preserve the token's incidental leading/trailing whitespace; only
            // the literal itself is quoted. A pure-digit literal contains no `'`,
            // so single-quoting needs no escaping.
            let lead = &raw[..raw.len() - raw.trim_start().len()];
            let tail = &raw[raw.trim_end().len()..];
            out.push_str(lead);
            out.push('\'');
            out.push_str(trimmed);
            out.push('\'');
            out.push_str(tail);
            *changed = true;
        } else {
            out.push_str(raw);
        }
    }

    while i < bytes.len() {
        let b = bytes[i];
        match b {
            // Quoted scalars: copy through verbatim, skipping their contents so a
            // structural char or digit run inside a string is never reinterpreted.
            b'\'' | b'"' => {
                if let Some(start) = scalar_start.take() {
                    flush(&mut out, flow, start, i, &mut changed);
                }
                let quote = b;
                let str_start = i;
                i += 1;
                while i < bytes.len() {
                    if bytes[i] == quote {
                        // A doubled single-quote (`''`) is an escaped quote inside
                        // a single-quoted YAML scalar, not the closing delimiter.
                        if quote == b'\'' && i + 1 < bytes.len() && bytes[i + 1] == b'\'' {
                            i += 2;
                            continue;
                        }
                        // A backslash-escaped quote inside a double-quoted scalar
                        // does not close it.
                        if quote == b'"' && bytes[i - 1] == b'\\' {
                            i += 1;
                            continue;
                        }
                        i += 1;
                        break;
                    }
                    i += 1;
                }
                out.push_str(&flow[str_start..i]);
            }
            // Structural characters end the current plain scalar and are copied
            // through. `:` separates a flow-mapping key from its value; `,`
            // separates entries; brackets/braces open or close a (possibly
            // nested) collection.
            b'[' | b']' | b'{' | b'}' | b',' | b':' => {
                if let Some(start) = scalar_start.take() {
                    flush(&mut out, flow, start, i, &mut changed);
                }
                out.push(b as char);
                i += 1;
            }
            _ => {
                if scalar_start.is_none() {
                    scalar_start = Some(i);
                }
                i += 1;
            }
        }
    }
    if let Some(start) = scalar_start.take() {
        flush(&mut out, flow, start, bytes.len(), &mut changed);
    }

    if changed {
        Some(out)
    } else {
        None
    }
}

/// Pre-quote bare integer literals beyond the `i64`/`u64` range so they parse as
/// STRING scalars and round-trip verbatim.
///
/// `serde_norway` (no arbitrary-precision) cannot represent such an integer: it
/// rejects `(u64::MAX, u128::MAX]` as a hard parse error and silently truncates
/// `> u128::MAX` to `f64` (`999…9` → `1e39` on the next re-emit) — corrupting an
/// imported numeric ID and breaking the SPEC guarantee that unknown fields
/// round-trip byte-for-byte. Quoting them up front makes them string-valued (the
/// type narrows from number to string, but no data is destroyed).
///
/// Conservative: only a canonical decimal integer beyond `i64`/`u64` is quoted —
/// whether it appears as a bare value (`key: <int>` / `- <int>` / `- key: <int>`)
/// or as a scalar inside a single-line flow collection (`key: [ … ]` /
/// `key: { … }`, including nested collections and mixed/multiple literals); block
/// scalars are tracked and never touched; anything already in range, quoted, or
/// not a bare integer is left exactly as written.
fn quote_oversized_integers(yaml: &str) -> std::borrow::Cow<'_, str> {
    if !has_long_digit_run(yaml, 19) {
        return std::borrow::Cow::Borrowed(yaml);
    }
    let mut out = String::with_capacity(yaml.len());
    let mut changed = false;
    let mut block_indent: Option<usize> = None;
    for line in yaml.split_inclusive('\n') {
        let content = line.trim_end_matches(['\r', '\n']);
        let term = &line[content.len()..];
        let indent = content.len() - content.trim_start().len();

        // Inside a block scalar: emit verbatim until a non-blank line dedents to
        // at or before the introducer's key indent.
        if let Some(key_indent) = block_indent {
            if content.trim().is_empty() || indent > key_indent {
                out.push_str(line);
                continue;
            }
            block_indent = None; // block ended; process this line normally
        }
        if introduces_block_scalar(content) {
            block_indent = Some(indent);
            out.push_str(line);
            continue;
        }
        match quote_int_value_in_line(content) {
            Some(rewritten) => {
                out.push_str(&rewritten);
                out.push_str(term);
                changed = true;
            }
            None => out.push_str(line),
        }
    }
    if changed {
        std::borrow::Cow::Owned(out)
    } else {
        std::borrow::Cow::Borrowed(yaml)
    }
}

impl Frontmatter {
    /// Parse a YAML frontmatter block (the text between the opening and closing
    /// `---` fences, exclusive) into a [`Frontmatter`].
    ///
    /// Lenient on unknown keys (they go to [`extra`](Frontmatter::extra));
    /// returns [`ParseError::MalformedYaml`] only on YAML that doesn't parse.
    pub fn parse(yaml: &str, file: &Path) -> Result<Self, ParseError> {
        // An empty (or whitespace-only) frontmatter block is a valid, empty
        // mapping — not a YAML error.
        let value: Value = if yaml.trim().is_empty() {
            Value::Mapping(Mapping::new())
        } else {
            // Preserve integer literals beyond i64/u64 range: serde_norway would
            // otherwise reject `(u64,u128]` or silently truncate `>u128` to f64,
            // corrupting imported numeric IDs. Quoting them up front makes them
            // round-trip verbatim as strings.
            let prepared = quote_oversized_integers(yaml);
            serde_norway::from_str(&prepared).map_err(|source| ParseError::MalformedYaml {
                file: file.to_path_buf(),
                source,
            })?
        };

        // Top-level frontmatter must be a mapping. A scalar or sequence at the
        // top level is malformed for our purposes; surface it as such.
        let map = match value {
            Value::Mapping(m) => m,
            Value::Null => Mapping::new(),
            other => {
                // serde_norway::Error has no public constructor, so let the
                // deserializer decide: a value that coerces to a Mapping (e.g. a
                // YAML-tagged mapping `!tag\n k: v`, where the tag is ambient) is
                // accepted as that mapping; a genuine scalar or sequence top
                // level fails to coerce and IS the malformed case. (Using a
                // match here, not `expect_err`, avoids a panic on the
                // tagged-mapping case, which deserializes to a Mapping just
                // fine.)
                match serde_norway::from_value::<Mapping>(other) {
                    Ok(m) => m,
                    Err(source) => {
                        return Err(ParseError::MalformedYaml {
                            file: file.to_path_buf(),
                            source,
                        });
                    }
                }
            }
        };

        let mut fm = Frontmatter::default();
        for (k, v) in map {
            let key = match k.as_str() {
                Some(s) => s.to_string(),
                // Non-string keys (`2026:`, `true:`, `3.14:`) are unusual but
                // valid YAML; per SPEC § "Unknown fields pass through" they must
                // not be corrupted on re-emit. Stringify them through the YAML
                // scalar emitter — `2026`, `true`, `3.14` — NOT the Rust `Debug`
                // formatter (which produced `Number(2026)`, `Bool(true)`, …), so
                // the key text survives. `extra` is `String`-keyed, so on the
                // write side the key re-emits as a quoted-string key carrying that
                // text (e.g. `'2026':`) — the type narrows from number to string,
                // but the data is no longer destroyed and ordinary string keys are
                // wholly unaffected.
                None => yaml_scalar_key(&k),
            };
            match key.as_str() {
                // Coerce scalar values rather than `v.as_str()` (which is None
                // for Number/Bool/Null). A bare scalar that YAML reads as a
                // non-string — `summary: 2026`, `id: 100`, `status: 0` — would
                // otherwise be set to None AND dropped (it is a matched arm, so
                // the raw value never reaches `extra`), and `to_yaml` then omits
                // the None field, so `dbmd format` (read_file -> write_file)
                // silently deletes the line from disk. `scalar_string` mirrors
                // the coercion `validate`/`store` already apply to these fields,
                // so a numeric/bool-looking scalar is preserved as its string
                // form and round-trips instead of being destroyed.
                //
                // A sequence/mapping value on a universal key (`status: [a, b]`,
                // a nested-mapping `summary:`) is NOT a valid scalar; rather than
                // let the matched arm consume-and-drop it (silent data loss on
                // the next re-emit), `scalar_string` returns None and we fall
                // through to preserving the raw value in `extra` so `to_yaml`
                // re-emits it verbatim. The universal accessors stay None (the
                // value was never a valid scalar for that field), but the
                // operator's bytes are never destroyed.
                "type" => match scalar_string(&v) {
                    Some(s) => fm.type_ = Some(s),
                    None => {
                        fm.extra.insert(key, v);
                    }
                },
                "meta-type" => match scalar_string(&v) {
                    Some(s) => fm.meta_type = Some(s),
                    None => {
                        fm.extra.insert(key, v);
                    }
                },
                "id" => match scalar_string(&v) {
                    Some(s) => fm.id = Some(s),
                    None => {
                        fm.extra.insert(key, v);
                    }
                },
                // Same preserve-don't-destroy rule as the scalar keys above,
                // applied to the two typed ones. A value that is not RFC3339
                // has nowhere to live in the typed `Option<DateTime>`, so it
                // rides in `extra` verbatim and `to_yaml` re-emits it byte-for-
                // byte; the typed accessor stays None because there is no
                // timestamp to offer.
                //
                // The READ path is deliberately tolerant while `set` (the write
                // path) stays strict: a store is whatever it already is —
                // date-only stamps are the single most common legacy spelling
                // in migrated stores — and refusing to PARSE one made every
                // file-touching command (`format`, `fm`, `link`, `rename`)
                // unusable on exactly the imperfect stores that most need them.
                // Reporting the defect is `validate`'s job, and it still does:
                // `FM_BAD_TIMESTAMP` is raised from the raw YAML value, never
                // from this parse (validate.rs — `is_iso8601` over
                // `scalar_string`), so leniency here costs no enforcement.
                "created" => match parse_timestamp(&v, "created", file) {
                    Ok(ts) => fm.created = ts,
                    Err(_) => {
                        fm.extra.insert(key, v);
                    }
                },
                "updated" => match parse_timestamp(&v, "updated", file) {
                    Ok(ts) => fm.updated = ts,
                    Err(_) => {
                        fm.extra.insert(key, v);
                    }
                },
                "summary" => match scalar_string(&v) {
                    Some(s) => fm.summary = Some(s),
                    None => {
                        fm.extra.insert(key, v);
                    }
                },
                "status" => match scalar_string(&v) {
                    Some(s) => fm.status = Some(s),
                    None => {
                        fm.extra.insert(key, v);
                    }
                },
                "tags" => match parse_tags_preserving(&v) {
                    Ok(tags) => fm.tags = tags,
                    // A `tags` value with a non-scalar item (`tags: [[vip]]`,
                    // `tags: [a, [b]]`) is preserved verbatim in `extra` rather
                    // than silently filtered down / erased on re-emit. The typed
                    // `tags` vec stays empty (no valid scalar list was present),
                    // so `to_yaml` won't ALSO emit a `tags:` from the vec.
                    Err(raw) => {
                        fm.extra.insert(key, raw);
                    }
                },
                _ => {
                    fm.extra.insert(key, v);
                }
            }
        }

        // Disambiguate the one YAML shape `serde_norway` cannot tell apart on its
        // own: an *inline scalar* wiki-link `field: [[x]]` and a *genuine 2D
        // array* `field:`\n`- - x` BOTH parse to the identical
        // `Seq[ Seq[String("x")] ]`. The parsed `Value` has lost which one the
        // source wrote, but the source text has not — so we resolve it here, the
        // only place the original spelling is still visible. For every `extra`
        // key the source wrote in the inline `[[…]]` form, store the canonical
        // quoted scalar `String("[[x]]")` instead of the ambiguous nested
        // sequence. `to_yaml`/`canonicalize_extra_value` then emit it inline and
        // round-trip it (SPEC § Linking, `company: [[…]]`), while a real nested
        // array — which never appears in inline-link source form — stays a
        // sequence and is preserved verbatim rather than silently retyped.
        for key in inline_scalar_link_keys(yaml) {
            if let Some(value) = fm.extra.get_mut(&key) {
                // The parsed value of an inline `key: [[x]]` is the one-element
                // outer `Seq[ Seq[String(x)] ]`; `unquoted_inline_link` reads the
                // inner `Seq[String(x)]`, so unwrap the lone outer item first.
                if let Value::Sequence(items) = value {
                    if items.len() == 1 {
                        if let Some(link) = unquoted_inline_link(&items[0]) {
                            *value = Value::String(wiki_link_literal(&link));
                        }
                    }
                }
            }
        }

        Ok(fm)
    }

    /// Serialize the frontmatter back to a YAML block (no `---` fences) in
    /// canonical key order. Round-trips [`extra`](Frontmatter::extra) verbatim.
    pub fn to_yaml(&self) -> String {
        // Build an order-preserving mapping in canonical key order:
        //   type, meta-type, id, created, updated, summary  (universal head)
        //   <type-specific extra, BTreeMap-sorted>
        //   status, tags                          (universal tail)
        // serde_norway::Mapping preserves insertion order, so one serialize call
        // emits the block in exactly this order with correct YAML quoting.
        let mut map = Mapping::new();

        if let Some(t) = &self.type_ {
            map.insert(Value::String("type".into()), Value::String(t.clone()));
        }
        if let Some(mt) = &self.meta_type {
            map.insert(Value::String("meta-type".into()), Value::String(mt.clone()));
        }
        if let Some(id) = &self.id {
            map.insert(Value::String("id".into()), Value::String(id.clone()));
        }
        if let Some(created) = &self.created {
            map.insert(
                Value::String("created".into()),
                Value::String(created.to_rfc3339()),
            );
        }
        if let Some(updated) = &self.updated {
            map.insert(
                Value::String("updated".into()),
                Value::String(updated.to_rfc3339()),
            );
        }
        if let Some(summary) = &self.summary {
            map.insert(
                Value::String("summary".into()),
                Value::String(summary.clone()),
            );
        }

        // Type-specific + custom fields, in BTreeMap (sorted) order. Each value
        // is canonicalized so a wiki-link round-trips to the form the writer and
        // `dbmd validate` agree on — critically, the SPEC-canonical *unquoted*
        // scalar `field: [[x]]` (which YAML parses to a nested `Seq[Seq[String]]`)
        // is re-emitted as a quoted scalar `'[[x]]'` instead of the bracket-less
        // block sequence `- - x` that a verbatim re-emit would produce and that
        // destroys the link. See [`canonicalize_extra_value`].
        for (k, v) in &self.extra {
            map.insert(Value::String(k.clone()), canonicalize_extra_value(v));
        }

        if let Some(status) = &self.status {
            map.insert(
                Value::String("status".into()),
                Value::String(status.clone()),
            );
        }
        if !self.tags.is_empty() {
            map.insert(
                Value::String("tags".into()),
                Value::Sequence(self.tags.iter().cloned().map(Value::String).collect()),
            );
        }

        if map.is_empty() {
            return String::new();
        }
        serde_norway::to_string(&Value::Mapping(map)).unwrap_or_default()
    }

    /// True if the file is content (under `sources/` or `records/`)
    /// and not an `index.md`. Used by validate to decide which files require a
    /// `summary`. Meta files (`DB.md`, `index.md`, `log.md`) return false.
    pub fn is_content_file(path: &Path) -> bool {
        // index.md is a meta file at every level, never content.
        if path.file_name().and_then(|n| n.to_str()) == Some("index.md") {
            return false;
        }
        // Content iff some path component is one of the two layer dirs. This
        // works for both store-relative (`sources/emails/x.md`) and absolute
        // (`/home/db/sources/emails/x.md`) paths. DB.md / log.md sit at the
        // root, under no layer, so they fall through to false.
        path.components().any(|c| {
            c.as_os_str()
                .to_str()
                .is_some_and(|s| LAYER_DIRS.contains(&s))
        })
    }

    /// Resolve the file's effective `id`: the explicit `id` field if present,
    /// otherwise derived from the store-relative path (filename without `.md`).
    pub fn effective_id(&self, store_relative_path: &Path) -> String {
        if let Some(id) = &self.id {
            if !id.is_empty() {
                return id.clone();
            }
        }
        // Derived id = filename without the `.md` extension.
        store_relative_path
            .file_stem()
            .and_then(|s| s.to_str())
            .unwrap_or_default()
            .to_string()
    }

    /// The effective `meta-type` for a record: the declared value, or `fact`
    /// when absent. Records only — sources carry no meta-type; callers apply
    /// this only to record-layer files.
    pub fn effective_meta_type(&self) -> &str {
        self.meta_type.as_deref().unwrap_or("fact")
    }

    /// Read a single frontmatter key as a raw YAML [`Value`], looking in the
    /// typed fields first and then [`extra`](Frontmatter::extra).
    pub fn get(&self, key: &str) -> Option<Value> {
        match key {
            "type" => self.type_.clone().map(Value::String),
            "meta-type" => self.meta_type.clone().map(Value::String),
            "id" => self.id.clone().map(Value::String),
            "created" => self.created.map(|d| Value::String(d.to_rfc3339())),
            "updated" => self.updated.map(|d| Value::String(d.to_rfc3339())),
            "summary" => self.summary.clone().map(Value::String),
            "status" => self.status.clone().map(Value::String),
            "tags" => {
                if self.tags.is_empty() {
                    None
                } else {
                    Some(Value::Sequence(
                        self.tags.iter().cloned().map(Value::String).collect(),
                    ))
                }
            }
            _ => self.extra.get(key).cloned(),
        }
    }

    /// Set a single frontmatter key from a string value, routing universal-
    /// contract keys to their typed fields and everything else to
    /// [`extra`](Frontmatter::extra). Used by `dbmd fm set`.
    pub fn set(&mut self, key: &str, value: &str) -> Result<(), ParseError> {
        match key {
            "type" => self.type_ = Some(value.to_string()),
            "meta-type" => self.meta_type = Some(value.to_string()),
            "id" => self.id = Some(value.to_string()),
            "created" => {
                self.created = Some(parse_rfc3339(value, "created", Path::new("<fm set>"))?)
            }
            "updated" => {
                self.updated = Some(parse_rfc3339(value, "updated", Path::new("<fm set>"))?)
            }
            "summary" => self.summary = Some(value.to_string()),
            "status" => self.status = Some(value.to_string()),
            "tags" => {
                // Accept either a YAML flow list (`[a, b]`) or a single scalar
                // tag. Anything that parses to a sequence becomes the tag list;
                // otherwise the whole string is one tag.
                self.tags = match serde_norway::from_str::<Value>(value) {
                    Ok(Value::Sequence(seq)) => parse_tags(&Value::Sequence(seq)),
                    _ => vec![value.to_string()],
                };
            }
            _ => {
                // A custom / type-specific field. The value is a scalar string by
                // default, but the spec's list-valued link fields (e.g.
                // `meeting.attendees`, SPEC § Linking) must serialize as a YAML
                // block sequence of quoted wiki-links — never the flow-form string
                // `"[[[a]], [[b]]]"`, which `dbmd validate` rejects as
                // `WIKI_LINK_FLOW_FORM_LIST`. When the value parses as a YAML
                // sequence whose every item is a clean single wiki-link, store the
                // canonical sequence so `to_yaml` emits block form. Everything else
                // — plain text, and a single inline `[[x]]` (which YAML reads as a
                // nested `Seq[Seq[String]]`, not a list of link strings) — stays a
                // verbatim scalar string, preserving the prior behavior.
                let stored = parse_link_list_value(value)
                    .unwrap_or_else(|| Value::String(value.to_string()));
                self.extra.insert(key.to_string(), stored);
            }
        }
        Ok(())
    }

    /// Extract every frontmatter field whose value is a wiki-link (scalar
    /// inline form or a block-sequence list), pairing each with its key. The
    /// validate engine checks these against `(link)` schema annotations.
    pub fn link_fields(&self) -> Vec<(String, WikiLink)> {
        let mut out = Vec::new();
        // `summary` may carry navigational wiki-links (spec encourages it).
        if let Some(summary) = &self.summary {
            for link in extract_wiki_links(summary, Path::new("")) {
                out.push(("summary".to_string(), link));
            }
        }
        // Every type-specific / custom field: a scalar wiki-link or a list of
        // wiki-links, in either the quoted (`"[[x]]"`) or the canonical unquoted
        // (`[[x]]`) form. See [`links_in_field_value`] for the YAML shapes.
        for (key, value) in &self.extra {
            for link in links_in_field_value(value) {
                out.push((key.clone(), link));
            }
        }
        out
    }
}

/// A wiki-link reference inside the store: `[[target]]` or `[[target|display]]`.
///
/// `target` is always recorded as written; [`is_full_path`](WikiLink::is_full_path)
/// flags whether it's a full store-relative path (the doctrine) versus a
/// short-form (a validation error).
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct WikiLink {
    /// The link target as written, without the `[[ ]]` and without `|display`.
    pub target: String,
    /// The optional `|display` text override.
    pub display: Option<String>,
    /// True when `target` is a full store-relative path (contains a `/` and
    /// resolves under a known layer); false for short-form targets like
    /// `sarah-chen` — which validate reports as `WIKI_LINK_SHORT_FORM`.
    pub is_full_path: bool,
    /// True when `target` carries a trailing `.md` extension — validate warns
    /// `WIKI_LINK_HAS_EXTENSION`; the canonical writers emit the bare form.
    pub has_md_extension: bool,
    /// Where the link appears: `(file, line, col)`, 1-based line and column.
    pub location: (PathBuf, u32, u32),
}

/// A standard markdown link `[text](url)` — an external reference, kept in a
/// stream separate from [`WikiLink`] so external targets are visible to the
/// toolkit without being conflated with in-store edges. Not graph-validated.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct MarkdownLink {
    /// The link text inside `[ ]`.
    pub text: String,
    /// The URL or path inside `( )`.
    pub url: String,
    /// Where the link appears: `(file, line, col)`, 1-based.
    pub location: (PathBuf, u32, u32),
}

/// A `##`/`###` section of a markdown body: the heading text plus the byte
/// slice of the body it spans (heading line through the line before the next
/// heading of equal-or-shallower depth).
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct Section {
    /// The heading text (without the leading `#`s).
    pub heading: String,
    /// Heading depth (number of leading `#`s).
    pub level: u8,
    /// The 1-based line where the heading appears.
    pub line: u32,
    /// The section body, from the heading line to the next sibling-or-shallower
    /// heading (exclusive), as a slice of the original body.
    pub body: String,
}

/// The parsed structured content of a store's `DB.md` config file.
///
/// All four parts are optional in the source; absent parts fall back to spec
/// defaults. Produced by [`parse_db_md`].
#[derive(Debug, Clone, Default, PartialEq)]
pub struct Config {
    /// Body of the `## Agent instructions` section — free-form prose passed to
    /// the agent's system prompt.
    pub agent_instructions: Option<String>,
    /// `## Policies` → `### Frozen pages`: store-relative paths the toolkit
    /// refuses to write (`POLICY_FROZEN_PAGE`).
    pub frozen_pages: Vec<PathBuf>,
    /// `## Policies` → `### Ignored types`: type names the curator never
    /// synthesizes (still readable as ambient context).
    pub ignored_types: Vec<String>,
    /// `## Schemas` → one entry per `### <type>` sub-section.
    pub schemas: BTreeMap<String, Schema>,
    /// `## Folders` → optional per-folder display + description, surfaced in the
    /// root + layer `index.md` rollups. Agent-authored; the tool never invents a
    /// folder's description (absent ⇒ the rollup shows counts only). Keyed by the
    /// store-relative, unix-slash folder path (e.g. `records/contacts`).
    pub folders: BTreeMap<String, FolderMeta>,
}

/// Agent-authored display + description for one type-folder, declared in
/// `DB.md ## Folders` and surfaced in the root/layer `index.md` rollups. Both
/// fields are optional: `display` overrides the rollup's derived folder name
/// (for casing the tool can't guess, e.g. acronyms like HubSpot); `description`
/// is the one-line "what's in here" the rollup shows. The tool only ever
/// *surfaces* these — it never composes a folder description from the folder's
/// contents (that would be the tool inventing the curator's judgment).
#[derive(Debug, Clone, Default, PartialEq, Eq)]
pub struct FolderMeta {
    /// Display-name override (absent ⇒ derived from the folder basename).
    pub display: Option<String>,
    /// One-line folder description shown in the rollup (absent ⇒ counts only).
    pub description: Option<String>,
}

impl Config {
    /// The `### Frozen pages` entry that matches a store-relative `target`, if
    /// any. The **single** frozen-page matcher every write surface must funnel
    /// through so the policy is enforced identically on `write` / `fm set` /
    /// `fm init` / `link` / `rename` / `format`.
    ///
    /// Comparison is normalized so a policy line and a write target match
    /// regardless of incidental spelling differences:
    /// - `/` path separators on every OS,
    /// - a single leading `./` dropped,
    /// - a trailing `.md` dropped on **both** sides — `parse_db_md` stores
    ///   frozen entries verbatim, so an operator who writes the natural
    ///   extensionless spelling (`records/decisions/q1`) must protect the file
    ///   (`records/decisions/q1.md`) exactly as the `.md` spelling does.
    ///
    /// Returns the matched config entry verbatim (its original spelling) so the
    /// caller can name it in the `POLICY_FROZEN_PAGE` refusal.
    pub fn frozen_match(&self, target: &Path) -> Option<PathBuf> {
        let want = normalize_frozen_path(target);
        self.frozen_pages
            .iter()
            .find(|frozen| {
                let pat = normalize_frozen_path(frozen);
                // A literal entry matches by exact normalized equality; an entry
                // carrying a `*`/`**` glob matches by segment-wise glob so a
                // pattern like `records/decisions/*` actually protects the
                // concrete files under it instead of silently failing open.
                if pat.contains('*') {
                    frozen_glob_matches(&pat, &want)
                } else {
                    pat == want
                }
            })
            .cloned()
    }

    /// True if `target` (store-relative) is a frozen page. Convenience wrapper
    /// over [`Config::frozen_match`] for callers that only need presence.
    pub fn is_frozen(&self, target: &Path) -> bool {
        self.frozen_match(target).is_some()
    }
}

/// Normalize a path for frozen-page comparison: `/` separators, a leading `./`
/// or `/` dropped, and a trailing `.md` dropped. Both the policy entry and the
/// write target pass through this before equality/glob, so the match is
/// separator-, `./`-, leading-`/`-, and `.md`-insensitive. Without the leading
/// `/` drop, an operator who wrote `/records/decisions/q1.md` normalized to a
/// path that never equals the target's `records/decisions/q1`, silently failing
/// the freeze OPEN.
fn normalize_frozen_path(p: &Path) -> String {
    use std::path::Component;
    // Keep only the `Normal` path segments, dropping `RootDir`/`Prefix` (a
    // leading `/` or drive prefix) and `CurDir` (`.`). This is what makes a
    // leading-slash entry (`/records/decisions/q1.md`) normalize to the same
    // `records/decisions/q1` as the store-relative target, instead of the
    // doubled-`//` prefix `Path::components` + naive join produced — which never
    // equalled the target and silently failed the freeze OPEN.
    let unix: String = p
        .components()
        .filter_map(|c| match c {
            Component::Normal(s) => s.to_str(),
            _ => None,
        })
        .collect::<Vec<_>>()
        .join("/");
    unix.strip_suffix(".md").unwrap_or(&unix).to_string()
}

/// Match a normalized frozen-page glob `pat` against a normalized target `path`,
/// segment by segment. `*` matches any run of characters *within a single path
/// segment* (never crossing `/`); `**` as a whole segment matches zero or more
/// whole segments. Both sides are already `normalize_frozen_path`-normalized, so
/// this only deals with `/`-joined segment text. Keeps the substrate dependency-
/// free (no glob crate) while making `records/decisions/*` actually freeze the
/// files beneath it instead of failing open.
fn frozen_glob_matches(pat: &str, path: &str) -> bool {
    // Collapse runs of consecutive `**` segments into a single `**` before
    // matching: `**/**` matches exactly the same set of paths as `**`, so the
    // duplicates carry no semantics — they only multiply the number of
    // (star-index, path-index) splits the matcher must consider. Dropping them
    // up front is the first half of keeping the match polynomial (the
    // two-pointer matcher below is the second); without it, a DB.md bullet like
    // `**/**/…/zzz` against a deep non-matching target made the old recursive
    // matcher explore exponentially many splits and hang the entire write path.
    let pat_segs: Vec<&str> = collapse_double_stars(pat.split('/'));
    let path_segs: Vec<&str> = path.split('/').collect();
    glob_segments(&pat_segs, &path_segs)
}

/// Drop every `**` segment that immediately follows another `**`, leaving at most
/// one `**` per run. Consecutive `**` are semantically identical to a single `**`
/// (each matches "zero or more whole segments"), so this never changes the set of
/// matched paths — it only removes the redundant pattern positions that otherwise
/// fuel catastrophic backtracking.
fn collapse_double_stars<'a>(segs: impl Iterator<Item = &'a str>) -> Vec<&'a str> {
    let mut out: Vec<&str> = Vec::new();
    for seg in segs {
        if seg == "**" && out.last() == Some(&"**") {
            continue;
        }
        out.push(seg);
    }
    out
}

/// Segment matcher for [`frozen_glob_matches`]. `**` consumes any number of path
/// segments; every other pattern segment must match exactly one path segment
/// (with `*` wildcards inside it).
///
/// Implemented as the classic linear wildcard match: a single forward scan with a
/// remembered "last `**`" backtrack point, never the two-way recursion the old
/// version used. The old `glob_segments(rest, path)` OR `glob_segments(pat,
/// &path[1..])` recursion had no memoization, so N consecutive `**` against a
/// deep target that ultimately fails to match explored an exponential number of
/// (star-index, path-index) splits — one DB.md frozen-page bullet could hang the
/// store's whole write path. This greedy scan with backtrack is O(pat × path) in
/// the worst case while matching exactly the same set of paths.
fn glob_segments(pat: &[&str], path: &[&str]) -> bool {
    let mut pi = 0usize; // cursor into pattern segments
    let mut si = 0usize; // cursor into path segments
                         // Backtrack point: where in the pattern the last `**` sat, and the path
                         // position to resume from if a later literal mismatch forces the `**` to
                         // swallow one more segment. `None` until we have seen a `**`.
    let mut star_pi: Option<usize> = None;
    let mut star_si = 0usize;

    while si < path.len() {
        if pi < pat.len() && pat[pi] == "**" {
            // Record this `**` as the resume point and tentatively let it match
            // zero segments (advance past it). If a later segment fails, we come
            // back here and let the `**` swallow one more path segment.
            star_pi = Some(pi);
            star_si = si;
            pi += 1;
        } else if pi < pat.len() && glob_segment_text(pat[pi], path[si]) {
            // Ordinary segment match: advance both cursors.
            pi += 1;
            si += 1;
        } else if let Some(sp) = star_pi {
            // Mismatch (or pattern exhausted) but an earlier `**` can absorb more:
            // resume just after that `**`, having it consume one extra segment.
            pi = sp + 1;
            star_si += 1;
            si = star_si;
        } else {
            // Mismatch with no `**` to fall back on.
            return false;
        }
    }

    // Path consumed; any trailing pattern must be all `**` (each matching zero
    // segments) for a full match.
    while pi < pat.len() && pat[pi] == "**" {
        pi += 1;
    }
    pi == pat.len()
}

/// Match a single glob segment against a single path segment. `*` matches any
/// run of characters within the segment; all other characters are literal.
fn glob_segment_text(pat: &str, seg: &str) -> bool {
    if !pat.contains('*') {
        return pat == seg;
    }
    // Split on `*` into literal fragments. The first fragment must be a prefix,
    // the last a suffix, and the middle fragments must appear in order.
    let parts: Vec<&str> = pat.split('*').collect();
    let mut pos = 0usize;
    for (idx, part) in parts.iter().enumerate() {
        if part.is_empty() {
            continue;
        }
        if idx == 0 {
            // Leading literal must be a prefix.
            if !seg[pos..].starts_with(part) {
                return false;
            }
            pos += part.len();
        } else if idx == parts.len() - 1 {
            // Trailing literal must be a suffix at or after the current cursor.
            return seg[pos..].ends_with(part);
        } else {
            // Interior literal: find it at or after the cursor.
            match seg[pos..].find(part) {
                Some(off) => pos += off + part.len(),
                None => return false,
            }
        }
    }
    true
}

/// A user-declared type schema parsed from a `DB.md` `### <type>` sub-section.
/// The store's `## Schemas` is the **only** source of schema enforcement — the
/// toolkit ships no built-in or implicit per-type schema (see SPEC § Schemas).
#[derive(Debug, Clone, Default, PartialEq)]
pub struct Schema {
    /// One [`FieldSpec`] per bulleted field line, in source order.
    pub fields: Vec<FieldSpec>,
    /// `- unique: <field>[, <field> …]` directives — each inner vec is one
    /// uniqueness constraint over the listed field(s) (compound when >1). Two
    /// records of this type whose listed values collide warn as
    /// `DUP_UNIQUE_KEY`.
    pub unique_keys: Vec<Vec<String>>,
    /// `- summary_template: <template>` directive — the `{field}` interpolation
    /// pattern `dbmd fm init` / `dbmd write` use to compose a default `summary`
    /// for this type. `None` falls back to the body's first paragraph.
    pub summary_template: Option<String>,
    /// `- shard: by-date | flat` directive — whether records of this type are
    /// date-sharded on disk (`records/<type>/<YYYY>/<MM>/…`) or kept flat.
    /// `None` = no directive declared, so the store's built-in default for the
    /// type applies ([`crate::store::Store::type_shards`]); `Some(true)` forces
    /// date-sharding (e.g. a custom event type the toolkit has no built-in for);
    /// `Some(false)` forces flat. This is the v0.2 generic-model way to declare
    /// sharding — the toolkit ships no implicit per-type behavior beyond the
    /// example-type defaults.
    pub shard: Option<bool>,
}

/// One field declaration inside a [`Schema`]: `- <name> (<modifiers>)`.
///
/// Modifiers are comma-separated inside the parens; this captures the
/// recognized ones as typed fields and stashes anything unrecognized in
/// [`unknown_modifiers`](FieldSpec::unknown_modifiers) (surfaced as `Info`).
#[derive(Debug, Clone, Default, PartialEq)]
pub struct FieldSpec {
    /// The field name.
    pub name: String,
    /// `required` modifier present.
    pub required: bool,
    /// The shape modifier (`string`/`int`/`bool`/`date`/`email`/`currency`/
    /// `url`), if any.
    pub shape: Option<Shape>,
    /// `link to <prefix>/` — the store-relative prefix a wiki-link target must
    /// start with. The trailing slash is required in the source syntax.
    pub link_prefix: Option<PathBuf>,
    /// `default <value>` — the value written when the field is absent.
    pub default: Option<Value>,
    /// `enum: <v1>, <v2>, ...` — the allowed values (must be the last modifier
    /// on the line because of its own commas).
    pub enum_values: Option<Vec<String>>,
    /// Any modifiers not in the recognized vocabulary, preserved verbatim;
    /// validate surfaces these as `Info`, never errors.
    pub unknown_modifiers: Vec<String>,
}

/// A recognized shape modifier for a schema field. Validate enforces the
/// corresponding value shape (`SCHEMA_SHAPE_MISMATCH` on violation).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Shape {
    /// Any scalar string.
    String,
    /// Integer.
    Int,
    /// Boolean.
    Bool,
    /// RFC3339 / ISO-8601 date.
    Date,
    /// `<local>@<domain>` email address.
    Email,
    /// A currency amount.
    Currency,
    /// A URL.
    Url,
}

/// The result of splitting a raw file into its frontmatter block and body.
///
/// `body` is the verbatim remainder after the closing `---` fence — the writer
/// preserves it byte-for-byte so operator edits are never reflowed.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ParsedFile {
    /// The raw frontmatter YAML (between the fences, exclusive of them).
    pub frontmatter_yaml: String,
    /// The verbatim body (everything after the closing `---`).
    pub body: String,
}

/// Split a file's full text into its frontmatter block and body. The
/// frontmatter block must be the very first thing in the file, delimited by
/// `---` on its own line at start and end. Returns
/// [`ParseError::MissingFrontmatter`] if absent.
pub fn split_frontmatter(text: &str, file: &Path) -> Result<ParsedFile, ParseError> {
    // Tolerate a single leading UTF-8 BOM (U+FEFF) before the opening fence,
    // matching `store::frontmatter_block` and `index::extract_frontmatter_block`
    // which already strip it. Without this, a BOM-prefixed file (common from
    // Windows / exported markdown dropped into `sources/`) gets walked and
    // indexed by `dbmd index` yet hard-fails every write/edit surface that
    // routes through `read_file` (`fm get/set`, `format`, `link`, `write`). The
    // BOM is dropped from the emitted body so the canonical writer never carries
    // it forward.
    let text = text.strip_prefix('\u{feff}').unwrap_or(text);

    // The opening fence must be the very first line: `---`, no leading
    // whitespace, nothing before it. Trailing whitespace on the fence line is
    // tolerated via `trim_end()` (which strips spaces/tabs as well as CR/LF) so
    // this matches `index::extract_frontmatter_block` and
    // `validate::split_frontmatter`, both of which use `trim_end()`. Without this
    // agreement a fence written `--- ` (a single trailing space — invisible in an
    // editor, easily produced by hand edits or exporters) was indexed and
    // validated clean by those scanners yet hard-failed every write/edit surface
    // routed through `read_file` (`fm get/set`, `format`, `link`, `write`) — the
    // same cross-scanner drift class already fixed for the UTF-8 BOM above.
    let mut lines = text.split_inclusive('\n');
    let first = lines.next().unwrap_or("");
    if first.trim_end() != "---" {
        return Err(ParseError::MissingFrontmatter {
            file: file.to_path_buf(),
        });
    }

    // Scan for the closing fence line. Track byte offsets so we can slice the
    // YAML (between fences, exclusive) and the body (verbatim, after the
    // closing fence's line terminator).
    let opening_len = first.len();
    let mut offset = opening_len;
    for line in lines {
        if line.trim_end() == "---" {
            let yaml = &text[opening_len..offset];
            let body_start = offset + line.len();
            let body = &text[body_start..];
            return Ok(ParsedFile {
                frontmatter_yaml: yaml.to_string(),
                body: body.to_string(),
            });
        }
        offset += line.len();
    }

    // Opening fence present but no closing fence: malformed frontmatter block.
    Err(ParseError::MissingFrontmatter {
        file: file.to_path_buf(),
    })
}

/// Read a file from disk and parse it into typed [`Frontmatter`] plus the
/// verbatim body string.
pub fn read_file(path: &Path) -> Result<(Frontmatter, String), ParseError> {
    let bytes = crate::fsx::read_bounded_nofollow(path, MAX_DBMD_FILE_BYTES).map_err(|error| {
        ParseError::Io(std::io::Error::new(
            error.kind(),
            format!(
                "could not read a bounded regular db.md file ({} MiB cap): {error}",
                MAX_DBMD_FILE_BYTES / (1024 * 1024)
            ),
        ))
    })?;
    let text = String::from_utf8(bytes).map_err(|error| {
        ParseError::Io(std::io::Error::new(std::io::ErrorKind::InvalidData, error))
    })?;
    let parsed = split_frontmatter(&text, path)?;
    let fm = Frontmatter::parse(&parsed.frontmatter_yaml, path)?;
    Ok((fm, parsed.body))
}

/// Atomically write a markdown file from frontmatter + body: emit the
/// frontmatter in canonical key order, then the body verbatim, via a
/// temp-file-rename so a reader never sees a half-written file. Preserves the
/// operator-edited body exactly as given.
pub fn write_file(path: &Path, frontmatter: &Frontmatter, body: &str) -> Result<(), ParseError> {
    let contents = render_file(frontmatter, body);

    // One durable, atomic write for all primary data (see `crate::fsx`):
    // temp-file + fsync + rename + parent-fsync. Content records are primary
    // data, so they get the durable path (unlike the rebuildable index).
    crate::fsx::write_atomic(path, contents.as_bytes())?;
    Ok(())
}

/// Atomically create a markdown file from frontmatter + body, refusing with
/// [`std::io::ErrorKind::AlreadyExists`] if the destination already exists.
///
/// This is the create-new sibling of [`write_file`]: same canonical rendering
/// and durable temp-file path, but backed by [`crate::fsx::write_atomic_new`] so
/// two concurrent creators for the same path cannot both succeed.
pub fn write_file_new(
    path: &Path,
    frontmatter: &Frontmatter,
    body: &str,
) -> Result<(), ParseError> {
    let contents = render_file(frontmatter, body);
    crate::fsx::write_atomic_new(path, contents.as_bytes())?;
    Ok(())
}

/// Render canonical file bytes without performing I/O. Transactional callers
/// use this to stage the exact bytes that [`write_file`] would commit.
pub fn render_file(frontmatter: &Frontmatter, body: &str) -> String {
    let yaml = frontmatter.to_yaml();
    // `to_yaml` already terminates each block with a newline. Compose the file
    // as: opening fence, frontmatter YAML, closing fence, then body verbatim.
    let mut contents = String::with_capacity(yaml.len() + body.len() + 8);
    contents.push_str("---\n");
    contents.push_str(&yaml);
    contents.push_str("---\n");
    contents.push_str(body);
    contents
}

/// Extract every wiki-link from a body (and inline frontmatter), returning the
/// structured [`WikiLink`] stream with short-form / `.md`-extension flags and
/// `(file, line, col)` locations set.
pub fn extract_wiki_links(body: &str, file: &Path) -> Vec<WikiLink> {
    static RE: std::sync::OnceLock<regex::Regex> = std::sync::OnceLock::new();
    let re = RE.get_or_init(|| {
        // [[target]] or [[target|display]]; target/display exclude brackets and
        // (for target) the `|` separator so nested forms don't over-match.
        regex::Regex::new(r"\[\[([^\[\]|]+?)(?:\|([^\[\]]*))?\]\]").expect("valid wiki-link regex")
    });

    let mut out = Vec::new();
    for (line_idx, line) in body.lines().enumerate() {
        // Running (byte, char) cursor: derive each match's column in ONE linear
        // pass over the line instead of recomputing it from the line start per
        // match. `captures_iter` yields non-overlapping matches in increasing
        // byte order, so advancing the char count by the gap since the previous
        // match keeps the whole line O(line_len) rather than O(matches × len).
        let mut cursor = ColCursor::new();
        for caps in re.captures_iter(line) {
            let whole = caps.get(0).expect("group 0 always present");
            let col = cursor.column_at(line, whole.start());
            let target = caps.get(1).map(|m| m.as_str()).unwrap_or("").to_string();
            let display = caps.get(2).map(|m| m.as_str().to_string());
            out.push(WikiLink {
                is_full_path: target_is_full_path(&target),
                has_md_extension: target_has_md_extension(&target),
                target,
                display,
                location: (file.to_path_buf(), (line_idx as u32) + 1, col),
            });
        }
    }
    out
}

/// Extract every standard markdown link `[text](url)` from a body into a
/// separate stream, kept distinct from wiki-links.
pub fn extract_markdown_links(body: &str, file: &Path) -> Vec<MarkdownLink> {
    static RE: std::sync::OnceLock<regex::Regex> = std::sync::OnceLock::new();
    let re = RE.get_or_init(|| {
        // [text](url). `text` excludes brackets so a wiki-link `[[x]]` (which
        // has `]]`, not `](`) never matches; `url` excludes `)` and whitespace.
        regex::Regex::new(r"\[([^\[\]]*)\]\(([^)\s]*)\)").expect("valid markdown-link regex")
    });

    let mut out = Vec::new();
    for (line_idx, line) in body.lines().enumerate() {
        // One linear column cursor per line (see `extract_wiki_links`): avoids the
        // O(matches × line_len) recompute on a link-dense line.
        let mut cursor = ColCursor::new();
        for caps in re.captures_iter(line) {
            let whole = caps.get(0).expect("group 0 always present");
            let col = cursor.column_at(line, whole.start());
            out.push(MarkdownLink {
                text: caps.get(1).map(|m| m.as_str()).unwrap_or("").to_string(),
                url: caps.get(2).map(|m| m.as_str()).unwrap_or("").to_string(),
                location: (file.to_path_buf(), (line_idx as u32) + 1, col),
            });
        }
    }
    out
}

/// Detect the frontmatter wiki-link-list mis-encoding: a wiki-link *list*
/// written so YAML parses it as nested sequences instead of a clean list of
/// strings. Returns the offending keys so validate can emit
/// `WIKI_LINK_FLOW_FORM_LIST`.
///
/// The subtlety is that `[[x]]` is YAML for "a list containing `[x]`", so the
/// shapes nest:
///
/// - **Scalar inline** `company: [[records/x]]` → `Seq[ Seq[String] ]`
///   (double-nested). This is the spec's scalar wiki-link form — NOT flagged.
/// - **Flow list** `attendees: [[[a]], [[b]]]` → `Seq[ Seq[Seq[String]], … ]`
///   (triple-nested). The list mis-encoding — flagged.
/// - **Unquoted block list** (`- [[a]]` per line) → also triple-nested, so it
///   is flagged too; the canonical list form must quote each item
///   (`- "[[a]]"`), which parses to a clean `Seq[String, …]` and is NOT flagged.
///
/// So the discriminator is nesting depth: a *list* mis-encoding has at least one
/// item that is itself a sequence-of-sequences, whereas a scalar inline link's
/// single item is a sequence-of-scalars.
pub fn detect_flow_form_link_lists(frontmatter_yaml: &str) -> Vec<String> {
    let value: Value = match serde_norway::from_str(frontmatter_yaml) {
        Ok(v) => v,
        // Malformed YAML is FM_MALFORMED_YAML's job, not ours; report nothing.
        Err(_) => return Vec::new(),
    };
    let Value::Mapping(map) = value else {
        return Vec::new();
    };

    let mut out = Vec::new();
    for (k, v) in &map {
        if let Value::Sequence(items) = v {
            // Triple-nesting: some outer item is a sequence that itself holds a
            // sequence. Scalar inline `[[x]]` is only double-nested, so it
            // never matches.
            let is_link_list = items.iter().any(|item| match item {
                Value::Sequence(inner) => inner.iter().any(|x| matches!(x, Value::Sequence(_))),
                _ => false,
            });
            if is_link_list {
                if let Some(key) = k.as_str() {
                    out.push(key.to_string());
                }
            }
        }
    }
    out
}

/// One extracted section plus its verbatim line span: `[start, end)` as
/// 0-based indices into `body.split_inclusive('\n')` — exactly the slice the
/// section's `body` field was concatenated from. The span is what the
/// section editors (`dbmd section set` / `append`) splice against, so the
/// read views (`sections`, `outline`) and the write path share one boundary
/// rule (including the H1 terminator, which ends a span without ever being a
/// section itself).
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct SectionSpan {
    /// The extracted section.
    pub section: Section,
    /// First line of the span — the heading line — 0-based.
    pub start: usize,
    /// One past the last line of the span, 0-based.
    pub end: usize,
}

/// Extract the `##`/`###` sections of a markdown body with their verbatim
/// line spans. [`extract_sections`] is the span-less projection of this.
pub fn extract_section_spans(body: &str) -> Vec<SectionSpan> {
    // Keep each line's start so we can slice the body verbatim (exact newlines).
    let lines: Vec<&str> = body.split_inclusive('\n').collect();

    // First pass: classify heading levels (0 = not a heading), honoring fenced
    // code blocks so a `## x` inside a ``` fence is not treated as a heading.
    let mut levels: Vec<u8> = Vec::with_capacity(lines.len());
    let mut fence: Option<(u8, usize)> = None;
    for line in &lines {
        let content = line.trim_end_matches(['\n', '\r']);
        if let Some(f) = fence {
            if is_closing_fence(content, f) {
                fence = None;
            }
            levels.push(0);
            continue;
        }
        if let Some(opened) = opening_fence(content) {
            fence = Some(opened);
            levels.push(0);
            continue;
        }
        levels.push(heading_level(content));
    }

    // Second pass: emit `##`+ headings; each section body runs from its heading
    // line to the next heading at an equal-or-shallower level (exclusive).
    let mut sections = Vec::new();
    for (i, &lvl) in levels.iter().enumerate() {
        if lvl < 2 {
            continue;
        }
        let heading_line = lines[i].trim_end_matches(['\n', '\r']);
        let heading = heading_text(heading_line, lvl);

        let mut end = lines.len();
        for (j, &other) in levels.iter().enumerate().skip(i + 1) {
            if other != 0 && other <= lvl {
                end = j;
                break;
            }
        }

        sections.push(SectionSpan {
            section: Section {
                heading,
                level: lvl,
                line: (i + 1) as u32,
                body: lines[i..end].concat(),
            },
            start: i,
            end,
        });
    }
    sections
}

/// Extract the `##`/`###` sections of a markdown body into a flat list with
/// body slices.
pub fn extract_sections(body: &str) -> Vec<Section> {
    extract_section_spans(body)
        .into_iter()
        .map(|s| s.section)
        .collect()
}

/// Extract the `##`/`###` sections of a **whole file** (frontmatter + body),
/// returning each [`Section`] with `line` numbered against the *source file*,
/// not the body.
///
/// [`extract_sections`] numbers headings 1-based within the body it is handed —
/// the right frame for callers that already track the frontmatter offset
/// (`validate` adds `fm_end_line`). But the single-file views (`dbmd sections`,
/// `dbmd outline`) present `Section::line` as a source line an agent can jump to;
/// because every db.md file opens with a frontmatter block, the body-relative
/// number is off by the block's length (`opening fence + frontmatter lines +
/// closing fence`) for every file. This helper does the offset once, in the
/// parser, so those surfaces report true file lines. A file with no leading
/// frontmatter block is treated as all-body (offset 0), so the function never
/// fails just because a file lacks frontmatter.
pub fn extract_sections_in_file(text: &str) -> Vec<Section> {
    // Tolerate a leading BOM the same way `split_frontmatter` does, so the line
    // count and the body slice agree with the read path.
    let text = text.strip_prefix('\u{feff}').unwrap_or(text);

    // Find the body and how many source lines precede it. The body begins right
    // after the closing fence; the number of lines consumed by the frontmatter
    // block (both fences + the YAML between) is the offset to add to each
    // body-relative heading line.
    let (body, offset) = match split_frontmatter(text, Path::new("<sections>")) {
        Ok(parsed) => {
            // Lines before the body = total lines in `text` minus lines in body.
            let total_lines = count_lines(text);
            let body_lines = count_lines(&parsed.body);
            (parsed.body, total_lines.saturating_sub(body_lines))
        }
        // No frontmatter block: the whole text is body, no offset.
        Err(_) => (text.to_string(), 0),
    };

    let mut sections = extract_sections(&body);
    for s in &mut sections {
        s.line += offset;
    }
    sections
}

/// Count the number of lines a string spans for line-number offsetting: one line
/// per `\n`, plus one more for a final line with no trailing newline. An empty
/// string is zero lines.
fn count_lines(s: &str) -> u32 {
    if s.is_empty() {
        return 0;
    }
    let newlines = s.bytes().filter(|&b| b == b'\n').count() as u32;
    if s.ends_with('\n') {
        newlines
    } else {
        newlines + 1
    }
}

/// Parse a store's `DB.md` file into a [`Config`]: the `## Agent instructions`
/// prose, `## Policies` (`### Frozen pages` + `### Ignored types`), and
/// `## Schemas` (`### <type>` field-bullet blocks). Unrecognized sections are
/// ignored; absent sections leave their [`Config`] fields at default.
pub fn parse_db_md(text: &str, file: &Path) -> Result<Config, ParseError> {
    // The structured sections live in the body (after frontmatter). DB.md must
    // still start with a valid `---` block (`type: db-md`); if it's missing we
    // surface MissingFrontmatter like any other file.
    let parsed = split_frontmatter(text, file)?;
    let _frontmatter = Frontmatter::parse(&parsed.frontmatter_yaml, file)?;
    let sections = extract_sections(&parsed.body);

    let mut config = Config::default();
    // Track which H2 region each H3 belongs to as we walk the flat list.
    let mut current_h2: Option<String> = None;

    for section in &sections {
        match section.level {
            2 => {
                let name = section.heading.trim().to_ascii_lowercase();
                current_h2 = Some(name.clone());
                if name == "agent instructions" {
                    let prose = section_prose(&section.body);
                    if !prose.is_empty() {
                        config.agent_instructions = Some(prose);
                    }
                } else if name == "folders" {
                    // `## Folders` carries its bullets directly under the H2 (no
                    // `### <type>` sub-sections), like `## Agent instructions`.
                    for b in bullet_lines(&section.body) {
                        if let Some((path, meta)) = parse_folder_bullet(&b) {
                            config.folders.insert(path, meta);
                        }
                    }
                }
            }
            3 => {
                let h2 = current_h2.as_deref().unwrap_or("");
                let h3 = section.heading.trim().to_ascii_lowercase();
                match (h2, h3.as_str()) {
                    ("policies", "frozen pages") => {
                        config.frozen_pages = bullet_lines(&section.body)
                            .into_iter()
                            .map(|b| PathBuf::from(extract_path_bullet(&b)))
                            .collect();
                    }
                    ("policies", "ignored types") => {
                        config.ignored_types = bullet_lines(&section.body)
                            .into_iter()
                            .flat_map(|b| extract_type_list_bullet(&b))
                            .collect();
                    }
                    ("schemas", _) => {
                        // The H3 heading text (as written) is the type name.
                        let type_name = section.heading.trim().to_string();
                        let mut schema = Schema::default();
                        for b in bullet_lines(&section.body) {
                            match parse_schema_bullet(&b) {
                                SchemaBullet::Field(f) => schema.fields.push(f),
                                SchemaBullet::Unique(k) if !k.is_empty() => {
                                    schema.unique_keys.push(k)
                                }
                                SchemaBullet::SummaryTemplate(t) if !t.is_empty() => {
                                    schema.summary_template = Some(t)
                                }
                                SchemaBullet::Shard(Some(b)) => schema.shard = Some(b),
                                // Empty `unique:`/`summary_template:`, or a `shard:`
                                // with an unrecognized value — ignored.
                                SchemaBullet::Unique(_)
                                | SchemaBullet::SummaryTemplate(_)
                                | SchemaBullet::Shard(None) => {}
                            }
                        }
                        config.schemas.insert(type_name, schema);
                    }
                    _ => {}
                }
            }
            _ => {}
        }
    }

    Ok(config)
}

/// One parsed bullet inside a `### <type>` schema block: an ordinary field, or a
/// reserved directive (`unique:` / `summary_template:` / `shard:`). The names
/// `unique`, `summary_template`, and `shard` are reserved and cannot be used as
/// field names.
#[derive(Debug)]
enum SchemaBullet {
    /// An ordinary `- <name> (<modifiers>)` field.
    Field(FieldSpec),
    /// `- unique: <field>[, <field> …]` — a (possibly compound) uniqueness key.
    Unique(Vec<String>),
    /// `- summary_template: <template>` — the default-`summary` pattern.
    SummaryTemplate(String),
    /// `- shard: by-date | flat` — date-shard records of this type, or keep them
    /// flat. `None` = an unrecognized value, ignored like an unknown modifier.
    Shard(Option<bool>),
}

/// Classify one `## Schemas` bullet as a directive or a field. The directive
/// forms are `- unique: a, b, …` and `- summary_template: …`; the keyword check
/// guards against false positives — a field like `- status (enum: a, b)` has a
/// `(` before any `:`, so its head isn't a bare reserved keyword and it parses
/// as a [`FieldSpec`].
fn parse_schema_bullet(bullet_line: &str) -> SchemaBullet {
    let line = bullet_line.trim();
    let line = line
        .strip_prefix("- ")
        .or_else(|| line.strip_prefix("* "))
        .or_else(|| line.strip_prefix("+ "))
        .or_else(|| line.strip_prefix('-'))
        .unwrap_or(line)
        .trim();

    if let Some((head, rest)) = line.split_once(':') {
        match head.trim().to_ascii_lowercase().as_str() {
            "unique" => {
                let fields = rest
                    .split(',')
                    .map(|f| f.trim().to_string())
                    .filter(|f| !f.is_empty())
                    .collect();
                return SchemaBullet::Unique(fields);
            }
            "summary_template" => {
                return SchemaBullet::SummaryTemplate(rest.trim().to_string());
            }
            "shard" => {
                // `by-date` (synonyms: date/sharded/true) enables date-sharding;
                // `flat` (none/false) forces flat; anything else is ignored.
                let v = match rest.trim().to_ascii_lowercase().as_str() {
                    "by-date" | "date" | "sharded" | "true" => Some(true),
                    "flat" | "none" | "false" => Some(false),
                    _ => None,
                };
                return SchemaBullet::Shard(v);
            }
            _ => {}
        }
    }

    SchemaBullet::Field(parse_field_spec(bullet_line))
}

/// Parse one `## Folders` bullet — `- <path>[|<display>] — <description>` — into
/// the folder path (store-relative, unix-slash, no trailing slash) and its
/// [`FolderMeta`]. The optional `|<display>` overrides the rollup's derived
/// folder name (mirroring the wiki-link `|display` convention); the text after
/// the first em-dash (`—`), or ` - `, is the description. Backticks around the
/// path are tolerated (matching the `### Frozen pages` spelling). Returns `None`
/// for a bullet with no usable path.
fn parse_folder_bullet(bullet_line: &str) -> Option<(String, FolderMeta)> {
    let line = bullet_line.trim();
    let line = line
        .strip_prefix("- ")
        .or_else(|| line.strip_prefix("* "))
        .or_else(|| line.strip_prefix("+ "))
        .or_else(|| line.strip_prefix('-'))
        .unwrap_or(line)
        .trim();

    // Split off the description at the first em-dash (preferred, matching the
    // rollup's own ` — ` separator) or a ` - ` fallback.
    let (pathspec, description) = match line.find('—') {
        Some(i) => (line[..i].trim(), Some(line[i + '—'.len_utf8()..].trim())),
        None => match line.find(" - ") {
            Some(i) => (line[..i].trim(), Some(line[i + 3..].trim())),
            None => (line, None),
        },
    };

    // Optional `|display` override lives on the path side.
    let (path_raw, display) = match pathspec.split_once('|') {
        Some((p, d)) => (p.trim(), Some(d.trim())),
        None => (pathspec, None),
    };

    // Normalize the path: drop surrounding backticks, a leading `./`, a trailing `/`.
    let path = path_raw.trim().trim_matches('`').trim();
    let path = path.strip_prefix("./").unwrap_or(path);
    let path = path.strip_suffix('/').unwrap_or(path).trim();
    if path.is_empty() {
        return None;
    }

    let non_empty = |s: &str| {
        let t = s.trim();
        (!t.is_empty()).then(|| t.to_string())
    };
    Some((
        path.to_string(),
        FolderMeta {
            display: display.and_then(non_empty),
            description: description.and_then(non_empty),
        },
    ))
}

/// Parse a single `## Schemas` field-bullet line — `- <name> (<modifiers>)` —
/// into a [`FieldSpec`], capturing recognized modifiers and stashing the rest
/// in [`FieldSpec::unknown_modifiers`].
pub fn parse_field_spec(bullet_line: &str) -> FieldSpec {
    // Strip the leading bullet marker (`- ` / `* ` / `+ `) and surrounding ws.
    let line = bullet_line.trim();
    let line = line
        .strip_prefix("- ")
        .or_else(|| line.strip_prefix("* "))
        .or_else(|| line.strip_prefix("+ "))
        .or_else(|| line.strip_prefix('-'))
        .unwrap_or(line)
        .trim();

    // Split `<name> (<modifiers>)` — the canonical paren form — OR the natural
    // mis-spelling `<name>: <modifiers>` (colon instead of parens). The two
    // delimiters are interchangeable for the field head; whichever appears FIRST
    // wins, so a paren form whose modifiers contain a colon (`status (enum: a,
    // b)`) still parses by parens (the `(` precedes the `:`), while a bare
    // `title: string, required` parses by colon instead of being swallowed whole
    // into the field name with every modifier silently dropped.
    let paren = line.find('(');
    let colon = line.find(':');
    // Choose the head delimiter. The paren form wins when its `(` precedes any
    // `:` (so `status (enum: a, b)` parses by parens, the colon being inside the
    // modifiers); otherwise a `:` before the paren — or with no paren at all —
    // selects the colon form `<name>: <modifiers>`, the natural mis-spelling that
    // must NOT be swallowed whole into the field name with every modifier lost.
    let use_paren = matches!((paren, colon), (Some(p), c) if c.is_none_or(|c| p < c));
    let (name, modifiers) = if use_paren {
        let open = paren.expect("use_paren implies a paren");
        let name = line[..open].trim().to_string();
        let after = &line[open + 1..];
        let mods = match after.rfind(')') {
            Some(close) => &after[..close],
            None => after, // tolerate a missing close paren
        };
        (name, mods.trim())
    } else if let Some(c) = colon {
        // Colon form: everything after the first colon is the modifier list,
        // parsed identically to the parenthesized modifiers below.
        let name = line[..c].trim().to_string();
        (name, line[c + 1..].trim())
    } else {
        // Neither delimiter: a free-form optional field of any shape — name only.
        (line.to_string(), "")
    };

    let mut spec = FieldSpec {
        name,
        ..FieldSpec::default()
    };

    if modifiers.is_empty() {
        return spec;
    }

    // Modifiers are comma-separated. `enum` and `default` are special: their own
    // values may contain commas, so each is a *greedy* clause that runs from its
    // keyword to the start of the next recognized greedy clause (or end of line).
    // This lets `default North America, EMEA fallback` keep its comma and lets a
    // `default …` written after an `enum …` still be recognized, instead of the
    // value being truncated at the first comma or absorbed into the enum list.
    let raw: Vec<&str> = modifiers.split(',').collect();
    let mut i = 0;
    while i < raw.len() {
        let token = raw[i].trim();
        if token.is_empty() {
            i += 1;
            continue;
        }
        let lower = token.to_ascii_lowercase();

        if lower == "required" {
            spec.required = true;
            i += 1;
        } else if let Some(shape) = shape_from_str(&lower) {
            spec.shape = Some(shape);
            i += 1;
        } else if let Some(rest) = lower.strip_prefix("link to ") {
            // The trailing slash is required in the source; store the prefix
            // without it so `Path::starts_with` comparisons are clean.
            let prefix = token["link to ".len()..].trim().trim_end_matches('/');
            let _ = rest; // lowercase form only used for the keyword match
            spec.link_prefix = Some(PathBuf::from(prefix));
            i += 1;
        } else if token.len() >= "default ".len() && lower.starts_with("default ") {
            // Greedy `default <value>`: the value is this token (after the
            // keyword) plus every following comma-token up to the next greedy
            // clause, rejoined with the commas the split removed — so a comma
            // inside the default value is preserved. Original case is kept.
            let end = next_greedy_clause(&raw, i + 1);
            let mut value = token["default ".len()..].to_string();
            for tok in &raw[i + 1..end] {
                value.push(',');
                value.push_str(tok);
            }
            spec.default = Some(Value::String(value.trim().to_string()));
            i = end;
        } else if lower == "enum" || lower.starts_with("enum:") {
            // Greedy `enum` (bare `enum, a, b` or `enum: a, b`): the values run
            // from here to the next greedy clause (e.g. a trailing `default …`),
            // NOT unconditionally to end-of-line — so a `default` after `enum` is
            // parsed instead of swallowed as a bogus enum member.
            let end = next_greedy_clause(&raw, i + 1);
            // Rejoin this clause's tokens (trimmed so the `enum` head sits at the
            // start), drop the leading `enum`/`enum:` head, then re-split the
            // remainder into values.
            let joined = raw[i..end].join(",");
            let joined = joined.trim();
            let after_kw = match joined.find(':') {
                // `enum: a, b` — values follow the colon.
                Some(colon) => &joined[colon + 1..],
                // bare `enum, a, b` — values follow the keyword itself.
                None => joined.get("enum".len()..).unwrap_or(""),
            };
            let values: Vec<String> = after_kw
                .split(',')
                .map(|v| v.trim().to_string())
                .filter(|v| !v.is_empty())
                .collect();
            spec.enum_values = Some(values);
            i = end;
        } else {
            // Unrecognized modifier — captured verbatim, surfaced as Info.
            spec.unknown_modifiers.push(token.to_string());
            i += 1;
        }
    }

    spec
}

// ── Private helpers ─────────────────────────────────────────────────────────

/// Parse a frontmatter timestamp value into a `DateTime<FixedOffset>`. A `null`
/// is treated as absent; anything else must be an RFC3339 string.
fn parse_timestamp(
    value: &Value,
    key: &str,
    file: &Path,
) -> Result<Option<DateTime<FixedOffset>>, ParseError> {
    match value {
        Value::Null => Ok(None),
        Value::String(s) => parse_rfc3339(s, key, file).map(Some),
        other => Err(ParseError::BadTimestamp {
            file: file.to_path_buf(),
            key: key.to_string(),
            value: format!("{other:?}"),
        }),
    }
}

/// Parse an RFC3339 timestamp string, mapping failure to [`ParseError::BadTimestamp`].
fn parse_rfc3339(s: &str, key: &str, file: &Path) -> Result<DateTime<FixedOffset>, ParseError> {
    DateTime::parse_from_rfc3339(s.trim()).map_err(|_| ParseError::BadTimestamp {
        file: file.to_path_buf(),
        key: key.to_string(),
        value: s.to_string(),
    })
}

/// Coerce a YAML scalar value to its string form for the universal-contract
/// fields (`type`/`id`/`summary`/`status`). Mirrors `validate::scalar_string`
/// and `store::yaml_scalar_string` so the four modules agree on one coercion
/// rule: a bare numeric/bool scalar (`id: 100`, `summary: 2026`, `status: 0`)
/// is preserved as its string form rather than being read as None and silently
/// dropped on the next `to_yaml` re-emit. Returns `None` only for genuinely
/// non-scalar values (sequences, mappings, null), which were never a valid
/// shape for these fields.
fn scalar_string(value: &Value) -> Option<String> {
    match value {
        Value::String(s) => Some(s.clone()),
        Value::Number(n) => Some(n.to_string()),
        Value::Bool(b) => Some(b.to_string()),
        _ => None,
    }
}

/// Read a `tags` value into a flat `Vec<String>`. Accepts a sequence of scalars
/// (the canonical form) or a single scalar (coerced to a one-element list).
fn parse_tags(value: &Value) -> Vec<String> {
    match value {
        Value::Sequence(items) => items
            .iter()
            .filter_map(|v| match v {
                Value::String(s) => Some(s.clone()),
                Value::Number(n) => Some(n.to_string()),
                Value::Bool(b) => Some(b.to_string()),
                _ => None,
            })
            .collect(),
        Value::String(s) => vec![s.clone()],
        _ => Vec::new(),
    }
}

/// Read a `tags` value into a flat `Vec<String>` **without losing data**: a
/// sequence of clean scalars (the canonical form) or a single scalar coerce to a
/// string list. Any other shape — a sequence with a non-scalar item
/// (`tags: [[vip]]` → `Seq[Seq[String]]`, `tags: [a, [b]]`), or a mapping — is
/// rejected as `Err(value.clone())` so the caller preserves the raw value in
/// `extra` rather than silently filtering items out / erasing the field on the
/// next re-emit. This is the `tags` analog of routing a non-scalar universal
/// value to pass-through instead of the destroy path.
fn parse_tags_preserving(value: &Value) -> Result<Vec<String>, Value> {
    match value {
        Value::Sequence(items) => {
            let mut out = Vec::with_capacity(items.len());
            for item in items {
                match item {
                    Value::String(s) => out.push(s.clone()),
                    Value::Number(n) => out.push(n.to_string()),
                    Value::Bool(b) => out.push(b.to_string()),
                    // A non-scalar item (nested sequence/mapping/null) means this
                    // is not a clean tag list; preserve the whole value verbatim.
                    _ => return Err(value.clone()),
                }
            }
            Ok(out)
        }
        Value::String(s) => Ok(vec![s.clone()]),
        Value::Number(n) => Ok(vec![n.to_string()]),
        Value::Bool(b) => Ok(vec![b.to_string()]),
        // A mapping / null `tags` value is not a list; preserve it verbatim.
        _ => Err(value.clone()),
    }
}

/// Render a non-string YAML mapping key as the scalar text YAML would emit for
/// it (`2026`, `true`, `3.14`, …), so a numeric/bool/float frontmatter key
/// preserves its key *text* on round-trip instead of being rewritten to its Rust
/// `Debug` form (`Number(2026)`, `Bool(true)`, `'Null'`). The key re-emits as a
/// string-typed key carrying the original text (`'2026':`) — the type narrows to
/// string, but the operator's data is no longer corrupted, and ordinary string
/// keys are wholly unaffected. Falls back to `Debug` only for a key shape that
/// cannot be a scalar (a sequence/mapping key — not expressible in our
/// `String`-keyed `extra`), which never occurs in practice.
fn yaml_scalar_key(key: &Value) -> String {
    match key {
        Value::String(s) => s.clone(),
        Value::Number(n) => n.to_string(),
        Value::Bool(b) => b.to_string(),
        Value::Null => "null".to_string(),
        // Non-scalar key: not representable as a plain `extra` string key; keep
        // the defensive Debug form so nothing panics (unreachable in practice).
        other => format!("{other:?}"),
    }
}

/// Parse a single `[[target|display]]` string into a [`WikiLink`] with no
/// location, or `None` if the string is not a bare wiki-link. Used for
/// frontmatter-valued links where there is no body position to report.
fn parse_wiki_link_str(s: &str) -> Option<WikiLink> {
    let s = s.trim();
    let inner = s.strip_prefix("[[")?.strip_suffix("]]")?;
    // Reject anything with further brackets (e.g. the nested flow-form item),
    // which is not a clean single wiki-link.
    if inner.contains('[') || inner.contains(']') {
        return None;
    }
    let (target, display) = match inner.split_once('|') {
        Some((t, d)) => (t.to_string(), Some(d.to_string())),
        None => (inner.to_string(), None),
    };
    Some(WikiLink {
        is_full_path: target_is_full_path(&target),
        has_md_extension: target_has_md_extension(&target),
        target,
        display,
        location: (PathBuf::new(), 0, 0),
    })
}

/// Extract every wiki-link from a single frontmatter field value, accepting the
/// two canonical forms the spec defines (SPEC § Linking):
///
/// - a **scalar** wiki-link field, in either the quoted (`f: "[[x]]"`) or the
///   canonical unquoted inline (`f: [[x]]`) form, and
/// - a **list** field whose items are quoted wiki-link strings
///   (`- "[[x]]"`).
///
/// YAML eats the brackets of an unquoted `[[x]]`, leaving a flow-list-in-a-list,
/// so the parsed [`Value`] shapes are not what one would naively expect:
///
/// | source                         | parsed `Value`                     | here |
/// |--------------------------------|------------------------------------|------|
/// | `f: "[[x]]"`       (quoted)    | `String("[[x]]")`                  | link |
/// | `f: [[x]]`         (unquoted)  | `Seq[ Seq[String("x")] ]`          | link |
/// | `f:`\n`  - "[[x]]"`(quoted)    | `Seq[ String("[[x]]"), … ]`        | link |
/// | `f:`\n`  - [[x]]`  (unquoted)  | `Seq[ Seq[Seq[String("x")]], … ]`  | —    |
///
/// The last row — an *unquoted list* — parses identically to the flow-form list
/// `f: [[a], [b]]` and is a mis-encoding the canonical writer never emits;
/// `dbmd validate` reports it as `WIKI_LINK_FLOW_FORM_LIST` (see
/// [`detect_flow_form_link_lists`]). It is deliberately NOT surfaced here, so an
/// edge enumerator only ever sees the valid canonical forms.
///
/// The unquoted scalar (`Seq[Seq[String]]`, one element) is told apart from a
/// plain one-item flow list (`f: [x]` → `Seq[String]`, one fewer nesting level)
/// by [`unquoted_inline_link`] requiring its argument to be a `Sequence`.
fn links_in_field_value(value: &Value) -> Vec<WikiLink> {
    // Quoted scalar: `field: "[[x]]"`.
    if let Value::String(s) = value {
        return parse_wiki_link_str(s).into_iter().collect();
    }
    let Value::Sequence(items) = value else {
        return Vec::new();
    };
    // Unquoted scalar inline form `field: [[x]]` → `Seq[ Seq[String(x)] ]`.
    // (A quoted single-item list `["[[x]]"]` is `Seq[String]`, so its lone item
    // is a `String`, not a `Sequence`, and falls through to the list path below.)
    if items.len() == 1 {
        if let Some(link) = unquoted_inline_link(&items[0]) {
            return vec![link];
        }
    }
    // Otherwise a list of quoted wiki-link strings; non-string items (the
    // unquoted-list mis-encoding) are left for validate to flag.
    items
        .iter()
        .filter_map(|item| parse_wiki_link_str(item.as_str()?))
        .collect()
}

/// Canonicalize one `extra` frontmatter value for emission by [`Frontmatter::to_yaml`].
///
/// The read path ([`Frontmatter::parse`]) stores every unknown key's raw parsed
/// [`Value`] verbatim, so a SPEC-canonical *unquoted* inline scalar wiki-link
/// (`company: [[records/companies/northstar]]`) lands in `extra` as the nested
/// shape YAML produces for it — `Seq[ Seq[String("records/companies/northstar")] ]`.
/// Re-emitting that verbatim yields the block sequence
///
/// ```text
/// company:
/// - - records/companies/northstar
/// ```
///
/// which has lost the `[[ ]]` brackets entirely: the link is destroyed, and every
/// reader (validate, graph, backlinks) stops seeing the edge. This normalizes such
/// a value back into the canonical emitted form before it is written:
///
/// - a **scalar** wiki-link (quoted `String("[[x]]")` or unquoted `Seq[Seq[String]]`,
///   one element) → a quoted scalar `Value::String("[[x]]")`, which serde_norway emits
///   inline as `'[[x]]'` — the form the finding confirms survives a round-trip and
///   that [`links_in_field_value`] reads back as the same scalar link;
/// - a **list** of wiki-links (in any spelling [`links_in_field_value`] accepts) →
///   a block `Value::Sequence` of quoted-link strings (`- "[[x]]"`), matching the
///   `set` write-in path and the canonical list form;
/// - everything else → returned verbatim (the common no-op for non-link values).
///
/// `|display` is preserved in both link branches. This is the single point that
/// keeps all three curator-loop writers (`format`, `fm set`, `link`) from
/// corrupting a pre-existing canonical link, since they all funnel through
/// `to_yaml`.
fn canonicalize_extra_value(value: &Value) -> Value {
    match value {
        // Scalar wiki-link, quoted form: `field: "[[x]]"` → `String("[[x]]")`.
        // Re-emit as a quoted scalar so it stays a string (never the brackets-as-
        // YAML nested sequence). Non-link strings are returned untouched.
        Value::String(s) => match parse_wiki_link_str(s) {
            Some(link) => Value::String(wiki_link_literal(&link)),
            None => value.clone(),
        },
        Value::Sequence(items) => {
            // NOTE: we deliberately do NOT collapse a one-element
            // `Seq[ Seq[String(x)] ]` to the scalar `String("[[x]]")` here. That
            // shape is ambiguous — `serde_norway` parses BOTH an inline scalar
            // wiki-link `field: [[x]]` AND a genuine 2D array `field:`\n`- - x`
            // to exactly that value, so collapsing it silently retyped a real
            // nested array (`matrix: [["cell"]]`) into the string `'[[cell]]'`
            // and the file stopped round-tripping. The two cases ARE
            // distinguishable, but only from the source text, so the genuine
            // inline-link case is resolved at parse time
            // ([`Frontmatter::parse`] → [`inline_scalar_link_keys`]), where it is
            // stored as a `String("[[x]]")` and handled by the arm above. By the
            // time a `Seq[Seq[String]]` reaches here it is a real nested array and
            // must pass through verbatim (SPEC § "Unknown fields pass through").
            // List of wiki-links: re-emit as a block sequence of quoted-link
            // strings, the canonical list form `to_yaml` renders block-style and
            // `links_in_field_value` accepts. Only canonicalize when *every* item
            // is a clean single wiki-link; a list with any non-link item is left
            // verbatim so unrelated sequences (and the unquoted-list mis-encoding
            // validate flags) are untouched.
            let mut links = Vec::with_capacity(items.len());
            for item in items {
                match link_from_flow_list_item(item) {
                    Some(link) => links.push(link),
                    None => return value.clone(),
                }
            }
            if links.is_empty() {
                return value.clone();
            }
            Value::Sequence(
                links
                    .iter()
                    .map(|l| Value::String(wiki_link_literal(l)))
                    .collect(),
            )
        }
        // Mappings, scalars other than strings, nulls: nothing to canonicalize.
        _ => value.clone(),
    }
}

/// Render a [`WikiLink`] back to its `[[target]]` / `[[target|display]]` literal,
/// the inner form the canonical writer emits and `links_in_field_value` accepts.
fn wiki_link_literal(link: &WikiLink) -> String {
    match &link.display {
        Some(d) => format!("[[{}|{}]]", link.target, d),
        None => format!("[[{}]]", link.target),
    }
}

/// Recognize the inner token of an unquoted scalar `[[x]]`: after YAML strips the
/// outer brackets, the inner `[x]` is a single-element sequence `Seq[String(x)]`.
/// Reconstructs `[[x]]` (preserving any `|display`) and parses it, or returns
/// `None` when `v` is not that shape. Requiring a `Sequence` here is what keeps a
/// plain one-item flow list (`field: [x]` → `Seq[String]`, not `Seq[Seq[String]]`)
/// from being mistaken for a wiki-link.
fn unquoted_inline_link(v: &Value) -> Option<WikiLink> {
    let Value::Sequence(items) = v else {
        return None;
    };
    if items.len() != 1 {
        return None;
    }
    let s = items[0].as_str()?;
    // A clean unquoted wiki-link has no further brackets inside it.
    if s.contains('[') || s.contains(']') {
        return None;
    }
    parse_wiki_link_str(&format!("[[{s}]]"))
}

/// Scan raw frontmatter YAML for top-level keys whose value is written in the
/// **inline scalar wiki-link** form `key: [[target]]` (optionally
/// `[[target|display]]`).
///
/// This is the one disambiguation the parsed [`Value`] cannot supply on its own:
/// `serde_norway` parses BOTH
///
/// ```yaml
/// field: [[x]]
/// ```
///
/// and
///
/// ```yaml
/// field:
/// - - x
/// ```
///
/// to the identical `Seq[ Seq[String("x")] ]`. Only the source text says which one
/// the operator wrote. [`Frontmatter::parse`] calls this and rewrites the inline
/// cases to the canonical scalar `String("[[x]]")`, leaving every genuine nested
/// array a sequence (preserved verbatim per SPEC § "Unknown fields pass through").
///
/// Conservative by construction: a key is reported only when, on a single
/// top-level (zero-indent) line, the value after the first `:` is *exactly* one
/// `[[…]]` token (whitespace and an optional trailing `# comment` aside) with no
/// nested brackets inside. A quoted value (`field: "[[x]]"`), a flow list
/// (`field: [[a], [b]]`), a block sequence, or any indented/multi-token value is
/// left for the normal parse path. Duplicate keys (last-wins in YAML) are handled
/// by the caller looking up the final stored value.
fn inline_scalar_link_keys(yaml: &str) -> Vec<String> {
    let mut keys = Vec::new();
    for line in yaml.lines() {
        // Only top-level keys: an indented line is a nested mapping/sequence
        // entry, never a top-level `key: [[x]]` scalar.
        if line.starts_with(' ') || line.starts_with('\t') {
            continue;
        }
        let Some((raw_key, raw_val)) = line.split_once(':') else {
            continue;
        };
        let key = raw_key.trim();
        if key.is_empty() {
            continue;
        }
        // Drop a trailing `# comment` (YAML allows one after a plain scalar on the
        // same line). A `#` inside the bracketed link target is not a comment, but
        // such a target is rejected below anyway (it would not be a clean link).
        let val = match raw_val.split_once(" #") {
            Some((before, _)) => before.trim(),
            None => raw_val.trim(),
        };
        // The value must be exactly one bracket-delimited `[[…]]` token: starts
        // with `[[`, ends with `]]`, and the inner text carries no further
        // brackets (which would make it a flow list / nested collection, not a
        // single inline wiki-link).
        let Some(inner) = val.strip_prefix("[[").and_then(|s| s.strip_suffix("]]")) else {
            continue;
        };
        if inner.contains('[') || inner.contains(']') {
            continue;
        }
        // Confirm it is actually a parseable wiki-link, not e.g. an empty `[[]]`.
        if parse_wiki_link_str(val).is_some() {
            keys.push(key.to_string());
        }
    }
    keys
}

/// Decide whether a `dbmd fm set` / `--fm` value string is a **list of
/// wiki-links** that should be stored as a YAML block sequence, returning the
/// canonical `Value::Sequence` of quoted-link strings when so.
///
/// The value path of every write surface stringifies its argument; without this
/// a required list-of-links field (`meeting.attendees`) was unwritable in valid
/// form — passing `[[[a]], [[b]]]` stored a single scalar string that mis-parses
/// and trips `WIKI_LINK_FLOW_FORM_LIST` / `WIKI_LINK_BROKEN`. This recognizes the
/// two list spellings an agent naturally types and normalizes both to the block
/// form the canonical writer emits and `dbmd validate` accepts:
///
/// - flow list of quoted links — `["[[a]]", "[[b]]"]`
/// - flow list of unquoted links — `[[[a]], [[b]]]` (YAML: `Seq[Seq[String], …]`)
///
/// Returns `None` (⇒ caller stores a verbatim scalar string) for everything that
/// is not unambiguously a list of clean wiki-links — plain text, a single inline
/// `[[x]]` (YAML reads it as a one-item `Seq[Seq[String]]`, kept scalar so it
/// renders inline), an empty list, or a list with any non-link item. A single
/// link must stay scalar; only genuine multi-item-or-explicit lists become
/// sequences, matching `links_in_field_value`'s acceptance rule so writer and
/// validator never disagree.
fn parse_link_list_value(value: &str) -> Option<Value> {
    let trimmed = value.trim();
    // Only a YAML *flow sequence* literal is a list candidate; anything not
    // wrapped in `[ … ]` is a scalar (a bare `[[x]]` is wrapped, and handled by
    // the single-inline-link guard below).
    if !(trimmed.starts_with('[') && trimmed.ends_with(']')) {
        return None;
    }
    let Ok(Value::Sequence(items)) = serde_norway::from_str::<Value>(trimmed) else {
        return None;
    };
    // A single inline `[[x]]` parses to `Seq[ Seq[String(x)] ]` (one item, itself
    // a sequence) — that is the unquoted *scalar* form, not a list. Keep it scalar
    // so it round-trips to the inline `field: [[x]]` rather than a one-item block
    // list. `links_in_field_value` reads it back as a scalar link either way.
    if items.len() == 1 && unquoted_inline_link(&items[0]).is_some() {
        return None;
    }
    // Every item must resolve to exactly one clean wiki-link, in any of the flow
    // spellings an agent types (see [`link_from_flow_list_item`]).
    let mut links = Vec::with_capacity(items.len());
    for item in &items {
        links.push(link_from_flow_list_item(item)?);
    }
    if links.is_empty() {
        return None;
    }
    // Normalize to a block sequence of quoted-link strings — the form `to_yaml`
    // renders block-style and `links_in_field_value` accepts. `|display` is
    // preserved.
    let normalized = links
        .iter()
        .map(|l| Value::String(wiki_link_literal(l)))
        .collect();
    Some(Value::Sequence(normalized))
}

/// Recognize one clean wiki-link from a single **item** of a YAML flow sequence,
/// across the spellings an agent types for a list. After top-level flow parsing,
/// a list item arrives in one of:
///
/// - quoted — `"[[x]]"` ⇒ `String("[[x]]")`
/// - unquoted in a flow list — `[[x]]` inside `[…]` ⇒ `Seq[ Seq[String(x)] ]`
///   (one level deeper than a bare unquoted scalar, because the surrounding list
///   adds a wrapper); unwrap the single-element wrapper, then read the inline
///   `Seq[String(x)]` with [`unquoted_inline_link`].
///
/// Returns `None` for any item that is not exactly one clean wiki-link, so the
/// caller falls back to a scalar string and never fabricates a partial list.
fn link_from_flow_list_item(item: &Value) -> Option<WikiLink> {
    match item {
        Value::String(s) => parse_wiki_link_str(s),
        Value::Sequence(inner) => {
            // Unquoted list item `[[x]]` → `Seq[ Seq[String(x)] ]`: peel the lone
            // wrapper to expose the inline-link shape `Seq[String(x)]`.
            //
            // Only this triple-nested shape is a wiki-link. We deliberately do
            // NOT fall back to `unquoted_inline_link(item)` on the bare double
            // nesting `Seq[String(x)]` (a plain one-element string list `[x]`):
            // that fallback fabricated a wiki-link out of an ordinary nested
            // string list — `groups: [[alpha], [beta]]` (data `[["alpha"],
            // ["beta"]]`) was rewritten to `- '[[alpha]]'` / `- '[[beta]]'`,
            // silently changing the field's type and manufacturing short-form
            // links the tool then flags as `WIKI_LINK_SHORT_FORM`. An unknown
            // nested string list must pass through verbatim (SPEC § "Unknown
            // fields pass through").
            if inner.len() == 1 {
                if let Some(link) = unquoted_inline_link(&inner[0]) {
                    return Some(link);
                }
            }
            None
        }
        _ => None,
    }
}

/// A target is a full store-relative path when its first path segment is one of
/// the three canonical layer dirs and at least one `/` separator follows. A
/// trailing `.md` does not affect this classification.
fn target_is_full_path(target: &str) -> bool {
    let target = target.trim();
    match target.split_once('/') {
        Some((head, _rest)) => LAYER_DIRS.contains(&head),
        None => false,
    }
}

/// True when the target carries a trailing `.md` extension (validate warns
/// `WIKI_LINK_HAS_EXTENSION`).
fn target_has_md_extension(target: &str) -> bool {
    target.trim().ends_with(".md")
}

/// A forward-only cursor that yields the 1-based character (Unicode scalar)
/// column of successive byte offsets within a single line in ONE linear pass.
///
/// The previous helper recomputed `line[..offset].chars().count()` from the line
/// start for every match, so a line with N matches cost O(N × line_len) — a
/// quadratic blowup on a link-dense line. Because regex matches arrive in
/// non-decreasing byte order, this cursor advances the char count only across the
/// gap since the last queried offset, giving O(line_len) total per line.
///
/// Offsets MUST be queried in non-decreasing order and must fall on UTF-8
/// character boundaries (regex match starts always do).
struct ColCursor {
    byte: usize,
    chars: u32,
}

impl ColCursor {
    fn new() -> Self {
        ColCursor { byte: 0, chars: 0 }
    }

    /// 1-based character column of `byte_offset` in `line`. `byte_offset` must be
    /// `>=` every previously queried offset (debug-asserted).
    fn column_at(&mut self, line: &str, byte_offset: usize) -> u32 {
        debug_assert!(byte_offset >= self.byte, "ColCursor queried out of order");
        self.chars += line[self.byte..byte_offset].chars().count() as u32;
        self.byte = byte_offset;
        self.chars + 1
    }
}

/// Index of the first comma-token in `raw[from..]` that *starts a greedy
/// modifier clause* (`enum`, `enum:…`, or `default …`), or `raw.len()` when none
/// remain. Used to bound a greedy `default`/`enum` value so it stops at the next
/// such clause instead of either truncating at the first comma or swallowing a
/// following greedy clause whole.
fn next_greedy_clause(raw: &[&str], from: usize) -> usize {
    let mut j = from;
    while j < raw.len() {
        let lower = raw[j].trim().to_ascii_lowercase();
        if lower == "enum" || lower.starts_with("enum:") || lower.starts_with("default ") {
            return j;
        }
        j += 1;
    }
    raw.len()
}

/// Map a lowercase shape keyword to its [`Shape`].
fn shape_from_str(s: &str) -> Option<Shape> {
    match s {
        "string" => Some(Shape::String),
        "int" => Some(Shape::Int),
        "bool" => Some(Shape::Bool),
        "date" => Some(Shape::Date),
        "email" => Some(Shape::Email),
        "currency" => Some(Shape::Currency),
        "url" => Some(Shape::Url),
        _ => None,
    }
}

/// The ATX heading level of a line (number of leading `#`), or 0 if not a
/// heading. Up to three leading spaces (CommonMark), requires a space/tab (or
/// end-of-line) after the `#` run, caps the run at six.
fn heading_level(line: &str) -> u8 {
    let indent = line.len() - line.trim_start_matches(' ').len();
    if indent > 3 {
        return 0;
    }
    let rest = &line[indent..];
    let hashes = rest.len() - rest.trim_start_matches('#').len();
    if hashes == 0 || hashes > 6 {
        return 0;
    }
    let after = &rest[hashes..];
    if after.is_empty() || after.starts_with(' ') || after.starts_with('\t') {
        hashes as u8
    } else {
        0
    }
}

/// The heading text after the `#` run, trimmed, with a trailing ATX *closing*
/// `#` sequence removed per CommonMark (`## Title ##` → `Title`).
///
/// CommonMark only treats a trailing run of `#` as a closing sequence when it is
/// **preceded by a space or tab** (or the content is empty). A `#` that abuts the
/// preceding word is literal heading text: `## C#` → `C#`, `## F#` → `F#`,
/// `## issue-123#` → `issue-123#`. The old unconditional `trim_end_matches('#')`
/// stripped those, corrupting `dbmd sections`/`outline` heading text and — via
/// `parse_db_md` using the heading verbatim as the schema type key — silently
/// binding a `### c#` schema to `type: c` instead of `type: c#`.
fn heading_text(line: &str, level: u8) -> String {
    let indent = line.len() - line.trim_start_matches(' ').len();
    let after_hashes = &line[indent + level as usize..];
    let trimmed = after_hashes.trim();

    // Peel a trailing run of `#`. It is a closing sequence only if what precedes
    // it (within `trimmed`) is empty or ends in a space/tab; otherwise the `#`s
    // are literal content.
    let without_hashes = trimmed.trim_end_matches('#');
    if without_hashes.len() == trimmed.len() {
        // No trailing `#` at all.
        return trimmed.to_string();
    }
    if without_hashes.is_empty() || without_hashes.ends_with([' ', '\t']) {
        // A genuine closing sequence (`## Title ##`, `## ##`): drop it and the
        // whitespace before it.
        without_hashes.trim_end().to_string()
    } else {
        // The `#` run abuts content (`## C#`): keep it as literal heading text.
        trimmed.to_string()
    }
}

/// If `line` opens a fenced code block, return `(fence byte, run length)`.
fn opening_fence(line: &str) -> Option<(u8, usize)> {
    let indent = line.len() - line.trim_start_matches(' ').len();
    if indent > 3 {
        return None;
    }
    let rest = &line[indent..];
    let byte = rest.bytes().next()?;
    if byte != b'`' && byte != b'~' {
        return None;
    }
    let run = rest.len() - rest.trim_start_matches(byte as char).len();
    if run < 3 {
        return None;
    }
    // A backtick fence's info string may not itself contain a backtick.
    if byte == b'`' && rest[run..].contains('`') {
        return None;
    }
    Some((byte, run))
}

/// True if `line` closes the currently open fence: same char, run at least as
/// long, nothing but trailing whitespace after.
fn is_closing_fence(line: &str, fence: (u8, usize)) -> bool {
    let (byte, open_len) = fence;
    let indent = line.len() - line.trim_start_matches(' ').len();
    if indent > 3 {
        return false;
    }
    let rest = &line[indent..];
    let run = rest.len() - rest.trim_start_matches(byte as char).len();
    if run < open_len {
        return false;
    }
    rest[run..].trim().is_empty()
}

/// The prose body of a section: everything after the heading line, trimmed.
fn section_prose(section_body: &str) -> String {
    match section_body.split_once('\n') {
        Some((_heading, rest)) => rest.trim().to_string(),
        None => String::new(),
    }
}

/// The bullet lines (`-`/`*`/`+`) of a section body, excluding the heading
/// line, each returned with its leading whitespace trimmed.
fn bullet_lines(section_body: &str) -> Vec<String> {
    section_body
        .lines()
        .skip(1) // the heading line
        .map(str::trim)
        .filter(|l| l.starts_with("- ") || l.starts_with("* ") || l.starts_with("+ "))
        .map(|l| l.to_string())
        .collect()
}

/// Cut a bullet's content at the first comment separator, returning only the
/// meaningful prefix. Recognizes the em-dash (` — `), en-dash (` – `), double-
/// hyphen (` -- `), and the plain single-ASCII-hyphen (` - `) spellings an
/// operator naturally types — without the single-hyphen form, a comment like
/// `records/decisions/q3.md - finalized` left the whole line (comment included)
/// as the frozen path, so the entry never matched and the freeze failed OPEN.
/// A store-relative path never contains a ` - ` (paths are `/`-joined, spaceless),
/// so this does not truncate legitimate path text.
fn strip_bullet_comment(content: &str) -> &str {
    let mut cut = content.len();
    for sep in [" — ", " -- ", " – ", " - "] {
        if let Some(idx) = content.find(sep) {
            cut = cut.min(idx);
        }
    }
    content[..cut].trim()
}

/// Strip the leading bullet marker, returning the trimmed content after it.
fn bullet_content(bullet: &str) -> &str {
    let t = bullet.trim();
    t.strip_prefix("- ")
        .or_else(|| t.strip_prefix("* "))
        .or_else(|| t.strip_prefix("+ "))
        .unwrap_or(t)
        .trim()
}

/// Extract a store-relative path from a Frozen-pages bullet. The path may be
/// wrapped in backticks and followed by an em-dash comment.
fn extract_path_bullet(bullet: &str) -> String {
    let content = bullet_content(bullet);
    // Prefer a backtick-delimited span if present.
    if let Some(start) = content.find('`') {
        if let Some(end_rel) = content[start + 1..].find('`') {
            return content[start + 1..start + 1 + end_rel].trim().to_string();
        }
    }
    // Otherwise take the text up to a comment separator, stripping quotes.
    strip_bullet_comment(content)
        .trim_matches('"')
        .trim_matches('\'')
        .trim()
        .to_string()
}

/// Extract a comma-separated type list from an Ignored-types bullet, stripping
/// backticks/quotes and any trailing em-dash comment.
fn extract_type_list_bullet(bullet: &str) -> Vec<String> {
    let content = strip_bullet_comment(bullet_content(bullet));
    content
        .split(',')
        .map(|t| {
            t.trim()
                .trim_matches('`')
                .trim_matches('"')
                .trim_matches('\'')
                .trim()
                .to_string()
        })
        .filter(|t| !t.is_empty())
        .collect()
}

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

    #[test]
    fn read_file_refuses_oversized_sparse_input_before_allocating() {
        let dir = tempfile::tempdir().unwrap();
        let path = dir.path().join("hostile.md");
        let file = std::fs::File::create(&path).unwrap();
        file.set_len(MAX_DBMD_FILE_BYTES + 1).unwrap();

        let err = read_file(&path).unwrap_err();
        assert!(
            matches!(err, ParseError::Io(ref io) if io.kind() == std::io::ErrorKind::InvalidData),
            "oversized file must fail at the metadata gate: {err:?}"
        );
    }
    use std::path::Path;
    use tempfile::tempdir;

    // ── Config::frozen_match (the single write-surface policy matcher) ───────

    #[test]
    fn frozen_match_is_md_insensitive_both_directions() {
        // A policy entry stored WITHOUT `.md` (the natural extensionless
        // spelling `parse_db_md` keeps verbatim) must still match a `.md`
        // write target — the regression every write surface had.
        let cfg = Config {
            frozen_pages: vec![PathBuf::from("records/decisions/q1")],
            ..Config::default()
        };
        assert_eq!(
            cfg.frozen_match(Path::new("records/decisions/q1.md")),
            Some(PathBuf::from("records/decisions/q1")),
            "extensionless policy entry must freeze the .md file"
        );
        assert!(cfg.is_frozen(Path::new("records/decisions/q1.md")));

        // The symmetric case: a policy entry WITH `.md` matches a bare target.
        let cfg = Config {
            frozen_pages: vec![PathBuf::from("records/decisions/q1.md")],
            ..Config::default()
        };
        assert_eq!(
            cfg.frozen_match(Path::new("records/decisions/q1")),
            Some(PathBuf::from("records/decisions/q1.md")),
        );
        // And the same-spelling cases still match.
        assert!(cfg.is_frozen(Path::new("records/decisions/q1.md")));
    }

    #[test]
    fn frozen_match_drops_leading_dot_slash() {
        let cfg = Config {
            frozen_pages: vec![PathBuf::from("records/decisions/q1.md")],
            ..Config::default()
        };
        assert!(cfg.is_frozen(Path::new("./records/decisions/q1.md")));
        assert!(cfg.is_frozen(Path::new("./records/decisions/q1")));
    }

    #[test]
    fn frozen_match_returns_none_for_unlisted_and_prefix_paths() {
        let cfg = Config {
            frozen_pages: vec![PathBuf::from("records/decisions/q1")],
            ..Config::default()
        };
        assert!(cfg
            .frozen_match(Path::new("records/decisions/q2.md"))
            .is_none());
        // A prefix is not a match: `q1` must not freeze `q1-draft`.
        assert!(cfg
            .frozen_match(Path::new("records/decisions/q1-draft.md"))
            .is_none());
        assert!(!cfg.is_frozen(Path::new("records/decisions/q11.md")));
    }

    // ── split_frontmatter ───────────────────────────────────────────────────

    #[test]
    fn split_frontmatter_separates_yaml_and_verbatim_body() {
        let text = "---\ntype: contact\nsummary: x\n---\n# Heading\n\nBody line.\n";
        let p = split_frontmatter(text, Path::new("f.md")).unwrap();
        assert_eq!(p.frontmatter_yaml, "type: contact\nsummary: x\n");
        // Body is everything after the closing fence's newline, byte-for-byte.
        assert_eq!(p.body, "# Heading\n\nBody line.\n");
    }

    #[test]
    fn split_frontmatter_preserves_body_without_trailing_newline() {
        let text = "---\ntype: x\n---\nno trailing newline";
        let p = split_frontmatter(text, Path::new("f.md")).unwrap();
        assert_eq!(p.body, "no trailing newline");
    }

    #[test]
    fn split_frontmatter_empty_body_when_nothing_after_fence() {
        let text = "---\ntype: x\n---\n";
        let p = split_frontmatter(text, Path::new("f.md")).unwrap();
        assert_eq!(p.body, "");
    }

    #[test]
    fn split_frontmatter_missing_opening_fence_errors() {
        let text = "# No frontmatter here\ntype: x\n";
        let err = split_frontmatter(text, Path::new("f.md")).unwrap_err();
        assert!(matches!(err, ParseError::MissingFrontmatter { .. }));
    }

    #[test]
    fn split_frontmatter_leading_content_before_fence_rejected() {
        // The opening fence must be the very first line; a blank line first is
        // not allowed.
        let text = "\n---\ntype: x\n---\nbody";
        let err = split_frontmatter(text, Path::new("f.md")).unwrap_err();
        assert!(matches!(err, ParseError::MissingFrontmatter { .. }));
    }

    #[test]
    fn split_frontmatter_unterminated_block_errors() {
        let text = "---\ntype: x\nsummary: y\n";
        let err = split_frontmatter(text, Path::new("f.md")).unwrap_err();
        assert!(matches!(err, ParseError::MissingFrontmatter { .. }));
    }

    // ── Frontmatter::parse ───────────────────────────────────────────────────

    #[test]
    fn parse_populates_typed_fields_and_routes_unknowns_to_extra() {
        let yaml = "type: contact\nid: sarah-chen\nsummary: Director of Ops\nstatus: active\ntags: [vip, renewal]\nemail: sarah@northstar.io\nrole: Director";
        let fm = Frontmatter::parse(yaml, Path::new("f.md")).unwrap();
        assert_eq!(fm.type_.as_deref(), Some("contact"));
        assert_eq!(fm.id.as_deref(), Some("sarah-chen"));
        assert_eq!(fm.summary.as_deref(), Some("Director of Ops"));
        assert_eq!(fm.status.as_deref(), Some("active"));
        assert_eq!(fm.tags, vec!["vip".to_string(), "renewal".to_string()]);
        // Type-specific fields are NOT promoted to typed slots.
        assert!(fm.type_.is_some() && !fm.extra.contains_key("type"));
        assert!(!fm.extra.contains_key("tags"));
        assert_eq!(
            fm.extra.get("email").and_then(|v| v.as_str()),
            Some("sarah@northstar.io")
        );
        assert_eq!(
            fm.extra.get("role").and_then(|v| v.as_str()),
            Some("Director")
        );
    }

    #[test]
    fn parse_reads_rfc3339_timestamps() {
        let yaml =
            "type: email\ncreated: 2026-05-27T08:00:00-07:00\nupdated: 2026-05-28T09:30:00-07:00";
        let fm = Frontmatter::parse(yaml, Path::new("f.md")).unwrap();
        let created = fm.created.expect("created parsed");
        // -07:00 offset is 7 * 3600 seconds west.
        assert_eq!(created.offset().utc_minus_local(), 7 * 3600);
        assert_eq!(created.to_rfc3339(), "2026-05-27T08:00:00-07:00");
        assert!(fm.updated.is_some());
    }

    #[test]
    fn parse_preserves_non_rfc3339_timestamp_verbatim() {
        // A date-only value is not a full RFC3339 timestamp, so the typed
        // accessor stays None — but the READ path must never destroy it or
        // refuse the file. It rides in `extra` and round-trips byte-for-byte,
        // exactly like a non-scalar `type`/`summary`. `validate` is what
        // reports it (FM_BAD_TIMESTAMP, raised from the raw YAML value).
        //
        // Regression: erroring here made `dbmd format` (and `fm`/`link`/
        // `rename`, all of which go through `read_file`) fail outright on any
        // store carrying a legacy date-only stamp — the common migrated shape.
        let yaml = "type: email\ncreated: 2026-05-27";
        let fm = Frontmatter::parse(yaml, Path::new("bad.md")).unwrap();
        assert!(
            fm.created.is_none(),
            "unparseable stamp offers no typed value"
        );
        assert_eq!(
            fm.extra.get("created").and_then(Value::as_str),
            Some("2026-05-27"),
            "the operator's bytes must survive the read"
        );
        assert!(
            fm.to_yaml().contains("created: 2026-05-27"),
            "and must re-emit verbatim; got:\n{}",
            fm.to_yaml()
        );
    }

    #[test]
    fn set_still_refuses_to_author_a_bad_timestamp() {
        // The read/write asymmetry is the point: tolerate what a store already
        // contains, never CREATE a malformed value. (`set_timestamp_validates_
        // rfc3339` covers the same boundary from the write side.)
        let mut fm = Frontmatter::parse("type: email\ncreated: 2026-05-27", Path::new("b.md"))
            .expect("read tolerates the legacy stamp");
        assert!(matches!(
            fm.set("created", "still-not-a-date").unwrap_err(),
            ParseError::BadTimestamp { .. }
        ));
    }

    #[test]
    fn parse_malformed_yaml_errors() {
        // Unclosed flow mapping is invalid YAML.
        let yaml = "type: contact\n  bad: : :\n- nope";
        let err = Frontmatter::parse(yaml, Path::new("bad.md")).unwrap_err();
        assert!(matches!(err, ParseError::MalformedYaml { .. }));
    }

    #[test]
    fn frontmatter_with_yaml_tag_on_mapping_does_not_panic() {
        // Regression: a YAML tag on the top-level mapping made the old
        // `expect_err` path PANIC, because a tagged mapping deserializes to a
        // `Mapping` just fine. It must now be handled — accepted as the inner
        // mapping, never a panic.
        let fm = Frontmatter::parse("!mytag\ntype: contact\nsummary: hi\n", Path::new("x.md"))
            .expect("tagged-mapping frontmatter must parse, not panic");
        assert_eq!(fm.type_.as_deref(), Some("contact"));
        // A genuine scalar/sequence top level is still malformed (and still
        // doesn't panic).
        assert!(Frontmatter::parse("- a\n- b\n", Path::new("x.md")).is_err());
    }

    #[test]
    fn parse_empty_block_is_empty_frontmatter() {
        let fm = Frontmatter::parse("", Path::new("f.md")).unwrap();
        assert_eq!(fm, Frontmatter::default());
    }

    #[test]
    fn parse_scalar_top_level_is_malformed() {
        // A bare scalar at the top level is not a frontmatter mapping.
        let err = Frontmatter::parse("just a string", Path::new("f.md")).unwrap_err();
        assert!(matches!(err, ParseError::MalformedYaml { .. }));
    }

    // ── to_yaml canonical order ──────────────────────────────────────────────

    #[test]
    fn to_yaml_emits_canonical_key_order() {
        let mut fm = Frontmatter {
            type_: Some("contact".into()),
            id: Some("sarah-chen".into()),
            summary: Some("Director of Ops".into()),
            status: Some("active".into()),
            tags: vec!["vip".into()],
            created: Some(DateTime::parse_from_rfc3339("2026-05-27T08:00:00-07:00").unwrap()),
            updated: Some(DateTime::parse_from_rfc3339("2026-05-28T09:30:00-07:00").unwrap()),
            ..Default::default()
        };
        // Two type-specific fields, inserted in NON-alphabetical order to prove
        // the writer sorts them (BTreeMap) between the universal head and tail.
        fm.extra
            .insert("role".into(), Value::String("Director".into()));
        fm.extra.insert(
            "company".into(),
            Value::String("[[records/companies/northstar]]".into()),
        );

        let yaml = fm.to_yaml();
        let keys: Vec<&str> = yaml
            .lines()
            .filter(|l| !l.starts_with(['-', ' ']) && l.contains(':'))
            .map(|l| l.split(':').next().unwrap())
            .collect();
        assert_eq!(
            keys,
            vec![
                "type", "id", "created", "updated", "summary", // universal head
                "company", "role",   // type-specific, sorted
                "status", // universal tail
                "tags",
            ],
            "canonical order violated; got:\n{yaml}"
        );
        // Timestamps round-trip as RFC3339 strings (YAML may quote them).
        assert!(
            yaml.contains("2026-05-27T08:00:00-07:00"),
            "created timestamp missing; got:\n{yaml}"
        );
        // The value re-parses to the same instant regardless of quoting.
        let reparsed = Frontmatter::parse(&yaml, Path::new("rt.md")).unwrap();
        assert_eq!(reparsed.created, fm.created);
        assert_eq!(reparsed.updated, fm.updated);
    }

    /// Format v0.4: a minted-form (lowercase ULID) `id` round-trips verbatim
    /// through parse → to_yaml → parse and holds its canonical head slot —
    /// directly after `type` (and after `meta-type` when one is present),
    /// before `created`. Pins the emit order for the id-carrying record shape
    /// `dbmd write` produces.
    #[test]
    fn ulid_id_roundtrips_verbatim_in_head_position() {
        let ulid = "01j5qc3v9k4ym8rwbn2tqe6f7d";
        let yaml = format!(
            "type: profile\nmeta-type: conclusion\nid: {ulid}\ncreated: 2026-05-27T08:00:00-07:00\nupdated: 2026-05-27T08:00:00-07:00\nsummary: x\n"
        );
        let fm = Frontmatter::parse(&yaml, Path::new("rt.md")).unwrap();
        assert_eq!(
            fm.id.as_deref(),
            Some(ulid),
            "id must parse into the typed field"
        );

        let emitted = fm.to_yaml();
        let keys: Vec<&str> = emitted
            .lines()
            .filter(|l| !l.starts_with(['-', ' ']) && l.contains(':'))
            .map(|l| l.split(':').next().unwrap())
            .collect();
        assert_eq!(
            keys,
            vec!["type", "meta-type", "id", "created", "updated", "summary"],
            "id must sit in the universal head; got:\n{emitted}"
        );
        assert!(
            emitted.contains(&format!("id: {ulid}")),
            "ULID must emit unquoted and verbatim; got:\n{emitted}"
        );
        let reparsed = Frontmatter::parse(&emitted, Path::new("rt.md")).unwrap();
        assert_eq!(reparsed.id.as_deref(), Some(ulid));
        assert_eq!(reparsed, fm, "round-trip must be lossless");
    }

    #[test]
    fn to_yaml_omits_absent_optional_fields() {
        let fm = Frontmatter {
            type_: Some("note".into()),
            ..Default::default()
        };
        let yaml = fm.to_yaml();
        assert!(yaml.contains("type: note"));
        assert!(!yaml.contains("status"));
        assert!(!yaml.contains("tags"));
        assert!(!yaml.contains("summary"));
    }

    // ── Regression: non-string scalar universal fields round-trip (finding #1) ─

    #[test]
    fn regression_parse_preserves_non_string_scalar_universal_fields() {
        // A hand/externally-authored file whose universal fields are bare
        // scalars YAML reads as Number/Bool — `id: 100`, `summary: 2026`,
        // `status: 0`, `type: 42` — must be PRESERVED as their string form, not
        // read as None. Before the fix, `v.as_str()` returned None for these and
        // the matched arm discarded the value entirely (never reaching `extra`).
        let yaml = "type: 42\nid: 100\nsummary: 2026\nstatus: 0";
        let fm = Frontmatter::parse(yaml, Path::new("x.md")).unwrap();
        assert_eq!(fm.type_.as_deref(), Some("42"), "type scalar dropped");
        assert_eq!(fm.id.as_deref(), Some("100"), "id scalar dropped");
        assert_eq!(
            fm.summary.as_deref(),
            Some("2026"),
            "summary scalar dropped"
        );
        assert_eq!(fm.status.as_deref(), Some("0"), "status scalar dropped");
        // The values must surface through the public `get` accessor too.
        assert_eq!(
            fm.get("summary")
                .and_then(|v| v.as_str().map(str::to_string)),
            Some("2026".to_string())
        );
    }

    #[test]
    fn regression_format_round_trip_does_not_delete_numeric_frontmatter() {
        // The exact finding-#1 trigger: `dbmd format` is read_file -> write_file.
        // A file whose `id`/`summary`/`status` are bare numeric scalars must
        // still carry those fields after the canonical re-emit. Before the fix,
        // the lines were silently deleted from disk (only `type` survived).
        let dir = tempdir().unwrap();
        let path = dir.path().join("x.md");
        let original = "---\ntype: contact\nid: 100\nsummary: 2026\nstatus: 0\n---\nbody\n";
        std::fs::write(&path, original).unwrap();

        // Re-emit through the canonical writer, exactly as `dbmd format` does.
        let (fm, body) = read_file(&path).unwrap();
        write_file(&path, &fm, &body).unwrap();

        let after = std::fs::read_to_string(&path).unwrap();
        // None of the four fields may vanish; they survive as string scalars.
        let reparsed = Frontmatter::parse(
            &split_frontmatter(&after, &path).unwrap().frontmatter_yaml,
            &path,
        )
        .unwrap();
        assert_eq!(reparsed.type_.as_deref(), Some("contact"));
        assert_eq!(reparsed.id.as_deref(), Some("100"), "id deleted by format");
        assert_eq!(
            reparsed.summary.as_deref(),
            Some("2026"),
            "summary deleted by format"
        );
        assert_eq!(
            reparsed.status.as_deref(),
            Some("0"),
            "status deleted by format"
        );
        // The body is preserved verbatim.
        assert_eq!(body, "body\n");
    }

    #[test]
    fn regression_format_round_trip_preserves_oversized_integer_frontmatter() {
        // Adversarial review #6: a bare integer literal beyond i64/u64 range must
        // survive `dbmd format` (read_file -> write_file) byte-for-byte. Before
        // the fix, serde_norway silently truncated `> u128::MAX` to f64 (`999…9`
        // -> `1e39`) and hard-rejected `(u64::MAX, u128::MAX]` — corrupting an
        // imported numeric ID and breaking the unknown-field round-trip contract.
        let dir = tempdir().unwrap();
        let path = dir.path().join("x.md");
        let big = "999999999999999999999999999999999999999"; // 39 digits, > u128::MAX
        let mid = "99999999999999999999"; // 20 digits, in (u64::MAX, u128::MAX]
        let original = format!(
            "---\ntype: contact\nsummary: x\naccount_number: {big}\nid_num: {mid}\n---\nbody\n"
        );
        std::fs::write(&path, &original).unwrap();

        // Two round-trips: the value must survive verbatim AND be idempotent.
        for _ in 0..2 {
            let (fm, body) = read_file(&path).expect("oversized-int frontmatter must parse");
            write_file(&path, &fm, &body).unwrap();
            let after = std::fs::read_to_string(&path).unwrap();
            assert!(
                after.contains(big),
                "39-digit integer corrupted by format:\n{after}"
            );
            assert!(
                after.contains(mid),
                "20-digit integer corrupted by format:\n{after}"
            );
            assert!(
                !after.to_lowercase().contains("1e39"),
                "integer was truncated to a float:\n{after}"
            );
            assert_eq!(body, "body\n", "body must be preserved verbatim");
        }
    }

    #[test]
    fn oversized_int_literal_detection_is_precise() {
        // In range (serde_norway handles losslessly) → never quoted.
        for ok in [
            "0",
            "42",
            "-17",
            "9223372036854775807",
            "18446744073709551615",
            "12.5",
            "007",
            "abc",
            "",
        ] {
            assert!(
                !is_oversized_int_literal(ok),
                "must NOT be flagged oversized: {ok:?}"
            );
        }
        // Beyond i64/u64 → quoted to preserve the literal.
        for big in [
            "18446744073709551616",                    // u64::MAX + 1
            "99999999999999999999",                    // 20 digits
            "999999999999999999999999999999999999999", // 39 digits
            "-9999999999999999999999",                 // very negative
        ] {
            assert!(
                is_oversized_int_literal(big),
                "must be flagged oversized: {big:?}"
            );
        }
    }

    #[test]
    fn regression_oversized_int_in_flow_sequence_round_trips() {
        // The single-line flow SEQUENCE form regressed: an oversized int inside
        // `ids: [123…]` reached serde_norway un-quoted and hard-failed the whole
        // block as MalformedYaml (`as u128`), making every read surface
        // (format / fm get/set / link / validate) unable to read the file at all.
        // It must now parse, preserve the literal verbatim, and be idempotent.
        let dir = tempdir().unwrap();
        let path = dir.path().join("f.md");
        let big = "123456789012345678901234567890"; // 30 digits, > u128::MAX
        let original = format!("---\ntype: note\nsummary: x\nids: [{big}]\n---\nbody\n");
        std::fs::write(&path, &original).unwrap();

        for _ in 0..2 {
            let (fm, body) = read_file(&path).expect("flow-sequence oversized int must parse");
            // The list value survives in `extra`, holding the literal as a string.
            let ids = fm.extra.get("ids").expect("ids field preserved");
            assert!(
                matches!(ids, Value::Sequence(_)),
                "ids should stay a sequence, got: {ids:?}"
            );
            write_file(&path, &fm, &body).unwrap();
            let after = std::fs::read_to_string(&path).unwrap();
            assert!(
                after.contains(big),
                "30-digit integer in flow sequence corrupted by format:\n{after}"
            );
            assert!(
                !after.to_lowercase().contains("1.234"),
                "integer was truncated to a float:\n{after}"
            );
            assert_eq!(body, "body\n", "body must be preserved verbatim");
        }
    }

    #[test]
    fn regression_oversized_int_in_flow_mapping_round_trips() {
        // The single-line flow MAPPING form regressed identically:
        // `meta: {ext: 123…}` hard-failed the block. It must now parse and the
        // oversized value must survive verbatim.
        let dir = tempdir().unwrap();
        let path = dir.path().join("m.md");
        let big = "123456789012345678901234567890";
        let original = format!("---\ntype: note\nsummary: x\nmeta: {{ext: {big}}}\n---\nbody\n");
        std::fs::write(&path, &original).unwrap();

        for _ in 0..2 {
            let (fm, body) = read_file(&path).expect("flow-mapping oversized int must parse");
            let meta = fm.extra.get("meta").expect("meta field preserved");
            assert!(
                matches!(meta, Value::Mapping(_)),
                "meta should stay a mapping, got: {meta:?}"
            );
            write_file(&path, &fm, &body).unwrap();
            let after = std::fs::read_to_string(&path).unwrap();
            assert!(
                after.contains(big),
                "oversized integer in flow mapping corrupted by format:\n{after}"
            );
            assert_eq!(body, "body\n", "body must be preserved verbatim");
        }
    }

    #[test]
    fn regression_oversized_int_in_mixed_flow_collection_round_trips() {
        // A flow collection mixing an oversized int with an in-range int and a
        // string: only the oversized int is quoted; the in-range int stays a
        // number, the string stays a string, and the whole thing parses.
        let dir = tempdir().unwrap();
        let path = dir.path().join("mix.md");
        let big = "123456789012345678901234567890";
        let original = format!(
            "---\ntype: note\nsummary: x\nvals: [{big}, 42, hello, \"world\"]\n---\nbody\n"
        );
        std::fs::write(&path, &original).unwrap();

        let (fm, body) = read_file(&path).expect("mixed flow collection must parse");
        let Value::Sequence(seq) = fm.extra.get("vals").expect("vals preserved") else {
            panic!("vals should be a sequence");
        };
        assert_eq!(seq.len(), 4, "all four entries preserved");
        // The oversized literal narrows to a string; the in-range int stays a
        // number; the bare and quoted strings stay strings.
        assert_eq!(seq[0].as_str(), Some(big), "oversized int -> string");
        assert_eq!(seq[1].as_i64(), Some(42), "in-range int stays a number");
        assert_eq!(seq[2].as_str(), Some("hello"));
        assert_eq!(seq[3].as_str(), Some("world"));

        write_file(&path, &fm, &body).unwrap();
        let after = std::fs::read_to_string(&path).unwrap();
        assert!(after.contains(big), "oversized int lost:\n{after}");
        assert_eq!(body, "body\n");
    }

    #[test]
    fn regression_multiple_oversized_ints_in_one_flow_line_round_trip() {
        // Two oversized literals on the same flow line — and a nested collection —
        // must each be quoted in the single left-to-right pass.
        let dir = tempdir().unwrap();
        let path = dir.path().join("multi.md");
        let a = "99999999999999999999"; // 20 digits
        let b = "123456789012345678901234567890"; // 30 digits
        let original =
            format!("---\ntype: note\nsummary: x\nm: {{a: {a}, nested: [{b}, 7]}}\n---\nbody\n");
        std::fs::write(&path, &original).unwrap();

        let (fm, body) = read_file(&path).expect("multi oversized flow must parse");
        write_file(&path, &fm, &body).unwrap();
        let after = std::fs::read_to_string(&path).unwrap();
        assert!(after.contains(a), "first oversized int lost:\n{after}");
        assert!(after.contains(b), "second oversized int lost:\n{after}");
        assert_eq!(body, "body\n");
    }

    #[test]
    fn regression_flow_with_only_in_range_and_strings_is_byte_exact() {
        // A flow collection with NO oversized int must round-trip byte-for-byte:
        // the pre-quoter must not touch in-range ints, strings, or floats. We
        // assert on the prepared-YAML stage so an unaffected line is left as the
        // borrowed input (no rewrite, no quoting drift).
        let yaml = "type: note\nids: [1, 2, 3]\nmeta: {ext: 42, name: bob}\nf: [1.5, 2.5]\n";
        let prepared = quote_oversized_integers(yaml);
        assert_eq!(
            prepared.as_ref(),
            yaml,
            "in-range flow collections must be left byte-exact"
        );
        // And it still parses cleanly with the expected numeric types intact.
        let fm = Frontmatter::parse(yaml, Path::new("n.md")).unwrap();
        let Value::Sequence(ids) = fm.extra.get("ids").unwrap() else {
            panic!("ids should be a sequence");
        };
        assert_eq!(ids[0].as_i64(), Some(1));
    }

    #[test]
    fn quote_oversized_ints_in_flow_skips_quoted_and_digit_strings() {
        // A quoted scalar whose contents happen to be a long digit run must NOT
        // be re-quoted or otherwise altered — it is already a string. A flow with
        // only such strings yields no change (None).
        let flow = "[\"123456789012345678901234567890\", '99999999999999999999']";
        assert_eq!(
            quote_oversized_ints_in_flow(flow),
            None,
            "already-quoted digit strings must be left untouched"
        );
        // A bare oversized int alongside a quoted one: only the bare one is quoted.
        let flow2 = "[123456789012345678901234567890, \"already\"]";
        let out = quote_oversized_ints_in_flow(flow2).expect("bare int should be quoted");
        assert_eq!(out, "['123456789012345678901234567890', \"already\"]");
    }

    // ── Regression: BOM-prefixed files parse like store/index (finding #19) ────

    #[test]
    fn regression_split_frontmatter_tolerates_leading_utf8_bom() {
        // A BOM-prefixed file (EF BB BF + `---\n...`) is walked and indexed by
        // `dbmd index` (store/index strip the BOM) but, before the fix, every
        // write/edit surface routed through `read_file` hard-failed with
        // MissingFrontmatter. `split_frontmatter` must now strip a single leading
        // U+FEFF and emit a BOM-free body.
        let text = "\u{feff}---\ntype: note\nsummary: x\n---\nbody\n";
        let parsed = split_frontmatter(text, Path::new("note.md")).unwrap();
        assert_eq!(parsed.frontmatter_yaml, "type: note\nsummary: x\n");
        // Body never carries the BOM forward into the canonical writer.
        assert_eq!(parsed.body, "body\n");
        assert!(!parsed.body.starts_with('\u{feff}'));
    }

    #[test]
    fn regression_read_file_parses_bom_prefixed_file() {
        // End-to-end through the same `read_file` path `dbmd fm get/set`,
        // `format`, `link`, and `write` use. Before the fix this returned
        // Err(MissingFrontmatter) on a file the catalog had already indexed.
        let dir = tempdir().unwrap();
        let path = dir.path().join("note.md");
        std::fs::write(&path, "\u{feff}---\ntype: note\nsummary: x\n---\nbody\n").unwrap();

        let (fm, body) = read_file(&path).expect("BOM-prefixed file must parse");
        assert_eq!(fm.type_.as_deref(), Some("note"));
        assert_eq!(fm.summary.as_deref(), Some("x"));
        assert_eq!(body, "body\n");
    }

    #[test]
    fn to_yaml_preserves_unquoted_scalar_wiki_link_round_trip() {
        // Regression (PRIMARY): the SPEC-canonical scalar wiki-link is the
        // *unquoted* inline `company: [[records/companies/northstar]]`
        // (SPEC § Linking, the worked `contact` example). YAML parses it to the
        // nested `Seq[Seq[String]]` shape. Before the fix, `to_yaml` re-emitted
        // it block-style as
        //     company:
        //     - - records/companies/northstar
        // — the `[[ ]]` brackets GONE — so a no-op re-emit (`dbmd format`, and
        // any `fm set` / `link` write) silently destroyed the link.
        let yaml = "type: contact\ncompany: [[records/companies/northstar]]";
        let fm = Frontmatter::parse(yaml, Path::new("c.md")).unwrap();
        // Sanity: `parse` now disambiguates the inline-link source form at read
        // time (the genuine `Seq[Seq[String]]` of a 2D array no longer gets
        // collapsed at emit), so the inline link is stored as the canonical
        // scalar `String("[[x]]")`.
        assert_eq!(
            fm.extra.get("company").and_then(|v| v.as_str()),
            Some("[[records/companies/northstar]]")
        );

        let out = fm.to_yaml();
        // The link must survive as a quoted inline scalar — brackets intact, and
        // never the bracket-less block sequence `- - records/...`.
        assert!(
            out.contains("[[records/companies/northstar]]"),
            "canonical writer dropped the wiki-link brackets; got:\n{out}"
        );
        assert!(
            !out.contains("- - "),
            "canonical writer emitted a nested block sequence (link corrupted); got:\n{out}"
        );

        // And it round-trips: re-parsing the emitted YAML still surfaces exactly
        // one link with the right target (the edge graph/backlinks rely on).
        let reparsed = Frontmatter::parse(&out, Path::new("c.md")).unwrap();
        let fields = reparsed.link_fields();
        let links: Vec<(&str, &str, Option<&str>)> = fields
            .iter()
            .map(|(k, l)| (k.as_str(), l.target.as_str(), l.display.as_deref()))
            .collect();
        assert_eq!(
            links,
            vec![("company", "records/companies/northstar", None)]
        );

        // A second re-emit is a fixed point — no progressive corruption across
        // repeated curator-loop writes.
        assert_eq!(
            reparsed.to_yaml(),
            out,
            "to_yaml is not idempotent on links"
        );
    }

    #[test]
    fn to_yaml_preserves_unquoted_scalar_link_with_display() {
        // The `|display` segment must survive the unquoted-inline round-trip too.
        let yaml = "type: contact\ncompany: [[records/companies/northstar|Northstar]]";
        let fm = Frontmatter::parse(yaml, Path::new("c.md")).unwrap();
        let out = fm.to_yaml();
        assert!(
            out.contains("[[records/companies/northstar|Northstar]]"),
            "display segment lost on round-trip; got:\n{out}"
        );
        let reparsed = Frontmatter::parse(&out, Path::new("c.md")).unwrap();
        let f = reparsed.link_fields();
        assert_eq!(f.len(), 1);
        assert_eq!(f[0].1.target, "records/companies/northstar");
        assert_eq!(f[0].1.display.as_deref(), Some("Northstar"));
    }

    #[test]
    fn to_yaml_does_not_mangle_link_list_or_plain_nested_sequence() {
        // A genuine quoted block list of links round-trips as a clean string
        // list — never collapsed to a scalar — and a plain nested sequence that
        // is NOT a wiki-link is left exactly as written (no false conversion).
        let yaml = "type: meeting\nattendees:\n  - \"[[records/contacts/elena]]\"\n  - \"[[records/contacts/sarah]]\"\nmatrix:\n  - - 1\n    - 2";
        let fm = Frontmatter::parse(yaml, Path::new("m.md")).unwrap();
        let out = fm.to_yaml();

        // Both attendee links survive as quoted strings.
        assert!(out.contains("[[records/contacts/elena]]"), "got:\n{out}");
        assert!(out.contains("[[records/contacts/sarah]]"), "got:\n{out}");

        let reparsed = Frontmatter::parse(&out, Path::new("m.md")).unwrap();
        let fields = reparsed.link_fields();
        let attendees: Vec<&str> = fields
            .iter()
            .filter(|(k, _)| k == "attendees")
            .map(|(_, l)| l.target.as_str())
            .collect();
        assert_eq!(
            attendees,
            vec!["records/contacts/elena", "records/contacts/sarah"]
        );
        // The non-link nested sequence is preserved verbatim, not touched.
        assert_eq!(reparsed.extra.get("matrix"), fm.extra.get("matrix"));
    }

    // ── read_file / write_file round-trip ────────────────────────────────────

    #[test]
    fn write_then_read_roundtrips_and_preserves_body_verbatim() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("sources/emails/x.md");
        let body = "# Subject\n\nHello,\n\nSee [[records/contacts/sarah-chen]].\n";
        let mut fm = Frontmatter {
            type_: Some("email".into()),
            summary: Some("renewal note".into()),
            created: Some(DateTime::parse_from_rfc3339("2026-05-27T08:00:00-07:00").unwrap()),
            ..Default::default()
        };
        fm.extra
            .insert("from".into(), Value::String("elena@northstar.io".into()));

        write_file(&path, &fm, body).unwrap();

        let (read_fm, read_body) = read_file(&path).unwrap();
        assert_eq!(read_body, body, "body must be preserved byte-for-byte");
        assert_eq!(read_fm.type_.as_deref(), Some("email"));
        assert_eq!(read_fm.summary.as_deref(), Some("renewal note"));
        assert_eq!(
            read_fm.extra.get("from").and_then(|v| v.as_str()),
            Some("elena@northstar.io")
        );
        // The on-disk file starts with a fence and ends with the verbatim body.
        let raw = std::fs::read_to_string(&path).unwrap();
        assert!(raw.starts_with("---\n"));
        assert!(raw.ends_with(body));
    }

    #[test]
    fn roundtrip_modify_summary_then_write_changes_only_summary() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("records/contacts/sarah.md");
        let body = "Long-form operator notes about Sarah.\n";
        let fm = Frontmatter {
            type_: Some("contact".into()),
            summary: Some("old summary".into()),
            ..Default::default()
        };
        write_file(&path, &fm, body).unwrap();

        // Read → modify summary → write back.
        let (mut fm2, body2) = read_file(&path).unwrap();
        fm2.summary = Some("new summary".into());
        write_file(&path, &fm2, &body2).unwrap();

        let (fm3, body3) = read_file(&path).unwrap();
        assert_eq!(fm3.summary.as_deref(), Some("new summary"));
        assert_eq!(fm3.type_.as_deref(), Some("contact"));
        assert_eq!(body3, body, "body unchanged across the round-trip");
    }

    #[test]
    fn roundtrip_preserves_handwritten_unquoted_scalar_wiki_link_on_disk() {
        // End-to-end analog of `dbmd format` on the verbatim SPEC worked example:
        // a hand-written file carrying the canonical UNQUOTED scalar link
        // `company: [[records/companies/northstar]]`, read from disk then written
        // back unchanged. Before the fix this no-op re-emit rewrote the on-disk
        // value to the bracket-less block sequence `company:\n- - records/...`,
        // and every reader (validate/graph/backlinks) then lost the edge.
        let dir = tempdir().unwrap();
        let path = dir.path().join("records/contacts/sarah-chen.md");
        let file = "---\ntype: contact\nid: sarah-chen\nsummary: Director of Ops\ncompany: [[records/companies/northstar]]\n---\n# Sarah Chen\n\nNotes.\n";
        std::fs::create_dir_all(path.parent().unwrap()).unwrap();
        std::fs::write(&path, file).unwrap();

        // Read → write back unchanged (the canonical no-op re-emit).
        let (fm, body) = read_file(&path).unwrap();
        write_file(&path, &fm, &body).unwrap();

        // On-disk bytes still carry the bracketed link, never `- - records/...`.
        let raw = std::fs::read_to_string(&path).unwrap();
        assert!(
            raw.contains("[[records/companies/northstar]]"),
            "on-disk wiki-link brackets were destroyed; got:\n{raw}"
        );
        assert!(
            !raw.contains("- - "),
            "on-disk value became a nested block sequence; got:\n{raw}"
        );

        // And the edge is still readable after the round-trip.
        let (fm2, _) = read_file(&path).unwrap();
        let fields = fm2.link_fields();
        let links: Vec<(&str, &str)> = fields
            .iter()
            .map(|(k, l)| (k.as_str(), l.target.as_str()))
            .collect();
        assert_eq!(links, vec![("company", "records/companies/northstar")]);
    }

    #[test]
    fn write_file_does_not_leave_temp_files_behind() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("records/x.md");
        let fm = Frontmatter {
            type_: Some("note".into()),
            ..Default::default()
        };
        write_file(&path, &fm, "body\n").unwrap();
        // The directory should contain only the target file, no `.x.md.tmp.*`.
        let entries: Vec<String> = std::fs::read_dir(path.parent().unwrap())
            .unwrap()
            .map(|e| e.unwrap().file_name().to_string_lossy().into_owned())
            .collect();
        assert_eq!(entries, vec!["x.md".to_string()]);
    }

    // ── is_content_file ──────────────────────────────────────────────────────

    #[test]
    fn is_content_file_recognizes_layers_and_excludes_meta() {
        assert!(Frontmatter::is_content_file(Path::new(
            "sources/emails/2026-05-22.md"
        )));
        assert!(Frontmatter::is_content_file(Path::new(
            "records/contacts/sarah-chen.md"
        )));
        // A synthesis profile the agent authored lives under `records/` (the
        // old `wiki/` layer is gone, so a `wiki/...` path is NOT content).
        assert!(Frontmatter::is_content_file(Path::new(
            "records/profiles/sarah-chen.md"
        )));
        assert!(!Frontmatter::is_content_file(Path::new(
            "wiki/people/sarah-chen.md"
        )));
        // Absolute paths under a layer are still content.
        assert!(Frontmatter::is_content_file(Path::new(
            "/home/db/records/companies/northstar.md"
        )));
        // index.md at any level is meta.
        assert!(!Frontmatter::is_content_file(Path::new(
            "records/contacts/index.md"
        )));
        assert!(!Frontmatter::is_content_file(Path::new("index.md")));
        // Root meta files.
        assert!(!Frontmatter::is_content_file(Path::new("DB.md")));
        assert!(!Frontmatter::is_content_file(Path::new("log.md")));
    }

    // ── effective_id ─────────────────────────────────────────────────────────

    #[test]
    fn effective_id_prefers_explicit_then_derives_from_path() {
        let with_id = Frontmatter {
            id: Some("explicit-id".into()),
            ..Default::default()
        };
        assert_eq!(
            with_id.effective_id(Path::new("records/profiles/sarah-chen.md")),
            "explicit-id"
        );
        let no_id = Frontmatter::default();
        assert_eq!(
            no_id.effective_id(Path::new("records/profiles/sarah-chen.md")),
            "sarah-chen"
        );
    }

    // ── get / set ────────────────────────────────────────────────────────────

    #[test]
    fn set_routes_universal_and_custom_keys() {
        let mut fm = Frontmatter::default();
        fm.set("type", "contact").unwrap();
        fm.set("summary", "hi").unwrap();
        fm.set("company", "[[records/companies/northstar]]")
            .unwrap();
        assert_eq!(fm.type_.as_deref(), Some("contact"));
        assert_eq!(fm.summary.as_deref(), Some("hi"));
        // Custom key landed in extra, not a typed slot.
        assert_eq!(
            fm.extra.get("company").and_then(|v| v.as_str()),
            Some("[[records/companies/northstar]]")
        );
        // get reads from both typed fields and extra.
        assert_eq!(
            fm.get("type").and_then(|v| v.as_str().map(String::from)),
            Some("contact".into())
        );
        assert_eq!(
            fm.get("company").and_then(|v| v.as_str().map(String::from)),
            Some("[[records/companies/northstar]]".into())
        );
        assert!(fm.get("nonexistent").is_none());
    }

    #[test]
    fn set_timestamp_validates_rfc3339() {
        let mut fm = Frontmatter::default();
        fm.set("created", "2026-05-27T08:00:00-07:00").unwrap();
        assert!(fm.created.is_some());
        let err = fm.set("updated", "not-a-date").unwrap_err();
        assert!(matches!(err, ParseError::BadTimestamp { .. }));
    }

    // ── extract_wiki_links ───────────────────────────────────────────────────

    #[test]
    fn extract_wiki_links_flags_full_path_short_form_and_extension() {
        let body = "See [[records/contacts/sarah-chen]] and [[sarah-chen]].\nAlso [[records/profiles/sarah-chen.md|Sarah]].\n";
        let links = extract_wiki_links(body, Path::new("doc.md"));
        assert_eq!(links.len(), 3);

        // Full path, no extension, no display.
        assert_eq!(links[0].target, "records/contacts/sarah-chen");
        assert!(links[0].is_full_path);
        assert!(!links[0].has_md_extension);
        assert_eq!(links[0].display, None);
        assert_eq!(links[0].location.1, 1, "first link on line 1");

        // Short form: not a full path.
        assert_eq!(links[1].target, "sarah-chen");
        assert!(!links[1].is_full_path, "bare target is short-form");

        // Full path WITH .md extension and a display override on line 2.
        assert_eq!(links[2].target, "records/profiles/sarah-chen.md");
        assert!(links[2].is_full_path);
        assert!(links[2].has_md_extension);
        assert_eq!(links[2].display.as_deref(), Some("Sarah"));
        assert_eq!(links[2].location.1, 2);
    }

    #[test]
    fn extract_wiki_links_reports_1_based_column_counting_chars() {
        // A multi-byte prefix (é is 2 bytes) must not skew the char column.
        let body = "café [[records/x/y]]";
        let links = extract_wiki_links(body, Path::new("d.md"));
        assert_eq!(links.len(), 1);
        // "café " is 5 chars, so the `[[` starts at char column 6 (1-based).
        assert_eq!(links[0].location.2, 6);
    }

    #[test]
    fn extract_wiki_links_columns_are_correct_for_multiple_links_on_one_line() {
        // Locks the single-pass column cursor (the O(n²)→O(n) fix): each `[[`
        // reports the right 1-based CHAR column even with multi-byte prefixes and
        // several links per line.
        let body = "café [[a]] · [[records/x/y]] end";
        let links = extract_wiki_links(body, Path::new("d.md"));
        assert_eq!(links.len(), 2);
        // "café " = 5 chars → first `[[` at col 6.
        assert_eq!(links[0].location.2, 6);
        // "café [[a]] · " = 5 + 5 (`[[a]]`) + 3 (` · `, `·` is 1 char) = 13 chars
        // → second `[[` at col 14.
        assert_eq!(links[1].location.2, 14);
    }

    #[test]
    fn extract_wiki_links_ignores_a_lone_path_without_brackets() {
        let links = extract_wiki_links(
            "records/contacts/sarah-chen is not a link",
            Path::new("d.md"),
        );
        assert!(links.is_empty());
    }

    // ── extract_markdown_links ───────────────────────────────────────────────

    #[test]
    fn extract_markdown_links_captures_external_and_not_wiki_links() {
        let body =
            "See [the thread](https://x.com/a) and [[records/contacts/sarah-chen]] internally.\n";
        let md = extract_markdown_links(body, Path::new("d.md"));
        assert_eq!(
            md.len(),
            1,
            "wiki-link must not be captured as a markdown link"
        );
        assert_eq!(md[0].text, "the thread");
        assert_eq!(md[0].url, "https://x.com/a");
        assert_eq!(md[0].location.1, 1);

        // And the wiki-link extractor must not pick up the markdown link.
        let wl = extract_wiki_links(body, Path::new("d.md"));
        assert_eq!(wl.len(), 1);
        assert_eq!(wl[0].target, "records/contacts/sarah-chen");
    }

    // ── link_fields ──────────────────────────────────────────────────────────

    #[test]
    fn link_fields_extracts_scalar_list_and_summary_links() {
        // The canonical list form quotes each item so YAML parses it as clean
        // strings; a scalar field may be quoted OR written in the canonical
        // unquoted inline form `company: [[x]]` (SPEC § Linking).
        let yaml = "type: meeting\nsummary: with [[records/contacts/elena]]\ncompany: \"[[records/companies/northstar]]\"\nattendees:\n  - \"[[records/contacts/elena]]\"\n  - \"[[records/contacts/sarah]]\"\nnotes: just plain text";
        let fm = Frontmatter::parse(yaml, Path::new("m.md")).unwrap();
        // Sanity: company really did parse as a scalar string here.
        assert!(fm.extra.get("company").and_then(|v| v.as_str()).is_some());
        let fields = fm.link_fields();

        // company (scalar) once, with the right target.
        let company: Vec<&str> = fields
            .iter()
            .filter(|(k, _)| k == "company")
            .map(|(_, l)| l.target.as_str())
            .collect();
        assert_eq!(company, vec!["records/companies/northstar"]);
        // attendees (block list) twice.
        let attendees: Vec<&str> = fields
            .iter()
            .filter(|(k, _)| k == "attendees")
            .map(|(_, l)| l.target.as_str())
            .collect();
        assert_eq!(
            attendees,
            vec!["records/contacts/elena", "records/contacts/sarah"]
        );
        // summary link surfaced.
        assert_eq!(fields.iter().filter(|(k, _)| k == "summary").count(), 1);
        // Plain-text field is not a link.
        assert_eq!(fields.iter().filter(|(k, _)| k == "notes").count(), 0);
    }

    #[test]
    fn link_fields_surfaces_canonical_unquoted_scalar_link() {
        // Regression: the canonical scalar wiki-link form is the *unquoted*
        // inline `company: [[records/companies/northstar]]` (SPEC § Linking).
        // YAML parses `[[x]]` as a flow-list-in-a-list (`Seq[Seq[String]]`), so
        // a naive `as_str()`-only walk drops it. link_fields() must still
        // surface exactly one link with the correct target.
        let yaml = "type: meeting\ncompany: [[records/companies/northstar]]";
        let fm = Frontmatter::parse(yaml, Path::new("m.md")).unwrap();
        // Sanity: `parse` disambiguates the inline-link source form at read time,
        // storing it as the canonical scalar `String("[[x]]")` (so a genuine
        // `Seq[Seq[String]]` 2D array is never collapsed/retyped). link_fields()
        // reads either spelling back as the same link.
        assert_eq!(
            fm.extra.get("company").and_then(|v| v.as_str()),
            Some("[[records/companies/northstar]]")
        );

        let fields = fm.link_fields();
        let links: Vec<(&str, &str, Option<&str>)> = fields
            .iter()
            .map(|(k, l)| (k.as_str(), l.target.as_str(), l.display.as_deref()))
            .collect();
        assert_eq!(
            links,
            vec![("company", "records/companies/northstar", None)]
        );

        // The `|display` segment survives the unquoted inline form too.
        let fm2 = Frontmatter::parse(
            "type: meeting\ncompany: [[records/companies/northstar|Northstar]]",
            Path::new("m.md"),
        )
        .unwrap();
        let f2 = fm2.link_fields();
        assert_eq!(f2.len(), 1);
        assert_eq!(f2[0].0, "company");
        assert_eq!(f2[0].1.target, "records/companies/northstar");
        assert_eq!(f2[0].1.display.as_deref(), Some("Northstar"));
    }

    #[test]
    fn link_fields_ignores_plain_one_item_flow_list() {
        // A plain one-item flow list `aliases: [foo]` parses to `Seq[String]`
        // — one nesting level shallower than an unquoted `[[foo]]` — and must
        // NOT be mistaken for a wiki-link.
        let yaml = "type: contact\naliases: [foo]";
        let fm = Frontmatter::parse(yaml, Path::new("c.md")).unwrap();
        assert_eq!(fm.link_fields(), Vec::new());
    }

    // ── detect_flow_form_link_lists ──────────────────────────────────────────

    #[test]
    fn detect_flow_form_flags_list_misencodings_not_scalars() {
        // The flow-form list mis-encoding (triple-nested) IS flagged; a scalar
        // inline wiki-link (double-nested) is NOT.
        let bad = "attendees: [[[records/x]], [[records/y]]]\nscalar_inline: [[records/z]]";
        let flagged = detect_flow_form_link_lists(bad);
        assert_eq!(flagged, vec!["attendees".to_string()]);

        // An UNquoted block list is also a mis-encoding (parses triple-nested).
        let unquoted_block = "attendees:\n  - [[records/x]]\n  - [[records/y]]";
        assert_eq!(
            detect_flow_form_link_lists(unquoted_block),
            vec!["attendees".to_string()]
        );

        // The canonical QUOTED block form parses to clean strings — NOT flagged.
        let good = "attendees:\n  - \"[[records/x]]\"\n  - \"[[records/y]]\"";
        assert!(detect_flow_form_link_lists(good).is_empty());

        // A plain scalar list of strings is not flagged.
        let plain = "tags: [a, b, c]";
        assert!(detect_flow_form_link_lists(plain).is_empty());
    }

    // ── extract_sections ─────────────────────────────────────────────────────

    #[test]
    fn extract_sections_levels_nesting_and_boundaries() {
        let body = "intro text\n## First\nalpha\n### Sub\nbeta\n## Second\ngamma\n";
        let secs = extract_sections(body);
        let headings: Vec<(&str, u8)> =
            secs.iter().map(|s| (s.heading.as_str(), s.level)).collect();
        assert_eq!(headings, vec![("First", 2), ("Sub", 3), ("Second", 2)]);

        // "First" (H2) body extends through its H3 child, stopping at "Second".
        let first = &secs[0];
        assert!(first.body.contains("alpha"));
        assert!(first.body.contains("### Sub"));
        assert!(first.body.contains("beta"));
        assert!(!first.body.contains("Second"));

        // "Sub" (H3) stops at the next equal-or-shallower heading ("Second").
        let sub = &secs[1];
        assert!(sub.body.contains("beta"));
        assert!(!sub.body.contains("gamma"));

        // 1-based line numbers within the body.
        assert_eq!(first.line, 2);
        assert_eq!(secs[2].line, 6);
    }

    #[test]
    fn extract_sections_ignores_headings_in_fenced_code() {
        let body = "## Real\n```\n## Fake heading in code\n```\nafter\n";
        let secs = extract_sections(body);
        assert_eq!(secs.len(), 1);
        assert_eq!(secs[0].heading, "Real");
        // The fenced "## Fake" is part of Real's body, not its own section.
        assert!(secs[0].body.contains("## Fake heading in code"));
    }

    // ── parse_field_spec ─────────────────────────────────────────────────────

    #[test]
    fn parse_field_spec_required_and_shape() {
        let f = parse_field_spec("- email (required, email)");
        assert_eq!(f.name, "email");
        assert!(f.required);
        assert_eq!(f.shape, Some(Shape::Email));
        assert!(f.unknown_modifiers.is_empty());
    }

    #[test]
    fn parse_field_spec_link_prefix_strips_trailing_slash() {
        let f = parse_field_spec("- company (required, link to records/companies/)");
        assert!(f.required);
        assert_eq!(f.link_prefix, Some(PathBuf::from("records/companies")));
        assert_eq!(f.shape, None);
    }

    #[test]
    fn parse_field_spec_default_preserves_case_and_value() {
        let f = parse_field_spec("- currency (default USD)");
        assert_eq!(f.name, "currency");
        assert_eq!(f.default, Some(Value::String("USD".into())));
    }

    #[test]
    fn parse_field_spec_enum_captures_comma_list_as_last_modifier() {
        let f = parse_field_spec("- status (required, enum: open, closed, pending)");
        assert!(f.required);
        assert_eq!(
            f.enum_values,
            Some(vec![
                "open".to_string(),
                "closed".to_string(),
                "pending".to_string()
            ])
        );
    }

    #[test]
    fn parse_field_spec_bare_enum_keyword_is_not_itself_a_value() {
        // `enum` with no colon: the values are the remaining tokens; the keyword
        // itself must NOT leak in as an allowed value.
        let f = parse_field_spec("- status (required, enum, open, closed)");
        assert!(f.required);
        assert_eq!(
            f.enum_values,
            Some(vec!["open".to_string(), "closed".to_string()])
        );
    }

    #[test]
    fn parse_field_spec_unknown_modifier_is_captured_not_errored() {
        let f = parse_field_spec("- weird (required, frobnicate, string)");
        assert!(f.required);
        assert_eq!(f.shape, Some(Shape::String));
        assert_eq!(f.unknown_modifiers, vec!["frobnicate".to_string()]);
    }

    #[test]
    fn parse_field_spec_no_parens_is_freeform_optional() {
        let f = parse_field_spec("- nickname");
        assert_eq!(f.name, "nickname");
        assert!(!f.required);
        assert_eq!(f.shape, None);
        assert!(f.link_prefix.is_none());
        assert!(f.enum_values.is_none());
        assert!(f.unknown_modifiers.is_empty());
    }

    // ── parse_schema_bullet (directives) ─────────────────────────────────────

    #[test]
    fn schema_bullet_unique_single_field() {
        match parse_schema_bullet("- unique: email") {
            SchemaBullet::Unique(fields) => assert_eq!(fields, vec!["email".to_string()]),
            other => panic!("expected Unique, got {other:?}"),
        }
    }

    #[test]
    fn schema_bullet_unique_compound_trims_and_splits() {
        match parse_schema_bullet("- unique: date, amount , vendor") {
            SchemaBullet::Unique(fields) => assert_eq!(
                fields,
                vec![
                    "date".to_string(),
                    "amount".to_string(),
                    "vendor".to_string()
                ]
            ),
            other => panic!("expected Unique, got {other:?}"),
        }
    }

    #[test]
    fn schema_bullet_summary_template_keeps_braces_and_inner_colons() {
        match parse_schema_bullet("- summary_template: {role} at {company} (x: y)") {
            SchemaBullet::SummaryTemplate(t) => assert_eq!(t, "{role} at {company} (x: y)"),
            other => panic!("expected SummaryTemplate, got {other:?}"),
        }
    }

    #[test]
    fn schema_bullet_field_with_enum_modifier_is_not_a_directive() {
        // A field whose modifiers contain a colon (`enum:`) parses as a field, not
        // a directive — its head has a `(` before any `:`.
        match parse_schema_bullet("- status (enum: open, closed)") {
            SchemaBullet::Field(f) => {
                assert_eq!(f.name, "status");
                assert_eq!(
                    f.enum_values,
                    Some(vec!["open".to_string(), "closed".to_string()])
                );
            }
            other => panic!("expected Field, got {other:?}"),
        }
    }

    #[test]
    fn parse_db_md_schema_captures_unique_and_summary_template() {
        let db = "---\ntype: db-md\nscope: x\nowner: y\n---\n\n## Schemas\n\n### contact\n- email (required, email)\n- unique: email\n- summary_template: {role} at {company}\n";
        let config = parse_db_md(db, Path::new("DB.md")).unwrap();
        let s = config.schemas.get("contact").expect("contact schema");
        assert_eq!(s.fields.len(), 1, "directives are not parsed as fields");
        assert_eq!(s.unique_keys, vec![vec!["email".to_string()]]);
        assert_eq!(s.summary_template.as_deref(), Some("{role} at {company}"));
    }

    #[test]
    fn schema_bullet_shard_directive_parses_values() {
        assert!(matches!(
            parse_schema_bullet("- shard: by-date"),
            SchemaBullet::Shard(Some(true))
        ));
        assert!(matches!(
            parse_schema_bullet("- shard: flat"),
            SchemaBullet::Shard(Some(false))
        ));
        // An unrecognized value is ignored (None), like an unknown modifier.
        assert!(matches!(
            parse_schema_bullet("- shard: weekly"),
            SchemaBullet::Shard(None)
        ));
        // A field whose name has a `(` before any `:` is still a field — the same
        // guard that keeps `- status (enum: a, b)` a field, not a directive.
        assert!(matches!(
            parse_schema_bullet("- shardiness (string)"),
            SchemaBullet::Field(_)
        ));
    }

    #[test]
    fn parse_db_md_schema_captures_shard_directive() {
        let db = "---\ntype: db-md\nscope: x\nowner: y\n---\n\n## Schemas\n\n### shipment\n- carrier (string)\n- shard: by-date\n\n### contact\n- shard: flat\n";
        let config = parse_db_md(db, Path::new("DB.md")).unwrap();
        let shipment = config.schemas.get("shipment").expect("shipment schema");
        assert_eq!(shipment.shard, Some(true));
        assert_eq!(
            shipment.fields.len(),
            1,
            "`shard:` is a directive, not a field"
        );
        assert_eq!(config.schemas.get("contact").unwrap().shard, Some(false));
    }

    // ── parse_db_md ──────────────────────────────────────────────────────────

    const CANONICAL_DB_MD: &str = "---\ntype: db-md\nscope: company\nowner: Sarah Chen\n---\n\n# Acme operations knowledge base\n\nCompany-scale institutional memory for Acme.\n\n## Agent instructions\n\nPrioritize creating `contact` records from new-sender emails. Use British English.\n\n## Policies\n\n### Frozen pages\n- `records/decisions/2026-q1-strategy.md` — finalized, do not modify.\n- `records/synthesis/2026-annual-plan.md` — signed-off plan.\n\n### Ignored types\n- `test`, `temp` — read but never synthesize.\n\n## Schemas\n\n### contact\n- name (required)\n- email (required, email)\n- company (required, link to records/companies/)\n- role (string)\n\n### expense\n- date (required, date)\n- amount (required)\n- currency (default USD)\n";

    #[test]
    fn parse_db_md_extracts_all_canonical_sections() {
        let config = parse_db_md(CANONICAL_DB_MD, Path::new("DB.md")).unwrap();

        // Agent instructions: free-form prose, heading line stripped.
        let ai = config
            .agent_instructions
            .expect("agent instructions present");
        assert!(ai.starts_with("Prioritize creating"));
        assert!(!ai.contains("## Agent instructions"));

        // Frozen pages: paths extracted from backticked bullets, comments dropped.
        assert_eq!(
            config.frozen_pages,
            vec![
                PathBuf::from("records/decisions/2026-q1-strategy.md"),
                PathBuf::from("records/synthesis/2026-annual-plan.md"),
            ]
        );

        // Ignored types: comma list, backticks/comment stripped.
        assert_eq!(
            config.ignored_types,
            vec!["test".to_string(), "temp".to_string()]
        );

        // Schemas: two types, each with its fields in source order.
        assert_eq!(config.schemas.len(), 2);
        let contact = config.schemas.get("contact").expect("contact schema");
        let names: Vec<&str> = contact.fields.iter().map(|f| f.name.as_str()).collect();
        assert_eq!(names, vec!["name", "email", "company", "role"]);
        assert!(contact.fields[0].required); // name
        assert_eq!(contact.fields[1].shape, Some(Shape::Email)); // email
        assert_eq!(
            contact.fields[2].link_prefix,
            Some(PathBuf::from("records/companies"))
        ); // company

        let expense = config.schemas.get("expense").expect("expense schema");
        let cur = expense
            .fields
            .iter()
            .find(|f| f.name == "currency")
            .unwrap();
        assert_eq!(cur.default, Some(Value::String("USD".into())));
    }

    #[test]
    fn parse_db_md_handles_malformed_and_unknown_modifiers() {
        // corpus-b shape: a `## Schemas` section with a malformed bullet, an
        // unknown modifier, and bullets that appear with NO `### <type>`
        // heading (so they belong to no schema and are dropped).
        let text = "---\ntype: db-md\n---\n\n## Schemas\n- orphan (required)\n\n### ticket\n- priority (required, mystery, enum: low, high)\n- broken (\n";
        let config = parse_db_md(text, Path::new("DB.md")).unwrap();

        // The orphan bullet under `## Schemas` with no `### type` heading is not
        // captured as a schema.
        assert_eq!(config.schemas.len(), 1);
        let ticket = config.schemas.get("ticket").expect("ticket schema");
        assert_eq!(ticket.fields.len(), 2);

        let priority = &ticket.fields[0];
        assert!(priority.required);
        assert_eq!(priority.unknown_modifiers, vec!["mystery".to_string()]);
        assert_eq!(
            priority.enum_values,
            Some(vec!["low".to_string(), "high".to_string()])
        );

        // A bullet with an unclosed paren still yields a usable name.
        let broken = &ticket.fields[1];
        assert_eq!(broken.name, "broken");
    }

    #[test]
    fn parse_db_md_missing_frontmatter_errors() {
        let text = "# No frontmatter\n\n## Agent instructions\nhi\n";
        let err = parse_db_md(text, Path::new("DB.md")).unwrap_err();
        assert!(matches!(err, ParseError::MissingFrontmatter { .. }));
    }

    #[test]
    fn parse_db_md_absent_sections_default_empty() {
        let text = "---\ntype: db-md\n---\n\n# Title only\n";
        let config = parse_db_md(text, Path::new("DB.md")).unwrap();
        assert_eq!(config, Config::default());
    }

    // ── fm set / --fm list-valued link fields (meeting.attendees & friends) ──

    /// `Frontmatter::set` is the value path every write surface (`fm set`,
    /// `write --fm`) funnels through. A list-of-wiki-links value (the SPEC's
    /// `meeting.attendees` shape) must serialize as a YAML **block sequence** of
    /// quoted links — readable back by [`links_in_field_value`] and accepted by
    /// `dbmd validate` — never the flow-form scalar string that trips
    /// `WIKI_LINK_FLOW_FORM_LIST`. Both the unquoted (`[[[a]], [[b]]]`) and
    /// quoted (`["[[a]]", "[[b]]"]`) spellings an agent types must normalize.
    #[test]
    fn set_list_of_wiki_links_becomes_block_sequence_both_spellings() {
        for value in [
            "[[[records/contacts/a]], [[records/contacts/b]]]",
            r#"["[[records/contacts/a]]", "[[records/contacts/b]]"]"#,
        ] {
            let mut fm = Frontmatter::default();
            fm.set("attendees", value).unwrap();

            // Stored as a 2-element sequence of clean quoted links.
            let stored = fm.extra.get("attendees").expect("attendees set");
            let Value::Sequence(items) = stored else {
                panic!("attendees must be a Sequence, got {stored:?} for input {value}");
            };
            assert_eq!(items.len(), 2, "input {value}");
            assert_eq!(items[0], Value::String("[[records/contacts/a]]".into()));
            assert_eq!(items[1], Value::String("[[records/contacts/b]]".into()));

            // The edge enumerator reads exactly the two links back (no stray
            // bracket targets, the flow-form-string symptom).
            let links: Vec<_> = links_in_field_value(stored)
                .into_iter()
                .map(|l| l.target)
                .collect();
            assert_eq!(
                links,
                vec!["records/contacts/a", "records/contacts/b"],
                "input {value}"
            );

            // And the canonical writer renders it block-style, not as a scalar.
            let yaml = fm.to_yaml();
            assert!(
                yaml.contains("attendees:\n"),
                "expected block list in:\n{yaml}"
            );
            assert!(
                !yaml.contains("attendees: '[["),
                "must not be a flow-form scalar string in:\n{yaml}"
            );
        }
    }

    /// A *single* inline wiki-link stays a scalar string (renders inline
    /// `field: [[x]]`), and a single link must never be widened to a one-item
    /// list — preserving the common `contact.company` / `expense.vendor` shape.
    #[test]
    fn set_single_inline_wiki_link_stays_scalar() {
        let mut fm = Frontmatter::default();
        fm.set("company", "[[records/companies/tideform]]").unwrap();
        assert_eq!(
            fm.extra.get("company"),
            Some(&Value::String("[[records/companies/tideform]]".into())),
        );
        // Still recognized as one link.
        let links: Vec<_> = links_in_field_value(fm.extra.get("company").unwrap())
            .into_iter()
            .map(|l| l.target)
            .collect();
        assert_eq!(links, vec!["records/companies/tideform"]);
    }

    /// Plain text and a non-link flow list are left as verbatim scalar strings —
    /// the list normalization only triggers when every item is a clean wiki-link.
    #[test]
    fn set_non_link_values_stay_scalar_strings() {
        let mut fm = Frontmatter::default();
        fm.set("location", "Video call (remote)").unwrap();
        assert_eq!(
            fm.extra.get("location"),
            Some(&Value::String("Video call (remote)".into())),
        );

        // A flow list whose items are NOT wiki-links must not be reinterpreted as
        // a link sequence; it stays the scalar string the agent passed.
        fm.set("note", "[draft, wip]").unwrap();
        assert_eq!(
            fm.extra.get("note"),
            Some(&Value::String("[draft, wip]".into()))
        );
    }

    // ── Regression: non-string YAML keys round-trip (no Rust Debug corruption) ─

    #[test]
    fn regression_non_string_yaml_keys_keep_their_text_on_round_trip() {
        // A numeric/bool/null/float frontmatter key is valid YAML and must NOT be
        // rewritten to its Rust `Debug` form (`Number(2026)`, `Bool(true)`,
        // `'Null'`). After the fix the key text survives (the key narrows to a
        // string-typed key, but the operator's data is no longer corrupted).
        let yaml = "type: note\n2026: planning notes\ntrue: yes-key\n3.14: f\n";
        let fm = Frontmatter::parse(yaml, Path::new("x.md")).unwrap();
        // Keys are stored as their scalar text, not the Debug string.
        assert!(fm.extra.contains_key("2026"), "numeric key text lost");
        assert!(fm.extra.contains_key("true"), "bool key text lost");
        assert!(fm.extra.contains_key("3.14"), "float key text lost");
        assert!(!fm.extra.keys().any(|k| k.starts_with("Number(")));
        assert!(!fm.extra.keys().any(|k| k.starts_with("Bool(")));

        // And a re-emit never produces the Debug forms on disk.
        let out = fm.to_yaml();
        assert!(!out.contains("Number("), "Debug-form key emitted:\n{out}");
        assert!(!out.contains("Bool("), "Debug-form key emitted:\n{out}");
        // The key text is still present (quoted, since it now reads as a string).
        assert!(out.contains("2026"), "numeric key dropped:\n{out}");
        assert!(out.contains("planning notes"), "value dropped:\n{out}");
    }

    // ── Regression: universal-key sequence/mapping values are preserved (#2) ───

    #[test]
    fn regression_universal_key_non_scalar_value_is_preserved_not_deleted() {
        // A universal key carrying a sequence/mapping (`status: [active, draft]`)
        // is not a valid scalar for that field. Before the fix, the matched arm
        // consumed-and-dropped it (scalar_string -> None) and `to_yaml` then
        // omitted the field — `dbmd format` silently DELETED it. It must now pass
        // through `extra` and re-emit verbatim.
        let yaml = "type: note\nstatus:\n  - active\n  - draft\nsummary:\n  a: 1\n  b: 2\n";
        let fm = Frontmatter::parse(yaml, Path::new("x.md")).unwrap();
        // The typed accessors stay None (no valid scalar), but the data lives in
        // extra so nothing is lost.
        assert!(fm.status.is_none());
        assert!(fm.summary.is_none());
        assert!(fm.extra.contains_key("status"), "status value destroyed");
        assert!(fm.extra.contains_key("summary"), "summary value destroyed");

        // A re-emit keeps both fields' data on disk.
        let out = fm.to_yaml();
        assert!(out.contains("status"), "status deleted on re-emit:\n{out}");
        assert!(out.contains("active"), "status items deleted:\n{out}");
        assert!(
            out.contains("summary"),
            "summary deleted on re-emit:\n{out}"
        );

        // Round-trips as a fixed point — repeated curator-loop writes don't lose
        // the data.
        let reparsed = Frontmatter::parse(&out, Path::new("x.md")).unwrap();
        assert!(reparsed.extra.contains_key("status"));
        assert!(reparsed.extra.contains_key("summary"));
    }

    // ── Regression: non-scalar tags items don't erase the tags field (#5) ──────

    #[test]
    fn regression_non_scalar_tags_value_is_preserved_not_erased() {
        // `tags: [[vip]]` (an authoring slip — wiki-link brackets around a tag)
        // parses to a nested sequence; before the fix `parse_tags` filtered the
        // non-scalar item out and `to_yaml` then omitted the now-empty tags vec,
        // silently DELETING the tags line. It must now survive the re-emit (the
        // key data is preserved; the field is never dropped).
        let yaml = "type: note\ntags: [[vip]]\n";
        let fm = Frontmatter::parse(yaml, Path::new("x.md")).unwrap();
        // The typed tags vec is empty (no clean scalar list), but the raw value
        // is preserved in extra so nothing is destroyed.
        assert!(fm.tags.is_empty());
        assert!(fm.extra.contains_key("tags"), "tags value destroyed");

        let out = fm.to_yaml();
        assert!(out.contains("tags"), "tags deleted on re-emit:\n{out}");
        // The `vip` text survives on disk in some form (never erased).
        assert!(out.contains("vip"), "tag content erased:\n{out}");

        // A clean tag list still parses to the typed vec (not regressed).
        let clean =
            Frontmatter::parse("type: note\ntags: [vip, renewal]\n", Path::new("x.md")).unwrap();
        assert_eq!(clean.tags, vec!["vip".to_string(), "renewal".to_string()]);
        assert!(!clean.extra.contains_key("tags"));
    }

    // ── Regression: plain nested string lists are NOT fabricated into links (#3) ─

    #[test]
    fn regression_plain_nested_string_list_is_not_turned_into_wiki_links() {
        // `groups: [[alpha], [beta]]` is the data [["alpha"],["beta"]] — an
        // unknown nested string list that must pass through verbatim. Before the
        // fix, canonicalize_extra_value fabricated `- '[[alpha]]'` / `- '[[beta]]'`
        // (short-form links the tool then flagged), changing the field's type.
        let yaml = "type: note\ngroups: [[alpha], [beta]]\n";
        let fm = Frontmatter::parse(yaml, Path::new("x.md")).unwrap();
        let before = fm.extra.get("groups").cloned();

        let out = fm.to_yaml();
        // No fabricated wiki-link brackets in the emitted YAML.
        assert!(!out.contains("[[alpha]]"), "fabricated a wiki-link:\n{out}");
        assert!(!out.contains("[[beta]]"), "fabricated a wiki-link:\n{out}");

        // The value is unchanged across the canonical re-emit.
        let reparsed = Frontmatter::parse(&out, Path::new("x.md")).unwrap();
        assert_eq!(
            reparsed.extra.get("groups"),
            before.as_ref(),
            "nested string list mutated by canonicalize_extra_value"
        );
        // And it surfaces no links.
        assert!(reparsed.link_fields().is_empty());
    }

    #[test]
    fn regression_genuine_nested_array_is_not_retyped_to_scalar_string() {
        // BUG: `dbmd format` silently retyped a genuine 2D array
        //     matrix:
        //     - - cell
        // (data `[["cell"]]`) into the scalar string `matrix: '[[cell]]'`. The
        // root cause is the irreducible YAML ambiguity: serde parses BOTH the
        // inline scalar wiki-link `field: [[x]]` AND the block nested-seq
        // `field:`\n`- - x` to the identical `Seq[Seq[String]]`. The old
        // `canonicalize_extra_value` collapsed every one-element `Seq[Seq[String]]`
        // to a string, destroying the array. The fix resolves the inline-link
        // case from the SOURCE text at parse time and leaves a genuine block
        // array verbatim.
        let yaml = "type: note\nsummary: nested\nmatrix:\n- - cell\n";
        let fm = Frontmatter::parse(yaml, Path::new("nested.md")).unwrap();

        // The block source form stays a nested sequence, NOT a string — the
        // inline-link disambiguation only fires for source written `key: [[x]]`.
        let stored = fm.extra.get("matrix").expect("matrix preserved");
        assert!(
            matches!(stored, Value::Sequence(items)
                if items.len() == 1 && matches!(&items[0], Value::Sequence(_))),
            "genuine 2D array was retyped at parse time; got {stored:?}"
        );

        let out = fm.to_yaml();
        // Emit must keep the array (a block nested sequence), never the bogus
        // scalar string `'[[cell]]'`.
        assert!(
            !out.contains("'[[cell]]'") && !out.contains("[[cell]]"),
            "genuine nested array retyped to a scalar wiki-link string; got:\n{out}"
        );
        assert!(
            out.contains("- - cell"),
            "nested array lost its 2D shape on emit; got:\n{out}"
        );

        // Full round-trip: re-parsing the emitted YAML yields the identical value
        // — the file's bytes are preserved, which is what BUG 2 was about. (The
        // read-side `link_fields` still treats a one-element `Seq[Seq[String]]` as
        // the inline-link shape it is indistinguishable from on disk; that is the
        // same irreducible ambiguity and is out of scope here — the fix's job is
        // that `format` no longer silently RETYPES the array to a string.)
        let reparsed = Frontmatter::parse(&out, Path::new("nested.md")).unwrap();
        assert_eq!(
            reparsed.extra.get("matrix"),
            fm.extra.get("matrix"),
            "nested array did not round-trip through format"
        );
        // The stored value is still a sequence after round-trip (never a string).
        assert!(
            matches!(reparsed.extra.get("matrix"), Some(Value::Sequence(_))),
            "nested array became a non-sequence after round-trip"
        );
    }

    #[test]
    fn inline_scalar_wiki_link_still_round_trips_after_nested_array_fix() {
        // The companion guarantee to the test above: the SPEC-canonical inline
        // scalar wiki-link `field: [[x]]` (SPEC.md:383) must still format to a
        // canonical inline `[[x]]` that round-trips and surfaces as one link —
        // the nested-array fix must not regress it.
        let yaml = "type: contact\ncompany: [[records/companies/northstar]]\n";
        let fm = Frontmatter::parse(yaml, Path::new("c.md")).unwrap();
        // Disambiguated at parse time to the canonical scalar string.
        assert_eq!(
            fm.extra.get("company").and_then(|v| v.as_str()),
            Some("[[records/companies/northstar]]")
        );

        let out = fm.to_yaml();
        assert!(
            out.contains("[[records/companies/northstar]]") && !out.contains("- - "),
            "inline wiki-link not canonical after the nested-array fix; got:\n{out}"
        );

        let reparsed = Frontmatter::parse(&out, Path::new("c.md")).unwrap();
        let fields = reparsed.link_fields();
        let links: Vec<(&str, &str)> = fields
            .iter()
            .map(|(k, l)| (k.as_str(), l.target.as_str()))
            .collect();
        assert_eq!(links, vec![("company", "records/companies/northstar")]);
        // Idempotent across repeated curator-loop writes.
        assert_eq!(
            reparsed.to_yaml(),
            out,
            "inline link is not a format fixed point"
        );
    }

    // ── Regression: fence-line trailing whitespace is tolerated (#4) ───────────

    #[test]
    fn regression_split_frontmatter_tolerates_trailing_whitespace_on_fences() {
        // A fence written `--- ` (trailing space — invisible in editors) is
        // indexed/validated clean by index.rs/validate.rs (both use `trim_end()`)
        // but, before the fix, hard-failed every read/edit surface routed through
        // `split_frontmatter`. All three must now agree.
        let text = "--- \ntype: note\nsummary: x\n---\t\nbody\n";
        let parsed = split_frontmatter(text, Path::new("f.md")).unwrap();
        assert_eq!(parsed.frontmatter_yaml, "type: note\nsummary: x\n");
        assert_eq!(parsed.body, "body\n");

        // End to end through read_file's parse.
        let fm = Frontmatter::parse(&parsed.frontmatter_yaml, Path::new("f.md")).unwrap();
        assert_eq!(fm.type_.as_deref(), Some("note"));
    }

    // ── Regression: CommonMark trailing-'#' heading rule (#6) ──────────────────

    #[test]
    fn regression_heading_text_keeps_abutting_hash_drops_closing_sequence() {
        // `## C#` → `C#` (the `#` abuts content, not a closing sequence).
        assert_eq!(heading_text("## C#", 2), "C#");
        assert_eq!(heading_text("## F#", 2), "F#");
        assert_eq!(heading_text("## issue-123#", 2), "issue-123#");
        // A genuine ATX closing sequence (space before the `#` run) is dropped.
        assert_eq!(heading_text("## Title ##", 2), "Title");
        assert_eq!(heading_text("## Title #", 2), "Title");
        // All-hashes content collapses to empty.
        assert_eq!(heading_text("## ##", 2), "");
        // No trailing hashes — unchanged.
        assert_eq!(heading_text("## Plain", 2), "Plain");
    }

    #[test]
    fn regression_extract_sections_keeps_csharp_heading_and_schema_type_binds() {
        // `dbmd sections` must report `C#`, not `C`.
        let secs = extract_sections("## C#\nbody\n");
        assert_eq!(secs.len(), 1);
        assert_eq!(secs[0].heading, "C#");

        // And a `### c#` schema must register under `c#`, not `c`.
        let db = "---\ntype: db-md\n---\n\n## Schemas\n\n### c#\n- name (required)\n";
        let config = parse_db_md(db, Path::new("DB.md")).unwrap();
        assert!(
            config.schemas.contains_key("c#"),
            "schema bound to wrong key"
        );
        assert!(!config.schemas.contains_key("c"));
    }

    // ── Regression: section line numbers offset by the frontmatter block (#7) ──

    #[test]
    fn regression_extract_sections_in_file_reports_source_line_numbers() {
        // A heading on file line 6 (after a 4-line frontmatter block + 1 body
        // line) must be reported as L6, not the body-relative L2.
        let text = "---\ntype: note\nsummary: x\n---\nbody line\n## Heading\nmore\n";
        let secs = extract_sections_in_file(text);
        assert_eq!(secs.len(), 1);
        assert_eq!(secs[0].heading, "Heading");
        assert_eq!(secs[0].line, 6, "section line not offset by frontmatter");

        // The body-relative helper is unchanged (validate relies on that frame).
        let body_secs = extract_sections("body line\n## Heading\nmore\n");
        assert_eq!(body_secs[0].line, 2);

        // No frontmatter: whole text is body, no offset.
        let plain = extract_sections_in_file("## Top\nx\n## Next\n");
        assert_eq!(plain[0].line, 1);
        assert_eq!(plain[1].line, 3);
    }

    // ── Regression: colon-form schema field bullet parses modifiers (#8) ───────

    #[test]
    fn regression_colon_form_field_bullet_parses_modifiers() {
        // `- title: string, required` is the natural mis-spelling of
        // `- title (string, required)`; before the fix the whole text became the
        // field name and every modifier was silently lost.
        let f = parse_field_spec("- title: string, required");
        assert_eq!(f.name, "title");
        assert!(f.required, "required modifier lost on colon-form");
        assert_eq!(f.shape, Some(Shape::String));

        // Through the schema-bullet classifier (the real path), it is a Field.
        match parse_schema_bullet("- title: string, required") {
            SchemaBullet::Field(f) => {
                assert_eq!(f.name, "title");
                assert!(f.required);
                assert_eq!(f.shape, Some(Shape::String));
            }
            other => panic!("expected Field, got {other:?}"),
        }

        // A paren form whose modifiers contain a colon still parses by parens.
        let g = parse_field_spec("- status (enum: open, closed)");
        assert_eq!(g.name, "status");
        assert_eq!(
            g.enum_values,
            Some(vec!["open".to_string(), "closed".to_string()])
        );
    }

    // ── Regression: comma inside a `default` value is preserved (#9) ───────────

    #[test]
    fn regression_default_value_preserves_internal_commas() {
        let f = parse_field_spec("- title (default Director, Operations)");
        assert_eq!(
            f.default,
            Some(Value::String("Director, Operations".into())),
            "comma-bearing default truncated"
        );

        let g = parse_field_spec("- region (default North America, EMEA fallback)");
        assert_eq!(
            g.default,
            Some(Value::String("North America, EMEA fallback".into()))
        );

        // A single-token default still works (no regression).
        let h = parse_field_spec("- currency (default USD)");
        assert_eq!(h.default, Some(Value::String("USD".into())));
    }

    // ── Regression: a `default` after `enum` is parsed, not swallowed (#10) ────

    #[test]
    fn regression_default_after_enum_is_parsed_not_an_enum_member() {
        let f = parse_field_spec("- status (enum: open, closed, default open)");
        assert_eq!(
            f.enum_values,
            Some(vec!["open".to_string(), "closed".to_string()]),
            "`default open` leaked into the enum list"
        );
        assert_eq!(
            f.default,
            Some(Value::String("open".into())),
            "default after enum was dropped"
        );

        // The bare `enum` keyword form, with a trailing default.
        let g = parse_field_spec("- status (enum, open, closed, default open)");
        assert_eq!(
            g.enum_values,
            Some(vec!["open".to_string(), "closed".to_string()])
        );
        assert_eq!(g.default, Some(Value::String("open".into())));
    }

    // ── Regression: frozen-page policy does not fail open (#11) ────────────────

    #[test]
    fn regression_frozen_match_handles_leading_slash() {
        let cfg = Config {
            frozen_pages: vec![PathBuf::from("/records/decisions/q1.md")],
            ..Config::default()
        };
        assert!(
            cfg.is_frozen(Path::new("records/decisions/q1.md")),
            "leading-slash entry failed open"
        );
        assert!(cfg.is_frozen(Path::new("records/decisions/q1")));
    }

    #[test]
    fn regression_frozen_match_supports_globs() {
        let cfg = Config {
            frozen_pages: vec![PathBuf::from("records/decisions/*")],
            ..Config::default()
        };
        assert!(
            cfg.is_frozen(Path::new("records/decisions/q1.md")),
            "glob entry failed to protect a concrete file"
        );
        assert!(cfg.is_frozen(Path::new("records/decisions/q2.md")));
        // The glob does not cross a `/` segment.
        assert!(!cfg.is_frozen(Path::new("records/decisions/sub/q1.md")));
        // `**` crosses segments.
        let deep = Config {
            frozen_pages: vec![PathBuf::from("records/**")],
            ..Config::default()
        };
        assert!(deep.is_frozen(Path::new("records/decisions/sub/q1.md")));
        assert!(deep.is_frozen(Path::new("records/x.md")));
        assert!(!deep.is_frozen(Path::new("sources/x.md")));
        // A `*.md`-style intra-segment glob.
        let suffix = Config {
            frozen_pages: vec![PathBuf::from("records/decisions/q*")],
            ..Config::default()
        };
        assert!(suffix.is_frozen(Path::new("records/decisions/q1.md")));
        assert!(!suffix.is_frozen(Path::new("records/decisions/draft.md")));
    }

    #[test]
    fn regression_frozen_glob_many_double_stars_does_not_backtrack_exponentially() {
        use std::time::Instant;

        // A DB.md frozen-page bullet with many consecutive `**` segments and a
        // literal tail (`zzz`), matched against a deep target that ends in a
        // DIFFERENT segment (`file.md`), is the catastrophic-backtracking case:
        // the old two-way `glob_segments` recursion explored an exponential
        // number of (star, path) splits before concluding "no match" — ~119s for
        // 15 stars — hanging the store's entire write path (every write/rename/
        // fm-set funnels through `frozen_match`). The two-pointer matcher + `**`
        // collapse make this polynomial.
        let pat = format!("{}/zzz", vec!["**"; 30].join("/"));
        let target_path = format!("records/{}/file.md", vec!["a"; 40].join("/"));
        let cfg = Config {
            frozen_pages: vec![PathBuf::from(&pat)],
            ..Config::default()
        };

        let start = Instant::now();
        let frozen = cfg.is_frozen(Path::new(&target_path));
        let elapsed = start.elapsed();

        // The tail `zzz` never matches the target's `file.md`, so it is NOT frozen…
        assert!(
            !frozen,
            "non-matching deep target wrongly reported frozen (semantics changed)"
        );
        // …and the decision must be near-instant, not exponential. The pre-fix
        // code took tens of seconds here; a generous ceiling still fails loudly
        // if the blow-up ever returns.
        assert!(
            elapsed.as_secs() < 1,
            "frozen glob took {elapsed:?} — catastrophic backtracking is back"
        );

        // Semantics preserved: the same many-`**` pattern with a tail that DOES
        // match still freezes the file (a real match still refuses the write).
        let pat_hit = format!("{}/file.md", vec!["**"; 30].join("/"));
        let cfg_hit = Config {
            frozen_pages: vec![PathBuf::from(&pat_hit)],
            ..Config::default()
        };
        assert!(
            cfg_hit.is_frozen(Path::new(&target_path)),
            "many-`**` pattern failed to freeze a genuinely-matching deep target"
        );
    }

    #[test]
    fn frozen_glob_double_star_collapse_preserves_match_set() {
        // Collapsing consecutive `**` must not change which paths match: `**/**`
        // matches exactly what `**` does. Interleaved `**` and literals still
        // match across segments, and a non-matching literal tail still fails.
        let collapsed = Config {
            frozen_pages: vec![PathBuf::from("records/**/**/**/q1.md")],
            ..Config::default()
        };
        assert!(collapsed.is_frozen(Path::new("records/decisions/q1.md")));
        assert!(collapsed.is_frozen(Path::new("records/a/b/c/q1.md")));
        assert!(collapsed.is_frozen(Path::new("records/q1.md")));
        assert!(!collapsed.is_frozen(Path::new("records/a/b/c/q2.md")));
        assert!(!collapsed.is_frozen(Path::new("sources/a/q1.md")));

        // `**` between two literals spans zero or more intermediate segments.
        let between = Config {
            frozen_pages: vec![PathBuf::from("records/**/draft.md")],
            ..Config::default()
        };
        assert!(between.is_frozen(Path::new("records/draft.md")));
        assert!(between.is_frozen(Path::new("records/a/b/draft.md")));
        assert!(!between.is_frozen(Path::new("records/a/b/final.md")));
    }

    #[test]
    fn regression_frozen_entry_single_hyphen_comment_is_stripped() {
        // `records/decisions/q3.md - finalized` (single ASCII hyphen comment, no
        // backticks): the comment must be stripped so the entry is just the path.
        let path = extract_path_bullet("- records/decisions/q3.md - finalized");
        assert_eq!(path, "records/decisions/q3.md");

        // End to end: such a bullet freezes the file.
        let cfg = Config {
            frozen_pages: vec![PathBuf::from(extract_path_bullet(
                "- records/decisions/q3.md - finalized",
            ))],
            ..Config::default()
        };
        assert!(
            cfg.is_frozen(Path::new("records/decisions/q3.md")),
            "single-hyphen-comment entry failed open"
        );
    }
}