rto_render/

docs.rs

//! The documentation-site renderer: ADR markdown → themed HTML pages, produced
//! deterministically so CI diffs are meaningful. Replaces the shell
//! `md2html.awk` stopgap with a real `CommonMark` parser (`pulldown-cmark`),
//! fixing the whole class of hand-rolled-parser bugs (backtick runs, tables,
//! heading edge cases) we hit before.
//!
//! Page chrome (theme, nav, back-link, footer) matches the previous site so the
//! switch is drop-in. This module is pure string generation; the `roteiro`
//! binary owns walking `docs/adr` and copying static assets.

use std::collections::BTreeMap;
use std::fmt::Write as _;

use pulldown_cmark::{CowStr, Event, HeadingLevel, Options, Parser, Tag, TagEnd, html};

/// A rendered ADR: its title (for the index) and the full themed HTML page.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct RenderedAdr {
    /// The ADR title (first `# ` heading, or the fallback passed to
    /// [`render_adr`]).
    pub title: String,
    /// The complete HTML document.
    pub html: String,
}

/// Where each source document is **actually published**: the file the site
/// serves, keyed by the source markdown's file name.
///
/// [`rewrite_doc_link`] used to derive a link's target from the link's own
/// spelling — `../BUILD_PLAN_V2.md` → `../BUILD_PLAN_V2.html` — which is correct
/// only while every document is served under its own stem. Site pages ended
/// that: a page is published as its declared `site-page:` slug, and a slug is
/// URL-safe by construction (`[a-z0-9-]+`), so `docs/BUILD_PLAN_V2.md` is served
/// as `build-plan-v2.html`. The rewrite then pointed four correct repository
/// links at a page that is never emitted — issue #446, live on roteiro.dev.
///
/// So the served name is *looked up* rather than guessed. The renderer is handed
/// the index of what the site emits, which is the only thing that knows the
/// answer.
///
/// Keyed by file name rather than by full path because the site mirrors the
/// repository's layout — `docs/*.md` at the root, `docs/adr/*.md` under `adr/` —
/// so a link's directory hops are already correct and only the final segment can
/// differ. A file name claimed by two published documents is recorded as
/// **ambiguous** and left unrewritten: guessing which one a link meant is how a
/// link silently points at the wrong page, which is worse than the 404 it
/// replaces.
#[derive(Debug, Default, Clone, PartialEq, Eq)]
pub struct PublishedPages(BTreeMap<String, Option<String>>);

impl PublishedPages {
    /// An empty index: every `.md` link falls back to its own stem.
    #[must_use]
    pub fn new() -> Self {
        Self::default()
    }

    /// Record that `source_file` (a markdown file name, e.g. `BUILD_PLAN_V2.md`)
    /// is served as `served_as` (e.g. `build-plan-v2.html`).
    ///
    /// A second, differing claim on one file name makes it ambiguous; see the
    /// type's documentation for why that is left unrewritten.
    pub fn publish(&mut self, source_file: &str, served_as: &str) {
        self.0
            .entry(source_file.to_owned())
            .and_modify(|slot| {
                if slot.as_deref() != Some(served_as) {
                    *slot = None;
                }
            })
            .or_insert_with(|| Some(served_as.to_owned()));
    }

    /// The file `source_file` is served as, or `None` when it is unknown or
    /// ambiguous.
    fn served(&self, source_file: &str) -> Option<&str> {
        self.0.get(source_file)?.as_deref()
    }
}

/// Where a link that leaves the site points instead: the repository's own web
/// view, at the commit the site was built from.
///
/// The Build Plan cites code as evidence for its claims — `[sync](../crates/…
/// /sync.rs)` — which is correct in a checkout and dead on roteiro.dev, because
/// `render docs` publishes documents and not source. Six such links were live on
/// the site (issue #456). This is the answer chosen for them: keep the link's
/// affordance and move its target to the one place the file is actually served.
///
/// # Pinned to a commit, not to a branch
///
/// `blob` carries a sha (`…/blob/<sha>`), not `…/blob/main`. GitHub serves a
/// blob by sha forever, so the link keeps resolving after the file is renamed or
/// deleted; a `main` link 404s on the next rename, and this is a *retired* plan
/// whose citations describe the code as it stood, so drifting them onto today's
/// `main` would be wrong even when it resolved. It is also the rule the vault
/// renderer already ships (`source_blob_base` + `head_commit_id`), and one repo
/// with two answers to "which commit does a source link mean" is its own defect.
///
/// The cost is stated rather than hidden: a site rendered from a commit that was
/// never pushed yields links the host has never heard of. That is a local
/// preview, not the published site — the Website workflow renders from a commit
/// GitHub already has.
///
/// # No mappable origin
///
/// Construction goes through [`SourceBase::new`], which yields `None` when the
/// caller has no blob base to offer — no `origin` remote, or a remote whose URL
/// does not map to a web view. Links are then left exactly as they are: still
/// correct in a checkout, still dead on the site. That is deliberate and is the
/// least-bad of the three: refusing to render would break `render docs` in any
/// repository without an `origin` (every test fixture, every fresh `git init`),
/// and demoting the link to plain text would destroy information to hide a
/// problem the reader can otherwise route around.
///
/// # Can the class recur?
///
/// Not while a base exists: the rule is structural, not a list of the six links
/// that were found. *Any* link climbing above the site root is re-aimed, so a new
/// citation added to any rendered document is handled the day it is written, and
/// `a_link_out_of_the_site_goes_to_the_repository` is what fails if that stops.
///
/// It recurs silently in exactly one case — a site rendered where no base can be
/// derived — and nothing catches that, because the output is the *authored* link
/// and there is no rendered-site link gate (issue #459) to notice. That case is
/// the one the deploy does not hit: the Website workflow checks out with an
/// `origin` on `github.com`, and `without_an_origin_remote_a_source_link_is_left_as_authored`
/// pins the behaviour rather than the absence of a gate.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct SourceBase {
    /// Web blob base, no trailing slash — e.g.
    /// `https://github.com/owner/repo/blob/<sha>`.
    blob: String,
    /// The rendered document's own directory, repository-relative and with no
    /// trailing slash: `docs`, `docs/adr`, `website/pages`. A link is resolved
    /// against this to get the path the repository serves.
    dir: String,
}

impl SourceBase {
    /// A source base for a document at repository-relative directory `dir`,
    /// served from `blob`. `None` when `blob` is `None` — see the type's
    /// documentation for why that leaves links alone rather than failing.
    #[must_use]
    pub fn new(blob: Option<&str>, dir: &str) -> Option<Self> {
        Some(Self {
            blob: blob?.trim_end_matches('/').to_owned(),
            dir: dir.trim_matches('/').to_owned(),
        })
    }

    /// The web URL for `path` — a link written relative to this document —
    /// carrying `frag` through unchanged (`#L12` is a GitHub line anchor, and
    /// the site has no better guess than the author's).
    ///
    /// `None` when `path` climbs out of the repository altogether, which no base
    /// can name.
    fn blob_url(&self, path: &str, frag: Option<&str>) -> Option<String> {
        let joined = format!("{}/{}", self.dir, path);
        let (up, segs) = resolve_relative(&joined);
        if up > 0 || segs.is_empty() {
            return None;
        }
        let repo_path = segs.join("/");
        Some(match frag {
            Some(frag) => format!("{}/{repo_path}#{frag}", self.blob),
            None => format!("{}/{repo_path}", self.blob),
        })
    }
}

/// One page in the site navigation bar: where it goes and what it is called.
///
/// Built by the caller from the authored site pages (`rto_spec::site_nav` puts
/// them in order), and passed to [`render_site_page`] whole so every page emits
/// the *same* bar. A per-page bar assembled independently is a bar that can
/// disagree with itself, which is how a page ends up unreachable from its
/// neighbours.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct NavEntry {
    /// Root-relative href (e.g. `modes.html`, or `./` for the landing page).
    pub href: String,
    /// Short label shown in the bar.
    pub label: String,
}

/// An entry in the ADR/docs index page.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct IndexEntry {
    /// Relative href (e.g. `0001-….html`).
    pub href: String,
    /// Display title.
    pub title: String,
}

/// Convert `CommonMark` `md` to an HTML fragment (GitHub tables + strikethrough,
/// and Roteiro `[[wiki-links]]` resolved). Resolves ADR links relative to the
/// ADR directory; use [`render_doc`] for root-level pages.
///
/// A fragment renderer has no site to be a part of, so it carries neither
/// [`PublishedPages`] nor [`SourceBase`]: a `.md` link is rewritten to its own
/// stem, which is right for an ADR and a guess for anything published under a
/// slug, and a link out of the site is left alone.
#[must_use]
pub fn markdown_to_html(md: &str) -> String {
    render_markdown(md, "", &PublishedPages::new(), None, 0)
}

/// Render `md` to HTML: resolve `[[wiki-links]]` (ADR links use `adr_prefix` as
/// their href prefix), rewrite ordinary `[…](*.md)` links to their rendered
/// `.html` targets and links out of the site to `source`, then run `CommonMark`
/// with GitHub tables/strikethrough. `depth` is the page's own depth below the
/// site root; see [`rewrite_doc_link`].
fn render_markdown(
    md: &str,
    adr_prefix: &str,
    pages: &PublishedPages,
    source: Option<&SourceBase>,
    depth: usize,
) -> String {
    let pre = rewrite_wiki_links(md, adr_prefix);
    let ids = heading_ids(&pre);
    let mut next_id = 0usize;
    // Rewrite link destinations pointing at local Markdown files to the HTML the
    // site actually serves (e.g. `adr/0001-….md` → `adr/0001-….html`), and give
    // every heading a stable `id` so it can be linked to.
    let parser = Parser::new_ext(&pre, options()).map(|event| match event {
        Event::Start(Tag::Link {
            link_type,
            dest_url,
            title,
            id,
        }) => Event::Start(Tag::Link {
            link_type,
            dest_url: rewrite_doc_link(&dest_url, pages, source, depth)
                .map_or(dest_url, CowStr::from),
            title,
            id,
        }),
        Event::Start(Tag::Heading {
            level,
            classes,
            attrs,
            ..
        }) => {
            let id = ids.get(next_id).cloned().map(CowStr::from);
            next_id += 1;
            Event::Start(Tag::Heading {
                level,
                id,
                classes,
                attrs,
            })
        }
        other => other,
    });
    let mut out = String::new();
    html::push_html(&mut out, parser);
    out
}

/// The `CommonMark` dialect the whole site is parsed with — [`rto_graph`]'s, not
/// a copy of it.
///
/// This used to build its own `Options` with the same three flags. It was the
/// same set by agreement rather than by construction, which is the arrangement
/// [`rto_graph::markdown_dialect`] exists to end: a different option set is a
/// different language, and two parsers that disagree about it do not fail — they
/// quietly disagree about where a heading's text ends, which is the defect #469
/// was. That crate could not finish the consolidation while this file was held
/// open by the work in #456/#457/#508; this is the remaining half.
///
/// What the dialect buys *this* renderer, and why the heading-attribute flag in
/// particular is load-bearing here: **heading attributes are how a URL outlives a
/// restructure.** A page split out of the old single-page site keeps the anchor
/// the old page published — the heading declares `{#modes}` and lands at
/// `#modes` — instead of silently becoming whatever the new heading text happens
/// to slugify to. External links point at those anchors and cannot be updated, so
/// the alternative is not a tidier URL; it is a dead one.
fn options() -> Options {
    // Keep this body a single delegation. It is the shape that invites a
    // "just for rendering" flag, and the shape where adding one leaves nothing
    // to notice it by — the duplicate that used to sit here is gone, so a
    // divergence introduced now is invisible rather than merely unnoticed.
    //
    // A flag this renderer sets and `rto-graph`'s extractors do not is a flag
    // that changes what a heading's text *is* on one surface and not the other.
    // That is the #469 defect again, with the evidence removed.
    //
    // So a renderer-only flag belongs in `markdown_dialect` or nowhere, and "or
    // nowhere" is not rhetoric: the claim that some future flag cannot reach
    // heading text is an argument, not an observation, and an argument belongs
    // in `rto-graph` beside the flag it licenses, where every surface reads it.
    //
    // `the_dialect_is_not_extended_here` fails if this body grows a flag.
    rto_graph::markdown_dialect()
}

/// The `id` for every heading in `md`, in document order.
///
/// An explicit `{#anchor}` wins; otherwise the id is [`rto_graph::slugify`] of
/// the heading text. Both branches go through [`rto_graph::heading_id_from`] —
/// the *same* function `rto_spec` builds the section's node key with, for ADRs,
/// blueprints and site pages alike — so an authored link to `site:modes#offline`
/// lands on the heading the graph says it does.
///
/// That sentence used to be here and was only half true: this honoured an
/// explicit `{#anchor}` and `rto_spec` slugified the heading text regardless, so
/// a heading that declared its own address had one id in the page and a
/// different key in the graph (#524). The claim is now enforced by construction
/// rather than asserted in prose, which is the difference that matters.
///
/// The two **document-level** rules stay here, because only a renderer emits
/// elements: a heading whose id would be empty (`## ###`) falls back to its
/// position, and a repeat gets a `-2`, `-3`, … suffix, because two elements
/// sharing an `id` means one of them is unreachable. `rto_spec` does neither, so
/// a page with two identically-titled headings would diverge again — latent
/// today, since no page has one.
///
/// Computed from a *first parse* rather than a line scan: heading text can be
/// spread over several inline events, and `#` inside a fenced block is not a
/// heading at all. Parsing twice costs a document-sized pass and cannot be wrong
/// about what the renderer will see, because it is the same parser.
///
/// # This rule is not GitHub's, and deliberately stays that way
///
/// A document in `docs/` is read in two places under two slug rules: GitHub
/// renders `**v0.10.x**` as `v010x` and this renders it as `v0-10-x`, so an
/// anchor hand-written against one is dead in the other. That is real, and it is
/// the *second* half of issue #457 — the six anchors that issue found were dead
/// under **both** rules, because the heading text had changed under them.
///
/// It is not fixed by aligning this rule to GitHub's, and that is not a
/// deferral. [`rto_graph::slugify`] is the *one* rule, shared on purpose:
/// `rto_spec` builds every section node key with it (`adr:0001#design`,
/// `site:modes#offline-mode`) and this builds the matching `id`, which is the
/// only reason a `[[doc#section]]` wiki-link resolves through one and lands
/// through the other. Re-keying it to GitHub's would re-key every section node in
/// the graph and break every authored wiki-link `roteiro check` gates — to fix
/// anchors in one retired document. rustdoc is a *third* rule already in play,
/// so there is no single rule to converge on in any case.
///
/// Note what does **not** protect this: #397's guard (`doc_anchor_fragments.rs`)
/// replicates rustdoc's rule in its own private `slugify` and never calls
/// [`rto_graph::slugify`], so changing this rule would not have made it fail. It
/// is not a guard on this code path, and treating it as one was the mistake worth
/// recording here.
///
/// So the divergence is left, documented, and the affected anchors were given
/// explicit ids that neither rule touches (see `docs/BUILD_PLAN.md`). **Nothing
/// currently checks an intra-document anchor** in either rendering — not
/// `roteiro check`, not this crate. Issue #459 is where that check belongs.
fn heading_ids(md: &str) -> Vec<String> {
    let mut ids: Vec<String> = Vec::new();
    let mut seen: BTreeMap<String, usize> = BTreeMap::new();
    let mut current: Option<(Option<String>, String)> = None;
    for event in Parser::new_ext(md, options()) {
        match event {
            Event::Start(Tag::Heading { id, .. }) => {
                current = Some((id.map(|i| i.to_string()), String::new()));
            }
            Event::Text(t) | Event::Code(t) => {
                if let Some((_, text)) = current.as_mut() {
                    text.push_str(&t);
                }
            }
            Event::End(TagEnd::Heading(_)) => {
                let Some((explicit, text)) = current.take() else {
                    continue;
                };
                // The shared rule, not a local copy of it: `rto_spec` builds the
                // section's node key with this same function, so the `id` emitted
                // here and the key a `[[doc#section]]` link resolves to cannot
                // disagree (#524).
                let base = rto_graph::heading_id_from(explicit.as_deref(), text.trim());
                let base = if base.is_empty() {
                    format!("section-{}", ids.len() + 1)
                } else {
                    base
                };
                let n = seen.entry(base.clone()).or_insert(0);
                *n += 1;
                ids.push(if *n == 1 { base } else { format!("{base}-{n}") });
            }
            _ => {}
        }
    }
    ids
}

/// Split a relative path into the number of hops it takes **above** its own
/// directory and the segments that remain, resolving `.` and `..` the way a
/// browser and a filesystem both do.
///
/// `../crates/x.rs` is `(1, ["crates", "x.rs"])`; `adr/../guide.md` is
/// `(0, ["guide.md"])`. The hop count is the whole escape test in
/// [`rewrite_doc_link`]: a link that climbs further than the page sits below the
/// site root is a link to something outside the site.
fn resolve_relative(path: &str) -> (usize, Vec<&str>) {
    let mut up = 0usize;
    let mut segs: Vec<&str> = Vec::new();
    for seg in path.split('/') {
        match seg {
            "" | "." => {}
            ".." => {
                if segs.pop().is_none() {
                    up += 1;
                }
            }
            s => segs.push(s),
        }
    }
    (up, segs)
}

/// Rewrite a relative link so it points at what the **site** serves, preserving
/// any `#fragment`. Returns `None` for external, protocol-relative, `mailto:`,
/// pure-anchor and root-relative links, and for anything the site already serves
/// under the spelling the link uses — all of which are left unchanged.
///
/// `depth` is how far below the site root the page being rendered sits: 0 for a
/// root-level page, 1 for an ADR under `adr/`. It is supplied by the entry point
/// rather than by the caller, because the entry point is the thing that knows.
///
/// Two rewrites live here, and the order between them is the interesting part:
///
/// * a `.md` link is aimed at the page the site publishes it as
///   ([`PublishedPages`]) — checked **first**, so a document that happens to be
///   reached by a path climbing out of its own directory still lands on its
///   published page rather than being treated as unpublished (issue #446);
/// * a link that climbs above the site root and is *not* published is aimed at
///   the repository's web view ([`SourceBase`]) — issue #456.
fn rewrite_doc_link(
    dest: &str,
    pages: &PublishedPages,
    source: Option<&SourceBase>,
    depth: usize,
) -> Option<String> {
    if dest.starts_with("http://")
        || dest.starts_with("https://")
        || dest.starts_with("//")
        || dest.starts_with("mailto:")
        || dest.starts_with('#')
        || dest.starts_with('/')
    {
        return None;
    }
    let (path, frag) = dest
        .split_once('#')
        .map_or((dest, None), |(p, f)| (p, Some(f)));
    // Only the final segment can differ between the repository and the site, so
    // the link's own directory hops are kept verbatim; see [`PublishedPages`].
    let (dir, file) = path.rsplit_once('/').map_or(("", path), |(d, f)| (d, f));
    // `strip_suffix` rather than `ends_with`: the extension is matched exactly as
    // it is written, which is the rule every document in this repository follows.
    let is_markdown = path.strip_suffix(".md").is_some();
    let sep = if dir.is_empty() { "" } else { "/" };

    // A page the site publishes, under the name it publishes it as (issue #446).
    // First, so a document reached by a path that climbs out of its own
    // directory still lands on its page rather than being read as unpublished.
    if let Some(served) = is_markdown.then(|| pages.served(file)).flatten() {
        return Some(match frag {
            Some(frag) => format!("{dir}{sep}{served}#{frag}"),
            None => format!("{dir}{sep}{served}"),
        });
    }

    // Not published, and climbing above the site root: no rewrite *within* the
    // site can make this resolve, so hand it to the repository (issue #456).
    // When there is no base to hand it to, fall through — everything below is
    // the behaviour that predates this, unchanged, so a repository with no
    // mappable `origin` renders exactly the site it rendered before.
    if resolve_relative(path).0 > depth
        && let Some(url) = source.and_then(|s| s.blob_url(path, frag))
    {
        return Some(url);
    }

    if !is_markdown {
        return None;
    }
    // Unknown or ambiguous: fall back to the stem rewrite this has always done,
    // which is right for every ADR (each is served under its own stem) and no
    // worse than before for anything else.
    let served = format!("{}.html", file.trim_end_matches(".md"));
    Some(match frag {
        Some(frag) => format!("{dir}{sep}{served}#{frag}"),
        None => format!("{dir}{sep}{served}"),
    })
}

/// Render one ADR markdown document to a themed HTML page. Leading YAML
/// frontmatter is stripped; the title is the first `# ` heading, or `fallback`
/// if there is none. ADR `[[…]]` links resolve to sibling ADR pages.
///
/// An ADR page is served one directory below the site root (`adr/`), which is
/// the `1` below: a `../…` link from an ADR still lands inside the site, and only
/// a second hop leaves it. See [`SourceBase`] for `source`.
#[must_use]
pub fn render_adr(
    markdown: &str,
    fallback_title: &str,
    pages: &PublishedPages,
    source: Option<&SourceBase>,
) -> RenderedAdr {
    let body = strip_frontmatter(markdown);
    let title = first_heading(body).unwrap_or_else(|| fallback_title.to_owned());
    let content = render_markdown(body, "", pages, source, 1);
    let nav = "<p class=\"nav\"><a href=\"../\">← Roteiro home</a> · \
               <a href=\"./\">All ADRs</a> · <a href=\"../build-plan.html\">Build Plan</a></p>";
    let html = page(&format!("{title} — Roteiro"), "../", nav, &content);
    RenderedAdr { title, html }
}

/// Render a root-level "lifetime doc" (e.g. the Build Plan) to a themed page.
/// Its `[[docs/adr/…]]` links resolve into the `adr/` subdirectory.
///
/// The page is served *at* the site root — the `0` below — so any `../…` link
/// leaves the site; see [`SourceBase`] for where those go.
#[must_use]
pub fn render_doc(
    markdown: &str,
    fallback_title: &str,
    pages: &PublishedPages,
    source: Option<&SourceBase>,
) -> RenderedAdr {
    let body = strip_frontmatter(markdown);
    let title = first_heading(body).unwrap_or_else(|| fallback_title.to_owned());
    let content = render_markdown(body, "adr/", pages, source, 0);
    let nav = "<p class=\"nav\"><a href=\"./\">← Roteiro home</a> · \
               <a href=\"adr/\">ADRs</a></p>";
    let html = page(&format!("{title} — Roteiro"), "./", nav, &content);
    RenderedAdr { title, html }
}

/// Render one **site page** — a document that declared itself published — to a
/// themed root-level page carrying the site navigation bar.
///
/// `nav` is the whole bar, in order; `current_href` is this page's own entry,
/// which is marked `aria-current="page"` and rendered unlinked so the reader can
/// see where they are. A `current_href` that matches nothing in `nav` simply
/// yields a bar with nothing marked, which is what a preview of an unlisted page
/// should look like rather than an error.
///
/// The title is the first `# ` heading, or `fallback_title`. `[[docs/adr/…]]`
/// links resolve into the `adr/` subdirectory, exactly as they do for the Build
/// Plan: a site page is a root-level document.
#[must_use]
pub fn render_site_page(
    markdown: &str,
    fallback_title: &str,
    nav: &[NavEntry],
    current_href: &str,
    pages: &PublishedPages,
    source: Option<&SourceBase>,
) -> RenderedAdr {
    let body = strip_frontmatter(markdown);
    let title = first_heading(body).unwrap_or_else(|| fallback_title.to_owned());
    let content = render_markdown(body, "adr/", pages, source, 0);
    let bar = render_nav(nav, current_href);
    let html = page(&format!("{title} — Roteiro"), "./", &bar, &content);
    RenderedAdr { title, html }
}

/// The site navigation bar: one link per page, the current one marked.
///
/// Plain anchors in a `<nav>`, styled by `website/public/style.css`. No script:
/// the explorer is deliberately vendored with no build step (ADR-0010), and a
/// navigation bar that needs JavaScript to be a navigation bar would be the
/// first thing on this site that does.
#[must_use]
pub fn render_nav(nav: &[NavEntry], current_href: &str) -> String {
    let mut out = String::from("<nav class=\"sitenav\">");
    for entry in nav {
        if entry.href == current_href {
            let _ = write!(
                out,
                "<span aria-current=\"page\">{}</span>",
                escape_html(&entry.label)
            );
        } else {
            let _ = write!(
                out,
                "<a href=\"{}\">{}</a>",
                escape_attr(&entry.href),
                escape_html(&entry.label)
            );
        }
    }
    out.push_str("</nav>");
    out
}

/// The marker whose contents [`replace_site_nav`] owns.
const SITENAV_OPEN: &str = "<nav class=\"sitenav\">";

/// Replace the `<nav class="sitenav">…</nav>` block in a **hand-written** page
/// with the bar the renderer computes, returning `None` when the page carries no
/// such block.
///
/// # Why this exists
///
/// `website/public/index.html` is the one page of roteiro.dev nothing renders —
/// it is copied verbatim — and it used to carry a *hand-maintained copy* of the
/// list the renderer derives from `site-order` (issue #508). Adding
/// `docs/SERVING.md` appeared in every rendered page's bar automatically and had
/// to be typed into the landing page by hand.
///
/// The failure that made it worth removing rather than remembering is silent and
/// points the wrong way: a new page is published, reachable, and linked from
/// every page **except the front one**. Nothing errors and `roteiro check`
/// passes, because everything that is there resolves.
///
/// **That is also why no link auditor could have caught it, and why none will
/// catch the next one of its shape.** The defect is a link that does *not*
/// exist; auditing what is there cannot find what is missing. This removes the
/// possibility instead — after this, the landing page has no independent list to
/// disagree with. What still guards the seam is
/// `the_landing_page_carries_the_bar_the_renderer_emits`, which now checks that
/// the replacement *happened*: a landing page whose marker was renamed away
/// keeps its stale bar silently, and that test is what fails.
///
/// # Why `None` rather than an error
///
/// A landing page with no `sitenav` is not claiming a bar, and a site is allowed
/// not to have one — every `render docs` fixture writes a one-line `index.html`.
/// The caller leaves such a page alone.
#[must_use]
pub fn replace_site_nav(html: &str, nav: &[NavEntry], current_href: &str) -> Option<String> {
    let open = html.find(SITENAV_OPEN)?;
    let close = html[open..].find("</nav>")? + open + "</nav>".len();
    let mut out = String::with_capacity(html.len());
    out.push_str(&html[..open]);
    out.push_str(&render_nav(nav, current_href));
    out.push_str(&html[close..]);
    Some(out)
}

/// Render the docs index: any `lifetime` docs (Build Plan, …) then the ADRs.
#[must_use]
pub fn render_adr_index(lifetime: &[IndexEntry], entries: &[IndexEntry]) -> String {
    let mut list = String::new();
    if !lifetime.is_empty() {
        list.push_str("<h1>Documentation</h1><ul>");
        for e in lifetime {
            let _ = write!(
                list,
                "<li><a href=\"{}\">{}</a></li>",
                escape_attr(&e.href),
                escape_html(&e.title)
            );
        }
        list.push_str("</ul>");
    }
    list.push_str("<h1>Architecture Decision Records</h1><ul>");
    for e in entries {
        let _ = write!(
            list,
            "<li><a href=\"{}\">{}</a></li>",
            escape_attr(&e.href),
            escape_html(&e.title)
        );
    }
    list.push_str("</ul>");
    let nav = "<p class=\"nav\"><a href=\"../\">← Roteiro home</a></p>";
    page("Documentation — Roteiro", "../", nav, &list)
}

/// Rewrite Roteiro `[[wiki-links]]` into Markdown, honouring code spans/fences:
/// `[[docs/adr/<slug>.md]]` (optionally `#section`) becomes a link to that ADR
/// page (`<adr_prefix><slug>.html`); any other `[[…]]` (code/file references,
/// for which the site has no page) becomes inline code so it renders cleanly
/// instead of leaking literal brackets.
fn rewrite_wiki_links(md: &str, adr_prefix: &str) -> String {
    let mut out = String::new();
    let mut in_fence = false;
    for line in md.lines() {
        let trimmed = line.trim_start();
        if trimmed.starts_with("```") || trimmed.starts_with("~~~") {
            in_fence = !in_fence;
            out.push_str(line);
            out.push('\n');
            continue;
        }
        if in_fence {
            out.push_str(line);
            out.push('\n');
            continue;
        }
        rewrite_line_outside_code(line, adr_prefix, &mut out);
        out.push('\n');
    }
    out
}

/// Rewrite wiki-links in one line, leaving `CommonMark` inline code spans
/// untouched. A code span opens with a run of *n* backticks and closes with the
/// next run of *exactly* *n* backticks; anything between (including `[[…]]`
/// examples) is emitted verbatim. Backtick runs with no matching close are
/// literal text and do not shield what follows.
fn rewrite_line_outside_code(line: &str, adr_prefix: &str, out: &mut String) {
    let bytes = line.as_bytes();
    let mut text_start = 0;
    let mut i = 0;
    while i < bytes.len() {
        if bytes[i] != b'`' {
            i += 1;
            continue;
        }
        let run_start = i;
        while i < bytes.len() && bytes[i] == b'`' {
            i += 1;
        }
        let run = i - run_start;
        if let Some(rel) = find_closing_run(&bytes[i..], run) {
            // Text before the opening delimiter is ordinary prose.
            rewrite_wiki_in(&line[text_start..run_start], adr_prefix, out);
            let code_end = i + rel + run;
            out.push_str(&line[run_start..code_end]); // span, delimiters included
            i = code_end;
            text_start = i;
        }
        // No close → treat the run as literal text; keep it in the pending
        // buffer (rewrite_wiki_in leaves backticks alone) and keep scanning.
    }
    rewrite_wiki_in(&line[text_start..], adr_prefix, out);
}

/// Byte offset (within `bytes`) of the next backtick run of *exactly* `run`
/// backticks, or `None`. Longer or shorter runs are skipped, per `CommonMark`.
fn find_closing_run(bytes: &[u8], run: usize) -> Option<usize> {
    let mut i = 0;
    while i < bytes.len() {
        if bytes[i] != b'`' {
            i += 1;
            continue;
        }
        let start = i;
        while i < bytes.len() && bytes[i] == b'`' {
            i += 1;
        }
        if i - start == run {
            return Some(start);
        }
    }
    None
}

/// Rewrite every `[[…]]` in one non-code text segment.
fn rewrite_wiki_in(seg: &str, adr_prefix: &str, out: &mut String) {
    let mut rest = seg;
    while let Some(open) = rest.find("[[") {
        out.push_str(&rest[..open]);
        let after = &rest[open + 2..];
        if let Some(close) = after.find("]]") {
            out.push_str(&wiki_target(&after[..close], adr_prefix));
            rest = &after[close + 2..];
        } else {
            out.push_str("[[");
            rest = after;
        }
    }
    out.push_str(rest);
}

/// Resolve one wiki-link's inner text to Markdown.
fn wiki_target(inner: &str, adr_prefix: &str) -> String {
    let inner = inner.trim();
    let path = inner.split_once('#').map_or(inner, |(p, _)| p.trim());
    if let Some(rest) = path.strip_prefix("docs/adr/")
        && let Some(stem) = rest.strip_suffix(".md")
    {
        return format!("[{}]({adr_prefix}{stem}.html)", adr_label(stem));
    }
    // Code/file reference — the site has no page for it; show it as code.
    format!("`{inner}`")
}

/// A display label for an ADR filename stem: `0001-build-…` → `ADR-0001`.
fn adr_label(stem: &str) -> String {
    let digits: String = stem.chars().take_while(char::is_ascii_digit).collect();
    if digits.is_empty() {
        stem.to_owned()
    } else {
        format!("ADR-{digits}")
    }
}

/// Wrap body HTML in the themed page chrome. `root` is the relative path to the
/// site root (e.g. `"../"` for pages under `adr/`).
fn page(title: &str, root: &str, nav: &str, body: &str) -> String {
    format!(
        "<!doctype html><html lang=\"en\"><head><meta charset=\"utf-8\">\
         <meta name=\"viewport\" content=\"width=device-width, initial-scale=1\">\
         <link rel=\"icon\" href=\"{root}favicon.svg\" type=\"image/svg+xml\">\
         <link rel=\"icon\" href=\"{root}favicon.ico\" type=\"image/x-icon\" sizes=\"16x16 32x32 48x48\">\
         <link rel=\"apple-touch-icon\" href=\"{root}apple-touch-icon.png\">\
         <link rel=\"stylesheet\" href=\"{root}style.css\">\
         <title>{title}</title></head><body>\
         {nav}{body}\
         <p class=\"backlink\"><a href=\"{root}\">← Back to roteiro.dev</a></p>\
         <footer>Dual-licensed MIT OR Apache-2.0 · The Roteiro Project Team</footer>\
         </body></html>",
        title = escape_html(title),
    )
}

/// Strip a leading `---`-delimited YAML frontmatter block.
fn strip_frontmatter(text: &str) -> &str {
    let Some(rest) = text.strip_prefix("---\n") else {
        return text;
    };
    match rest.find("\n---\n") {
        Some(end) => &rest[end + 5..],
        None => rest.strip_suffix("\n---").unwrap_or(text),
    }
}

/// The visible text of the document's first level-1 heading — what the reader
/// sees in the rendered `<h1>` — or `None` when the document has none.
///
/// Read from a **parse**, for the same reason [`heading_ids`] is: the heading's
/// raw line is source, not text. A line scan cannot tell `{#modes}` (a heading
/// attribute this renderer deliberately enables, see [`options`]) from the words
/// of the heading, so it read `# The five ways to run it {#modes}` back as a
/// title and put the markup in the `<title>` element of every page moved by the
/// site split — issue #460, live on roteiro.dev. The `<h1>` on the same page was
/// already right, because that side went through the parser.
///
/// The fix is *not* a second place that knows how to strip `{#…}`. A rule
/// spelled out twice is a rule that can disagree with itself, and this one
/// already disagrees once: the anchor is markup to the parser and text to the
/// scanner. Asking the parser removes the second opinion rather than aligning
/// it, and carries the rest of the dialect along for free — a fenced `# …` is
/// not a title, a setext underline is one, and inline markup (`` `code` ``,
/// emphasis, a link label) contributes its text and not its punctuation.
///
/// The parse stops at the first `</h1>`; nothing walks the rest of the document.
fn first_heading(body: &str) -> Option<String> {
    let mut text: Option<String> = None;
    for event in Parser::new_ext(body, options()) {
        match event {
            Event::Start(Tag::Heading {
                level: HeadingLevel::H1,
                ..
            }) => text = Some(String::new()),
            // Only accumulates once an H1 has opened; a code span is part of the
            // heading's text, exactly as it is for the heading's id.
            Event::Text(t) | Event::Code(t) => {
                if let Some(text) = text.as_mut() {
                    text.push_str(&t);
                }
            }
            Event::End(TagEnd::Heading(HeadingLevel::H1)) => break,
            _ => {}
        }
    }
    // An empty `#` heading names nothing, so it defers to the caller's fallback
    // rather than rendering `<title> — Roteiro</title>`.
    text.map(|t| t.trim().to_owned()).filter(|t| !t.is_empty())
}

fn escape_html(s: &str) -> String {
    s.replace('&', "&amp;")
        .replace('<', "&lt;")
        .replace('>', "&gt;")
}

fn escape_attr(s: &str) -> String {
    escape_html(s).replace('"', "&quot;")
}

#[cfg(test)]
mod tests {
    use super::{
        IndexEntry, NavEntry, PublishedPages, SourceBase, escape_html, markdown_to_html, options,
        render_adr, render_adr_index, render_doc, render_markdown, render_nav, render_site_page,
        replace_site_nav,
    };

    /// The site index most tests do not exercise: with it empty, a `.md` link
    /// falls back to its own stem, which is what every assertion below predates.
    fn no_pages() -> PublishedPages {
        PublishedPages::new()
    }

    fn nav() -> Vec<NavEntry> {
        vec![
            NavEntry {
                href: "./".into(),
                label: "Home".into(),
            },
            NavEntry {
                href: "modes.html".into(),
                label: "Modes & Co".into(),
            },
        ]
    }

    #[test]
    fn markdown_renders_headings_and_tables() {
        let html = markdown_to_html("# Title\n\n| a | b |\n|---|---|\n| 1 | 2 |\n");
        assert!(html.contains("<h1 id=\"title\">Title</h1>"), "{html}");
        assert!(html.contains("<table>"));
        assert!(html.contains("<td>1</td>"));
    }

    #[test]
    fn adr_wiki_links_become_sibling_page_links() {
        // An ADR-to-ADR wiki link resolves to the sibling .html; a code/file
        // reference becomes inline code; both stop leaking literal `[[ ]]`.
        let md = "See [[docs/adr/0001-build-roteiro.md]] and \
                  [[crates/rto-graph/src/store.rs#Store]] here.\n";
        let html = markdown_to_html(md);
        assert!(
            html.contains("<a href=\"0001-build-roteiro.html\">ADR-0001</a>"),
            "ADR wiki-link → sibling page: {html}"
        );
        assert!(
            html.contains("<code>crates/rto-graph/src/store.rs#Store</code>"),
            "code reference → inline code: {html}"
        );
        assert!(
            !html.contains("[["),
            "no literal wiki brackets leak: {html}"
        );
    }

    #[test]
    fn wiki_links_inside_code_are_left_literal() {
        // A documented example of the syntax, in backticks or a fence, must not
        // be rewritten.
        let inline = markdown_to_html("use `[[docs/adr/0001-x.md]]` in prose\n");
        assert!(
            inline.contains("<code>[[docs/adr/0001-x.md]]</code>"),
            "{inline}"
        );
        let fenced = markdown_to_html("```\n[[docs/adr/0001-x.md]]\n```\n");
        assert!(
            fenced.contains("[[docs/adr/0001-x.md]]"),
            "fence literal: {fenced}"
        );
    }

    #[test]
    fn multi_backtick_code_spans_are_honoured() {
        // A tight double-backtick span (`` ``…`` ``) and the Build Plan's
        // nested-backtick example must both survive verbatim — the previous
        // single-backtick split rewrote the wiki-link inside them.
        let tight = markdown_to_html("say ``[[docs/adr/0001-x.md]]`` please\n");
        assert!(
            tight.contains("<code>[[docs/adr/0001-x.md]]</code>"),
            "{tight}"
        );
        assert!(!tight.contains("<a "), "no link inside code span: {tight}");

        let nested = markdown_to_html("its `` `[[path#Symbol]]` `` example\n");
        assert!(
            nested.contains("<code>`[[path#Symbol]]`</code>"),
            "{nested}"
        );
        assert!(
            !nested.contains("<a "),
            "no link inside nested span: {nested}"
        );

        // An unterminated run is literal and does not shield a later real link.
        let stray = markdown_to_html("a ` stray tick then [[docs/adr/0001-x.md]]\n");
        assert!(
            stray.contains("<a href=\"0001-x.html\">ADR-0001</a>"),
            "unterminated backtick must not shield: {stray}"
        );
    }

    #[test]
    fn markdown_md_links_are_rewritten_to_html() {
        // Ordinary `[text](path.md)` links must point at the rendered `.html`,
        // preserving fragments; external and anchor links are left alone.
        let html = markdown_to_html(
            "See [ADR-1](adr/0001-x.md) and [§2](adr/0001-x.md#context) and \
             [home](https://x.dev) and [top](#intro).\n",
        );
        assert!(html.contains("href=\"adr/0001-x.html\""), "{html}");
        assert!(html.contains("href=\"adr/0001-x.html#context\""), "{html}");
        assert!(
            html.contains("href=\"https://x.dev\""),
            "external unchanged: {html}"
        );
        assert!(html.contains("href=\"#intro\""), "anchor unchanged: {html}");
        assert!(!html.contains(".md\""), "no raw .md hrefs remain: {html}");
    }

    #[test]
    fn render_doc_links_adrs_into_subdir() {
        // A root-level lifetime doc (Build Plan) resolves ADR links into `adr/`.
        let r = render_doc(
            "# Build Plan\n\nGoverned by [[docs/adr/0001-x.md]].\n",
            "Build Plan",
            &no_pages(),
            None,
        );
        assert_eq!(r.title, "Build Plan");
        assert!(
            r.html.contains("<a href=\"adr/0001-x.html\">ADR-0001</a>"),
            "root doc → adr/ prefix: {}",
            r.html
        );
        // Root-level chrome: assets/back-link relative to site root.
        assert!(r.html.contains("href=\"./style.css\""));
        // Full favicon set — root-relative from the site root.
        assert!(r.html.contains("href=\"./favicon.svg\""));
        assert!(r.html.contains("href=\"./favicon.ico\""));
        assert!(
            r.html
                .contains("rel=\"apple-touch-icon\" href=\"./apple-touch-icon.png\"")
        );
    }

    const ADR: &str = "---\nadr-id: \"0001\"\nstatus: Accepted\n---\n\n# ADR-0001: Example\n\n## Context\n\nSome `code` and a [link](https://x).\n";

    #[test]
    fn render_adr_strips_frontmatter_and_themes() {
        let r = render_adr(ADR, "fallback", &no_pages(), None);
        assert_eq!(r.title, "ADR-0001: Example");
        // Frontmatter is gone; heading + section rendered.
        assert!(!r.html.contains("adr-id"));
        assert!(
            r.html
                .contains("<h1 id=\"adr-0001-example\">ADR-0001: Example</h1>")
        );
        // The section anchor matches the section's node key (`adr:0001#context`),
        // so a link through the graph lands on the heading in the browser.
        assert!(r.html.contains("<h2 id=\"context\">Context</h2>"));
        assert!(r.html.contains("<code>code</code>"));
        // Themed chrome present.
        assert!(
            r.html
                .contains("<link rel=\"stylesheet\" href=\"../style.css\">")
        );
        // Full favicon set (SVG + `.ico` fallback for browsers without SVG-favicon
        // support, e.g. Safari) — root-relative from a sub-page.
        assert!(r.html.contains("href=\"../favicon.svg\""));
        assert!(r.html.contains("href=\"../favicon.ico\""));
        assert!(
            r.html
                .contains("rel=\"apple-touch-icon\" href=\"../apple-touch-icon.png\"")
        );
        assert!(r.html.contains("← Roteiro home"));
        assert!(r.html.contains("← Back to roteiro.dev"));
        assert!(r.html.starts_with("<!doctype html>"));
    }

    #[test]
    fn render_adr_falls_back_without_h1() {
        let r = render_adr(
            "no frontmatter, no heading\n",
            "slug-name",
            &no_pages(),
            None,
        );
        assert_eq!(r.title, "slug-name");
    }

    #[test]
    fn index_lists_entries_and_escapes() {
        let entries = [
            IndexEntry {
                href: "0001-x.html".into(),
                title: "First & <best>".into(),
            },
            IndexEntry {
                href: "0002-y.html".into(),
                title: "Second".into(),
            },
        ];
        let lifetime = [IndexEntry {
            href: "../build-plan.html".into(),
            title: "Build Plan".into(),
        }];
        let html = render_adr_index(&lifetime, &entries);
        assert!(html.contains("<a href=\"../build-plan.html\">Build Plan</a>"));
        assert!(html.contains("<a href=\"0001-x.html\">First &amp; &lt;best&gt;</a>"));
        assert!(html.contains("<a href=\"0002-y.html\">Second</a>"));
        // First entry precedes second (order preserved).
        assert!(html.find("0001-x").unwrap() < html.find("0002-y").unwrap());
        // Lifetime docs listed before the ADRs.
        assert!(html.find("build-plan").unwrap() < html.find("0001-x").unwrap());
    }

    #[test]
    fn an_explicit_anchor_survives_the_split_that_moved_its_section() {
        // The hazard this mechanism exists for. The old single-page site
        // published `#modes`, `#crossrepo`, `#remote-tier` — short, hand-chosen
        // ids that no heading text slugifies to. External links point at them and
        // cannot be updated, so a page that inherits a section must be able to
        // inherit its anchor verbatim.
        let html = markdown_to_html(
            "## The five ways to run it {#modes}\n\n## Cross-repo: a hub and its spokes {#crossrepo}\n",
        );
        assert!(
            html.contains("<h2 id=\"modes\">The five ways to run it</h2>"),
            "{html}"
        );
        assert!(
            html.contains("<h2 id=\"crossrepo\">Cross-repo: a hub and its spokes</h2>"),
            "{html}"
        );
        // The attribute is markup, not part of the heading's text.
        assert!(!html.contains("{#"), "no literal attribute leaks: {html}");
    }

    #[test]
    fn generated_anchors_match_the_graph_s_section_keys_and_stay_unique() {
        // `rto_spec` builds `<doc>#<slugify(heading)>` section keys from the same
        // function, so a link that resolves in the graph lands on the heading.
        let html = markdown_to_html("## Install & build\n\n## Install & build\n\n## ###\n");
        assert!(html.contains("id=\"install-build\""), "{html}");
        // A repeat is suffixed rather than duplicated: two elements sharing an
        // `id` makes one of them unreachable.
        assert!(html.contains("id=\"install-build-2\""), "{html}");
        // A heading that slugifies to nothing still gets a usable anchor.
        assert!(html.contains("id=\"section-3\""), "{html}");
    }

    #[test]
    fn inline_code_counts_as_heading_text() {
        // The old page's headings look like `What <code>init</code> sets up`.
        // Dropping the code span would slugify only the prose around it and give
        // the section an anchor nobody would guess.
        let html = markdown_to_html("### What `init` sets up\n");
        assert!(
            html.contains("<h3 id=\"what-init-sets-up\">"),
            "code span is part of the heading's text: {html}"
        );
    }

    #[test]
    fn a_hash_inside_a_fence_is_not_a_heading() {
        // The id list is computed from a real parse, so fenced content cannot
        // shift every subsequent heading's anchor by one.
        let html = markdown_to_html("```\n## Not a heading\n```\n\n## Real\n");
        assert!(html.contains("<h2 id=\"real\">Real</h2>"), "{html}");
    }

    #[test]
    fn a_heading_s_anchor_never_reaches_the_title() {
        // Issue #460, live on roteiro.dev: every page the site split moved
        // carries `{#…}` on its H1, and the title was read off the raw line.
        let r = render_site_page(
            "---\nsite-page: modes\n---\n\n# The five ways to run it {#modes}\n\nBody.\n",
            "fallback",
            &nav(),
            "modes.html",
            &no_pages(),
            None,
        );
        // The heading was always right; the title is the side that was wrong.
        assert!(
            r.html
                .contains("<h1 id=\"modes\">The five ways to run it</h1>"),
            "{}",
            r.html
        );
        assert_eq!(r.title, "The five ways to run it");
        assert!(
            r.html
                .contains("<title>The five ways to run it — Roteiro</title>"),
            "{}",
            r.html
        );
        // The most-seen string a page has: the tab, the bookmark, the search
        // result, the social preview. Nothing of the attribute survives anywhere.
        assert!(
            !r.html.contains("{#"),
            "no literal attribute leaks: {}",
            r.html
        );
    }

    #[test]
    fn the_same_holds_for_an_adr_and_for_a_root_level_doc() {
        // One extractor serves all three renderers, so all three are checked:
        // a fix that reached only the page the issue named would leave the ADR
        // index quoting `{#…}` back at the reader.
        let adr = render_adr("# ADR-0001: Example {#adr1}\n", "slug", &no_pages(), None);
        assert_eq!(adr.title, "ADR-0001: Example");
        assert!(
            adr.html
                .contains("<title>ADR-0001: Example — Roteiro</title>"),
            "{}",
            adr.html
        );
        let doc = render_doc(
            "# Roteiro — Build Plan {#plan}\n",
            "Build Plan",
            &no_pages(),
            None,
        );
        assert_eq!(doc.title, "Roteiro — Build Plan");
        assert!(!doc.html.contains("{#"), "{}", doc.html);
    }

    #[test]
    fn a_title_that_legitimately_spells_the_anchor_syntax_keeps_it() {
        // The other half of the rule, and the reason the fix is a parse and not
        // a strip: `{#…}` is an attribute only where the dialect says it is, and
        // a rule spelled out by hand does not know where that is. Inside a code
        // span it is prose, and a stripper blind to code spans mangles a page
        // whose subject *is* this syntax — which is most of the pages that
        // document it.
        let coded = render_doc(
            "# Why `{#anchor}` outlives a restructure\n",
            "fallback",
            &no_pages(),
            None,
        );
        assert_eq!(coded.title, "Why {#anchor} outlives a restructure");
        assert!(
            coded
                .html
                .contains("<title>Why {#anchor} outlives a restructure — Roteiro</title>"),
            "{}",
            coded.html
        );
        // Mid-heading and uncoded, it is still prose: an attribute block is
        // trailing or it is nothing.
        let mid = render_doc(
            "# Anchors are written {#id}, in prose\n",
            "fallback",
            &no_pages(),
            None,
        );
        assert_eq!(mid.title, "Anchors are written {#id}, in prose");
    }

    #[test]
    fn the_title_and_the_heading_never_disagree() {
        // The invariant underneath #460, stated directly. Where the attribute
        // block ends is the dialect's call, not this module's — braces the
        // parser eats are gone from *both* surfaces, braces it keeps are on
        // both. Reading the title from the same parse is what makes that true by
        // construction rather than by two rules that happen to match today.
        for md in [
            "# The five ways to run it {#modes}\n",
            "# Why `{#anchor}` outlives a restructure\n",
            "# Anchors are written {#id}, in prose\n",
            "# Install & build {#build}\n",
            "# What `init` sets up\n",
            "# Sets like {#1, #2}\n",
        ] {
            let r = render_doc(md, "fallback", &no_pages(), None);
            let inner = r
                .html
                .split_once("<h1")
                .and_then(|(_, rest)| rest.split_once('>'))
                .and_then(|(_, rest)| rest.split_once("</h1>"))
                .map(|(text, _)| text.to_owned())
                .unwrap_or_default();
            // The heading carries inline markup (`<code>`, emphasis); the title
            // is the words inside it. Dropping the tags — and nothing else, so
            // entities still have to match — is what makes them comparable.
            let mut heading = String::new();
            let mut depth = 0usize;
            for c in inner.chars() {
                match c {
                    '<' => depth += 1,
                    '>' => depth = depth.saturating_sub(1),
                    _ if depth == 0 => heading.push(c),
                    _ => {}
                }
            }
            assert_eq!(
                heading,
                escape_html(&r.title),
                "title and heading disagree for {md:?}: {}",
                r.html
            );
        }
    }

    #[test]
    fn the_title_is_the_heading_the_reader_sees() {
        // Inline markup contributes its text, not its punctuation — the same
        // rule the heading's own id already follows.
        let code = render_doc("# What `init` sets up\n", "fallback", &no_pages(), None);
        assert_eq!(code.title, "What init sets up");
        // A line scan called this document's title `Not a title`; the parser
        // knows a fenced hash is not a heading at all.
        let fenced = render_doc(
            "```\n# Not a title\n```\n\n# The real one\n",
            "fallback",
            &no_pages(),
            None,
        );
        assert_eq!(fenced.title, "The real one");
        // And a heading spelled the other way is still a heading: the page shows
        // an `<h1>`, so the tab has to show its words rather than the file stem.
        let setext = render_doc("Underlined\n==========\n", "fallback", &no_pages(), None);
        assert!(
            setext.html.contains("<h1 id=\"underlined\">"),
            "{}",
            setext.html
        );
        assert_eq!(setext.title, "Underlined");
    }

    #[test]
    fn a_document_with_no_h1_falls_back_and_the_fallback_is_used_verbatim() {
        // The fallback is the caller's string, not markdown: it is never parsed,
        // so it cannot be stripped and cannot leak markup it does not contain.
        // Callers pass a file stem or a declared slug.
        let none = render_site_page(
            "---\nsite-page: modes\n---\n\nNo heading at all.\n",
            "The five ways to run it",
            &nav(),
            "modes.html",
            &no_pages(),
            None,
        );
        assert_eq!(none.title, "The five ways to run it");
        assert!(
            none.html
                .contains("<title>The five ways to run it — Roteiro</title>"),
            "{}",
            none.html
        );
        // An H1 with nothing in it names nothing, so it defers to the fallback
        // rather than emitting `<title> — Roteiro</title>`.
        let empty = render_doc("#\n\nBody.\n", "build-plan", &no_pages(), None);
        assert_eq!(empty.title, "build-plan");
        // A lower heading is not the document's title.
        let sub = render_doc("## Only a section {#s}\n", "build-plan", &no_pages(), None);
        assert_eq!(sub.title, "build-plan");
    }

    #[test]
    fn a_site_page_carries_the_bar_with_itself_marked() {
        let r = render_site_page(
            "---\nsite-page: modes\n---\n\n# The five ways to run it\n\nSee [[docs/adr/0019-remote.md]].\n",
            "fallback",
            &nav(),
            "modes.html",
            &no_pages(),
            None,
        );
        assert_eq!(r.title, "The five ways to run it");
        // Frontmatter is chrome for the graph, not content for the reader.
        assert!(!r.html.contains("site-page"), "{}", r.html);
        // The current page is unlinked and marked; its neighbour is a link.
        assert!(
            r.html
                .contains("<span aria-current=\"page\">Modes &amp; Co</span>"),
            "{}",
            r.html
        );
        assert!(r.html.contains("<a href=\"./\">Home</a>"), "{}", r.html);
        // A root-level page: assets and ADR links resolve from the site root.
        assert!(r.html.contains("href=\"./style.css\""), "{}", r.html);
        assert!(
            r.html
                .contains("<a href=\"adr/0019-remote.html\">ADR-0019</a>"),
            "{}",
            r.html
        );
    }

    #[test]
    fn the_bar_is_plain_anchors_and_escapes_its_labels() {
        let bar = render_nav(&nav(), "nothing.html");
        assert!(bar.starts_with("<nav class=\"sitenav\">"), "{bar}");
        // Nothing marked when the current page is not in the bar — a preview of
        // an unlisted page, not an error.
        assert!(!bar.contains("aria-current"), "{bar}");
        assert!(bar.contains("Modes &amp; Co"), "escaped label: {bar}");
        // No script: the site has no build step and this must not introduce one.
        assert!(!bar.contains("<script"), "{bar}");
    }

    #[test]
    fn a_link_resolves_to_the_page_the_site_actually_serves() {
        // Issue #446: four ADRs link `../BUILD_PLAN_V2.md`, which is correct in
        // the repository. Published under a `site-page:` slug, that document is
        // served as `build-plan-v2.html` — so rewriting the link to its own stem
        // aims it at a page that is never emitted.
        let mut pages = PublishedPages::new();
        pages.publish("BUILD_PLAN_V2.md", "build-plan-v2.html");
        let html = render_markdown("See [V2](../BUILD_PLAN_V2.md).\n", "", &pages, None, 0);
        assert!(
            html.contains("href=\"../build-plan-v2.html\""),
            "served name, and the link's own hop kept: {html}"
        );
        // A fragment survives the substitution.
        let frag = render_markdown("[s](../BUILD_PLAN_V2.md#stage-21)\n", "", &pages, None, 0);
        assert!(
            frag.contains("href=\"../build-plan-v2.html#stage-21\""),
            "{frag}"
        );
        // An unpublished document still falls back to its stem, unchanged.
        let other = render_markdown("[x](../REVIEW_CHECKLIST.md)\n", "", &pages, None, 0);
        assert!(
            other.contains("href=\"../REVIEW_CHECKLIST.html\""),
            "{other}"
        );
    }

    #[test]
    fn a_file_name_two_documents_claim_is_left_alone() {
        // Guessing which one a link meant would silently point it at the wrong
        // page — worse than the 404 the lookup exists to remove.
        let mut pages = PublishedPages::new();
        pages.publish("GUIDE.md", "guide.html");
        pages.publish("GUIDE.md", "other-guide.html");
        let html = render_markdown("[g](GUIDE.md)\n", "", &pages, None, 0);
        assert!(html.contains("href=\"GUIDE.html\""), "unrewritten: {html}");
        // Re-publishing the *same* target is not a conflict.
        let mut same = PublishedPages::new();
        same.publish("GUIDE.md", "guide.html");
        same.publish("GUIDE.md", "guide.html");
        let html = render_markdown("[g](GUIDE.md)\n", "", &same, None, 0);
        assert!(html.contains("href=\"guide.html\""), "{html}");
    }

    #[test]
    fn the_dialect_is_not_extended_here() {
        // `options` exists to hold renderer-specific rationale, not to add
        // flags. This reads as a tautology against the body as written, and that
        // is exactly its job: it has no failure mode until someone gives the
        // body one, and that single edit is the only thing the comment beside it
        // can ask against rather than prevent.
        //
        // Note what it pins — the *dialect*, not the shape of the body. A
        // rewrite that still yields this option set is harmless and keeps
        // passing; every divergence that would change what a heading's text is
        // fails. That is the invariant worth holding, and it is a wider one than
        // "stay a single delegation".
        assert_eq!(options(), rto_graph::markdown_dialect());
    }

    /// A source base for a document in `dir`, at a fixed sha.
    fn source(dir: &str) -> SourceBase {
        SourceBase::new(Some("https://github.com/o/r/blob/abc123"), dir).expect("base")
    }

    #[test]
    fn a_link_out_of_the_site_goes_to_the_repository() {
        // Issue #456: the Build Plan cites code as evidence — correct in a
        // checkout, dead on the site, which publishes documents and not source.
        let base = source("docs");
        let html = render_markdown(
            "[sync](../crates/rto-graph/src/sync.rs) and [wf](../.github/workflows/website.yml)\n",
            "adr/",
            &no_pages(),
            Some(&base),
            0,
        );
        assert!(
            html.contains(
                "href=\"https://github.com/o/r/blob/abc123/crates/rto-graph/src/sync.rs\""
            ),
            "resolved against the document's own directory: {html}"
        );
        assert!(
            html.contains(
                "href=\"https://github.com/o/r/blob/abc123/.github/workflows/website.yml\""
            ),
            "a dotted directory is a directory, not a `.` segment: {html}"
        );
        // A line anchor is the author's, and travels.
        let frag = render_markdown(
            "[l](../crates/roteiro/src/init.rs#L12)\n",
            "adr/",
            &no_pages(),
            Some(&base),
            0,
        );
        assert!(
            frag.contains("blob/abc123/crates/roteiro/src/init.rs#L12\""),
            "{frag}"
        );
    }

    #[test]
    fn a_link_that_stays_inside_the_site_is_left_alone() {
        // The whole discrimination is the hop count: `ask.html` and `adr/` are
        // written *for* the site and are correct there, so rewriting them to the
        // repository would break links that work today.
        let base = source("docs");
        let html = render_markdown(
            "[a](ask.html), [d](adr/), [s](./style.css) and [r](/abs.html)\n",
            "adr/",
            &no_pages(),
            Some(&base),
            0,
        );
        assert!(!html.contains("github.com"), "none rewritten: {html}");
        for href in [
            "\"ask.html\"",
            "\"adr/\"",
            "\"./style.css\"",
            "\"/abs.html\"",
        ] {
            assert!(html.contains(href), "{href} kept verbatim: {html}");
        }
    }

    #[test]
    fn an_adr_may_climb_one_level_and_still_be_inside_the_site() {
        // An ADR page is served at `adr/<slug>.html`, so `../x` lands at the site
        // root. Treating that as an escape would send every ADR's back-link to
        // GitHub. The second hop does leave.
        let base = source("docs/adr");
        let inside = render_markdown("[b](../build-plan.html)\n", "", &no_pages(), Some(&base), 1);
        assert!(!inside.contains("github.com"), "{inside}");
        let outside = render_markdown("[c](../../Cargo.toml)\n", "", &no_pages(), Some(&base), 1);
        assert!(
            outside.contains("href=\"https://github.com/o/r/blob/abc123/Cargo.toml\""),
            "{outside}"
        );
    }

    #[test]
    fn a_published_page_beats_the_escape_rule() {
        // Order matters: #446's lookup runs first, so a document reached by a
        // path that climbs out of its own directory still lands on the page the
        // site publishes it as, rather than being handed to the repository.
        let mut pages = PublishedPages::new();
        pages.publish("BUILD_PLAN_V2.md", "build-plan-v2.html");
        let base = source("website/pages");
        let html = render_markdown(
            "[v2](../../docs/BUILD_PLAN_V2.md)\n",
            "adr/",
            &pages,
            Some(&base),
            0,
        );
        assert!(
            html.contains("href=\"../../docs/build-plan-v2.html\""),
            "still the site's page: {html}"
        );
        assert!(!html.contains("github.com"), "{html}");
    }

    #[test]
    fn without_a_source_base_the_link_is_left_as_authored() {
        // No `origin`, or one that maps to no web view. Leaving the link is the
        // deliberate choice: it stays correct in a checkout, and a rewrite that
        // silently produced a broken URL would be worse than the link it replaced.
        assert_eq!(SourceBase::new(None, "docs"), None);
        let html = render_markdown(
            "[s](../crates/rto-graph/src/sync.rs)\n",
            "adr/",
            &no_pages(),
            None,
            0,
        );
        assert!(
            html.contains("href=\"../crates/rto-graph/src/sync.rs\""),
            "{html}"
        );
    }

    #[test]
    fn the_bar_on_the_landing_page_is_replaced_rather_than_maintained() {
        // Issue #508. The stale copy is overwritten wholesale, so there is no
        // second list left to drift out of `site-order`.
        let stale = "<h1>Roteiro</h1>\n<nav class=\"sitenav\">\n<a href=\"old.html\">Old</a>\n\
                     </nav>\n<p>after</p>\n";
        let out = replace_site_nav(stale, &nav(), "./").expect("marker found");
        assert!(
            !out.contains("old.html"),
            "the hand-written list is gone: {out}"
        );
        assert!(
            out.contains("<a href=\"modes.html\">Modes &amp; Co</a>"),
            "the computed bar took its place: {out}"
        );
        assert!(
            out.starts_with("<h1>Roteiro</h1>\n") && out.ends_with("<p>after</p>\n"),
            "only the bar is touched: {out}"
        );
        // A page that claims no bar is left alone rather than failing: every
        // `render docs` fixture writes a one-line landing page.
        assert_eq!(replace_site_nav("<h1>Home</h1>\n", &nav(), "./"), None);
    }

    #[test]
    fn site_pages_render_deterministically() {
        let md = "---\nsite-page: a\n---\n\n# A\n\n## S\n";
        assert_eq!(
            render_site_page(md, "f", &nav(), "a.html", &no_pages(), None),
            render_site_page(md, "f", &nav(), "a.html", &no_pages(), None)
        );
    }

    #[test]
    fn rendering_is_deterministic() {
        assert_eq!(
            render_adr(ADR, "f", &no_pages(), None),
            render_adr(ADR, "f", &no_pages(), None)
        );
    }
}