Skip to main content

rto_render/
obsidian.rs

1//! The Obsidian-vault renderer: each graph node becomes a markdown note whose
2//! edges are `[[wikilinks]]`, so the provenance-tagged graph is browsable in
3//! Obsidian's graph view. Notes carry frontmatter `tags` (`roteiro/kind/*`,
4//! `roteiro/lang/*`, `roteiro/status/*`) so the graph is colourable/filterable —
5//! edge provenance is shown per-link in the body — surface the node's text as the
6//! knowledge base, show an ADR's status, and (when the repository's web host is
7//! known) a clickable **Source** link to the file.
8//!
9//! That text is the node's captured `meta.content` (a doc comment, PDF or image
10//! text) *except* where the caller supplies a full `body` — which it does for
11//! prose documents, because `meta.content` is an embedding budget and a note
12//! rendered from it is the document capped at 1500 characters and collapsed onto
13//! one line. See [`note_body`].
14//!
15//! A generated `_Home` note is the overview: what was
16//! scanned, counts by kind, provenance breakdown, ADR statuses, intent-debt (with
17//! the files it is densest in), an inventory of secret-**named** config keys and
18//! their redaction state, and the most depended-on symbols by directed call
19//! fan-in.
20//! Built from the same [`Explanation`] the query surface returns, so the vault
21//! and the CLI agree.
22
23use std::fmt::Write as _;
24
25use rto_graph::Explanation;
26
27/// Filename of the generated overview note (sorts first in the file list).
28pub const HOME_NOTE: &str = "_Home.md";
29
30/// A rendered vault note: its filename (with `.md`) and markdown content.
31#[derive(Debug, Clone, PartialEq, Eq)]
32pub struct VaultNote {
33    /// Filename including the `.md` extension.
34    pub filename: String,
35    /// Markdown content.
36    pub content: String,
37}
38
39/// Map a node key to a filesystem- and wikilink-safe note stem that is **unique
40/// per key even after case folding**.
41///
42/// A name is a lowercased, readable *hint* slugged from the key, followed by an
43/// unconditional 64-bit FNV-1a hash of the whole, exact key written as 16 hex
44/// digits — `<hint>-<16 hex digits>`. Characters outside `[a-z0-9._-]` collapse
45/// to a single `-` in the hint; the hash carries everything the hint threw away.
46///
47/// **The hash is the unconditional part, not the hint.** A key of nothing but
48/// separators slugs to an empty hint, and the name is then the bare 16 hex
49/// digits — no hint, and no `-` to join it to. The two forms cannot be confused
50/// for one another, which is what makes the exception safe rather than a second
51/// naming rule: a hinted name is at least 18 characters and contains a `-`,
52/// and a bare one is exactly 16 and contains none. That is argued again at the
53/// branch itself, and asserted by
54/// `every_name_carries_the_hash_however_short_the_key`.
55///
56/// # Why the hash is unconditional (issue #574)
57///
58/// It used to be applied only when the slug overran the filename limit, and the
59/// slug alone was lossy twice over. Measured on this repository — 8,239 nodes
60/// rendering to 8,135 notes, 104 of them silently overwritten:
61///
62/// | mechanism | lost | where |
63/// | --- | --- | --- |
64/// | every character outside the safe set becomes `-` and runs collapse, so `…cytoscape.min.js#$a` and `…cytoscape.min.js#a` are one name | 9 | everywhere |
65/// | macOS and Windows fold filename case, so `…#A` and `…#a` are two *names* but one *file* | 95 | macOS, Windows |
66///
67/// The second mechanism is the trap. A lossless-but-case-sensitive encoding
68/// fixes the 9, verifies clean on Linux CI, and still loses 95 notes on a Mac.
69/// So the requirement is stated after folding:
70///
71/// ```text
72/// lower(note_name(k1)) == lower(note_name(k2))  implies  k1 == k2
73/// ```
74///
75/// This matters more than lossiness in a cache would, because the note names are
76/// the vault's **only** stable interface: `reset_vault_dir` deletes and rebuilds
77/// the whole directory on every render, so the one thing that survives a render
78/// is a user's own note *outside* the vault linking in by name (issue #442).
79///
80/// # The trade taken
81///
82/// Two decisions, and what each bought:
83///
84/// **The hint is lowercased rather than case-preserved.** Case-preserving would
85/// also satisfy the requirement — the hash differs for `#A` and `#a`, so the two
86/// names differ in their suffix and stay distinct under folding. It was rejected
87/// because lowercasing makes `note_name(k) == note_name(k).to_lowercase()` an
88/// invariant of the function, and *that* collapses the folded property into the
89/// literal one: there is then no way to write a version of this that is green on
90/// Linux and lossy on macOS, which is the defect shape this repository keeps
91/// finding. The cost is that `parseHTTPHeader` reads as `parsehttpheader`. That
92/// is affordable precisely because the hint is a hint — once a 17-character
93/// suffix is mandatory the name is not something anyone types from memory, so
94/// its job is to be recognisable in a file list, not to be transcribed.
95///
96/// **Readability was spent, deliberately.** Every name grows by 17 characters and
97/// hand-writing a link now needs Obsidian's autocomplete. The alternatives that
98/// keep names short — hashing only the keys observed to collide — make the *set*
99/// of collisions platform-dependent, so one key would get one filename on macOS
100/// and another on Linux and a synced vault would churn. A name that is uglier
101/// everywhere beats a name that is different per platform.
102///
103/// The mapping is not reversible (the hint is lossy and the hash is one-way), but
104/// it does not need to be: every note's frontmatter carries `key:` verbatim, so
105/// name → key is recoverable from the vault itself, which is the direction a
106/// reader actually needs.
107///
108/// # What "unique" rests on
109///
110/// Equal names imply equal hashes, not equal keys — this is a 64-bit hash, not a
111/// proof. Over this repository's 8,239 keys there is no collision, and the
112/// birthday bound at that size is about 2e-12. Should one ever occur it is
113/// *reported*, not silent: `NoteNames` in the render path claims every filename
114/// case-insensitively and warns on a repeat. What is proved outright is the
115/// folding half — the output is lowercase by construction, so case folding is the
116/// identity on it.
117#[must_use]
118pub fn note_name(key: &str) -> String {
119    // Keep the whole stem well under the 255-byte filename limit (leaving room
120    // for ".md"). The hint is ASCII, so byte length equals char count and slicing
121    // is safe.
122    const MAX: usize = 200;
123    // '-' plus the 16 hex digits of the hash.
124    const SUFFIX: usize = 17;
125    const HINT: usize = MAX - SUFFIX;
126
127    let mut hint = String::with_capacity(key.len());
128    let mut prev_dash = false;
129    for c in key.chars() {
130        if c.is_ascii_alphanumeric() || matches!(c, '.' | '_' | '-') {
131            hint.push(c.to_ascii_lowercase());
132            prev_dash = false;
133        } else if !prev_dash {
134            hint.push('-');
135            prev_dash = true;
136        }
137    }
138    let hint = hint.trim_matches('-');
139    // Truncation is only ever cosmetic now: the hash, not the hint, is what keeps
140    // a 300-character grouped `use` distinct from its neighbour.
141    let hint = hint[..hint.len().min(HINT)].trim_end_matches('-');
142    let hash = fnv1a64(key.as_bytes());
143    if hint.is_empty() {
144        // A key of nothing but separators. Bare hex, and it cannot be confused
145        // with a hinted name: those are `<hint>-<16 hex>`, so at least 18
146        // characters, and this is exactly 16 with no `-` in it.
147        format!("{hash:016x}")
148    } else {
149        format!("{hint}-{hash:016x}")
150    }
151}
152
153/// FNV-1a (64-bit) — a dependency-free, deterministic hash carrying everything
154/// [`note_name`]'s hint discards. No cryptographic properties needed: nothing
155/// here defends against a chosen collision, only against an accidental one.
156///
157/// 64 bits rather than fewer because the cost of a collision is exactly the
158/// defect this suffix exists to fix — a note silently overwritten. At 8k keys a
159/// 32-bit hash collides about 0.8% of the time and a 48-bit one about 1e-5;
160/// 64 bits is 2e-12, and stays under 1e-10 for a workspace vault an order of
161/// magnitude larger.
162fn fnv1a64(bytes: &[u8]) -> u64 {
163    let mut hash: u64 = 0xcbf2_9ce4_8422_2325;
164    for &b in bytes {
165        hash ^= u64::from(b);
166        hash = hash.wrapping_mul(0x0000_0100_0000_01b3);
167    }
168    hash
169}
170
171/// Emit `value` as a YAML **double-quoted** scalar, `"`-delimited and escaped so
172/// it parses back to exactly `value`.
173///
174/// The one escaping rule for this module's frontmatter. It exists because the
175/// three hand-rolled variants it replaced disagreed with each other — `key:` and
176/// `project:` turned a `"` into an apostrophe, and `path:` escaped nothing — and
177/// two of the three could emit YAML that does not mean what it says:
178///
179/// | value | was emitted | parsed back as |
180/// | --- | --- | --- |
181/// | `foo\bar` | `"foo\bar"` | `foo<BS>ar` — `\b` is YAML's **backspace** escape |
182/// | `foo\dir` | `"foo\dir"` | *parse error* — `\d` is not a YAML escape |
183/// | `say"hi".rs` | `"say"hi".rs"` | *parse error* — the scalar ends at the `"` |
184///
185/// The first is the dangerous one: seven characters silently become six, and
186/// nothing anywhere reports it. The other two cost the reader every property on
187/// the note, because Obsidian parses this block as the note's properties and a
188/// block that does not parse yields no properties at all rather than an error.
189///
190/// All three inputs are legal path components on Linux and macOS. None occurs in
191/// this repository today, so this is a latent defect rather than an observed one.
192///
193/// Escapes, per YAML 1.2 §7.3.1: the two structural characters `\` and `"`, then
194/// anything a parser is not obliged to accept literally — C0 controls, `DEL`, the
195/// C1 range, and the three separators (`U+2028`, `U+2029`, `U+FEFF`) that some
196/// parsers treat as line breaks. Short escapes where YAML defines one, so the
197/// common cases stay readable, and `\uXXXX` otherwise.
198fn yaml_double_quoted(value: &str) -> String {
199    let mut out = String::with_capacity(value.len() + 2);
200    out.push('"');
201    for ch in value.chars() {
202        match ch {
203            '\\' => out.push_str(r"\\"),
204            '"' => out.push_str("\\\""),
205            '\n' => out.push_str(r"\n"),
206            '\r' => out.push_str(r"\r"),
207            '\t' => out.push_str(r"\t"),
208            '\u{0}' => out.push_str(r"\0"),
209            '\u{7}' => out.push_str(r"\a"),
210            '\u{8}' => out.push_str(r"\b"),
211            '\u{b}' => out.push_str(r"\v"),
212            '\u{c}' => out.push_str(r"\f"),
213            '\u{1b}' => out.push_str(r"\e"),
214            // Everything else a YAML parser may reject or fold: the rest of C0,
215            // DEL, the C1 range, and the separators that can read as line breaks.
216            c if (c < ' ')
217                || c == '\u{7f}'
218                || ('\u{80}'..='\u{9f}').contains(&c)
219                || matches!(c, '\u{2028}' | '\u{2029}' | '\u{feff}') =>
220            {
221                let _ = write!(out, "\\u{:04x}", c as u32);
222            }
223            c => out.push(c),
224        }
225    }
226    out.push('"');
227    out
228}
229
230/// Emit `value` in YAML **plain** (unquoted) style when that round-trips, and as
231/// [`yaml_double_quoted`] when it would not.
232///
233/// For the frontmatter fields that are written bare today — `kind`, `lang`,
234/// `status`. Those are constrained by *today's* producers (an ADR's status is
235/// validated against the house states; kinds and languages come from extraction),
236/// but `roteiro load` installs a caller-supplied graph artifact whose nodes carry
237/// whatever JSON they carry, so "the producer is careful" is not a property this
238/// renderer can rely on. A `status:` of `Accepted: superseded by 0012` emitted
239/// bare is a parse error, and `Accepted # pending` silently truncates to
240/// `Accepted`.
241///
242/// Escalating only when needed is what keeps the bytes of an existing vault
243/// unchanged — every `kind`, `lang` and `status` in this repository is plain-safe
244/// and stays bare. [`is_plain_safe`] is deliberately stricter than YAML's plain
245/// grammar for the same reason it is safe: a value it rejects is merely quoted.
246fn yaml_scalar(value: &str) -> String {
247    if is_plain_safe(value) {
248        value.to_owned()
249    } else {
250        yaml_double_quoted(value)
251    }
252}
253
254/// Whether `value` can be written as a bare YAML scalar and read back unchanged.
255///
256/// A conservative allowlist rather than YAML's actual plain-scalar grammar, which
257/// is subtle enough (indicator characters, `: ` and ` #` only in some positions,
258/// leading and trailing space, implicit typing) that implementing it is how the
259/// bug this replaces gets written a second time. Getting this wrong in the
260/// strict direction costs a pair of quotation marks; getting it wrong in the
261/// permissive direction costs the note's properties.
262///
263/// So: a leading ASCII letter, then letters, digits, `_`, `-`, `.` and `/` — which
264/// covers every kind, language and status this renderer emits — and never a word
265/// YAML resolves to a boolean or null. That last exclusion is not hypothetical:
266/// `no` is the ISO 639-1 code for Norwegian, and YAML 1.1 parsers read a bare `no`
267/// as `false`.
268fn is_plain_safe(value: &str) -> bool {
269    const NOT_STRINGS: [&str; 11] = [
270        "true", "false", "yes", "no", "on", "off", "null", "nil", "none", "y", "n",
271    ];
272    !value.is_empty()
273        && value.starts_with(|c: char| c.is_ascii_alphabetic())
274        && value
275            .chars()
276            .all(|c| c.is_ascii_alphanumeric() || matches!(c, '_' | '-' | '.' | '/'))
277        && !NOT_STRINGS.contains(&value.to_ascii_lowercase().as_str())
278}
279
280/// Which vault a note is being rendered into: a single project's, or one member
281/// of a **workspace** vault spanning several repositories.
282///
283/// This is the whole of the workspace-vault naming rule, in one place. Node keys
284/// are **repository-relative** (`file:README.md` names no repo), so every member
285/// of a workspace produces the same note name for its `README.md` and one would
286/// silently overwrite the rest. Qualifying the key with its project fixes that.
287///
288/// [`VaultScope::PROJECT`] (`project: None`) is not a degenerate case but the
289/// contract: it makes every name in this module reduce to exactly [`note_name`]
290/// of the bare key, with nothing qualified and no `project:` frontmatter.
291///
292/// That reduction is *still* the promise; what it no longer implies is stability
293/// against `main`. #570 could say "a single-project vault's names do not move",
294/// because the only thing moving them would have been workspace qualification.
295/// #574 moves them all, on purpose: the old names were not injective under
296/// filename case folding and the vault lost 104 notes to that. The promise here
297/// was always about **this axis** — turning workspace mode on must not rename a
298/// project's notes — and it holds unchanged. See [`note_name`] for the rename and
299/// what it bought.
300#[derive(Debug, Clone, Copy)]
301pub struct VaultScope<'a> {
302    /// The member project this note belongs to, qualifying its name as
303    /// `<project>::<key>` — the same form ADR-0009's cross-repo links already use.
304    /// `None` ⇒ a single-project vault, and names are unqualified exactly as
305    /// before.
306    pub project: Option<&'a str>,
307    /// The workspace's member project names. An external-ref placeholder whose
308    /// target names one of these is a cross-repo edge the vault can actually
309    /// follow, so it is rendered as a link straight to that member's note. Empty
310    /// for a single-project vault.
311    pub members: &'a std::collections::BTreeSet<String>,
312}
313
314/// The empty member set backing [`VaultScope::PROJECT`] — a single-project vault
315/// has no other members to resolve a cross-repo reference against.
316static NO_MEMBERS: std::collections::BTreeSet<String> = std::collections::BTreeSet::new();
317
318impl VaultScope<'_> {
319    /// A single-project vault: names are unqualified, and no cross-repo reference
320    /// resolves. Every name this produces is byte-identical to [`note_name`] of
321    /// the bare key — see the type's documentation for why that reduction is
322    /// load-bearing, and for what it does *not* promise.
323    pub const PROJECT: Self = Self {
324        project: None,
325        members: &NO_MEMBERS,
326    };
327}
328
329impl Default for VaultScope<'_> {
330    fn default() -> Self {
331        Self::PROJECT
332    }
333}
334
335impl VaultScope<'_> {
336    /// Whether an external-ref placeholder `key` is one this vault resolves for
337    /// itself — its target names a member, so every edge to it points at the real
338    /// note and the placeholder need not be rendered at all.
339    ///
340    /// The single rule behind both halves of that: [`link_target`] redirects
341    /// exactly the keys this accepts, and the caller skips writing exactly the
342    /// notes this accepts. They cannot disagree.
343    #[must_use]
344    pub fn redirects_external_ref(&self, key: &str) -> bool {
345        key.strip_prefix("extref:")
346            .and_then(rto_graph::parse_qualified)
347            .is_some_and(|(project, _)| self.members.contains(project))
348    }
349}
350
351/// The note name for a node `key` owned by `scope`'s project.
352///
353/// In a single-project vault (`scope.project == None`) this *is* [`note_name`].
354/// In a workspace vault it is [`note_name`] of the project-qualified key
355/// `<project>::<key>` — reusing ADR-0009's qualified form rather than inventing a
356/// second one, which is what lets a cross-repo external-ref target (already
357/// stored qualified) map to its note by the very same call.
358#[must_use]
359pub fn scoped_note_name(scope: &VaultScope<'_>, key: &str) -> String {
360    match scope.project {
361        None => note_name(key),
362        Some(project) => note_name(&format!("{project}::{key}")),
363    }
364}
365
366/// The note an edge pointing at `key` should link to.
367///
368/// Almost always [`scoped_note_name`]. The exception is the one cross-repo edge
369/// the graph already models: a spoke's inferred link to a hub is stored as an
370/// edge to a **local external-ref placeholder** (`extref:<project>::<key>`,
371/// [`rto_graph::external_ref_key`]) because store integrity requires both ends of
372/// an edge in one store. A workspace vault holds both repos' notes, so when the
373/// placeholder's target names a member the link is pointed at the **real** note
374/// instead of the stand-in.
375///
376/// This invents no edge. It renders the edge that is there, following the
377/// placeholder exactly as [`rto_graph::Workspace::follow_external_ref`] does at
378/// query time — the cross-repo graph has only ever been *rendered* one repo at a
379/// time.
380fn link_target(scope: &VaultScope<'_>, key: &str) -> String {
381    if scope.redirects_external_ref(key) {
382        // `note_name(qualified)` is by construction the same string
383        // `scoped_note_name` produces for that member's own copy of the node.
384        // `strip_prefix`, not `trim_start_matches`: the latter strips the prefix
385        // repeatedly, which would mangle a target that legitimately starts with it.
386        return note_name(key.strip_prefix("extref:").unwrap_or(key));
387    }
388    scoped_note_name(scope, key)
389}
390
391/// Render a node's [`Explanation`] into an Obsidian note: YAML frontmatter (with
392/// `tags` for the graph view and an ADR's `status`), a clickable **Source** link
393/// (when `source_base` — a web "blob" base like
394/// `https://github.com/org/repo/blob/<sha>` — is known and the node has a path),
395/// the content as the knowledge base, and its edges as provenance-labelled
396/// wikilinks.
397///
398/// `body` is the node's **full source text**, which only the caller can fetch:
399/// this function is a pure function of the `Explanation`, and an `Explanation`
400/// carries no repository, store or blob. When it is `Some`, it replaces
401/// `meta.content` in the note's `## Content` section — see [`note_body`] for why
402/// replacing is the only correct combination of the two.
403#[must_use]
404pub fn render_note(ex: &Explanation, source_base: Option<&str>, body: Option<&str>) -> VaultNote {
405    render_note_scoped(ex, source_base, body, &VaultScope::PROJECT)
406}
407
408/// [`render_note`], for one member of a **workspace** vault: identical except
409/// that the note's own name and every link it emits are resolved through `scope`
410/// (see [`VaultScope`]).
411///
412/// With [`VaultScope::PROJECT`] this is [`render_note`] byte for byte, which is
413/// how the single-project vault's compatibility promise is kept by construction
414/// rather than by a parallel code path that has to be kept in step.
415#[must_use]
416pub fn render_note_scoped(
417    ex: &Explanation,
418    source_base: Option<&str>,
419    body: Option<&str>,
420    scope: &VaultScope<'_>,
421) -> VaultNote {
422    let meta = &ex.meta;
423    let status = meta.get("status").and_then(|v| v.as_str());
424    let content = note_body(meta.get("content").and_then(|v| v.as_str()), body);
425
426    let mut c = String::new();
427    c.push_str("---\n");
428    let _ = writeln!(c, "key: {}", yaml_double_quoted(&ex.node.key));
429    let _ = writeln!(c, "kind: {}", yaml_scalar(ex.node.kind.as_str()));
430    // Which member this note came from. Absent in a single-project vault, where
431    // it would be one constant repeated on every note — and where adding it would
432    // change every note's bytes.
433    if let Some(project) = scope.project {
434        let _ = writeln!(c, "project: {}", yaml_double_quoted(project));
435    }
436    if let Some(path) = &ex.node.path {
437        let _ = writeln!(c, "path: {}", yaml_double_quoted(path));
438    }
439    if let Some(lang) = &ex.node.lang {
440        let _ = writeln!(c, "lang: {}", yaml_scalar(lang));
441    }
442    if let Some(status) = status {
443        let _ = writeln!(c, "status: {}", yaml_scalar(status));
444    }
445    // Nested tags group in Obsidian's tag pane and colour the graph view.
446    c.push_str("tags:\n");
447    let _ = writeln!(c, "  - roteiro/kind/{}", tag_slug(&ex.node.kind));
448    // Colours the graph view by member, which is the one thing a workspace vault
449    // is for and a per-project vault has no use for.
450    if let Some(project) = scope.project {
451        let _ = writeln!(c, "  - roteiro/project/{}", tag_slug(project));
452    }
453    if let Some(lang) = &ex.node.lang {
454        let _ = writeln!(c, "  - roteiro/lang/{}", tag_slug(lang));
455    }
456    if let Some(status) = status {
457        let _ = writeln!(c, "  - roteiro/status/{}", tag_slug(status));
458    }
459    c.push_str("---\n\n");
460
461    let _ = writeln!(c, "# {}", ex.node.name);
462    if let Some(status) = status {
463        let _ = writeln!(c, "\n> **Status:** {status}");
464    }
465
466    // A clickable link to the file this node comes from. An absolute URL, so it
467    // works from the downloaded vault too (which has no repo files beside it).
468    if let (Some(base), Some(path)) = (source_base, ex.node.path.as_deref()) {
469        let _ = writeln!(
470            c,
471            "\n**Source:** [`{path}`]({}/{path})",
472            base.trim_end_matches('/')
473        );
474    }
475
476    // The knowledge base: the full source text, or the captured doc comment /
477    // prose / PDF / image text.
478    if let Some(content) = content.map(str::trim).filter(|s| !s.is_empty()) {
479        c.push_str("\n## Content\n\n");
480        c.push_str(content);
481        c.push('\n');
482    }
483
484    if !ex.outgoing.is_empty() {
485        c.push_str("\n## Outgoing\n\n");
486        for e in &ex.outgoing {
487            let _ = writeln!(
488                c,
489                "- {} ({}){} → [[{}]]",
490                e.kind,
491                e.provenance,
492                confidence(e.confidence),
493                link_target(scope, &e.node)
494            );
495        }
496    }
497    if !ex.incoming.is_empty() {
498        c.push_str("\n## Incoming\n\n");
499        for e in &ex.incoming {
500            let _ = writeln!(
501                c,
502                "- [[{}]] {} ({}){} →",
503                link_target(scope, &e.node),
504                e.kind,
505                e.provenance,
506                confidence(e.confidence)
507            );
508        }
509    }
510
511    VaultNote {
512        filename: format!("{}.md", scoped_note_name(scope, &ex.node.key)),
513        content: c,
514    }
515}
516
517/// Choose the text a note shows: the caller's full `body` when it has one, else
518/// the node's stored `content`.
519///
520/// The two are **not** complementary, they are the same text at two fidelities,
521/// so a note shows one of them and never both. `meta.content` is an embedding
522/// budget — extraction caps it (1500 chars) and collapses every whitespace run to
523/// a single space, which is right for a store that ships with the graph and wrong
524/// for a note: a 23 KB document arrives as one 1500-character line with every
525/// heading, table and code fence flattened into it. Where the caller can supply
526/// the source, that is what a reader wants; appending the capped rendering
527/// underneath it would only restate its first 6% badly.
528fn note_body<'a>(content: Option<&'a str>, body: Option<&'a str>) -> Option<&'a str> {
529    body.or(content)
530}
531
532/// `" (0.82)"` for an inferred edge's confidence, else empty.
533fn confidence(c: Option<f64>) -> String {
534    c.map_or_else(String::new, |c| format!(" ({c:.2})"))
535}
536
537/// A tag-safe slug: lowercase, non-alphanumeric runs → `-`. Keeps Obsidian tags
538/// (`roteiro/kind/adr-section`) valid and stable.
539fn tag_slug(s: &str) -> String {
540    let mut out = String::with_capacity(s.len());
541    let mut prev_dash = false;
542    for ch in s.chars() {
543        if ch.is_ascii_alphanumeric() {
544            out.push(ch.to_ascii_lowercase());
545            prev_dash = false;
546        } else if !prev_dash {
547            out.push('-');
548            prev_dash = true;
549        }
550    }
551    out.trim_matches('-').to_owned()
552}
553
554/// One ADR in the overview, with its lifecycle status.
555#[derive(Debug, Clone)]
556pub struct AdrEntry {
557    /// The ADR node key (`adr:<id>`).
558    pub key: String,
559    /// The ADR title.
560    pub name: String,
561    /// Lifecycle status (`Accepted`, …), if recorded.
562    pub status: Option<String>,
563}
564
565/// The `_Home` overview's config-secret inventory figures.
566///
567/// Counts and file paths only — deliberately not the key names, which belong in
568/// `roteiro config-secrets` where the caveat can be stated at length. A vault note
569/// is read casually and out of context, which is exactly the wrong place for a
570/// list that looks like a secret scan's output.
571#[derive(Debug, Clone, Default)]
572pub struct ConfigSecretSummary {
573    /// Config keys whose **name** matched the secret-name heuristic.
574    pub secret_named: usize,
575    /// Of those, how many had their value redacted before persistence.
576    pub redacted: usize,
577    /// Of those, how many are declared in code with no literal value.
578    pub declared: usize,
579    /// Of those, how many carry an unredacted value. Expected to be zero.
580    pub unredacted: usize,
581    /// Distinct files carrying at least one secret-named key, ordered and capped
582    /// by the caller.
583    pub files: Vec<String>,
584}
585
586/// One file in the `_Home` overview's intent-debt density table.
587#[derive(Debug, Clone)]
588pub struct DensityEntry {
589    /// Repository-relative path, used for both the wikilink and the label.
590    pub path: String,
591    /// Retained markers in the file.
592    pub markers: u32,
593    /// The file's length in lines — the denominator.
594    pub lines: u32,
595    /// Markers per 1,000 lines.
596    pub per_kloc: f64,
597}
598
599/// One node in the `_Home` overview's directed-coupling table.
600#[derive(Debug, Clone)]
601pub struct CouplingEntry {
602    /// The node key, for the wikilink.
603    pub key: String,
604    /// The symbol name.
605    pub name: String,
606    /// Distinct callers.
607    pub fan_in: u32,
608    /// Distinct callees.
609    pub fan_out: u32,
610}
611
612/// Aggregate figures for the vault's `_Home` overview note.
613#[derive(Debug, Clone, Default)]
614pub struct VaultSummary {
615    /// Name of the scanned project (repository directory).
616    pub project: String,
617    /// Total node and edge counts.
618    pub total_nodes: usize,
619    /// Total edge count.
620    pub total_edges: usize,
621    /// `(kind, count)` for each node kind, most-frequent first.
622    pub node_counts: Vec<(String, usize)>,
623    /// `(provenance, edge count)` — `derived` / `authored` / `inferred`.
624    pub edge_provenance: Vec<(String, usize)>,
625    /// The ADRs, with status.
626    pub adrs: Vec<AdrEntry>,
627    /// `(category, count)` of intent-debt markers.
628    pub debt: Vec<(String, usize)>,
629    /// The files where that debt is most **concentrated**, already ranked and
630    /// capped by the caller. Empty when the graph has no markers, or when no
631    /// file carrying one has a recorded length.
632    pub densest_files: Vec<DensityEntry>,
633    /// Secret-named config keys and their redaction state. `None` when the graph
634    /// holds no secret-named config key — the section is then absent rather than
635    /// rendering a row of zeroes, which would read as a clean bill of health this
636    /// lens cannot give.
637    pub config_secrets: Option<ConfigSecretSummary>,
638    /// The most depended-on symbols by **directed** call fan-in, already ranked
639    /// and capped by the caller. Empty when the graph has no `calls` edges.
640    pub most_called: Vec<CouplingEntry>,
641    /// Web root of the repository (`https://host/owner/repo`), if derivable from
642    /// the git remote — for a "Repository" link in the overview, and the
643    /// **clone-from** column of a workspace vault's manifest (#442 part 2).
644    ///
645    /// The web root rather than the raw `origin` fetch URL on purpose: a vault is
646    /// made to be handed to someone, and `git@host:owner/repo.git` is only
647    /// actionable for a reader who already has SSH access to that host.
648    ///
649    /// It is **where the code lives, not a guaranteed clone URL**, and the
650    /// manifest says so. `repo_web_root` normalises a remote to `https://host/…`,
651    /// which clones on the common forges and may not on an unusual one, and says
652    /// nothing about whether the reader can read a private repository. A manifest
653    /// that promised "clone from here" would be making a claim it cannot check.
654    pub repo_url: Option<String>,
655    /// Hex commit the graph was rendered from, for a permalink note — and, in a
656    /// workspace vault, the commit this member is pinned at.
657    ///
658    /// With [`Self::repo_url`] it is what makes a workspace vault **replicable**
659    /// rather than merely browsable: *"here is my workspace"* is far less useful
660    /// than *"here is my workspace **at these commits**"*, and it is what lets a
661    /// reader tell a stale vault from a current one instead of guessing.
662    pub commit: Option<String>,
663    /// The **enabled** `[ingest]` toggles this member was extracted under, by
664    /// name (`prose`, `pdf`, …), and its `[debt] ignore` globs.
665    ///
666    /// The manifest's third leg, after clone URL and commit: those two get a
667    /// reader the same *source*, and these two are what decide whether the same
668    /// source produces the same *vault*. `[ingest] prose` off means notes with no
669    /// captured content; a `[debt] ignore` glob means the debt figures on this
670    /// page are already filtered. Without them "reproducible" means "you can
671    /// obtain the code", which is a weaker claim than the section makes.
672    pub settings: RenderedUnder,
673    /// This member's stored analyzer findings (ADR-0012), ordered most severe
674    /// first by the caller. Rendered in a workspace vault's `_Home`; read
675    /// together with [`Self::coverage`], which is what says whether an empty
676    /// list means anything at all.
677    pub findings: Vec<FindingEntry>,
678    /// Whether an analyzer has ever run against this member — the context that
679    /// makes an empty [`Self::findings`] readable rather than reassuring. See
680    /// [`Coverage`].
681    pub coverage: Coverage,
682}
683
684/// Render the vault's overview note: what was scanned, the structure by kind,
685/// the provenance breakdown, the decisions (ADRs) and their status, the
686/// intent-debt summary, and how to navigate. The entry point for the vault.
687#[must_use]
688pub fn render_home(s: &VaultSummary) -> VaultNote {
689    let mut c = String::new();
690    c.push_str("---\ntags:\n  - roteiro/home\n---\n\n");
691    let _ = writeln!(c, "# {} — knowledge graph", s.project);
692    c.push_str(
693        "\n*A browsable snapshot of this codebase as one **knowledge graph**, \
694         generated by [Roteiro](https://roteiro.dev). Every symbol, document and \
695         decision is a note, linked to the things it relates to.*\n",
696    );
697    c.push_str(HOW_TO_READ);
698    let _ = writeln!(
699        c,
700        "\n**{} nodes**, **{} edges** across the project.",
701        s.total_nodes, s.total_edges
702    );
703    write_repo_line(&mut c, s);
704    write_summary_sections(&mut c, s, &VaultScope::PROJECT, 2);
705    c.push_str(NAVIGATING);
706
707    VaultNote {
708        filename: HOME_NOTE.to_owned(),
709        content: c,
710    }
711}
712
713/// The "how to read a note" paragraph. Shared verbatim by the single-project and
714/// workspace overviews — the notes themselves are identical in both, so a reader
715/// who learns the format once has learned it for either.
716const HOW_TO_READ: &str = "\n**How to read it.** Open any note to see what a thing is, the intent or \
717     docs behind it (its **Content**), where it lives (its **Source** link), \
718     and how it connects (**Outgoing**/**Incoming** links). Each link is \
719     labelled with how the fact was established — `derived` (extracted from \
720     code), `authored` (human intent: ADRs, blueprints, annotations), or \
721     `inferred` (a scored suggestion). Open Obsidian's **graph view** to see \
722     the whole thing at once.\n";
723
724/// The closing navigation section.
725const NAVIGATING: &str = "\n## Navigating this vault\n\n\
726     - Open the **graph view** to see the whole codebase; notes are coloured/\
727     filterable by their `roteiro/kind/*`, `roteiro/lang/*` and \
728     `roteiro/status/*` tags.\n\
729     - Each note carries its captured **content** (doc comments, prose, PDF/\
730     image text) and its provenance-labelled incoming/outgoing links.\n\
731     - Start from an ADR above, or search the tag pane for a kind.\n";
732
733/// `**Repository:** …` — the web root and the commit the graph was rendered from.
734fn write_repo_line(c: &mut String, s: &VaultSummary) {
735    if let Some(repo) = &s.repo_url {
736        let _ = write!(c, "\n**Repository:** [{repo}]({repo})");
737        if let Some(commit) = &s.commit {
738            let short = &commit[..commit.len().min(12)];
739            let _ = write!(c, " · rendered at commit `{short}`");
740        }
741        c.push('\n');
742    }
743}
744
745/// Every aggregate the overview carries for **one project**: structure by kind,
746/// provenance, ADRs, intent debt (and where it is densest), the config-secret
747/// inventory and directed call coupling.
748///
749/// Factored out of [`render_home`] so a workspace vault's per-member section is
750/// *the same code*, not a reimplementation that can drift: the promise in issue
751/// #442 is that today's per-project view stays a **subset** of the workspace one
752/// rather than a casualty of it. `level` is the markdown heading depth — 2 for a
753/// single-project `_Home`, 3 inside a member's section — and `scope` decides
754/// whether the wikilinks point at bare or project-qualified notes.
755fn write_summary_sections(c: &mut String, s: &VaultSummary, scope: &VaultScope<'_>, level: usize) {
756    let hd = &"#".repeat(level);
757    let sub = &"#".repeat(level + 1);
758    write_structure(c, s, hd);
759    write_decisions(c, s, scope, hd);
760    write_debt(c, s, scope, hd, sub);
761    write_config_secrets(c, s, scope, hd);
762    write_coupling(c, s, scope, hd);
763}
764
765/// `Structure` (nodes by kind) and `Provenance` (edges by how they were established).
766fn write_structure(c: &mut String, s: &VaultSummary, hd: &str) {
767    let _ = write!(c, "\n{hd} Structure\n\n| Kind | Count |\n| --- | --- |\n");
768    for (kind, n) in &s.node_counts {
769        let _ = writeln!(c, "| {kind} | {n} |");
770    }
771
772    if !s.edge_provenance.is_empty() {
773        let _ = write!(
774            c,
775            "\n{hd} Provenance\n\n| Provenance | Edges |\n| --- | --- |\n"
776        );
777        for (prov, n) in &s.edge_provenance {
778            let _ = writeln!(c, "| {prov} | {n} |");
779        }
780    }
781}
782
783/// `Decisions (ADRs)` — the recorded decisions and their lifecycle status.
784fn write_decisions(c: &mut String, s: &VaultSummary, scope: &VaultScope<'_>, hd: &str) {
785    let _ = write!(c, "\n{hd} Decisions (ADRs)\n\n");
786    if s.adrs.is_empty() {
787        c.push_str("*No ADRs found.*\n");
788    } else {
789        for adr in &s.adrs {
790            let status = adr.status.as_deref().unwrap_or("—");
791            let _ = writeln!(
792                c,
793                "- **{status}** — [[{}|{}]]",
794                scoped_note_name(scope, &adr.key),
795                adr.name
796            );
797        }
798    }
799}
800
801/// `Intent debt` — the marker categories, and the files the debt is densest in.
802fn write_debt(c: &mut String, s: &VaultSummary, scope: &VaultScope<'_>, hd: &str, sub: &str) {
803    let _ = write!(c, "\n{hd} Intent debt\n\n");
804    if s.debt.is_empty() {
805        c.push_str("*None recorded.*\n");
806    } else {
807        c.push_str("| Category | Count |\n| --- | --- |\n");
808        for (cat, n) in &s.debt {
809            let _ = writeln!(c, "| {cat} | {n} |");
810        }
811    }
812
813    if !s.densest_files.is_empty() {
814        let _ = write!(
815            c,
816            "\n{sub} Densest files (markers per 1,000 lines)\n\n\
817             *Where the debt above is concentrated, rather than where there is \
818             most of it — a raw count ranks the biggest file first by \
819             construction. The denominator is file length: every line, blanks and \
820             comments included, not source lines of code. Prose matches (`for \
821             now`, `tbd`) count too, so a design document can rank high.*\n\n"
822        );
823        c.push_str("| File | Markers | Lines | Per 1k |\n| --- | --- | --- | --- |\n");
824        for e in &s.densest_files {
825            let _ = writeln!(
826                c,
827                "| [[{}\\|{}]] | {} | {} | {:.2} |",
828                scoped_note_name(scope, &format!("file:{}", e.path)),
829                e.path,
830                e.markers,
831                e.lines,
832                e.per_kloc
833            );
834        }
835    }
836}
837
838/// `Config keys named like secrets` — an inventory and its unconditional caveat.
839fn write_config_secrets(c: &mut String, s: &VaultSummary, scope: &VaultScope<'_>, hd: &str) {
840    if let Some(cs) = &s.config_secrets {
841        let _ = write!(c, "\n{hd} Config keys named like secrets\n\n");
842        let _ = writeln!(
843            c,
844            "**{}** secret-named config key(s): {} redacted before storage, {} \
845             declared in code without a value, {} unredacted.",
846            cs.secret_named, cs.redacted, cs.declared, cs.unredacted
847        );
848        if cs.unredacted > 0 {
849            let _ = writeln!(
850                c,
851                "\n> [!warning] {} key(s) carry an **unredacted** value. Extraction \
852                 always redacts, so these came from an import layer — inspect the \
853                 importing tool, not this repository.",
854                cs.unredacted
855            );
856        }
857        if !cs.files.is_empty() {
858            c.push_str("\nIn:\n");
859            for path in &cs.files {
860                let _ = writeln!(
861                    c,
862                    "- [[{}\\|{path}]]",
863                    scoped_note_name(scope, &format!("file:{path}"))
864                );
865            }
866        }
867        // The caveat is unconditional and comes last, so it is the final thing read
868        // in this section. A vault note is browsed out of context; this is exactly
869        // where "config keys named like secrets" would otherwise be misread as a
870        // secret scan that came back clean.
871        c.push_str(
872            "\n*An inventory of config keys whose **names** look secret, not a secret \
873             scan. Values are redacted before they are stored, so this reports that \
874             such keys exist and were redacted — never a value. It cannot see a \
875             hardcoded credential in source code, cannot judge whether a value is \
876             valid, and cannot tell a real secret from a placeholder. A credential \
877             under an innocuous key name (`dsn`, `endpoint`) does not appear here at \
878             all, so this section being small says nothing about whether this \
879             repository leaks secrets.*\n",
880        );
881    }
882}
883
884/// `Most depended-on (call fan-in)` — directed call coupling, capped.
885fn write_coupling(c: &mut String, s: &VaultSummary, scope: &VaultScope<'_>, hd: &str) {
886    if !s.most_called.is_empty() {
887        let _ = write!(
888            c,
889            "\n{hd} Most depended-on (call fan-in)\n\n\
890             *Distinct callers and callees over `calls` edges — direction kept, so \
891             \"everything calls this\" and \"this calls everything\" are not the same \
892             row. Call targets are resolved by simple name, so a short, generically-\
893             named function can absorb every call to that name: read a large fan-in on \
894             one as a question, not a finding.*\n\n"
895        );
896        c.push_str("| Symbol | Called by | Calls |\n| --- | --- | --- |\n");
897        for e in &s.most_called {
898            let _ = writeln!(
899                c,
900                "| [[{}\\|{}]] | {} | {} |",
901                scoped_note_name(scope, &e.key),
902                e.name,
903                e.fan_in,
904                e.fan_out
905            );
906        }
907    }
908}
909
910/// One cross-repo edge the workspace vault can actually follow: a spoke's node
911/// linking to a hub's, through the external-ref placeholder ADR-0009 persists.
912///
913/// Collected by the caller, which has every member's store open; the renderer
914/// only lays them out. Nothing here is a new edge — these are the `inferred`
915/// links `roteiro links` already reports, rendered for the first time.
916#[derive(Debug, Clone)]
917pub struct CrossLink {
918    /// The member the edge starts in.
919    pub from_project: String,
920    /// The source node's key, within `from_project`.
921    pub from_key: String,
922    /// The source node's display name.
923    pub from_name: String,
924    /// The edge kind (`links`, …).
925    pub kind: String,
926    /// Confidence, for an `inferred` edge.
927    pub confidence: Option<f64>,
928    /// Whether this link was **declared** (`[[links]]` in the source repo's
929    /// config, ADR-0009) rather than inferred by key matching.
930    ///
931    /// The distinction is the whole of ADR-0009's `authored → gold,
932    /// inferred → slate`: a declaration is a statement of intent by someone who
933    /// knows the topology, a match is a candidate. Until #573 the vault could not
934    /// draw it, because nothing persisted an authored cross-repo edge — so this
935    /// section carried a blanket caveat saying every row was a candidate.
936    pub authored: bool,
937    /// The project-qualified target, `<project>::<key>` (ADR-0009).
938    pub to_qualified: String,
939    /// Whether `to_qualified`'s project is a member of this workspace — and so
940    /// whether the link resolves to a note in this vault, or dangles because the
941    /// target repository is outside it.
942    pub resolves: bool,
943}
944
945/// The settings a member's notes were rendered under — the ones that change what
946/// the vault *contains*, not the whole merged config.
947///
948/// Deliberately a short list rather than the effective configuration in full.
949/// #442 asks the manifest to record "effective settings", and dumping every
950/// resolved key into a shareable artifact would re-open the redaction question
951/// this vault already has to warn about, to record settings that cannot change
952/// what a reader sees. These two can.
953#[derive(Debug, Clone, Default)]
954pub struct RenderedUnder {
955    /// Enabled `[ingest]` toggles by name, in declaration order. Empty means
956    /// every toggle was off — which is a real state and renders as such, not as
957    /// an absent row.
958    pub ingest: Vec<String>,
959    /// `[debt] ignore` globs. Non-empty means the debt figures on this page are
960    /// **already filtered**, and a reader comparing them against a fresh
961    /// `roteiro debt` without the same config will not match.
962    pub debt_ignore: Vec<String>,
963}
964
965/// One analyzer finding, as the vault renders it (ADR-0012).
966///
967/// A render-facing copy rather than `rto_graph::Finding`, matching [`AdrEntry`]
968/// and [`CouplingEntry`]: the renderer takes the fields it prints and stays free
969/// of the findings model. It deliberately does **not** carry `meta` — that is
970/// whatever the analyzer emitted, kept verbatim, and a shareable artifact is the
971/// worst place to reproduce "whatever the tool said" unread.
972#[derive(Debug, Clone)]
973pub struct FindingEntry {
974    /// The rule, advisory or check id the analyzer fired (`RUSTSEC-2026-0031`).
975    pub rule: String,
976    /// The severity the analyzer assigned — **a tool judgement, not a
977    /// confidence** — rendered as the analyzer's word rather than the vault's.
978    pub severity: String,
979    /// One-line summary.
980    pub title: String,
981    /// The analyzer's full message.
982    pub message: String,
983    /// Repository-relative path the finding is about, if the analyzer located one.
984    pub path: Option<String>,
985    /// The analyzer that produced it, so a reader can tell one tool's opinion
986    /// from another's rather than reading a merged list as a single verdict.
987    pub analyzer: String,
988}
989
990/// What a member's analyzer coverage actually is — the distinction the vault
991/// must never blur.
992///
993/// An empty findings list means one of two completely different things, and
994/// printing both as "no findings" is the failure `roteiro security status`
995/// records as `no-analyzer-on-record`: **nothing has been analyzed** is not
996/// **nothing is wrong**. A shareable artifact is the worst place to conflate
997/// them, because its reader is the one person who cannot check.
998///
999/// [`Coverage::NotRun`] is the `Default` deliberately. A `VaultSummary` built
1000/// from `Default` has had no analyzer run against it, and defaulting the other
1001/// way would render "no findings" for a member nobody looked at — the exact
1002/// conflation this type exists to prevent, arrived at by omission.
1003#[derive(Debug, Clone, Default)]
1004pub enum Coverage {
1005    /// No analyzer has ever run against this member. **Not** a clean result.
1006    #[default]
1007    NotRun,
1008    /// At least one analyzer ran, as `(analyzer, version)` per run — so an empty
1009    /// findings list is attributable to a tool that actually looked.
1010    Ran(Vec<(String, String)>),
1011}
1012
1013/// Aggregate figures for a **workspace** vault's `_Home` overview: the members,
1014/// each with exactly the aggregates a single-project `_Home` carries, plus the
1015/// cross-repo links between them.
1016#[derive(Debug, Clone, Default)]
1017pub struct WorkspaceSummary {
1018    /// When this vault was rendered, RFC 3339 UTC.
1019    ///
1020    /// A vault is **read-only and point-in-time**. Stamping it is what gives
1021    /// that property teeth: with the per-member commits, a reader can tell
1022    /// whether what they are looking at still describes the workspace, rather
1023    /// than assuming it does.
1024    pub generated_at: String,
1025    /// The workspace name (`--workspace-name`).
1026    pub name: String,
1027    /// One entry per member repository, in stable name order. Each is the very
1028    /// same [`VaultSummary`] a per-project vault would render.
1029    pub members: Vec<VaultSummary>,
1030    /// Cross-repo links between members, already ordered and capped by the caller.
1031    pub cross_links: Vec<CrossLink>,
1032    /// Cross-repo links found in total, which `cross_links` may be a capped view
1033    /// of — so the section can say what it is not showing.
1034    pub cross_links_total: usize,
1035    /// How many of `cross_links_total` were **declared** (`[[links]]`) rather
1036    /// than inferred.
1037    ///
1038    /// Counted before the cap, not from `cross_links`: that vector is truncated
1039    /// to [`WORKSPACE_CROSS_LINK_ROWS`](crate::WORKSPACE_CROSS_LINK_ROWS) rows,
1040    /// so counting it would describe the rows on screen while reading as a
1041    /// statement about the workspace — a caption that quietly changes meaning
1042    /// once a workspace grows past the cap.
1043    pub cross_links_authored: usize,
1044}
1045
1046/// Render a **workspace** vault's overview: the members and their scale, the
1047/// cross-repo links between them, and then each member's own aggregates —
1048/// structure, provenance, ADRs, intent debt, config-secret inventory and call
1049/// coupling — under its own heading.
1050///
1051/// The per-member sections are rendered by the same [`write_summary_sections`]
1052/// the single-project `_Home` uses, so the existing view is a **subset** of this
1053/// one: someone who came for their repository's coupling and debt tables finds
1054/// them, rather than a workspace total that averages them away.
1055#[must_use]
1056pub fn render_workspace_home(ws: &WorkspaceSummary) -> VaultNote {
1057    let members: std::collections::BTreeSet<String> =
1058        ws.members.iter().map(|m| m.project.clone()).collect();
1059
1060    let mut c = String::new();
1061    c.push_str("---\ntags:\n  - roteiro/home\n  - roteiro/workspace\n---\n\n");
1062    let _ = writeln!(c, "# {} — workspace knowledge graph", ws.name);
1063    c.push_str(
1064        "\n*A browsable snapshot of a whole **workspace** as one **knowledge \
1065         graph**, generated by [Roteiro](https://roteiro.dev). Every symbol, \
1066         document and decision in every member repository is a note, linked to the \
1067         things it relates to — including across repositories.*\n",
1068    );
1069    c.push_str(HOW_TO_READ);
1070    // The example is *rendered* by `note_name` rather than spelled out. A
1071    // hand-written spelling of this sentence survived #574 unchanged, so every
1072    // vault v2.0.0 built stated the pre-#574 naming rule — on the first page a
1073    // reader opens — while `note_name` was writing something else. This is the
1074    // one copy that lives in the crate defining the rule, so it can simply ask:
1075    // a derived example cannot drift, and a spelled one already has.
1076    //
1077    // The *key* it names has to be real too, or the fix trades one false
1078    // sentence in `_Home` for another: `<member>::file:README.md` was fabricated
1079    // from the member list, and workspace membership does not require a README.
1080    // A cross-repo link's **source** end is the strongest key available here —
1081    // `from_project` is a member by definition and `from_key` is a node in that
1082    // member's own store, which the Cross-repo links table below already links
1083    // to by name. The *target* end will not do: `resolves == false` means the
1084    // target repository is outside this vault, so `to_qualified` names no note
1085    // here — the same false claim one remove away.
1086    //
1087    // With no cross-repo links there is no key this function can prove is a
1088    // node, so the sentence says nothing rather than inventing one. The rule it
1089    // states is complete without an example; only the illustration is lost.
1090    let example = ws.cross_links.first().map_or_else(String::new, |l| {
1091        let key = format!("{}::{}", l.from_project, l.from_key);
1092        format!(" Here, `{key}` is the note `{}.md`.", note_name(&key))
1093    });
1094    let _ = writeln!(
1095        c,
1096        "\n**Every note is keyed `<project>::<key>`**, because a node key is \
1097         repository-relative: the same path or symbol can occur in more than one \
1098         member, and without the project the second note would overwrite the \
1099         first. A note's *filename* is derived from that key — a readable \
1100         lowercase hint, then a hash of the whole key — so no filename contains \
1101         `::`.{example} Filter the graph view by a member's `roteiro/project/*` \
1102         tag to see one repository at a time."
1103    );
1104
1105    let total_nodes: usize = ws.members.iter().map(|m| m.total_nodes).sum();
1106    let total_edges: usize = ws.members.iter().map(|m| m.total_edges).sum();
1107    let _ = writeln!(
1108        c,
1109        "\n**{total_nodes} nodes**, **{total_edges} edges** across **{}** member \
1110         repositor{}.",
1111        ws.members.len(),
1112        if ws.members.len() == 1 { "y" } else { "ies" }
1113    );
1114
1115    c.push_str("\n## Members\n\n| Project | Nodes | Edges | Repository | Commit |\n| --- | --- | --- | --- | --- |\n");
1116    for m in &ws.members {
1117        let repo = m
1118            .repo_url
1119            .as_ref()
1120            .map_or_else(|| "—".to_owned(), |u| format!("[{u}]({u})"));
1121        let commit = m.commit.as_ref().map_or_else(
1122            || "—".to_owned(),
1123            |c| format!("`{}`", &c[..c.len().min(12)]),
1124        );
1125        let _ = writeln!(
1126            c,
1127            "| [[#{}\\|{}]] | {} | {} | {repo} | {commit} |",
1128            m.project, m.project, m.total_nodes, m.total_edges
1129        );
1130    }
1131    // `[[#Heading]]`, the form the member rows above already use, rather than a
1132    // `](#slug)` anchor: the slug a renderer computes and the slug a reader's
1133    // tool computes are exactly what issue #524 is about, and a manifest is a
1134    // poor place to find out they disagree.
1135    c.push_str(
1136        "\n*The `Repository` and `Commit` columns say where each member came from \
1137         and what was read. The manifest that makes those reconstructable — and \
1138         states what this vault carries before you share it — is \
1139         [[#Reproducing this vault]], below.*\n",
1140    );
1141
1142    write_cross_links(&mut c, ws);
1143    write_findings(&mut c, ws);
1144    write_manifest(&mut c, ws);
1145
1146    for m in &ws.members {
1147        let _ = writeln!(c, "\n## {}", m.project);
1148        let _ = writeln!(
1149            c,
1150            "\n**{} nodes**, **{} edges** in this member.",
1151            m.total_nodes, m.total_edges
1152        );
1153        write_repo_line(&mut c, m);
1154        let scope = VaultScope {
1155            project: Some(&m.project),
1156            members: &members,
1157        };
1158        write_summary_sections(&mut c, m, &scope, 3);
1159    }
1160
1161    c.push_str(NAVIGATING);
1162
1163    VaultNote {
1164        filename: HOME_NOTE.to_owned(),
1165        content: c,
1166    }
1167}
1168
1169/// The `### Rendered under` table — the settings half of the manifest.
1170///
1171/// A clone URL and a commit get a reader the same **source**; these get them the
1172/// same **vault**. Rendering the same commits with `[ingest] prose` off, or with
1173/// a different `[debt] ignore`, produces a different document from the same
1174/// code — so a manifest that omits them promises a reproducibility it cannot
1175/// deliver, which is worse than promising less.
1176fn write_rendered_under(c: &mut String, ws: &WorkspaceSummary) {
1177    c.push_str("\n### Rendered under\n\n");
1178    c.push_str(
1179        "*These change what a vault **contains**, so re-rendering the commits \
1180         above under different ones will not reproduce this document.*\n\n",
1181    );
1182    c.push_str("| Member | `[ingest]` | `[debt] ignore` |\n| --- | --- | --- |\n");
1183    for m in &ws.members {
1184        let list = |v: &[String]| {
1185            if v.is_empty() {
1186                // "none" as a word, not an empty cell: a blank reads as unknown,
1187                // and these two are the difference between reproducing this
1188                // vault and something that merely resembles it.
1189                "*none*".to_owned()
1190            } else {
1191                // `[debt] ignore` is user config landing in a table cell: a
1192                // glob carrying a backtick or a `|` reshapes the row. Not
1193                // analyzer-sourced, but untrusted for the same reason — the
1194                // vault did not write it.
1195                v.iter()
1196                    .map(|x| table_cell(x))
1197                    .collect::<Vec<_>>()
1198                    .join(", ")
1199            }
1200        };
1201        let _ = writeln!(
1202            c,
1203            "| {} | {} | {} |",
1204            inline_untrusted(&m.project),
1205            list(&m.settings.ingest),
1206            list(&m.settings.debt_ignore)
1207        );
1208    }
1209}
1210
1211/// The `## Security findings` section: the stored analyzer output (ADR-0012) for
1212/// every member.
1213///
1214/// # Why this is here at all, stated once
1215///
1216/// ADR-0012 keeps findings out of `nodes`/`edges` **specifically** so they cannot
1217/// reach `export_factset`, and records the alternative as rejected because it
1218/// would "silently publish tool output into an artifact". This section is that
1219/// same publication, made **deliberately and visibly** rather than by accident:
1220/// a workspace vault is a hand-over document, and the owner ruled that it should
1221/// answer *"what is wrong with this workspace"* as well as *"what is in it"*.
1222/// The exclusions section says so at the point a reader is about to share it.
1223///
1224/// # The distinction this section exists to preserve
1225///
1226/// A member with no findings is rendered **two different ways** depending on
1227/// [`Coverage`], because "no analyzer has run" and "an analyzer ran and found
1228/// nothing" are opposite facts that look identical when both are printed as an
1229/// empty list. `roteiro security status` names that failure `no-analyzer-on-record`
1230/// and refuses to let it read as clean; a shareable artifact must refuse harder,
1231/// because its reader is the one person who cannot go and check.
1232fn write_findings(c: &mut String, ws: &WorkspaceSummary) {
1233    c.push_str("\n## Security findings\n\n");
1234
1235    let total: usize = ws.members.iter().map(|m| m.findings.len()).sum();
1236    let unanalyzed = ws
1237        .members
1238        .iter()
1239        .filter(|m| matches!(m.coverage, Coverage::NotRun))
1240        .count();
1241
1242    let _ = writeln!(
1243        c,
1244        "*Stored analyzer output (ADR-0012). **{total} finding(s)** across \
1245         {} member(s){}. A severity is the **analyzer's** judgement, not \
1246         Roteiro's, and a finding is a tool's claim rather than a confirmed \
1247         defect — read one as something to check, not something proven.*\n",
1248        ws.members.len(),
1249        if unanalyzed > 0 {
1250            format!(", with {unanalyzed} never analyzed at all")
1251        } else {
1252            String::new()
1253        }
1254    );
1255
1256    for m in &ws.members {
1257        match (&m.coverage, m.findings.is_empty()) {
1258            // Never analyzed. Said loudly, and *not* in the same breath as a
1259            // member that came back clean — the whole point of the split.
1260            (Coverage::NotRun, _) => {
1261                let _ = writeln!(c, "### {} — **not analyzed**\n", m.project);
1262                c.push_str(
1263                    "*No analyzer has run against this member, so nothing here is \
1264                     a statement about it. An empty list is the absence of a \
1265                     question, not a reassuring answer.*\n\n",
1266                );
1267            }
1268            // Analyzed, nothing found — attributable to a tool that looked.
1269            (Coverage::Ran(runs), true) => {
1270                let _ = writeln!(
1271                    c,
1272                    "### {} — no findings\n\n*{} ran and reported none.*\n",
1273                    m.project,
1274                    describe_runs(runs)
1275                );
1276            }
1277            (Coverage::Ran(runs), false) => {
1278                let _ = writeln!(
1279                    c,
1280                    "### {} — {} finding(s)\n\n*Reported by {}.*\n",
1281                    m.project,
1282                    m.findings.len(),
1283                    describe_runs(runs)
1284                );
1285                for f in &m.findings {
1286                    let at = f
1287                        .path
1288                        .as_deref()
1289                        .map_or_else(String::new, |p| format!(" · {}", inline_code(p)));
1290                    let _ = writeln!(
1291                        c,
1292                        "**{}** · {}{at} — {}\n",
1293                        inline_code(&f.severity),
1294                        inline_code(&f.rule),
1295                        // Prose, not a span: the title is the sentence a reader
1296                        // scans, so it is escaped rather than monospaced.
1297                        inline_untrusted(&f.title)
1298                    );
1299                    write_analyzer_message(c, f);
1300                }
1301            }
1302        }
1303    }
1304}
1305
1306/// An untrusted value as a **complete code span**, delimiters included.
1307///
1308/// A code span renders its content literally — no emphasis, no links, no HTML —
1309/// so the only thing that can escape one is a backtick run long enough to close
1310/// it. Widening the delimiter past the longest run inside the value is therefore
1311/// the whole defence, and it is the *right* defence: backslash-escaping inside a
1312/// span protects nothing and does show, turning `vendor/**` into `vendor\*\*`
1313/// on the page.
1314///
1315/// Whitespace is still collapsed — a newline ends the line the span sits on
1316/// however well the span itself is delimited.
1317///
1318/// **Not sufficient inside a table cell.** GFM splits a row on `|` *before*
1319/// inline parsing, so a pipe inside a code span still ends the cell. Use
1320/// [`table_cell`] there.
1321fn inline_code(s: &str) -> String {
1322    let collapsed = s.split_whitespace().collect::<Vec<_>>().join(" ");
1323    let longest_run = collapsed
1324        .split(|ch| ch != '`')
1325        .map(str::len)
1326        .max()
1327        .unwrap_or(0);
1328    let tick = "`".repeat(longest_run + 1);
1329    // CommonMark strips one leading and trailing space from a span, which is how
1330    // content that itself starts or ends with a backtick is expressed.
1331    let pad = if collapsed.starts_with('`') || collapsed.ends_with('`') {
1332        " "
1333    } else {
1334        ""
1335    };
1336    format!("{tick}{pad}{collapsed}{pad}{tick}")
1337}
1338
1339/// One line of analyzer-sourced text, made safe to interpolate into the note.
1340///
1341/// The message gets a fence; **these fields did not**, and they come from the
1342/// same place. A `title` carrying a newline ends the line it was placed on and
1343/// starts a new block — one that can open a heading, a list or a table row — so
1344/// the "quoted, not absorbed" property held for one field out of four. A `rule`
1345/// or `path` carrying a backtick escapes the code span it is wrapped in.
1346///
1347/// Four steps, in order:
1348///
1349/// 1. **Collapse every whitespace run to one space.** This is what removes the
1350///    structural attack: Markdown block constructs need a line start, and after
1351///    this there are no line starts left inside the value.
1352/// 2. **HTML-escape `&`, `<`, `>`.** Some of these values land inside raw HTML
1353///    (`<sub>`, `<details><summary>`), where a backslash escapes nothing — an
1354///    analyzer named `</sub><script>` would close the tag and keep going. This
1355///    step is why there is **one** helper rather than a Markdown one and an HTML
1356///    one: two helpers means two contexts to keep straight, and the reason this
1357///    function exists at all is that the first version secured one context and
1358///    missed its neighbour.
1359/// 3. **Backslash-escape the remaining Markdown punctuation.** A link in a
1360///    document handed to someone can point anywhere, and a stray backtick or
1361///    pipe silently reshapes the line it lands in. `<` and `>` are absent from
1362///    that set because step 2 already turned them into entities.
1363/// 4. **Defuse bare URLs.** Escaping `[` stops explicit `[text](url)` syntax and
1364///    nothing else: Obsidian and GFM *linkify* a plain `https://…`, so analyzer
1365///    text could still hand the reader a clickable link to anywhere. Replacing
1366///    the scheme's colon with `&#58;` renders identically and matches no
1367///    linkifier — the reader still sees the URL and can copy it deliberately.
1368///
1369/// Safe in both contexts, so a caller never has to know which one it is in.
1370fn inline_untrusted(s: &str) -> String {
1371    let collapsed = s.split_whitespace().collect::<Vec<_>>().join(" ");
1372    let mut out = String::with_capacity(collapsed.len());
1373    for ch in collapsed.chars() {
1374        match ch {
1375            '&' => out.push_str("&amp;"),
1376            '<' => out.push_str("&lt;"),
1377            '>' => out.push_str("&gt;"),
1378            // `&` is deliberately absent below: the entities written above must
1379            // survive intact, and backslash-escaping their `&` would render them
1380            // literally.
1381            '\\' | '`' | '*' | '_' | '[' | ']' | '|' | '#' => {
1382                out.push('\\');
1383                out.push(ch);
1384            }
1385            _ => out.push(ch),
1386        }
1387    }
1388    defuse_autolinks(&out)
1389}
1390
1391/// Stop a bare URL in untrusted text from being turned into a clickable link.
1392///
1393/// Escaping `[` covers explicit link syntax; it does nothing about linkification,
1394/// which is on by default in Obsidian and GFM. `https&#58;//evil.example` renders
1395/// as `https://evil.example` and matches no autolinker, so the URL stays
1396/// readable and copyable while ceasing to be a thing the reader can click by
1397/// accident in a document someone handed them.
1398fn defuse_autolinks(s: &str) -> String {
1399    s.replace("https://", "https&#58;//")
1400        .replace("http://", "http&#58;//")
1401}
1402
1403/// [`inline_code`] for a value that lands in a **table cell**.
1404///
1405/// A code span is not enough there: GFM splits a row on `|` before it parses
1406/// inline content, so a pipe inside the span still ends the cell and shifts
1407/// every column after it. `\|` is the documented escape, and it is applied to
1408/// the finished span so the delimiter widening still holds.
1409///
1410/// The distinction is not cosmetic — it is a third rendering context, and this
1411/// file has now been wrong about a context twice.
1412fn table_cell(s: &str) -> String {
1413    inline_code(s).replace('|', "\\|")
1414}
1415
1416/// One finding's full analyzer message, collapsed and **verbatim**.
1417///
1418/// Collapsed because the messages are long — a single GHSA advisory runs to
1419/// paragraphs, and seventeen of them inline turn `_Home` into a wall of text
1420/// nobody reads, which loses the findings as surely as omitting them. `<details>`
1421/// keeps every byte in the file (the ruling was to include them in full) while
1422/// letting the reader see the list first.
1423///
1424/// **Verbatim, in a fence, rather than as Markdown.** An analyzer message is
1425/// tool output travelling into a document that gets handed to people: rendered
1426/// as Markdown it can open headings that restructure the note, or links that
1427/// point anywhere. A fence is what makes it text the vault *quotes* rather than
1428/// text the vault *becomes* — and it preserves the advisory's own line structure,
1429/// which flattening to one line destroys.
1430///
1431/// The fence is sized to beat the longest backtick run in the message, for the
1432/// reason [`crate::docs`] handles multi-backtick spans: a three-backtick fence
1433/// around a message that itself contains one ends the block early and spills the
1434/// remainder into the note as prose.
1435fn write_analyzer_message(c: &mut String, f: &FindingEntry) {
1436    let message = f.message.trim();
1437    if message.is_empty() {
1438        let _ = writeln!(
1439            c,
1440            "<sub>reported by `{}`</sub>\n",
1441            inline_untrusted(&f.analyzer)
1442        );
1443        return;
1444    }
1445    let longest_run = message
1446        .split(|ch| ch != '`')
1447        .map(str::len)
1448        .max()
1449        .unwrap_or(0);
1450    let fence = "`".repeat(longest_run.max(2) + 1);
1451    let _ = writeln!(
1452        c,
1453        "<details><summary><sub>what `{}` said</sub></summary>\n\n\
1454         {fence}text\n{message}\n{fence}\n\n</details>\n",
1455        inline_untrusted(&f.analyzer)
1456    );
1457}
1458
1459/// `analyzer version` for each run that produced a member's findings, joined —
1460/// so "no findings" names the tool that looked rather than asserting a state of
1461/// the world.
1462fn describe_runs(runs: &[(String, String)]) -> String {
1463    if runs.is_empty() {
1464        // `Coverage::Ran` with no runs should not occur; say something true
1465        // rather than rendering an empty clause that reads as a name.
1466        return "an analyzer".to_owned();
1467    }
1468    runs.iter()
1469        .map(|(a, v)| {
1470            // `inline_untrusted`, not `inline_code`: this same value is also
1471            // interpolated into `<summary>`/`<sub>` by `write_analyzer_message`,
1472            // and keeping one treatment for it means there is no second context
1473            // to get wrong later. That is the mistake this pair of helpers was
1474            // introduced to fix, so it is not worth re-creating for monospace.
1475            let (a, v) = (inline_untrusted(a), inline_untrusted(v));
1476            // An ingested report often carries no version the producer stated,
1477            // and the store keeps that as `unknown`. Printing "`osv-scanner`
1478            // unknown" reads as a version string; saying nothing reads as the
1479            // absence it is.
1480            if v.is_empty() || v == "unknown" {
1481                format!("`{a}`")
1482            } else {
1483                format!("`{a}` {v}")
1484            }
1485        })
1486        .collect::<Vec<_>>()
1487        .join(", ")
1488}
1489
1490/// The `## Reproducing this vault` section — the **manifest** half of #442.
1491///
1492/// A vault that says *"here is my workspace"* is far less useful than one that
1493/// says *"here is my workspace **at these commits**"*. With an origin and a
1494/// commit per member, a reader can clone, check out, and hold exactly what this
1495/// vault describes; without them they have a picture and no way back to the
1496/// thing pictured.
1497///
1498/// It is also where the reader is told what a shared vault **does not** contain,
1499/// at the moment they are most likely to share it. Each exclusion is a decision
1500/// recorded on #442 rather than an oversight, so each is named with its reason.
1501fn write_manifest(c: &mut String, ws: &WorkspaceSummary) {
1502    c.push_str("\n## Reproducing this vault\n\n");
1503    let _ = writeln!(
1504        c,
1505        "*Rendered **{}**. A vault is read-only and point-in-time: it describes \
1506         these repositories at these commits, and does not change when they do.*\n",
1507        ws.generated_at
1508    );
1509    // "Repository", not "Clone from". The value is the **web root derived from
1510    // the `origin` remote**, which is the right thing to show a reader — an
1511    // `git@host:owner/repo.git` is only actionable for someone who already has
1512    // SSH to that host — but it is not guaranteed to be a working clone URL for
1513    // every forge or for a private repository. Naming the column "Clone from"
1514    // promised something this cannot deliver, and the manifest's whole value is
1515    // that its promises hold.
1516
1517    let any_origin = ws.members.iter().any(|m| m.repo_url.is_some());
1518    if any_origin {
1519        // Stated here rather than above the `if`: it says "each repository
1520        // below", and above the branch it would be followed immediately by
1521        // "no member has an `origin` remote".
1522        c.push_str(
1523            "*Each repository below is the web root derived from that member's \
1524             `origin` remote — where the code lives, not a guaranteed clone URL. \
1525             A private repository, or a forge with a different clone path, still \
1526             needs whatever access you would normally use.*\n\n",
1527        );
1528        c.push_str("| Member | Repository | At commit |\n| --- | --- | --- |\n");
1529        for m in &ws.members {
1530            let _ = writeln!(
1531                c,
1532                "| {} | {} | {} |",
1533                // The name is prose, not a span — `inline_untrusted` already
1534                // backslash-escapes `|`, so it is table-safe without becoming
1535                // monospaced. The two values below *are* spans, and go through
1536                // `table_cell`: a remote URL and a sha come from git, not from
1537                // us, and this table sits thirty lines above the one that
1538                // already knew that.
1539                inline_untrusted(&m.project),
1540                m.repo_url
1541                    .as_deref()
1542                    .map_or_else(|| "*(no `origin` remote)*".to_owned(), table_cell),
1543                m.commit
1544                    .as_deref()
1545                    .map_or_else(|| "*(unknown)*".to_owned(), table_cell),
1546            );
1547        }
1548    } else {
1549        // Every member is local-only. Say so rather than rendering a table of
1550        // "no origin" rows, which reads as a fault rather than as a workspace
1551        // that was never pushed anywhere.
1552        c.push_str(
1553            "*No member has an `origin` remote, so this vault cannot be \
1554             reconstructed from it — it describes repositories that exist only \
1555             where it was rendered.*\n",
1556        );
1557    }
1558
1559    write_rendered_under(c, ws);
1560
1561    c.push_str(
1562        "\n### Before you share this\n\n\
1563         Stated here because this is the point at which a vault is handed to \
1564         someone.\n\n\
1565         - **It lists this workspace's known security findings** — see *Security \
1566           findings* above. That is deliberate: a hand-over document should say \
1567           what is wrong as well as what is there. But it means this file is a \
1568           list of **unpatched weaknesses and where they are**, and unlike a \
1569           local store it cannot be un-shared once sent. Treat it as you would \
1570           the analyzer reports themselves.\n\
1571         - **Config keys are redacted by *name*, and that is narrower than it \
1572           looks.** The check matches ten well-known key names and inspects no \
1573           values, so a secret in a value whose key is not named like one — \
1574           `DATABASE_URL=postgres://user:pw@host` — is **not** redacted. \
1575           Tolerable in a local store; materially different in an artifact whose \
1576           purpose is to be handed on.\n\
1577         - **Agent memory** (ADR-0013) is the one thing deliberately left out. It \
1578           is per-developer and uncommitted, it records prose that can carry \
1579           pasted tokens, stack traces and customer names, and it has **no \
1580           redaction chokepoint** at all — so unlike the two above, there is no \
1581           version of it that is safe to include.\n",
1582    );
1583}
1584
1585/// The `## Cross-repo links` section: the edges that only a workspace vault can
1586/// show, and the honest statement of what is missing from them.
1587fn write_cross_links(c: &mut String, ws: &WorkspaceSummary) {
1588    c.push_str("\n## Cross-repo links\n\n");
1589    if ws.cross_links.is_empty() {
1590        c.push_str(
1591            "*None. These are the `inferred` cross-repo links `roteiro links \
1592             --infer --write` persists (ADR-0009); a workspace whose members have \
1593             never been inferred over has none recorded yet.*\n",
1594        );
1595        return;
1596    }
1597    // The caveat is per-row now, because the two provenances are no longer the
1598    // same claim: an **authored** row was declared by someone who knows the
1599    // topology, an **inferred** row is a scored guess. Saying "these are all
1600    // candidates" over a table containing declarations would understate the
1601    // declarations exactly as saying nothing would overstate the matches.
1602    let authored = ws.cross_links_authored;
1603    // `saturating_sub`: `WorkspaceSummary` is public, so a caller can hand us a
1604    // count larger than the total. In release that subtraction wraps and the
1605    // caption reports billions of inferred links — a rendering function should
1606    // not be the place an inconsistent input becomes nonsense. The debug
1607    // assertion says which caller was wrong, in the build that can afford to.
1608    debug_assert!(
1609        authored <= ws.cross_links_total,
1610        "cross_links_authored ({authored}) exceeds cross_links_total ({})",
1611        ws.cross_links_total
1612    );
1613    let inferred = ws.cross_links_total.saturating_sub(authored);
1614    let _ = writeln!(
1615        c,
1616        "*A spoke's config key and the hub key it corresponds to, across \
1617         repositories — the one thing a per-project vault structurally cannot \
1618         show. **{authored} declared** (`[[links]]`, ADR-0009 — a statement of \
1619         intent) and **{inferred} inferred** (`roteiro links --infer --write` — \
1620         read those as candidate correspondences).*\n"
1621    );
1622    c.push_str("| From | | To | Kind |\n| --- | --- | --- | --- |\n");
1623    for l in &ws.cross_links {
1624        let from_scope = VaultScope {
1625            project: Some(&l.from_project),
1626            members: &NO_MEMBERS,
1627        };
1628        let to = if l.resolves {
1629            format!("[[{}\\|{}]]", note_name(&l.to_qualified), l.to_qualified)
1630        } else {
1631            // Outside this workspace: there is no note to link to, and a wikilink
1632            // to a note that does not exist reads in Obsidian as one that is
1633            // merely unwritten.
1634            format!("`{}` *(outside this workspace)*", l.to_qualified)
1635        };
1636        // `declared` rather than a confidence score: an authored link carries no
1637        // score by construction, so an empty cell there would read as "confidence
1638        // unknown" instead of "not that kind of claim".
1639        let how = if l.authored {
1640            " *(declared)*".to_owned()
1641        } else {
1642            confidence(l.confidence)
1643        };
1644        let _ = writeln!(
1645            c,
1646            "| [[{}\\|{}]] | {} | {to} | {}{how} |",
1647            scoped_note_name(&from_scope, &l.from_key),
1648            l.from_name,
1649            l.from_project,
1650            l.kind,
1651        );
1652    }
1653    if ws.cross_links_total > ws.cross_links.len() {
1654        let _ = writeln!(
1655            c,
1656            "\n*Showing {} of {} — the full report is `roteiro links --matrix`.*",
1657            ws.cross_links.len(),
1658            ws.cross_links_total
1659        );
1660    }
1661    c.push_str(
1662        "\n*Shown in one direction only. The edge lives in the spoke's store, \
1663         pointing at a local placeholder for the hub's node, so the hub's own note \
1664         carries no matching **Incoming** entry — Obsidian's **Backlinks** pane \
1665         still shows it, because the link is in the vault.*\n",
1666    );
1667}
1668
1669#[cfg(test)]
1670mod tests {
1671    use super::{
1672        AdrEntry, ConfigSecretSummary, CouplingEntry, Coverage, CrossLink, DensityEntry,
1673        FindingEntry, HOME_NOTE, RenderedUnder, VaultScope, VaultSummary, WorkspaceSummary,
1674        note_name, render_home, render_note, render_note_scoped, render_workspace_home,
1675        scoped_note_name,
1676    };
1677    use rto_graph::{EdgeRef, Explanation, NodeSummary};
1678
1679    /// The shape of a name, pinned once so a change to it is a deliberate edit
1680    /// here rather than a diff spread over twenty other assertions.
1681    ///
1682    /// Everything else in this module composes `note_name` instead of repeating
1683    /// its output, because those tests are about *which key a link points at* and
1684    /// were never about the spelling.
1685    #[test]
1686    fn note_name_is_a_lowercase_hint_and_a_hash_of_the_whole_key() {
1687        assert_eq!(
1688            note_name("sym:rust:src/a.rs#Store"),
1689            "sym-rust-src-a.rs-store-b4cbf6633003361f"
1690        );
1691        assert_eq!(note_name("adr:0001"), "adr-0001-559a2e837953b2ff");
1692        assert_eq!(
1693            note_name("file:src/main.rs"),
1694            "file-src-main.rs-4a72627453f6780e"
1695        );
1696        // Deterministic: the suffix is a pure function of the key, so a vault
1697        // renders the same names on every machine and every run.
1698        assert_eq!(note_name("adr:0001"), note_name("adr:0001"));
1699    }
1700
1701    /// **The property `note_name` exists to have** (issue #574): distinct keys
1702    /// give distinct notes *on a case-folding filesystem*, which is where the
1703    /// vault was losing them.
1704    ///
1705    /// Asserted over lowercased names, not names. On macOS and Windows two names
1706    /// differing only in case are one file, so a name set that is distinct as
1707    /// strings can still be a vault with notes missing — and Linux CI cannot see
1708    /// it. Folding here makes the assertion say what the filesystem says, on
1709    /// every platform.
1710    ///
1711    /// The keys are the two mechanisms that were actually losing notes, taken
1712    /// from this repository's own render rather than invented: the vendored
1713    /// `cytoscape.min.js` bundle whose minified single-letter symbols differ only
1714    /// by a sigil or by case, and a pair of grouped Rust `use` keys differing
1715    /// only by a trailing comma. `render_cli` runs the same assertion end to end
1716    /// over a rendered vault; this is the unit-level statement of it.
1717    #[test]
1718    fn distinct_keys_give_distinct_notes_even_after_case_folding() {
1719        const JS: &str = "sym:javascript:crates/roteiro/src/assets/cytoscape.min.js";
1720        let keys: Vec<String> = [
1721            // Slug lossiness: the sigil and the letter both slugged to the same
1722            // thing (9 notes lost this way, on every platform).
1723            format!("{JS}#$a"),
1724            format!("{JS}#a"),
1725            format!("{JS}#$o"),
1726            format!("{JS}#o"),
1727            // Case folding: distinct names, one file (95 notes lost this way, and
1728            // only on macOS and Windows).
1729            format!("{JS}#A"),
1730            format!("{JS}#O"),
1731            format!("{JS}#S"),
1732            format!("{JS}#s"),
1733            // Real source symbols, same shape.
1734            "sym:rust:crates/rto-exec/src/sandbox_store.rs#Store".into(),
1735            "sym:rust:crates/rto-exec/src/sandbox_store.rs#store".into(),
1736            // A trailing comma is the whole difference between these two.
1737            "import:rust:crate::engine::{ChatRequest,Engine,ModelInfo,}".into(),
1738            "import:rust:crate::engine::{ChatRequest,Engine,ModelInfo}".into(),
1739            // Nothing but separators: no hint at all, so the name is bare hash.
1740            "::".into(),
1741            "##".into(),
1742            // Over the length bound, differing only past the truncation point —
1743            // the case truncation alone used to merge.
1744            format!("import:rust:{}A", "a::b::c,".repeat(60)),
1745            format!("import:rust:{}a", "a::b::c,".repeat(60)),
1746        ]
1747        .into();
1748
1749        let folded: std::collections::BTreeSet<String> =
1750            keys.iter().map(|k| note_name(k).to_lowercase()).collect();
1751        assert_eq!(
1752            folded.len(),
1753            keys.len(),
1754            "two keys share a note after case folding; the vault would hold one \
1755             file for both and report two"
1756        );
1757    }
1758
1759    /// Case folding is the identity on a note name, so the assertion above is not
1760    /// weaker than the filesystem it stands in for.
1761    ///
1762    /// This is the reason the hint is lowercased rather than case-preserved: it
1763    /// makes "distinct names" and "distinct files on macOS" the same statement,
1764    /// so there is no version of this module that passes on Linux and loses notes
1765    /// on a Mac. Without it, the two assertions could drift apart and only the
1766    /// weaker one would ever run in CI.
1767    #[test]
1768    fn a_note_name_is_already_lowercase() {
1769        for key in [
1770            "sym:rust:src/a.rs#Store",
1771            "file:README.md",
1772            "app::file:CHANGELOG.md",
1773            "sym:javascript:a.js#ABC",
1774        ] {
1775            let name = note_name(key);
1776            assert_eq!(name, name.to_lowercase(), "`{key}` kept case in its name");
1777        }
1778    }
1779
1780    /// `_Home` is a name in the same namespace as every note, and it is not
1781    /// derived from a key — so nothing must be able to collide with it. The
1782    /// mandatory suffix gives that for free: every generated name either ends in
1783    /// `-<16 hex>` or *is* 16 hex digits, and `_home` is neither.
1784    #[test]
1785    fn no_key_can_claim_the_home_note() {
1786        for key in ["_Home", "file:_Home", "_home", "::_Home::"] {
1787            assert_ne!(
1788                format!("{}.md", note_name(key)).to_lowercase(),
1789                HOME_NOTE.to_lowercase(),
1790                "`{key}` would overwrite the overview note"
1791            );
1792        }
1793    }
1794
1795    #[test]
1796    fn render_note_emits_frontmatter_and_wikilinks() {
1797        let ex = Explanation {
1798            schema: rto_graph::SCHEMA,
1799            node: NodeSummary {
1800                key: "sym:rust:a.rs#main".into(),
1801                kind: "fn".into(),
1802                name: "main".into(),
1803                path: Some("a.rs".into()),
1804                lang: Some("rust".into()),
1805            },
1806            meta: serde_json::Value::Null,
1807            outgoing: vec![EdgeRef {
1808                kind: "calls".into(),
1809                provenance: "derived",
1810                confidence: None,
1811                node: "sym:rust:a.rs#helper".into(),
1812            }],
1813            incoming: vec![EdgeRef {
1814                kind: "references".into(),
1815                provenance: "authored",
1816                confidence: None,
1817                node: "adr:0001".into(),
1818            }],
1819        };
1820        let note = render_note(&ex, None, None);
1821        assert_eq!(
1822            note.filename,
1823            format!("{}.md", note_name("sym:rust:a.rs#main"))
1824        );
1825        assert!(note.content.contains("kind: fn"));
1826        // No source base → no Source link.
1827        assert!(!note.content.contains("**Source:**"));
1828        assert!(note.content.contains("# main"));
1829        assert!(note.content.contains(&format!(
1830            "- calls (derived) → [[{}]]",
1831            note_name("sym:rust:a.rs#helper")
1832        )));
1833        assert!(note.content.contains(&format!(
1834            "- [[{}]] references (authored) →",
1835            note_name("adr:0001")
1836        )));
1837        // Tags for the graph view.
1838        assert!(note.content.contains("- roteiro/kind/fn"));
1839        assert!(note.content.contains("- roteiro/lang/rust"));
1840    }
1841
1842    #[test]
1843    fn note_name_bounds_long_keys_deterministically() {
1844        let long = format!("import:rust:{}", "a::b::c,".repeat(60));
1845        let a = note_name(&long);
1846        let b = note_name(&long);
1847        assert_eq!(a, b, "deterministic");
1848        assert!(
1849            a.len() <= 205,
1850            "bounded under the filename limit: {}",
1851            a.len()
1852        );
1853        assert_ne!(
1854            note_name(&format!("{long}x")),
1855            a,
1856            "different keys stay distinct after truncation"
1857        );
1858        // Truncation must not leave a doubled separator before the suffix — the
1859        // hint is trimmed after cutting, not before.
1860        assert!(!a.contains("--"), "{a}");
1861    }
1862
1863    /// A short key is bounded too, and every name carries the suffix — the hash
1864    /// is no longer reached for only when the hint overruns.
1865    ///
1866    /// That gating was the defect (#574): two keys short enough to skip the hash
1867    /// had nothing left to tell them apart once the slug had flattened them.
1868    #[test]
1869    fn every_name_carries_the_hash_however_short_the_key() {
1870        for key in ["a", "adr:0001", "file:README.md"] {
1871            let name = note_name(key);
1872            let (hint, hash) = name.rsplit_once('-').expect("a suffixed name");
1873            assert!(!hint.is_empty(), "{name}");
1874            assert_eq!(hash.len(), 16, "{name}");
1875            assert!(
1876                hash.chars().all(|c| c.is_ascii_hexdigit()),
1877                "the suffix is the key's hash, not part of the hint: {name}"
1878            );
1879        }
1880        // A key with no hint at all is the bare hash, which cannot be mistaken
1881        // for a hinted name (those are at least 18 characters).
1882        let bare = note_name("::");
1883        assert_eq!(bare.len(), 16, "{bare}");
1884        assert!(!bare.contains('-'), "{bare}");
1885    }
1886
1887    #[test]
1888    fn render_note_surfaces_content_and_status() {
1889        let ex = Explanation {
1890            schema: rto_graph::SCHEMA,
1891            node: NodeSummary {
1892                key: "adr:0001".into(),
1893                kind: "adr".into(),
1894                name: "Build Roteiro".into(),
1895                path: Some("docs/adr/0001.md".into()),
1896                lang: None,
1897            },
1898            meta: serde_json::json!({ "status": "Accepted", "content": "The decision text." }),
1899            outgoing: vec![],
1900            incoming: vec![],
1901        };
1902        let note = render_note(&ex, Some("https://github.com/org/repo/blob/abc123"), None);
1903        assert!(note.content.contains("status: Accepted"));
1904        assert!(note.content.contains("- roteiro/status/accepted"));
1905        assert!(note.content.contains("> **Status:** Accepted"));
1906        assert!(note.content.contains("## Content\n\nThe decision text."));
1907        // A clickable link to the actual ADR file on the repository host.
1908        assert!(
1909            note.content.contains(
1910                "**Source:** [`docs/adr/0001.md`](https://github.com/org/repo/blob/abc123/docs/adr/0001.md)"
1911            ),
1912            "{}",
1913            note.content
1914        );
1915    }
1916
1917    /// The structured document a prose note is supposed to reproduce: headings, a
1918    /// table and a fenced code block, none of which survive whitespace collapse.
1919    const DOC: &str = "# Working offline\n\nRoteiro is **offline-capable**.\n\n| Host | What |\n| --- | --- |\n| `example.com` | models |\n\n```sh\nroteiro model pull\n```\n";
1920
1921    fn prose_note(content: Option<&str>) -> Explanation {
1922        Explanation {
1923            schema: rto_graph::SCHEMA,
1924            node: NodeSummary {
1925                key: "file:docs/OFFLINE_SETUP.md".into(),
1926                kind: "file".into(),
1927                name: "OFFLINE_SETUP.md".into(),
1928                path: Some("docs/OFFLINE_SETUP.md".into()),
1929                lang: None,
1930            },
1931            meta: content.map_or(
1932                serde_json::Value::Null,
1933                |c| serde_json::json!({ "content": c }),
1934            ),
1935            outgoing: vec![],
1936            incoming: vec![],
1937        }
1938    }
1939
1940    /// The whole readability defect, in one assertion pair: a note built from
1941    /// `meta.content` alone is the document whitespace-collapsed onto one line,
1942    /// and a note built from the source is the document.
1943    ///
1944    /// The newline count is the claim. A character count alone would pass on a
1945    /// note that had merely grown longer while staying flat, which is exactly the
1946    /// failure being fixed — `meta.content` is capped *and* collapsed, and only
1947    /// the collapse is what makes it unreadable.
1948    #[test]
1949    fn a_supplied_body_supersedes_the_collapsed_stored_content() {
1950        // What extraction stores: the same text, whitespace-collapsed.
1951        let collapsed = DOC.split_whitespace().collect::<Vec<_>>().join(" ");
1952        let ex = prose_note(Some(&collapsed));
1953
1954        let note = render_note(&ex, None, Some(DOC));
1955        assert!(
1956            note.content.contains(DOC.trim()),
1957            "the source document is reproduced verbatim: {}",
1958            note.content
1959        );
1960        assert!(
1961            !note.content.contains(&collapsed),
1962            "the collapsed rendering is replaced, not appended: {}",
1963            note.content
1964        );
1965        assert!(
1966            note.content.contains("\n| Host | What |\n"),
1967            "a table needs its own lines to be a table: {}",
1968            note.content
1969        );
1970        assert!(
1971            note.content.contains("\n```sh\n"),
1972            "a fenced block needs its own lines to be a fence: {}",
1973            note.content
1974        );
1975
1976        // The flat control: the same node with no body is the one-line note.
1977        let flat = render_note(&ex, None, None);
1978        assert!(
1979            flat.content.contains(&collapsed),
1980            "without a body the stored content is still shown: {}",
1981            flat.content
1982        );
1983        assert!(
1984            content_lines(&note.content) > content_lines(&flat.content),
1985            "structure restored: {} line(s) with a body vs {} without",
1986            content_lines(&note.content),
1987            content_lines(&flat.content)
1988        );
1989        assert_eq!(
1990            content_lines(&flat.content),
1991            1,
1992            "the defect: the stored content is a single line"
1993        );
1994    }
1995
1996    /// A doc comment is a summary of a definition, not a document, and its note is
1997    /// correct as it stands. The caller supplies no body for these, so this pins
1998    /// the unchanged path — the fix must not depend on every node gaining one.
1999    #[test]
2000    fn a_note_with_no_body_is_unchanged() {
2001        let ex = Explanation {
2002            schema: rto_graph::SCHEMA,
2003            node: NodeSummary {
2004                key: "sym:rust:a.rs#main".into(),
2005                kind: "fn".into(),
2006                name: "main".into(),
2007                path: Some("a.rs".into()),
2008                lang: Some("rust".into()),
2009            },
2010            meta: serde_json::json!({ "content": "Entry point." }),
2011            outgoing: vec![],
2012            incoming: vec![],
2013        };
2014        assert!(
2015            render_note(&ex, None, None)
2016                .content
2017                .contains("## Content\n\nEntry point.")
2018        );
2019    }
2020
2021    /// Lines in the note's `## Content` section.
2022    fn content_lines(note: &str) -> usize {
2023        let body = note
2024            .split_once("## Content\n\n")
2025            .map_or("", |(_, rest)| rest);
2026        let body = body.split_once("\n## ").map_or(body, |(head, _)| head);
2027        body.trim_end().lines().count()
2028    }
2029
2030    #[test]
2031    fn render_note_shows_inferred_confidence() {
2032        let ex = Explanation {
2033            schema: rto_graph::SCHEMA,
2034            node: NodeSummary {
2035                key: "file:a.md".into(),
2036                kind: "file".into(),
2037                name: "a.md".into(),
2038                path: Some("a.md".into()),
2039                lang: None,
2040            },
2041            meta: serde_json::Value::Null,
2042            outgoing: vec![EdgeRef {
2043                kind: "related".into(),
2044                provenance: "inferred",
2045                confidence: Some(0.82),
2046                node: "file:b.md".into(),
2047            }],
2048            incoming: vec![],
2049        };
2050        let note = render_note(&ex, None, None);
2051        assert!(
2052            note.content.contains(&format!(
2053                "related (inferred) (0.82) → [[{}]]",
2054                note_name("file:b.md")
2055            )),
2056            "{}",
2057            note.content
2058        );
2059    }
2060
2061    /// The single-project `_Home` fixture, extracted so the test that reads its
2062    /// rendering stays about the rendering. Every field is populated: the
2063    /// sections it drives are individually omitted when empty, so a partial
2064    /// fixture would silently stop exercising them.
2065    fn home_summary() -> VaultSummary {
2066        VaultSummary {
2067            project: "demo".into(),
2068            total_nodes: 3,
2069            total_edges: 2,
2070            node_counts: vec![("fn".into(), 2), ("adr".into(), 1)],
2071            edge_provenance: vec![("derived".into(), 1), ("authored".into(), 1)],
2072            adrs: vec![AdrEntry {
2073                key: "adr:0001".into(),
2074                name: "First".into(),
2075                status: Some("Accepted".into()),
2076            }],
2077            debt: vec![("todo".into(), 4)], // roteiro:ignore
2078            densest_files: vec![DensityEntry {
2079                path: "src/small.rs".into(),
2080                markers: 3,
2081                lines: 120,
2082                per_kloc: 25.0,
2083            }],
2084            config_secrets: Some(ConfigSecretSummary {
2085                secret_named: 4,
2086                redacted: 3,
2087                declared: 1,
2088                unredacted: 0,
2089                files: vec![".env".into()],
2090            }),
2091            most_called: vec![CouplingEntry {
2092                key: "sym:rust:a.rs#helper".into(),
2093                name: "helper".into(),
2094                fan_in: 7,
2095                fan_out: 1,
2096            }],
2097            repo_url: Some("https://github.com/org/repo".into()),
2098            commit: Some("abcdef0123456789".into()),
2099            findings: vec![],
2100            coverage: Coverage::NotRun,
2101            settings: RenderedUnder::default(),
2102        }
2103    }
2104
2105    #[test]
2106    fn render_home_summarises_the_graph() {
2107        let summary = home_summary();
2108        let note = render_home(&summary);
2109        assert_eq!(note.filename, HOME_NOTE);
2110        assert!(note.content.contains("# demo — knowledge graph"));
2111        assert!(note.content.contains("**3 nodes**, **2 edges**"));
2112        assert!(note.content.contains("| fn | 2 |"));
2113        assert!(note.content.contains("| derived | 1 |"));
2114        assert!(note.content.contains(&format!(
2115            "**Accepted** — [[{}|First]]",
2116            note_name("adr:0001")
2117        )));
2118        assert!(note.content.contains("| todo | 4 |")); // roteiro:ignore
2119        // Directed coupling: the two fans are separate columns, and the wikilink's
2120        // own `|` is escaped so it cannot break the table it sits in.
2121        assert!(
2122            note.content.contains(&format!(
2123                "| [[{}\\|helper]] | 7 | 1 |",
2124                note_name("sym:rust:a.rs#helper")
2125            )),
2126            "{}",
2127            note.content
2128        );
2129        assert!(
2130            note.content.contains("resolved by simple name"),
2131            "the precision caveat travels with the figures"
2132        );
2133        // Density: the count and the denominator are both shown, so the ratio can
2134        // be checked rather than taken on trust, and the wikilink's own `|` is
2135        // escaped so it cannot break the table it sits in.
2136        assert!(
2137            note.content.contains(&format!(
2138                "| [[{}\\|src/small.rs]] | 3 | 120 | 25.00 |",
2139                note_name("file:src/small.rs")
2140            )),
2141            "{}",
2142            note.content
2143        );
2144        assert!(
2145            note.content.contains("not source lines of code"),
2146            "the denominator caveat travels with the figures"
2147        );
2148        // Config secrets: counts and files, and no key names — a vault note is
2149        // browsed out of context, which is the wrong place for a list that would
2150        // read as a secret scan's output.
2151        assert!(
2152            note.content.contains(
2153                "**4** secret-named config key(s): 3 redacted before storage, 1 \
2154                 declared in code without a value, 0 unredacted."
2155            ),
2156            "{}",
2157            note.content
2158        );
2159        assert!(
2160            note.content
2161                .contains(&format!("- [[{}\\|.env]]", note_name("file:.env"))),
2162            "{}",
2163            note.content
2164        );
2165        assert!(
2166            note.content.contains("not a secret scan")
2167                && note.content.contains("cannot see a hardcoded credential"),
2168            "the limitation travels with the figures: {}",
2169            note.content
2170        );
2171        assert!(
2172            !note.content.contains("[!warning]"),
2173            "no warning when nothing is unredacted: {}",
2174            note.content
2175        );
2176        // A repository link + short-commit permalink note.
2177        assert!(
2178            note.content
2179                .contains("**Repository:** [https://github.com/org/repo](https://github.com/org/repo) · rendered at commit `abcdef012345`"),
2180            "{}",
2181            note.content
2182        );
2183    }
2184
2185    #[test]
2186    fn render_home_omits_density_for_a_graph_with_no_markers() {
2187        // A clean repository has no markers, so there is no density to rank. An
2188        // empty table under a heading reads as "measured, and there is nothing";
2189        // the section is absent instead. Same rule as the coupling table below.
2190        let note = render_home(&VaultSummary {
2191            project: "clean".into(),
2192            total_nodes: 1,
2193            ..VaultSummary::default()
2194        });
2195        assert!(
2196            !note.content.contains("Densest files"),
2197            "no heading without rows: {}",
2198            note.content
2199        );
2200        // The intent-debt section itself still renders — density is an addition
2201        // to it, not a replacement.
2202        assert!(note.content.contains("## Intent debt"));
2203        assert!(note.content.contains("*None recorded.*"));
2204    }
2205
2206    #[test]
2207    fn render_home_omits_config_secrets_rather_than_rendering_zeroes() {
2208        // A row of zeroes under this heading would read as "scanned, and clean" —
2209        // a conclusion the lens cannot support, since a credential under an
2210        // innocuous key name never appears in it. The section is absent instead.
2211        let note = render_home(&VaultSummary {
2212            project: "clean".into(),
2213            total_nodes: 1,
2214            ..VaultSummary::default()
2215        });
2216        assert!(
2217            !note.content.contains("named like secrets"),
2218            "no heading without figures: {}",
2219            note.content
2220        );
2221    }
2222
2223    #[test]
2224    fn render_home_warns_loudly_about_an_unredacted_value() {
2225        // Extraction cannot produce this state, so if it appears something else
2226        // put an unredacted value in the store — and the note must say where to
2227        // look rather than implicating the repository.
2228        let note = render_home(&VaultSummary {
2229            project: "imported".into(),
2230            total_nodes: 1,
2231            config_secrets: Some(ConfigSecretSummary {
2232                secret_named: 1,
2233                redacted: 0,
2234                declared: 0,
2235                unredacted: 1,
2236                files: vec!["imported.env".into()],
2237            }),
2238            ..VaultSummary::default()
2239        });
2240        assert!(
2241            note.content.contains("[!warning]") && note.content.contains("**unredacted**"),
2242            "{}",
2243            note.content
2244        );
2245        assert!(
2246            note.content.contains("came from an import layer"),
2247            "and it points at the importing tool, not the repository: {}",
2248            note.content
2249        );
2250    }
2251
2252    #[test]
2253    fn render_home_omits_coupling_for_a_graph_with_no_calls() {
2254        // A prose-only vault has no `calls` edges. An empty table under a heading
2255        // reads as "measured, and there is nothing" — the section is absent instead.
2256        let note = render_home(&VaultSummary {
2257            project: "docs".into(),
2258            total_nodes: 1,
2259            ..VaultSummary::default()
2260        });
2261        assert!(
2262            !note.content.contains("Most depended-on"),
2263            "no heading without rows: {}",
2264            note.content
2265        );
2266        // The rest of the overview is unaffected.
2267        assert!(note.content.contains("# docs — knowledge graph"));
2268    }
2269
2270    // ---- Workspace vaults (issue #442 part 1) --------------------------------
2271
2272    /// A `Explanation` for `key`, with one outgoing edge to `to`.
2273    fn node_linking_to(key: &str, name: &str, to: &str) -> Explanation {
2274        Explanation {
2275            schema: rto_graph::SCHEMA,
2276            node: NodeSummary {
2277                key: key.into(),
2278                kind: "config_key".into(),
2279                name: name.into(),
2280                path: Some("config.toml".into()),
2281                lang: None,
2282            },
2283            meta: serde_json::Value::Null,
2284            outgoing: vec![EdgeRef {
2285                kind: "links".into(),
2286                provenance: "inferred",
2287                confidence: Some(0.91),
2288                node: to.into(),
2289            }],
2290            incoming: vec![],
2291        }
2292    }
2293
2294    fn members(names: &[&str]) -> std::collections::BTreeSet<String> {
2295        names.iter().map(|s| (*s).to_owned()).collect()
2296    }
2297
2298    /// **Rewritten deliberately under #574.** #570 landed this as "a project
2299    /// scope leaves every note name exactly as it was", and read that two ways at
2300    /// once: `PROJECT` reduces to `note_name`, *and* `note_name` itself does not
2301    /// move. #574 breaks the second half on purpose — the old names were not
2302    /// injective under filename case folding and this repository's vault lost 104
2303    /// notes to it — so the two halves are separated here rather than having
2304    /// expected values quietly updated underneath the old title.
2305    ///
2306    /// What survives is the half #570 was actually about, and it is unweakened:
2307    /// **turning workspace mode on must not rename a project's notes.** Names may
2308    /// move when `note_name` changes, for a reason argued at `note_name`; they may
2309    /// never move because a repository happens to sit inside a configured
2310    /// workspace, because that would happen by inference rather than by a release.
2311    ///
2312    /// The other half of #570's promise — that a project render is byte-identical
2313    /// apart from names — is now [`render_note_is_the_project_scoped_render_byte_for_byte`]
2314    /// and `render_cli`'s end-to-end pair.
2315    #[test]
2316    fn a_project_scope_never_qualifies_a_name() {
2317        // A user's own notes live outside the vault and link into it *by name*
2318        // (#442), so a rename breaks them silently, with no error and nothing to
2319        // grep for. Whatever workspace mode does, `VaultScope::PROJECT` must
2320        // reduce to `note_name` of the bare key.
2321        for key in [
2322            "file:README.md",
2323            "adr:0001",
2324            "sym:rust:src/a.rs#Store",
2325            "extref:other::file:README.md",
2326            "cfgkey:config.toml#serve.addr",
2327        ] {
2328            assert_eq!(
2329                scoped_note_name(&VaultScope::PROJECT, key),
2330                note_name(key),
2331                "single-project name moved for `{key}`"
2332            );
2333            // And the qualified form really is a different name, so the assertion
2334            // above is not vacuously true of every scope.
2335            let ms = members(&["app"]);
2336            assert_ne!(
2337                scoped_note_name(
2338                    &VaultScope {
2339                        project: Some("app"),
2340                        members: &ms,
2341                    },
2342                    key
2343                ),
2344                note_name(key),
2345                "qualification must move the name for `{key}`, or nothing above holds"
2346            );
2347        }
2348    }
2349
2350    #[test]
2351    fn render_note_is_the_project_scoped_render_byte_for_byte() {
2352        let ex = node_linking_to("cfgkey:config.toml#addr", "addr", "sym:rust:a.rs#A");
2353        assert_eq!(
2354            render_note(&ex, Some("https://h/b"), Some("body")),
2355            render_note_scoped(&ex, Some("https://h/b"), Some("body"), &VaultScope::PROJECT),
2356            "the unscoped entry point must stay the scoped one at PROJECT, so the \
2357             two cannot drift apart"
2358        );
2359    }
2360
2361    #[test]
2362    fn each_member_gets_its_own_note_for_the_same_key() {
2363        // The collision the whole feature exists for: node keys are
2364        // repository-relative, so every member's `README.md` is `file:README.md`.
2365        let ms = members(&["api", "sdk"]);
2366        let names: Vec<String> = ["api", "sdk"]
2367            .iter()
2368            .map(|p| {
2369                scoped_note_name(
2370                    &VaultScope {
2371                        project: Some(p),
2372                        members: &ms,
2373                    },
2374                    "file:README.md",
2375                )
2376            })
2377            .collect();
2378        assert_eq!(
2379            names,
2380            [
2381                note_name("api::file:README.md"),
2382                note_name("sdk::file:README.md")
2383            ]
2384        );
2385        assert_ne!(names[0], names[1], "two members must not share one note");
2386    }
2387
2388    /// The two names this feature has, pinned together in one place.
2389    ///
2390    /// They are easy to conflate and were, in this PR, described inconsistently
2391    /// in two doc comments — the **key** is `<project>::<key>` (ADR-0009's
2392    /// cross-repo form, which is why cross-repo links resolve), and the **note
2393    /// name** is [`note_name`] of that key, in which `::` has become `-`. A
2394    /// reader told the wrong one goes looking for a file with `::` in it.
2395    ///
2396    /// Asserting both here means the next description that drifts has something
2397    /// to disagree with, rather than waiting for a reviewer to read two comments
2398    /// side by side.
2399    #[test]
2400    fn the_qualified_key_and_the_note_name_are_different_strings() {
2401        let ms = members(&["app"]);
2402        let scope = VaultScope {
2403            project: Some("app"),
2404            members: &ms,
2405        };
2406        // The key: project-qualified, `::` intact — this is what the graph and
2407        // ADR-0009's external refs use.
2408        let qualified = "app::file:README.md";
2409        // The note name: `note_name` of exactly that key, `::` slugged to `-`,
2410        // the whole hint lowercased, and the key's own hash appended.
2411        assert_eq!(
2412            scoped_note_name(&scope, "file:README.md"),
2413            "app-file-readme.md-a114bde6dcaba1c1"
2414        );
2415        assert_eq!(note_name(qualified), "app-file-readme.md-a114bde6dcaba1c1");
2416        assert!(
2417            !scoped_note_name(&scope, "file:README.md").contains("::"),
2418            "no note name ever contains `::`"
2419        );
2420        // And on disk the stem gains the extension, which is the string a reader
2421        // actually looks for.
2422        let note = render_note_scoped(
2423            &node_with("file:README.md", Some("README.md"), None),
2424            None,
2425            None,
2426            &scope,
2427        );
2428        assert_eq!(note.filename, "app-file-readme.md-a114bde6dcaba1c1.md");
2429    }
2430
2431    /// `_Home` must *show* a name, not spell the form out.
2432    ///
2433    /// The test above pins the distinction in the code. It did not stop the
2434    /// distinction being described wrongly in the same file, because it guards
2435    /// the function and not the sentences: `render_workspace_home` went on
2436    /// writing the pre-#574 form into the `_Home` of every workspace vault
2437    /// v2.0.0 built, and nothing here disagreed with it.
2438    ///
2439    /// So this asserts the property that made that possible is gone — the
2440    /// paragraph now contains a string `note_name` actually produced for a key
2441    /// the workspace really holds, which a hand-written spelling cannot
2442    /// satisfy. It is not a tautology despite both sides calling `note_name`:
2443    /// what it rejects is the *shape* of the old copy, a form written out by
2444    /// hand next to the function that could have rendered it.
2445    ///
2446    /// That the *key* is real is the other half, and the reason the example is
2447    /// drawn from `cross_links` rather than invented from the member list —
2448    /// a name rendered for a node the vault does not hold is a true sentence
2449    /// about a note nobody can open. The empty case is
2450    /// `the_workspace_home_claims_no_example_note_when_it_has_no_real_key`.
2451    fn finding(rule: &str, severity: &str, message: &str) -> FindingEntry {
2452        FindingEntry {
2453            rule: rule.to_owned(),
2454            severity: severity.to_owned(),
2455            title: "a title".to_owned(),
2456            message: message.to_owned(),
2457            path: Some("Cargo.lock".to_owned()),
2458            analyzer: "osv-scanner".to_owned(),
2459        }
2460    }
2461
2462    fn ws_with(members: Vec<VaultSummary>) -> WorkspaceSummary {
2463        WorkspaceSummary {
2464            generated_at: "2026-08-22T10:00:00Z".to_owned(),
2465            name: "platform".into(),
2466            members,
2467            cross_links: vec![],
2468            cross_links_total: 0,
2469            cross_links_authored: 0,
2470        }
2471    }
2472
2473    /// An analyzer message is **tool output** being embedded in a document that
2474    /// gets handed to people. It must be quoted, never absorbed: a message
2475    /// carrying its own fence would otherwise close the block early and spill the
2476    /// rest into the note as vault prose — headings and links included.
2477    #[test]
2478    fn an_analyzer_message_cannot_break_out_of_its_fence() {
2479        let mut api = member_summary("api", 1);
2480        api.coverage = Coverage::Ran(vec![("osv-scanner".into(), "1.9.0".into())]);
2481        api.findings = vec![finding(
2482            "GHSA-x",
2483            "high",
2484            "before
2485```
2486## not a vault heading
2487```
2488after",
2489        )];
2490        let c = render_workspace_home(&ws_with(vec![api])).content;
2491
2492        // The fence opened must be longer than any run inside the message, so the
2493        // message's own ``` is content rather than a terminator.
2494        assert!(
2495            c.contains("````text"),
2496            "fence widened past the message's own: {c}"
2497        );
2498        assert!(
2499            c.contains("## not a vault heading"),
2500            "and the text is still all there: {c}"
2501        );
2502        // Nothing between the finding and the next section may be loose prose.
2503        let after = c.split("````text").nth(1).expect("a fenced block");
2504        let body = after.split("````").next().expect("a closing fence");
2505        assert!(
2506            body.contains("## not a vault heading"),
2507            "the heading is INSIDE the fence, not outside it: {body}"
2508        );
2509    }
2510
2511    /// The fence secured the **message**. Every other analyzer-sourced field on a
2512    /// finding is interpolated into a line of prose, and they come from the same
2513    /// place — so a title carrying a newline used to end its line and start a new
2514    /// block, which can open a heading the vault never wrote.
2515    #[test]
2516    fn every_analyzer_sourced_field_is_quoted_not_absorbed() {
2517        let mut api = member_summary("api", 1);
2518        api.coverage = Coverage::Ran(vec![("osv-scanner".into(), "1.9.0".into())]);
2519        api.findings = vec![FindingEntry {
2520            // Each field carries a different escape hatch.
2521            rule: "R-1`x".to_owned(),
2522            severity: "high".to_owned(),
2523            // Newlines, a heading, a link, and a list marker — the last of which
2524            // no escape set covers, so only the whitespace collapse stops it.
2525            title: "broken\n\n## an injected heading\n\n- a list item\n\n[a link](http://evil)"
2526                .to_owned(),
2527            message: "fine".to_owned(),
2528            path: Some("src/a`b.rs".to_owned()),
2529            analyzer: "osv-scanner".to_owned(),
2530        }];
2531        let c = render_workspace_home(&ws_with(vec![api])).content;
2532
2533        assert!(
2534            !c.contains("\n## an injected heading"),
2535            "a newline in a title must not start a block: {c}"
2536        );
2537        assert!(
2538            !c.contains("[a link](http://evil)"),
2539            "and a link must not survive as a link: {c}"
2540        );
2541        // The text is still readable — escaped, not deleted. Losing it would be a
2542        // different failure: a finding whose title vanished is a finding nobody
2543        // acts on.
2544        assert!(c.contains("an injected heading"), "the words remain: {c}");
2545        // A widened delimiter, not a backslash: inside a span the backslash
2546        // would show, and `vendor/**` would reach the reader as `vendor\*\*`.
2547        assert!(
2548            c.contains("``R-1`x``"),
2549            "the span widens past the backtick inside it: {c}"
2550        );
2551
2552        // The structural property, asserted directly rather than inferred from
2553        // the absence of a heading: the whole finding stays on **one line**. An
2554        // escape set can neutralise `#` and `[`, and a newline would still end
2555        // the line and start a block — and nothing escapes a `-` list marker.
2556        let line = c
2557            .lines()
2558            .find(|l| l.contains("R-1"))
2559            .unwrap_or_else(|| panic!("the finding renders: {c}"));
2560        assert!(
2561            line.contains("a link") && line.contains("a list item"),
2562            "every part of the title stays on the finding's own line: {line}"
2563        );
2564    }
2565
2566    /// The analyzer name is interpolated into **raw HTML** (`<sub>`,
2567    /// `<details><summary>`), where a backslash escapes nothing. Fixing the
2568    /// Markdown surface and leaving this one is how the first version of the
2569    /// escaping shipped.
2570    #[test]
2571    fn analyzer_text_cannot_break_out_of_the_html_it_sits_in() {
2572        let mut api = member_summary("api", 1);
2573        let hostile = "osv</sub><script>alert(1)</script>".to_owned();
2574        api.coverage = Coverage::Ran(vec![(hostile.clone(), "1.0".into())]);
2575        api.findings = vec![FindingEntry {
2576            rule: "R-1".to_owned(),
2577            severity: "high".to_owned(),
2578            title: "t".to_owned(),
2579            message: "m".to_owned(),
2580            path: None,
2581            analyzer: hostile,
2582        }];
2583        let c = render_workspace_home(&ws_with(vec![api])).content;
2584
2585        assert!(
2586            !c.contains("<script>"),
2587            "a tag in analyzer text must not survive as a tag: {c}"
2588        );
2589        assert!(
2590            !c.contains("osv</sub>"),
2591            "and must not close the element it was placed inside: {c}"
2592        );
2593        // Escaped, not dropped: the reader still sees which analyzer said it.
2594        assert!(c.contains("&lt;script&gt;"), "the text remains, inert: {c}");
2595    }
2596
2597    /// `[debt] ignore` is user config, not analyzer output — but the vault did
2598    /// not write it either, and it lands in a table cell where a `|` reshapes
2599    /// the row.
2600    #[test]
2601    fn a_config_glob_cannot_reshape_the_settings_table() {
2602        let mut api = member_summary("api", 1);
2603        api.settings = RenderedUnder {
2604            ingest: vec!["prose".into()],
2605            debt_ignore: vec!["a`b|c".into()],
2606        };
2607        let c = render_workspace_home(&ws_with(vec![api])).content;
2608        // Scoped to the settings table. The manifest table above it also has a
2609        // row starting `| api |`, and a bare `.find` matched *that* one — so the
2610        // first version of this test asserted against a row containing no glob
2611        // at all, and stayed green with the escaping removed.
2612        let section = c
2613            .split("### Rendered under")
2614            .nth(1)
2615            .unwrap_or_else(|| panic!("the settings section: {c}"));
2616        let row = section
2617            .lines()
2618            .find(|l| l.starts_with("| api |"))
2619            .unwrap_or_else(|| panic!("a settings row: {section}"));
2620        // Count **cell delimiters**, not `|` characters: an escaped `\|` is
2621        // content, and GFM does not split on it. Counting raw pipes would fail
2622        // on the correct output and pass on some wrong ones.
2623        let delimiters = row
2624            .char_indices()
2625            .filter(|&(i, ch)| ch == '|' && !row[..i].ends_with('\\'))
2626            .count();
2627        assert_eq!(
2628            delimiters, 4,
2629            "the row keeps exactly its own four cell delimiters: {row}"
2630        );
2631    }
2632
2633    /// The manifest is a table too. `table_cell` was written for the settings
2634    /// table and not applied to this one, thirty lines above it — a remote URL
2635    /// and a project name both come from git rather than from us.
2636    #[test]
2637    fn a_remote_url_cannot_reshape_the_manifest_row() {
2638        let mut api = member_summary("api", 1);
2639        api.repo_url = Some("https://host/a|b/c".to_owned());
2640        api.commit = Some("dead`beef".to_owned());
2641        let c = render_workspace_home(&ws_with(vec![api])).content;
2642
2643        let section = c
2644            .split("## Reproducing this vault")
2645            .nth(1)
2646            .unwrap_or_else(|| panic!("the manifest: {c}"));
2647        let row = section
2648            .lines()
2649            .find(|l| l.starts_with("| api |"))
2650            .unwrap_or_else(|| panic!("a manifest row: {section}"));
2651        let delimiters = row
2652            .char_indices()
2653            .filter(|&(i, ch)| ch == '|' && !row[..i].ends_with('\\'))
2654            .count();
2655        assert_eq!(
2656            delimiters, 4,
2657            "the row keeps exactly its own four cell delimiters: {row}"
2658        );
2659        assert!(
2660            row.contains("``dead`beef``"),
2661            "and a backtick in a sha widens its span: {row}"
2662        );
2663    }
2664
2665    /// Escaping `[` stops explicit link syntax and nothing else. Obsidian and
2666    /// GFM linkify a bare URL, so analyzer text could still hand the reader
2667    /// something clickable in a document they were given.
2668    #[test]
2669    fn a_bare_url_in_analyzer_text_is_not_left_clickable() {
2670        let mut api = member_summary("api", 1);
2671        api.coverage = Coverage::Ran(vec![("osv-scanner".into(), "1.9.0".into())]);
2672        api.findings = vec![FindingEntry {
2673            rule: "R-1".to_owned(),
2674            severity: "high".to_owned(),
2675            title: "see https://evil.example/path for details".to_owned(),
2676            message: "m".to_owned(),
2677            path: None,
2678            analyzer: "osv-scanner".to_owned(),
2679        }];
2680        let c = render_workspace_home(&ws_with(vec![api])).content;
2681
2682        assert!(
2683            !c.contains("https://evil.example"),
2684            "a bare URL must not survive in linkifiable form: {c}"
2685        );
2686        // Rendered identically for a reader, and still copyable — defusing it
2687        // must not amount to hiding it.
2688        assert!(c.contains("https&#58;//evil.example/path"), "{c}");
2689    }
2690
2691    /// The two empty states are opposite facts and must never render alike.
2692    #[test]
2693    fn an_unanalyzed_member_never_reads_as_one_that_came_back_clean() {
2694        let mut looked = member_summary("api", 1);
2695        looked.coverage = Coverage::Ran(vec![("osv-scanner".into(), "1.9.0".into())]);
2696        let never = member_summary("sdk", 1); // Coverage::NotRun by default
2697        let c = render_workspace_home(&ws_with(vec![looked, never])).content;
2698
2699        assert!(c.contains("### api — no findings"), "{c}");
2700        assert!(
2701            c.contains("osv-scanner` 1.9.0 ran and reported none"),
2702            "{c}"
2703        );
2704        assert!(c.contains("### sdk — **not analyzed**"), "{c}");
2705        assert!(
2706            !c.contains("### sdk — no findings"),
2707            "an unanalyzed member must never be rendered as clean: {c}"
2708        );
2709        assert!(c.contains("1 never analyzed at all"), "counted too: {c}");
2710    }
2711
2712    /// The severity a finding carries is the **analyzer's** word. An unrecognised
2713    /// level is rendered as given rather than mapped onto a known rung, because
2714    /// mapping it would be inventing a judgement the tool did not make.
2715    #[test]
2716    fn an_unrecognised_severity_keeps_the_analyzers_own_label() {
2717        let mut api = member_summary("api", 1);
2718        api.coverage = Coverage::Ran(vec![("semgrep".into(), "1.2.3".into())]);
2719        api.findings = vec![finding("R-1", "WARNING", "msg")];
2720        let c = render_workspace_home(&ws_with(vec![api])).content;
2721        assert!(c.contains("**`WARNING`**"), "{c}");
2722    }
2723
2724    /// #442 part 2: the vault says how to reconstruct the workspace it describes,
2725    /// and what it deliberately leaves out.
2726    ///
2727    /// A vault that says *"here is my workspace"* is far less useful than one
2728    /// that says *"here is my workspace at these commits"* — and the exclusions
2729    /// are stated at the point a reader is most likely to share it, because that
2730    /// is when they matter.
2731    #[test]
2732    fn the_workspace_home_says_how_to_reproduce_itself_and_what_it_omits() {
2733        let mut api = member_summary("api", 7);
2734        api.repo_url = Some("https://github.com/acme/api".to_owned());
2735        api.commit = Some("4e0d5a6afd0b1c2d".to_owned());
2736        // Deliberately **mixed**: one member reproducible, one not. A fixture where
2737        // every member has a remote never exercises the row that says so, and the
2738        // gap is the thing a reader most needs to see.
2739        let mut sdk = member_summary("sdk", 4);
2740        sdk.repo_url = None;
2741        sdk.commit = None;
2742        let ws = WorkspaceSummary {
2743            generated_at: "2026-08-22T10:00:00Z".to_owned(),
2744            name: "platform".into(),
2745            members: vec![api, sdk],
2746            cross_links: vec![],
2747            cross_links_total: 0,
2748            cross_links_authored: 0,
2749        };
2750        let c = render_workspace_home(&ws).content;
2751
2752        assert!(c.contains("2026-08-22T10:00:00Z"), "stamped: {c}");
2753        assert!(c.contains("read-only and point-in-time"), "{c}");
2754        assert!(c.contains("https://github.com/acme/api"), "clone-from: {c}");
2755        assert!(c.contains("4e0d5a6afd0b1c2d"), "pinned commit: {c}");
2756        // A member with no remote is named as such rather than omitted — a gap in
2757        // the manifest is the reader's problem to know about, not ours to hide.
2758        assert!(c.contains("no `origin` remote"), "{c}");
2759
2760        // The share-time warning, each item with its reason. Two of the three are
2761        // about what the vault *carries*, not what it omits — the owner ruled
2762        // findings in, so this section warns rather than reassures.
2763        assert!(
2764            c.contains("cannot be un-shared"),
2765            "including findings has a consequence, stated where it is acted on: {c}"
2766        );
2767        assert!(
2768            c.contains("Agent memory"),
2769            "the one genuine exclusion is still named: {c}"
2770        );
2771        assert!(
2772            c.contains("no redaction chokepoint"),
2773            "with its reason: {c}"
2774        );
2775        // The limit of what *is* included, which is narrower than it looks.
2776        assert!(
2777            c.contains("DATABASE_URL"),
2778            "the redaction gap is shown, not described: {c}"
2779        );
2780    }
2781
2782    /// The manifest records the settings a re-render must match, because a clone
2783    /// URL and a commit get a reader the same **source** and not the same
2784    /// **vault**: `[ingest] prose` off produces notes with no captured content,
2785    /// and a `[debt] ignore` glob means the debt figures on the page are already
2786    /// filtered.
2787    #[test]
2788    fn the_manifest_records_the_settings_a_re_render_would_have_to_match() {
2789        let mut api = member_summary("api", 1);
2790        api.settings = RenderedUnder {
2791            ingest: vec!["prose".into(), "pdf".into()],
2792            debt_ignore: vec!["vendor/**".into()],
2793        };
2794        // A member with everything off is a real state, and must not render as a
2795        // blank cell that reads as "unknown".
2796        let sdk = member_summary("sdk", 1);
2797        let c = render_workspace_home(&ws_with(vec![api, sdk])).content;
2798
2799        assert!(c.contains("### Rendered under"), "{c}");
2800        assert!(c.contains("`prose`, `pdf`"), "the enabled toggles: {c}");
2801        assert!(c.contains("`vendor/**`"), "the debt filter: {c}");
2802        assert!(
2803            c.contains("| sdk | *none* | *none* |"),
2804            "all-off is stated, not left blank: {c}"
2805        );
2806        assert!(
2807            c.contains("will not reproduce this document"),
2808            "and the section says why it is there: {c}"
2809        );
2810    }
2811
2812    /// A workspace of purely local repositories cannot be reconstructed, and the
2813    /// vault says so instead of rendering a table of blanks — which would read as
2814    /// a fault rather than as repositories that were never pushed anywhere.
2815    #[test]
2816    fn a_vault_with_no_remotes_says_it_cannot_be_reconstructed() {
2817        // `member_summary` gives every member a remote, so it must be cleared —
2818        // the fixture has to actually be the case it names.
2819        let mut api = member_summary("api", 1);
2820        api.repo_url = None;
2821        api.commit = None;
2822        let ws = WorkspaceSummary {
2823            generated_at: "2026-08-22T10:00:00Z".to_owned(),
2824            name: "local".into(),
2825            members: vec![api],
2826            cross_links: vec![],
2827            cross_links_total: 0,
2828            cross_links_authored: 0,
2829        };
2830        let c = render_workspace_home(&ws).content;
2831        assert!(c.contains("cannot be reconstructed"), "{c}");
2832        assert!(
2833            !c.contains("| Member | Repository |"),
2834            "no empty table: {c}"
2835        );
2836    }
2837
2838    /// #573: the cross-repo section distinguishes a **declaration** from a
2839    /// **match**, and says how many of each.
2840    ///
2841    /// Before authored links could be persisted, every row was a candidate and
2842    /// the section said so in one blanket caveat. That caveat is now false for
2843    /// declared rows, and an edge that exists but renders as a guess leaves
2844    /// ADR-0009's `authored → gold` path just as unreachable as no edge at all —
2845    /// so the rendering is part of the contract, not decoration.
2846    #[test]
2847    fn the_cross_repo_section_separates_declared_links_from_inferred_ones() {
2848        let link = |authored: bool, key: &str| CrossLink {
2849            from_project: "sdk".into(),
2850            from_key: format!("cfgkey:config.toml#{key}"),
2851            from_name: key.into(),
2852            kind: "references".into(),
2853            // A declaration carries no score by construction; a match does.
2854            confidence: if authored { None } else { Some(0.91) },
2855            to_qualified: format!("api::cfgkey:config.toml#{key}"),
2856            resolves: true,
2857            authored,
2858        };
2859        let ws = WorkspaceSummary {
2860            generated_at: "2026-08-22T10:00:00Z".to_owned(),
2861            name: "platform".into(),
2862            members: vec![member_summary("api", 7), member_summary("sdk", 4)],
2863            cross_links: vec![link(true, "addr"), link(false, "port")],
2864            // Totals deliberately **larger** than the two rows shown: the caption
2865            // is a statement about the workspace, and `cross_links` is a capped
2866            // view of it. Counting the rows would give 1 and 1 here and read as
2867            // correct — which is exactly the bug, invisible until a workspace
2868            // outgrows the cap.
2869            cross_links_total: 9,
2870            cross_links_authored: 4,
2871        };
2872        let c = render_workspace_home(&ws).content;
2873
2874        assert!(
2875            c.contains("**4 declared**"),
2876            "the caption counts the workspace, not the rows on screen: {c}"
2877        );
2878        assert!(c.contains("**5 inferred**"), "{c}");
2879        assert!(
2880            !c.contains("not authored facts"),
2881            "the blanket caveat is false once a declared row can appear: {c}"
2882        );
2883        // Per row: a declaration is marked as one, and a match keeps its score.
2884        assert!(c.contains("references *(declared)*"), "{c}");
2885        assert!(c.contains("references (0.91)"), "{c}");
2886    }
2887
2888    #[test]
2889    fn the_workspace_home_names_an_example_note_name_actually_produces() {
2890        let ws = WorkspaceSummary {
2891            generated_at: "2026-08-22T10:00:00Z".to_owned(),
2892            name: "platform".into(),
2893            members: vec![member_summary("api", 7), member_summary("sdk", 4)],
2894            cross_links: vec![CrossLink {
2895                from_project: "sdk".into(),
2896                from_key: "cfgkey:config.toml#addr".into(),
2897                from_name: "addr".into(),
2898                kind: "links".into(),
2899                confidence: Some(0.91),
2900                to_qualified: "api::cfgkey:config.toml#addr".into(),
2901                resolves: true,
2902                authored: false,
2903            }],
2904            cross_links_total: 1,
2905            cross_links_authored: 0,
2906        };
2907        let note = render_workspace_home(&ws);
2908
2909        // The source end of the first cross-repo link, rendered through the real
2910        // function. `from_project` is a member and `from_key` is one of its own
2911        // nodes, so this is a note the render writes rather than one the
2912        // sentence assumes.
2913        let expected = format!("{}.md", note_name("sdk::cfgkey:config.toml#addr"));
2914        assert!(
2915            note.content.contains(&expected),
2916            "the naming paragraph must show a real name ({expected}), not a \
2917             hand-written form:\n{}",
2918            note.content
2919        );
2920        // And the key form it is derived *from* is still stated, because that is
2921        // the half a reader needs to look a note up by its frontmatter.
2922        assert!(
2923            note.content.contains("`<project>::<key>`"),
2924            "{}",
2925            note.content
2926        );
2927        // No filename anywhere in the vault carries `::`.
2928        assert!(!expected.contains("::"), "{expected}");
2929    }
2930
2931    /// With no cross-repo links there is no key the renderer can prove is a
2932    /// node, so it must say nothing rather than fabricate one.
2933    ///
2934    /// The example this replaced was `<first member>::file:README.md`, invented
2935    /// from the member list — and membership does not require a README, so
2936    /// `_Home` could assert a note that was never written. That is the very
2937    /// defect this PR exists to fix, one remove away, so the empty case gets an
2938    /// assertion of its own rather than an assumption.
2939    #[test]
2940    fn the_workspace_home_claims_no_example_note_when_it_has_no_real_key() {
2941        let ws = WorkspaceSummary {
2942            generated_at: "2026-08-22T10:00:00Z".to_owned(),
2943            name: "platform".into(),
2944            members: vec![member_summary("api", 7), member_summary("sdk", 4)],
2945            cross_links: vec![],
2946            cross_links_total: 0,
2947            cross_links_authored: 0,
2948        };
2949        let note = render_workspace_home(&ws);
2950
2951        assert!(
2952            !note.content.contains("is the note "),
2953            "no cross-repo link means no provable key, so no `Here, X is the \
2954             note Y` claim:\n{}",
2955            note.content
2956        );
2957        // The fabricated form specifically: never emitted, with or without links.
2958        assert!(
2959            !note.content.contains("::file:README.md"),
2960            "{}",
2961            note.content
2962        );
2963        // The rule itself is still stated — only the illustration is absent.
2964        assert!(
2965            note.content.contains("`<project>::<key>`")
2966                && note.content.contains("no filename contains `::`"),
2967            "{}",
2968            note.content
2969        );
2970    }
2971
2972    #[test]
2973    fn a_member_note_declares_which_member_it_came_from() {
2974        let ms = members(&["api"]);
2975        let ex = node_linking_to("cfgkey:config.toml#addr", "addr", "sym:rust:a.rs#A");
2976        let note = render_note_scoped(
2977            &ex,
2978            None,
2979            None,
2980            &VaultScope {
2981                project: Some("api"),
2982                members: &ms,
2983            },
2984        );
2985        assert_eq!(
2986            note.filename,
2987            format!("{}.md", note_name("api::cfgkey:config.toml#addr"))
2988        );
2989        assert!(
2990            note.content.contains("project: \"api\""),
2991            "{}",
2992            note.content
2993        );
2994        assert!(
2995            note.content.contains("- roteiro/project/api"),
2996            "the tag is what filters the graph view to one repository: {}",
2997            note.content
2998        );
2999        // A within-member edge is qualified to the same member, not left bare.
3000        assert!(
3001            note.content
3002                .contains(&format!("→ [[{}]]", note_name("api::sym:rust:a.rs#A"))),
3003            "{}",
3004            note.content
3005        );
3006    }
3007
3008    #[test]
3009    fn a_project_note_declares_no_project() {
3010        let ex = node_linking_to("cfgkey:config.toml#addr", "addr", "sym:rust:a.rs#A");
3011        let note = render_note(&ex, None, None);
3012        assert!(!note.content.contains("project:"), "{}", note.content);
3013        assert!(
3014            !note.content.contains("roteiro/project/"),
3015            "a per-project vault would carry one constant on every note — and \
3016             adding it would change every note's bytes: {}",
3017            note.content
3018        );
3019    }
3020
3021    #[test]
3022    fn a_cross_repo_edge_links_straight_to_the_other_members_note() {
3023        // ADR-0009: the spoke's edge points at a *local placeholder* for the hub's
3024        // node, because store integrity needs both ends in one store. A workspace
3025        // vault holds both, so the link goes to the real note. No new edge — the
3026        // resolver already follows this placeholder at query time.
3027        let ms = members(&["spoke", "hub"]);
3028        let scope = VaultScope {
3029            project: Some("spoke"),
3030            members: &ms,
3031        };
3032        let ex = node_linking_to(
3033            "cfgkey:config.toml#addr",
3034            "addr",
3035            &rto_graph::external_ref_key("hub::cfgkey:config.toml#addr"),
3036        );
3037        let note = render_note_scoped(&ex, None, None, &scope);
3038        assert!(
3039            note.content.contains(&format!(
3040                "→ [[{}]]",
3041                note_name("hub::cfgkey:config.toml#addr")
3042            )),
3043            "the edge must land on the hub's own note: {}",
3044            note.content
3045        );
3046        assert!(
3047            !note.content.contains("extref"),
3048            "and never on the placeholder: {}",
3049            note.content
3050        );
3051        // The same rule decides that the placeholder is not written as a note, so
3052        // the two halves cannot disagree.
3053        assert!(
3054            scope.redirects_external_ref(&rto_graph::external_ref_key(
3055                "hub::cfgkey:config.toml#addr"
3056            ))
3057        );
3058    }
3059
3060    #[test]
3061    fn a_cross_repo_edge_out_of_the_workspace_keeps_its_placeholder() {
3062        // The target repo is not in this vault, so there is no note to point at.
3063        // Redirecting anyway would produce a link that resolves to nothing —
3064        // Obsidian shows that as merely unwritten, which is a worse lie than a
3065        // placeholder that honestly says "elsewhere".
3066        let ms = members(&["spoke"]);
3067        let scope = VaultScope {
3068            project: Some("spoke"),
3069            members: &ms,
3070        };
3071        let key = rto_graph::external_ref_key("elsewhere::cfgkey:config.toml#addr");
3072        assert!(!scope.redirects_external_ref(&key));
3073        let ex = node_linking_to("cfgkey:config.toml#addr", "addr", &key);
3074        let note = render_note_scoped(&ex, None, None, &scope);
3075        assert!(
3076            note.content.contains(&format!(
3077                "→ [[{}]]",
3078                note_name("spoke::extref:elsewhere::cfgkey:config.toml#addr")
3079            )),
3080            "{}",
3081            note.content
3082        );
3083    }
3084
3085    #[test]
3086    fn a_single_project_vault_never_redirects_an_external_ref() {
3087        // No members ⇒ nothing to resolve against, so today's vault keeps rendering
3088        // the placeholder exactly as it does now.
3089        let key = rto_graph::external_ref_key("hub::cfgkey:config.toml#addr");
3090        assert!(!VaultScope::PROJECT.redirects_external_ref(&key));
3091        assert_eq!(
3092            scoped_note_name(&VaultScope::PROJECT, &key),
3093            note_name(&key)
3094        );
3095    }
3096
3097    fn member_summary(project: &str, fan_in: u32) -> VaultSummary {
3098        VaultSummary {
3099            project: project.to_owned(),
3100            total_nodes: 3,
3101            total_edges: 2,
3102            node_counts: vec![("fn".into(), 2)],
3103            edge_provenance: vec![("derived".into(), 2)],
3104            adrs: vec![AdrEntry {
3105                key: "adr:0001".into(),
3106                name: "First".into(),
3107                status: Some("Accepted".into()),
3108            }],
3109            debt: vec![("todo".into(), 4)], // roteiro:ignore
3110            densest_files: vec![DensityEntry {
3111                path: "src/small.rs".into(),
3112                markers: 3,
3113                lines: 120,
3114                per_kloc: 25.0,
3115            }],
3116            config_secrets: None,
3117            most_called: vec![CouplingEntry {
3118                key: "sym:rust:a.rs#helper".into(),
3119                name: "helper".into(),
3120                fan_in,
3121                fan_out: 1,
3122            }],
3123            repo_url: Some(format!("https://github.com/org/{project}")),
3124            commit: Some("abcdef0123456789".into()),
3125            findings: vec![],
3126            coverage: Coverage::NotRun,
3127            settings: RenderedUnder::default(),
3128        }
3129    }
3130
3131    #[test]
3132    fn the_workspace_home_keeps_every_members_own_aggregates() {
3133        // The promise in issue #442: the existing per-project `_Home` view is a
3134        // *subset* of the workspace one, not a casualty of it. Someone who came for
3135        // their repository's coupling and debt tables must still find them —
3136        // not a workspace total that averages them away.
3137        let ws = WorkspaceSummary {
3138            generated_at: "2026-08-22T10:00:00Z".to_owned(),
3139            name: "platform".into(),
3140            members: vec![member_summary("api", 7), member_summary("sdk", 4)],
3141            cross_links: vec![],
3142            cross_links_total: 0,
3143            cross_links_authored: 0,
3144        };
3145        let note = render_workspace_home(&ws);
3146        assert_eq!(note.filename, HOME_NOTE);
3147        assert!(
3148            note.content
3149                .contains("# platform — workspace knowledge graph")
3150        );
3151        // Summed, and the members listed.
3152        assert!(
3153            note.content
3154                .contains("**6 nodes**, **4 edges** across **2** member")
3155        );
3156        assert!(note.content.contains("| [[#api\\|api]] | 3 | 2 |"));
3157
3158        for project in ["api", "sdk"] {
3159            assert!(
3160                note.content.contains(&format!("\n## {project}\n")),
3161                "each member gets its own section"
3162            );
3163        }
3164        // Today's sections, one level deeper, once per member.
3165        for section in [
3166            "### Structure",
3167            "### Provenance",
3168            "### Decisions (ADRs)",
3169            "### Intent debt",
3170            "#### Densest files",
3171            "### Most depended-on",
3172        ] {
3173            assert_eq!(
3174                note.content.matches(section).count(),
3175                2,
3176                "`{section}` must appear once per member: {}",
3177                note.content
3178            );
3179        }
3180        // And every link inside a member's section resolves within that member.
3181        assert!(note.content.contains(&format!(
3182            "**Accepted** — [[{}|First]]",
3183            note_name("api::adr:0001")
3184        )));
3185        assert!(note.content.contains(&format!(
3186            "**Accepted** — [[{}|First]]",
3187            note_name("sdk::adr:0001")
3188        )));
3189        assert!(note.content.contains(&format!(
3190            "[[{}\\|helper]] | 7 |",
3191            note_name("api::sym:rust:a.rs#helper")
3192        )));
3193        assert!(note.content.contains(&format!(
3194            "[[{}\\|src/small.rs]]",
3195            note_name("sdk::file:src/small.rs")
3196        )));
3197    }
3198
3199    #[test]
3200    fn the_workspace_home_renders_cross_repo_links_and_marks_the_ones_it_cannot_follow() {
3201        let ws = WorkspaceSummary {
3202            generated_at: "2026-08-22T10:00:00Z".to_owned(),
3203            name: "platform".into(),
3204            members: vec![member_summary("api", 7), member_summary("sdk", 4)],
3205            cross_links: vec![
3206                CrossLink {
3207                    from_project: "sdk".into(),
3208                    from_key: "cfgkey:config.toml#addr".into(),
3209                    from_name: "addr".into(),
3210                    kind: "links".into(),
3211                    confidence: Some(0.91),
3212                    to_qualified: "api::cfgkey:config.toml#addr".into(),
3213                    resolves: true,
3214                    authored: false,
3215                },
3216                CrossLink {
3217                    from_project: "sdk".into(),
3218                    from_key: "cfgkey:config.toml#other".into(),
3219                    from_name: "other".into(),
3220                    kind: "links".into(),
3221                    confidence: None,
3222                    to_qualified: "absent::cfgkey:config.toml#other".into(),
3223                    resolves: false,
3224                    authored: false,
3225                },
3226            ],
3227            cross_links_total: 2,
3228            cross_links_authored: 0,
3229        };
3230        let note = render_workspace_home(&ws);
3231        // Resolvable: a link to the other member's note, with its confidence.
3232        assert!(
3233            note.content.contains(&format!(
3234                "| [[{}\\|addr]] | sdk | [[{}\\|api::cfgkey:config.toml#addr]] | links (0.91) |",
3235                note_name("sdk::cfgkey:config.toml#addr"),
3236                note_name("api::cfgkey:config.toml#addr"),
3237            )),
3238            "{}",
3239            note.content
3240        );
3241        // Outside the workspace: stated as such, never as a wikilink — Obsidian
3242        // renders a link to a missing note as one that is merely unwritten.
3243        assert!(
3244            note.content
3245                .contains("`absent::cfgkey:config.toml#other` *(outside this workspace)*"),
3246            "{}",
3247            note.content
3248        );
3249        assert!(
3250            !note.content.contains("[[absent-"),
3251            "a dangling wikilink would read as a note someone forgot to write: {}",
3252            note.content
3253        );
3254    }
3255
3256    #[test]
3257    fn the_workspace_home_says_when_it_has_truncated_the_cross_links() {
3258        // A capped table that does not say it is capped reads as the whole set.
3259        let ws = WorkspaceSummary {
3260            generated_at: "2026-08-22T10:00:00Z".to_owned(),
3261            name: "platform".into(),
3262            members: vec![member_summary("api", 7)],
3263            cross_links: vec![CrossLink {
3264                from_project: "api".into(),
3265                from_key: "cfgkey:config.toml#addr".into(),
3266                from_name: "addr".into(),
3267                kind: "links".into(),
3268                confidence: None,
3269                to_qualified: "api::cfgkey:config.toml#addr".into(),
3270                resolves: true,
3271                authored: false,
3272            }],
3273            cross_links_total: 40,
3274            cross_links_authored: 0,
3275        };
3276        let note = render_workspace_home(&ws);
3277        assert!(note.content.contains("Showing 1 of 40"), "{}", note.content);
3278        assert!(note.content.contains("roteiro links --matrix"));
3279    }
3280
3281    #[test]
3282    fn a_workspace_with_no_cross_repo_links_says_why_rather_than_showing_nothing() {
3283        let ws = WorkspaceSummary {
3284            generated_at: "2026-08-22T10:00:00Z".to_owned(),
3285            name: "platform".into(),
3286            members: vec![member_summary("api", 7)],
3287            cross_links: vec![],
3288            cross_links_total: 0,
3289            cross_links_authored: 0,
3290        };
3291        let note = render_workspace_home(&ws);
3292        assert!(note.content.contains("## Cross-repo links"));
3293        assert!(
3294            note.content.contains("links --infer --write"),
3295            "an empty section must name what would fill it, or it reads as \
3296             \"these repos are unrelated\": {}",
3297            note.content
3298        );
3299        // Singular, because getting this wrong on a one-member workspace is the
3300        // kind of thing nobody notices until it ships.
3301        assert!(note.content.contains("**1** member repository."));
3302    }
3303
3304    // ---- YAML frontmatter escaping -------------------------------------------
3305
3306    /// Parse a note's frontmatter block with a **real** YAML parser and return
3307    /// `field`'s value, or the parse error.
3308    ///
3309    /// Every assertion below goes through this rather than checking the emitted
3310    /// bytes. An escaper that is wrong in a self-consistent way passes a
3311    /// byte-comparison — that is precisely how `"foo\bar"` survived: it looks
3312    /// exactly like what was asked for, and means something else.
3313    fn frontmatter_field(note: &str, field: &str) -> Result<Option<String>, String> {
3314        let block = note
3315            .strip_prefix("---\n")
3316            .and_then(|rest| rest.split_once("\n---\n"))
3317            .map(|(block, _)| block)
3318            .expect("note must open with a frontmatter block");
3319        let docs = yaml_rust2::YamlLoader::load_from_str(block).map_err(|e| e.to_string())?;
3320        Ok(docs[0][field].as_str().map(ToOwned::to_owned))
3321    }
3322
3323    /// A node whose key, path and language are whatever the test needs.
3324    fn node_with(key: &str, path: Option<&str>, lang: Option<&str>) -> Explanation {
3325        Explanation {
3326            schema: rto_graph::SCHEMA,
3327            node: NodeSummary {
3328                key: key.into(),
3329                kind: "fn".into(),
3330                name: "n".into(),
3331                path: path.map(ToOwned::to_owned),
3332                lang: lang.map(ToOwned::to_owned),
3333            },
3334            meta: serde_json::Value::Null,
3335            outgoing: vec![],
3336            incoming: vec![],
3337        }
3338    }
3339
3340    /// The three measured failure modes of the escaping this replaced, each
3341    /// asserted on the **parsed** value.
3342    ///
3343    /// Before the fix: `foo\bar` parsed back as `foo<BS>ar` (silently six
3344    /// characters, not seven), and the other two made the whole block
3345    /// unparseable — which in Obsidian costs the note *every* property, with no
3346    /// error shown.
3347    #[test]
3348    fn a_backslash_or_quote_in_a_path_still_parses_back_to_itself() {
3349        for path in [
3350            r"foo\bar",     // `\b` was YAML's backspace escape: silent corruption
3351            r"foo\dir",     // `\d` is not a YAML escape at all: parse error
3352            "say\"hi\".rs", // an unescaped `"` ended the scalar early: parse error
3353            r"a\\b",
3354            "trailing-backslash\\",
3355        ] {
3356            let note = render_note(&node_with("file:x", Some(path), None), None, None);
3357            assert_eq!(
3358                frontmatter_field(&note.content, "path"),
3359                Ok(Some(path.to_owned())),
3360                "path {path:?} must round-trip"
3361            );
3362        }
3363    }
3364
3365    /// `key:` is not hypothetical for this: node keys already carry `:` and `#`,
3366    /// and a symbol name can contain a quotation mark.
3367    #[test]
3368    fn a_node_key_round_trips_whatever_punctuation_it_carries() {
3369        for key in [
3370            "sym:rust:src/a.rs#Store",
3371            r"sym:rust:src\weird.rs#Thing",
3372            "sym:rust:a.rs#say\"hi\"",
3373            "cfgkey:config.toml#serve.addr",
3374        ] {
3375            let note = render_note(&node_with(key, None, None), None, None);
3376            assert_eq!(
3377                frontmatter_field(&note.content, "key"),
3378                Ok(Some(key.to_owned())),
3379                "key {key:?} must round-trip"
3380            );
3381        }
3382        // The old rule turned a `"` into an apostrophe, so the note reported a key
3383        // that was not the node's key — parseable, and wrong.
3384        let note = render_note(
3385            &node_with("sym:rust:a.rs#say\"hi\"", None, None),
3386            None,
3387            None,
3388        );
3389        assert!(
3390            !note.content.contains("say'hi'"),
3391            "a quotation mark must be escaped, not rewritten: {}",
3392            note.content
3393        );
3394    }
3395
3396    /// A member directory name is a path component, so it reaches the same rule.
3397    #[test]
3398    fn a_member_project_name_round_trips() {
3399        let ms: std::collections::BTreeSet<String> =
3400            std::iter::once(r"odd\name".to_owned()).collect();
3401        let note = render_note_scoped(
3402            &node_with("file:x", None, None),
3403            None,
3404            None,
3405            &VaultScope {
3406                project: Some(r"odd\name"),
3407                members: &ms,
3408            },
3409        );
3410        assert_eq!(
3411            frontmatter_field(&note.content, "project"),
3412            Ok(Some(r"odd\name".to_owned()))
3413        );
3414    }
3415
3416    /// The **bare** fields are the other half of the same class, and were missed
3417    /// by the review that found the quoted ones: `status` is written unquoted, and
3418    /// `roteiro load` installs a caller-supplied artifact whose nodes carry
3419    /// whatever they carry.
3420    #[test]
3421    fn a_bare_field_is_quoted_only_when_being_bare_would_change_it() {
3422        let with_status = |status: &str| {
3423            let mut ex = node_with("adr:0001", None, None);
3424            ex.meta = serde_json::json!({ "status": status });
3425            render_note(&ex, None, None)
3426        };
3427
3428        // Would be a parse error bare; would silently truncate bare.
3429        for status in [
3430            "Accepted: superseded by 0012",
3431            "Accepted # pending",
3432            "{draft}",
3433            "",
3434        ] {
3435            let note = with_status(status);
3436            assert_eq!(
3437                frontmatter_field(&note.content, "status"),
3438                Ok(Some(status.to_owned())),
3439                "status {status:?} must round-trip"
3440            );
3441        }
3442
3443        // …and a safe one stays bare, which is what keeps an existing vault's
3444        // bytes unchanged.
3445        let note = with_status("Accepted");
3446        assert!(
3447            note.content.contains("\nstatus: Accepted\n"),
3448            "a plain-safe status must not gain quotes: {}",
3449            note.content
3450        );
3451    }
3452
3453    /// `no` is Norwegian, and a bare `no` reads as `false` to a YAML **1.1**
3454    /// parser.
3455    ///
3456    /// The only assertion here that pins emitted bytes, and deliberately so:
3457    /// `yaml-rust2` implements YAML 1.2, whose core schema resolves a bare `no`
3458    /// to the *string* `no`, so a round-trip through this test's own oracle
3459    /// cannot see the problem — it passes either way. The exposure is to the
3460    /// parser on the other side, and Obsidian's is not this one. Quoting costs
3461    /// two characters on a value that never occurs here; guessing which YAML
3462    /// version every downstream reader implements does not seem like the better
3463    /// bet.
3464    #[test]
3465    fn a_language_that_spells_a_yaml_boolean_is_quoted() {
3466        let note = render_note(&node_with("file:x", None, Some("no")), None, None);
3467        assert!(
3468            note.content.contains("\nlang: \"no\"\n"),
3469            "a bare `no` is `false` to a 1.1 parser and must be quoted: {}",
3470            note.content
3471        );
3472        assert_eq!(
3473            frontmatter_field(&note.content, "lang"),
3474            Ok(Some("no".to_owned())),
3475            "and it must still read back as the string: {}",
3476            note.content
3477        );
3478        // And an ordinary language is untouched.
3479        let rust = render_note(&node_with("file:x", None, Some("rust")), None, None);
3480        assert!(rust.content.contains("\nlang: rust\n"), "{}", rust.content);
3481    }
3482
3483    /// Control characters and the separators some parsers fold as line breaks.
3484    #[test]
3485    fn control_characters_cannot_break_out_of_the_block() {
3486        for path in [
3487            "a\nb",
3488            "a\tb",
3489            "a\u{0}b",
3490            "a\u{2028}b",
3491            "a\u{7f}b",
3492            "a\u{85}b",
3493        ] {
3494            let note = render_note(&node_with("file:x", Some(path), None), None, None);
3495            assert_eq!(
3496                frontmatter_field(&note.content, "path"),
3497                Ok(Some(path.to_owned())),
3498                "path {path:?} must round-trip"
3499            );
3500            // A raw newline would end the scalar and inject a sibling key.
3501            assert_eq!(
3502                note.content.matches("\npath: ").count(),
3503                1,
3504                "the value must stay on one line: {}",
3505                note.content
3506            );
3507        }
3508    }
3509
3510    /// The escaping is *only* an escaping: for a value with nothing to escape it
3511    /// must emit the same bytes it always did, or #442's promise that a
3512    /// single-project vault is byte-identical does not hold.
3513    #[test]
3514    fn an_ordinary_value_is_emitted_exactly_as_before() {
3515        let note = render_note(
3516            &node_with("sym:rust:src/a.rs#Store", Some("src/a.rs"), Some("rust")),
3517            None,
3518            None,
3519        );
3520        assert!(
3521            note.content
3522                .contains("\nkey: \"sym:rust:src/a.rs#Store\"\n")
3523        );
3524        assert!(note.content.contains("\nkind: fn\n"));
3525        assert!(note.content.contains("\npath: \"src/a.rs\"\n"));
3526        assert!(note.content.contains("\nlang: rust\n"));
3527    }
3528
3529    /// The plain-style decision is checked against a real parser rather than
3530    /// against itself: whatever `is_plain_safe` accepts must actually round-trip
3531    /// bare, and whatever it rejects must round-trip quoted.
3532    #[test]
3533    fn the_plain_style_decision_agrees_with_a_real_yaml_parser() {
3534        for value in [
3535            "fn",
3536            "config_key",
3537            "rust",
3538            "Accepted",
3539            "a.b",
3540            "a/b",
3541            "a-b_c",
3542            "no",
3543            "yes",
3544            "true",
3545            "null",
3546            "y",
3547            "N",
3548            "",
3549            " lead",
3550            "trail ",
3551            "a: b",
3552            "a #c",
3553            "{x}",
3554            "[x]",
3555            "*x",
3556            "&x",
3557            "!x",
3558            "#x",
3559            ">x",
3560            "|x",
3561            "%x",
3562            "@x",
3563            "`x",
3564            "\"x",
3565            "'x",
3566            ",x",
3567            "123",
3568            "1.5",
3569            "-x",
3570            ".x",
3571            "a\\b",
3572        ] {
3573            let emitted = super::yaml_scalar(value);
3574            let doc = format!("v: {emitted}");
3575            let parsed = yaml_rust2::YamlLoader::load_from_str(&doc)
3576                .unwrap_or_else(|e| panic!("{value:?} emitted {emitted:?}: {e}"));
3577            assert_eq!(
3578                parsed[0]["v"].as_str(),
3579                Some(value),
3580                "{value:?} emitted as {emitted:?} did not round-trip"
3581            );
3582        }
3583    }
3584}